Method, apparatus, and recording medium for encoding / decoding image

Abstract

Disclosed is an image decoding method using neural network-based bi-prediction. The image decoding method comprises the steps of: determining at least one reference picture for a target block; generating a plurality of reference blocks from the at least one reference picture; configuring input data by using the plurality of reference blocks; and generating a prediction block for the target block from the input data by using a trained neural network.

Description

Method, device and recording medium for video encoding / decoding

[0001] The present invention relates to a method, device, and recording medium for image encoding / decoding. In one aspect, the present invention provides a method for predicting a target block using neural network-based positive prediction.

[0002] With the continuous development of the information and communication industry, services providing video through broadcasting and the Internet have spread worldwide.

[0003] Users demand higher resolution and higher quality video. To meet these demands, video encoding / decoding technologies tailored to these needs are required. Video encoding technology can create compressed video by compressing the video representing the images into a smaller amount of data. Video decoding technology can use the compressed video to create reconstructed images.

[0004] When it comes to video encoding / decoding, various technologies exist, including segmentation, prediction, transformation, quantization, filtering, and entropy encoding / decoding. By introducing, modifying, improving, and combining these diverse technologies, video and images can be compressed, transmitted, and stored more effectively.

[0005] One purpose of the present disclosure is to improve encoding efficiency by providing an image encoding / decoding technology using neural network-based positive prediction.

[0006] Another purpose of the present disclosure is to improve encoding efficiency by providing an image encoding / decoding technology using a neural network-based super-resolution technique.

[0007] One aspect of the present disclosure provides a method for decoding an image. The method comprises the steps of: determining at least one reference picture for a target block; generating a plurality of reference blocks from the at least one reference picture; configuring input data using the plurality of reference blocks; and generating a predicted block for the target block from the input data using a trained neural network.

[0008] Another aspect of the present disclosure provides a method for encoding an image. The method comprises: determining at least one reference picture for a target block; generating a plurality of reference blocks from the at least one reference picture; configuring input data using the plurality of reference blocks; and generating a predicted block for the target block from the input data using a trained neural network.

[0009] Another aspect of the present invention provides a method for transmitting a bitstream encoded by the above image encoding method.

[0010] Another aspect of the present invention provides a computer-readable recording medium for storing a bitstream encoded by the image encoding method.

[0011] Figure 1 illustrates a system for video coding according to one embodiment.

[0012] Figure 2 shows a segmentation structure of an image according to one embodiment.

[0013] Figure 3 illustrates the structure of intra prediction according to one embodiment.

[0014] Figure 4 shows the structure of inter prediction to explain the inter prediction process according to one embodiment.

[0015] Figure 5 shows the order in which spatial candidates are added to the candidate list according to one embodiment.

[0016] Figure 6 illustrates multiple in-loop filters according to an example.

[0017] Figure 7 shows the structure of entropy encoding and entropy decoding according to an example.

[0018] FIG. 8 is an exemplary diagram illustrating a method for selecting one of the first to fourth methods for applying a neural network technique according to one embodiment of the present disclosure.

[0019] FIG. 9 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the first method.

[0020] Figure 10 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the second method.

[0021] Figure 11 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the third method.

[0022] Figure 12 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the fourth method.

[0023] FIG. 13 is an exemplary diagram illustrating neural network-based quantity prediction using a prediction block model according to one embodiment of the present disclosure.

[0024] FIG. 14 is an exemplary diagram illustrating neural network-based quantity prediction using an encoded information combination model according to one embodiment of the present disclosure.

[0025] FIG. 15 is an exemplary diagram illustrating neural network-based quantile prediction using an extended prediction block model according to one embodiment of the present disclosure.

[0026] FIG. 16 is an exemplary diagram illustrating a method for generating a neural network model to be used for quantity prediction through a process of iterative training according to one embodiment of the present disclosure.

[0027] FIG. 17 is a block diagram illustrating an encoding device for applying a super-resolution technique according to one embodiment of the present disclosure.

[0028] Figure 18 is a diagram illustrating an interpolation filter used for downsampling in a reference picture resampling technique.

[0029] FIG. 19 is a block diagram illustrating a decoding device for applying a super-resolution technique according to one embodiment of the present disclosure.

[0030] FIG. 20 is an exemplary flowchart illustrating a method for selecting an upsampling method according to one embodiment of the present disclosure.

[0031] FIG. 21 is an exemplary flowchart illustrating a method for selecting an input data configuration method according to one embodiment of the present disclosure.

[0032] FIG. 22 is an exemplary flowchart illustrating a method for selecting a neural network model according to one embodiment of the present disclosure.

[0033] Fig. 23 is a diagram illustrating an interpolation filter used for upsampling of a luminance component in a reference picture resampling technique.

[0034] Fig. 24 is a diagram illustrating an interpolation filter used for upsampling of chrominance components in a reference picture resampling technique.

[0035] FIGS. 25 to 30 are exemplary diagrams showing neural network structures that can be used for upsampling using a single image neural network-based super-resolution technique according to one embodiment of the present disclosure.

[0036] FIGS. 31 to 34 are exemplary diagrams showing neural network structures that can be used for upsampling using a multi-image neural network-based super-resolution technique according to one embodiment of the present disclosure.

[0037] FIG. 35 and FIG. 36 are diagrams illustrating the configuration of an input picture input to a neural network according to one embodiment of the present disclosure.

[0038] The present invention is capable of various modifications. Furthermore, the present invention may have various embodiments. Specific embodiments are described in detail in the drawings and detailed description.

[0039] It should be understood that the specific examples are not intended to limit the invention to specific embodiments, and that all modifications, equivalents, and substitutes falling within the spirit and scope of the invention are intended to be encompassed within the scope of the invention.

[0040] The embodiments are described in sufficient detail to enable those skilled in the art to practice them. It should be understood that the various embodiments, while different from each other, are not necessarily mutually exclusive. For example, it should be understood that the shapes, structures, and characteristics described in connection with one embodiment may be applied to or implemented in other embodiments without departing from the spirit and scope of the present invention. Furthermore, it should be understood that the positions or arrangements of components within one embodiment may be modified without departing from the spirit and scope of the present invention. Accordingly, the following detailed description is not intended to be limiting, and the scope of the exemplary embodiments, if properly described, is defined only by the appended claims and all equivalents to those claimed by such claims.

[0041] A detailed description of the embodiments described below may refer to the drawings for the embodiments. Any description described in the drawings or the descriptions shown in the drawings may be considered part of the detailed description. In the drawings, similar reference numerals may designate the same or similar functions throughout various aspects. The dependencies between components may not be limited to those depicted in the drawings.

[0042] In the embodiments, a singular expression may include, and may be limited to, and / or restricted by, a plural expression, unless the context clearly excludes a plural expression. That is, expressions such as “at least one” and “one or more” in the embodiments may be replaced with “plural.” Terms such as “ / ,” “and / or,” “at least one of,” and “one or more of” described for a plurality of items may mean 1) one item of the plurality of items, 2) some of the plurality of items, 3) a combination of some of the plurality of items, or 4) a combination of the plurality of items. Furthermore, a plural expression may be replaced with a singular expression. The plural may mean an integer greater than or equal to 1, 2, 3, 4, or 5.

[0043] In the embodiments, terms related to numbers, such as "first" and "second," may be used to describe various components. These terms are used only to distinguish one component from another and do not limit the components. For example, without departing from the scope of the present invention, the first component could be referred to as the second component, and similarly, the second component could also be referred to as the first component.

[0044] When a first component transmits (or provides) information to a second component, it can mean that the first component directly transmits information to the second component, or it can mean that the first component transmits information to the second component via another third component. Here, the information that the second component receives (or obtains) can be information transmitted by the first component, or information generated by applying a specific process to information transmitted by the first component.

[0045] The components of the embodiments may be depicted independently to represent different characteristic functions, and this does not imply that each component corresponds to a separate hardware or software configuration unit. That is, the components of the embodiments may be distinguished and listed for convenience of description. Two or more components described in the embodiments may be regarded as a single component. Furthermore, a single component described in the embodiments may be separated into multiple components that perform the functions of the aforementioned component. Embodiments in which such components are integrated and embodiments in which such components are separated are also included in the scope of the present invention, as long as they do not depart from the essence of the present invention.

[0046] The terms used in the embodiments are used only to describe specific embodiments and are not intended to limit the present invention. In the embodiments, terms such as "comprise" or "have" indicate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the embodiments. These terms do not preclude the presence or addition of other features, numbers, steps, operations, components, parts, or combinations thereof that are not explicitly described in the embodiments. In other words, the description of "comprising" a specific component of an embodiment does not exclude other components other than the specific component, and means that additional components may also be included in the scope of the embodiments of the present invention or the technical idea of ​​the present invention.

[0047] Some of the components of the embodiments may be optional, not essential components for performing the essential functions of the present invention. These optional components may be used to improve performance. The embodiments may be implemented as a structure that includes only the essential components required to implement the essence of the embodiments, excluding the optional components. Such a structure is also within the scope of the embodiments.

[0048] Hereinafter, embodiments are described in detail with reference to the attached drawings to enable those skilled in the art to easily implement the embodiments. In describing the embodiments, if a detailed description of a related known configuration or function is judged to obscure the gist of the present specification, such detailed description will be omitted. Furthermore, identical reference numerals are used for identical components in the drawings, and redundant descriptions of identical components will be omitted.

[0049]

[0050] Interchange between terms in the examples

[0051] Below, terms listed on a single line may be used with the same meaning in the embodiments and may be used interchangeably in the embodiments.

[0052] - 'one or more', 'at least one'

[0053] - 'two or more', 'a plurality of', 'multiple', 'multiple'. (In embodiments, 'one or more' or 'at least one' may be further limited to 'two or more', 'plural', or 'multiple'.)

[0054] - 'Information', 'Signal'

[0055] - 'value', 'predefined value', 'specific value', 'threshold', 'threshold value', 'baseline value', 'reference value'

[0056] - 'statistical value', 'statistics value'

[0057] - 'indicator', 'index', 'index', 'flag', 'information'

[0058] - 'encoder', 'encoding apparatus'

[0059] - 'decoder', 'decoding apparatus'

[0060] - 'Entropy encoding', 'encoding', 'encoding'

[0061] - 'Entropy decryption', 'decoding', 'decoding'

[0062] - 'coding', 'encoding and / or decoding'

[0063] - 'video', 'moving picture', 'image', 'picture', 'frame', 'screen'

[0064] - 'Reference picture', 'Reference video'

[0065] - 'Reference Picture List (RPL),' 'Reference Image List'

[0066] - 'original', 'input', 'source'

[0067] - 'Block', 'Unit', 'Signal'

[0068] - 'square', 'square shape'

[0069] - 'pixel', 'pixels', 'samples', 'pels'

[0070] - 'region', 'area', 'part', 'segment'

[0071] - 'partition', 'split', 'divide'

[0072] - 'quad', 'quarternary'

[0073] - 'luma component', 'luma', 'luminance component', 'luminance', 'Y'

[0074] - 'chroma component', 'chroma', 'chrominance', 'chrominance component', 'Cb and Cr', 'Cb or Cr', 'Cb', 'Cr', 'U and V', 'U or V', 'U', 'V'

[0075] - 'target', 'current' (e.g. target block and current block, or target image and current image)

[0076] - 'neighbor', 'neighboring', 'adjacent', 'neighbor / neighboring' (e.g., neighboring block, adjacent block, and surrounding block)

[0077] - 'collocated', 'collected'

[0078] - 'reconstruction', 'reconstruction', 'decoding'

[0079] - 'reconstructed', 'reconstructed', 'decoded'

[0080] - 'difference', 'difference', 'difference', 'error', 'residual', 'residual'

[0081] - Largest Coding Unit (LCU), Coding Tree Unit (CTU)

[0082] - 'inter', 'inter-screen'

[0083] - 'Inter prediction', 'inter prediction', 'motion compensation'

[0084] - 'Inter mode', 'Inter prediction mode', 'Inter-screen mode', 'Inter-screen prediction mode'

[0085] - 'Motion vector', 'Predicted motion vector', 'Advanced Motion Vector Prediction (AMVP)'

[0086] - 'list', 'candidate list'

[0087] - 'Spatial candidate', 'Spatial merge candidate'

[0088] - 'Temporal candidate', 'Temporal merge candidate'

[0089] - 'Prediction motion vector candidate', 'motion vector predictor'

[0090] - 'Prediction method', 'Prediction mode'

[0091] - 'Intra', 'Intra'

[0092] - 'Intra prediction', 'Intra prediction'

[0093] - 'Intra mode', 'Intra prediction mode'

[0094] - 'Dequantization', 'scaling'

[0095] - 'Quantization matrix', 'Scaling list'

[0096] - 'Quantization matrix coefficients', 'matrix coefficients'

[0097] - 'Transform coefficient level', 'quantized level', 'quantized coefficient', 'quantized transform coefficient', 'quantized transform coefficient level'

[0098] - 'Dequantized coefficient', 'dequantized transform coefficient'

[0099] - 'Scanning type', 'Scanning direction'

[0100] - 'Directional mode', 'Angle mode', 'Angular mode', 'Intra prediction mode'

[0101] - '(mode) number of intra prediction mode', '(mode) index of intra prediction mode', '(mode) value of intra prediction mode', '(mode) angle of intra prediction mode', '(mode) direction of intra prediction mode', '(mode) number of intra prediction direction', '(mode) index of intra prediction direction', '(mode) value of intra prediction direction', '(mode) angle of intra prediction direction'

[0102] - 'Merge Mode', 'Movement Merge Mode'

[0103] - 'Geometric Partitioning Mode (GPM)', 'Triangle Partitioning Mode'

[0104] In addition to the terms exemplified above, terms having the same meaning according to common knowledge in the technical field may be used interchangeably in the embodiments.

[0105]

[0106] The range of information and values ​​of information described in the examples

[0107] In embodiments, information may include constants, flags, indices, variables, coding parameters, elements, syntax elements, motion information, attributes, entities, objects, and data. That is, the term 'information' may be replaced with 'data', 'flag', 'index', 'variable', 'element', 'syntax element', 'motion information', 'attribute', or 'object'.

[0108] Information can have one of multiple values. 'n-th value' can mean the nth value among multiple values.

[0109] For example, the first value could represent '0' or (logical) false. The second value could represent '1' or (logical) true. Alternatively, the first value could represent '1' or (logical) true. The second value could represent '0' or (logical) false.

[0110] A flag may be information having a value of either '0' or '1'. In embodiments, the flag values ​​'0' and '1' may be replaced with '1' and '0', respectively. For example, information indicating whether a specific process is performed or whether a specific process is applied may be considered a flag.

[0111] When a variable such as i or j is used to represent a row, column, or index, the variable can be an integer greater than or equal to 0 and less than or equal to n - 1. Alternatively, the variable can be an integer greater than or equal to 1 and less than or equal to n. Here, n can be the number of rows, the number of columns, or the number of entities pointed to by the index.

[0112]

[0113] Coding related concepts

[0114] Below, concepts related to coding are described. The descriptions disclosed below can be applied to embodiments.

[0115] Predefined value: A predefined value may refer to a value commonly used in an encoding device and a decoding device. For example, a predefined value may be interpreted as a fixed value. Alternatively, the predefined value may be a value shared by an encoding device and a decoding device through signaling. Alternatively, the predefined value may be a value derived through the same procedure in an encoding device and a decoding device so that the encoding device and the decoding device have a common value. Alternatively, the predefined value may be a common value in an encoding device and a decoding device. The above description of a predefined value may also be applied to predefined information. In the above descriptions, 'value' may be replaced with 'information'.

[0116] - The values ​​derived through the same procedure in the encoding device and decoding device may include values ​​derived through the same procedure for the same value and / or the same information in the encoding device and decoding device.

[0117] - The values ​​derived through the same procedure in the encoding device and decoding device may include values ​​derived using the same conditional statement for the same value and / or the same information in the encoding device and decoding device.

[0118] - The above description of the predefined values ​​can also be applied to predefined information. In the above descriptions, 'value' can be replaced with 'information'.

[0119] Availability: The availability of certain modes for a specific target may mean that a selected mode among the specific modes is used for the specific target. Other modes within the specific mode category may be unavailable. Unavailable modes may not be used for the specific target. The description of a specific mode above may also apply to other specific information. In the descriptions above, "mode" may be replaced with "information."

[0120] Adjacency: The 'direction' of the 'second entity' with respect to the 'first entity' may refer to the 'second entity' that is adjacent to the 'direction' corner / face of the first entity. For example, the 'top left block' with respect to the 'target block' may be a block adjacent to the top left of the target block. Here, the 'first entity' may be a target unit, a target block, or a target sample. The 'direction' may be one of left-above, above, right-above, left, right, left-below, below, and right-below. The 'second entity' may be a unit, a block, or a sample. For the directions of left-top, right-top, left-bottom, and right-bottom, the corners of the first entity and the corners of the second entity may be diagonally adjacent. For the directions of top, left, right and bottom, one side of the first object and one side of the second object can be in contact with each other.

[0121] - For example, the block adjacent to the upper left of the target block may be the block adjacent to the upper left of the block adjacent to the target block. The block adjacent to the upper right of the target block may be the block adjacent to the right of the block adjacent to the upper right of the target block. The block adjacent to the lower left of the target block may be the block adjacent to the lower left of the block adjacent to the target block.

[0122] Coding: Coding can mean encoding and / or decoding of images.

[0123] Signal: A signal can represent information about an image, unit, or block. A specific signal can represent a specific image, a specific unit, or a specific block.

[0124] Video: A video can refer to a single picture that constitutes a video, or it can refer to the video itself. For example, "encoding and / or decoding a video" can mean "encoding and / or decoding a video," or it can mean "encoding and / or decoding one of the pictures that constitute the video."

[0125] - A picture can mean the entire picture, or it can mean a part of a picture, such as a block.

[0126] Target Image: The target image may be an encoding target image, which is the target of encoding, and / or a decoding target image, which is the target of decoding. Furthermore, the target image may be an input image processed by an encoding device, or a restored image processed by a decoding device. The target image may be an image including a target block.

[0127] Subpicture: A picture can be divided into one or more subpictures.

[0128] - A subpicture may be a square or rectangular area within a picture. A subpicture may contain one or more CTUs.

[0129] - A subpicture may include one or more slices and / or one or more tiles. For example, a subpicture may consist of one or more slice rows and one or more slice columns. Alternatively, each subpicture may consist of one or more tile rows and one or more tile columns.

[0130] - A subpicture may include one or more slices that collectively cover a rectangular area within the picture. Accordingly, the boundary of each subpicture may always be the boundary of a slice. Additionally, each vertical subpicture boundary may always be a vertical tile boundary.

[0131] Slice: A slice may contain one or more tiles within a picture. A slice may consist of one or more rows of tiles and one or more columns of tiles.

[0132] Tile: A tile can be a square or rectangular area within a picture. A tile can contain one or more CTUs. A picture can be divided into one or more tile rows and one or more tile columns.

[0133] CTU: An image can be divided into multiple coding tree units (CTUs).

[0134] - A CTU may include one Y coding tree block (CTB) and at least one of a Cb CTB and a Cr CTB related to the Y CTB, and may include information about each CTB. The information may include syntax elements.

[0135] - Each CTU can be partitioned using one or more partitioning methods to form sub-units such as coding units (CUs), prediction units (PUs), and transform units (TUs). The one or more partitioning methods can include quad tree (QT) partitioning, binary tree (BT) partitioning, and ternary tree (TT) partitioning. Additionally, each CTU can be partitioned using multi-type tree (MTT) partitioning that uses a combination of multiple partitioning methods.

[0136] CTB: CTB can refer to one of Y CTB, Cb CTB, and Cr CTB.

[0137] Unit: A unit can be determined for specific processing in coding. A unit can contain information about a specific region within an image. For specific coding processing, an image can be recursively divided into multiple parts. A unit can represent the region to which a specific processing is applied and information about the region.

[0138] - The type of a unit may indicate a specific processing to be applied to the unit. Depending on the type of the unit, a specific processing may be applied to the unit. A 'specific' unit may be a unit for processing designated as 'specific' in coding. For example, the unit may be at least one of an original unit, a CTU, a coding unit, a prediction unit, a residual unit, a reconstructed residual unit, a transformation unit, and a reconstructed unit.

[0139] - A unit may include samples having a two-dimensional shape or arrangement. In this respect, a 'unit' may also mean a 'block'. For example, a block may be at least one of an original block, a CTB, a coding block (CB), a prediction block (PB), a residual block, a reconstructed residual block, a transform block (TB), and a reconstructed block. For example, a division of a unit may mean a division of a block corresponding to the unit.

[0140] - A unit can contain syntax elements. In other words, a block and its syntax elements can be combined to form a unit.

[0141] - A block is an MxN array of samples. Here, M and N can represent positive integer values, and a block can commonly represent a two-dimensional sample array. The current block can represent an encoding target block that is the target of encoding during encoding, and a decoding target block that is the target of decoding during decoding. In addition, the current block can be at least one of a coding block, a prediction block, a residual block, a transform block, and a restoration block. The block can have various sizes and shapes. For example, the shape of the block can be one or more of a tetragon, a rectangular block, a square block, a rectangle whose width is different from its height (that is, an oblong block), a trapezoid, a triangle, a right-angled triangle, and a pentagon. Here, the width and height of the rectangle can be different from each other. In addition, the shape of the block can include other geometric shapes that can be expressed in two dimensions. For example, the shape of a block may be a quadrilateral or a pentagon, which is defined by excluding the area of ​​a right triangle from the area of ​​a rectangle. Here, the right vertex of the right triangle may be one of the vertices of the rectangle. Furthermore, the shape of a block may be a combination of two or more of the aforementioned shapes. Furthermore, the shape of a block may be the remainder of one of the aforementioned shapes after excluding another shape.

[0142] - In embodiments, a rectangle may be limited to a non-square rectangle. When the shape of a particular object is described as a rectangle in an embodiment, such description may additionally imply that the width and height of the particular object are different from each other.

[0143] - In embodiments, a block may be limited to at least one of a vertically oriented block and a horizontally oriented block. A vertically oriented block may mean a block whose vertical length is greater than its horizontal length. A horizontally oriented block may mean a block whose horizontal length is greater than its vertical length.

[0144] - A unit may include a luma component block (i.e., a Y block) and two chroma component blocks (i.e., at least one of a Cb block and a Cr block), and may include information about each block. The information may include syntax elements.

[0145] - Unit information may include unit type, unit size, unit depth, unit encoding order, and unit decoding order.

[0146] Target Unit: A target unit may be a block, an encoding target unit, which is a target of encoding, and / or a decoding target unit, which is a target of decoding. A target unit may be a specific area within a target picture to which one or more specific coding processes are applied. A unit of a specific type may be generated by applying a specific process to a target unit. Alternatively, a target unit may represent a unit having a specific type for a specific coding process.

[0147] Depth: A block can be hierarchically divided into multiple sub-blocks, each with its own depth, according to a tree structure. The multiple sub-blocks created by block division can be called partitions.

[0148] - The depth of a block can indicate the level of the node corresponding to the block when the blocks that make up the image are expressed in a tree structure. Alternatively, the depth of a block can indicate the number of partitions applied until the block is determined. The depth of a block can increase by 1 as the block is further partitioned.

[0149] - In a tree structure, the root node can be considered to have the smallest level, and the leaf node can be considered to have the largest level. The root node can be the topmost node in the tree structure and corresponds to the first undivided block. The level of the root node can be 0 or 1. When the level of the root node is 0, a node with a level of 1 can represent a block determined by dividing the first block once. A node with a level of n can represent a block determined by dividing the first block n times. A leaf node can be the lowest node in the tree structure. A leaf node can be a node that cannot be further divided. The depth of a leaf node can be a predefined maximum depth. For example, the maximum depth can be a positive integer such as 3. The root node can mean a CTU. A leaf node can mean at least one of a CU, a PU, and a TU.

[0150] - Depth can have a type depending on the type of partition. QT depth can represent the depth for quadtree partitioning. BT depth can represent the depth for binary partitioning. TT depth can represent the depth for ternary partitioning.

[0151] Sample: A sample can be a base unit that constitutes a block. A sample can be composed of one or more bits. The bit depth can be the number of bits that constitute a sample. A sample can be numbered from 0 to 2 depending on the bit depth. Bd It can be expressed as values ​​up to -1.

[0152] PU: PU may denote a basic unit for prediction-related processing. For example, prediction-related processing may include inter-prediction, intra-prediction, intra-block copy (IBC) prediction, intra-compensation, and motion compensation.

[0153] - A PU can be divided into multiple sub-PUs, each of which has a smaller size than the PU itself. These multiple sub-PUs can also serve as the basis for prediction-related processing. In other words, a prediction unit partition generated by splitting a prediction unit can also be a prediction unit.

[0154] TU: A TU may be a basic unit for processing related to a residual block. The processing related to the residual block may include at least one of a transform, an inverse transform, quantization, inverse quantization, transform coefficient encoding, transform coefficient decoding, entropy encoding, and entropy decoding. - One TU may be split into a plurality of sub-transform units having a size smaller than the size of the TU. The plurality of sub-TUs may also be basic units for processing related to the residual block. In other words, a transform unit partition generated by splitting a transform unit may also be a transform unit.

[0155] - The transformation may include one or more of a primary transformation and a secondary transformation, and the inverse transformation may include one or more of a primary inverse transformation and a secondary inverse transformation.

[0156] Parameter set: A parameter set may correspond to header information among the structures within a bitstream.

[0157] - The parameter set may include at least one of a video parameter set (VPS), a sequence parameter set (SPS), a picture parameter set (PPS), an adaptation parameter set (APS), and a decoding parameter set (DPS).

[0158] - Information signaled through a parameter set can be applied to pictures referencing the parameter set. For example, information within a VPS can be applied to pictures referencing the VPS. Information within an SPS can be applied to pictures referencing the SPS. Information within a PPS can be applied to pictures referencing the PPS. A parameter set can refer to a higher-order parameter set. For example, a PPS can refer to an SPS. An SPS can refer to a VPS.

[0159] - Additionally, the parameter set may include tile group information, slice header information, and tile header information. A tile group may mean a group or slice including multiple tiles.

[0160] MPM (Most Probable Mode): MPM can indicate the intra prediction mode that is likely to be used for intra prediction for the target block.

[0161] - One or more different MPMs can be determined based on coding parameters related to the target block and properties of objects related to the target block.

[0162] - One or more MPMs may be determined based on the intra prediction mode of a reference block. There may be multiple reference blocks. Depending on which intra prediction modes are used for one or more reference blocks, one or more different MPMs may be determined. The reference blocks may include spatial neighboring blocks.

[0163] MPM List: An MPM list may contain one or more MPMs. The number of MPMs in an MPM list may be predefined.

[0164] MPM Index: The MPM index can indicate an MPM among one or more MPMs in the MPM list to be used for intra prediction for the target block.

[0165] MPM Usage Directive: The MPM usage directive can indicate whether the MPM list is used for prediction on the target block.

[0166] Prediction mode: The prediction mode may be information indicating a prediction method for a target block, such as a mode used for intra prediction or a mode used for inter prediction. The prediction mode may refer to one of the prediction-related modes described in the embodiments. In addition, the prediction mode may include at least one of an intra mode, an inter mode, and an intra block copy mode.

[0167] Reference image list: The reference image list may be a list containing one or more reference images used for prediction for the target block.

[0168] - There may be multiple reference image lists. Multiple reference image lists may include List 0 (L0), List 1 (L1), etc.

[0169] - One or more reference image lists may be used for inter prediction for a target block. Parts such as 'L0' and 'L1' in the names of information related to inter prediction may refer to reference image lists related to the information.

[0170] Reference picture: A reference picture may be an image referenced for prediction of a target block. Alternatively, the reference picture may be an image containing a reference block. The reference picture may include a previous image of the target image, a target image, and a subsequent image of the target image.

[0171] Reference image index: The reference image index may be an index indicating one reference image among one or more reference images in the reference image list that is used for prediction of the target block.

[0172] Reference Block: A reference block may be a block referenced for encoding / decoding a target block, such as prediction or filtering. For example, a reference block may include reference samples used to derive prediction samples, and may also refer to a block that provides information used for decoding the target block.

[0173] Reference sample: A reference sample may be a sample that is referenced for encoding / decoding of a target block, such as prediction and filtering.

[0174] Inter prediction indicator: The inter prediction indicator can indicate the direction of inter prediction for the target block. The inter prediction can be one of uni-directional prediction and bi-directional prediction. Alternatively, the inter prediction indicator can indicate the number of reference pictures used when generating a prediction block of the target block. Alternatively, the inter prediction indicator can indicate the number of prediction blocks used for inter prediction for the target block. The reference direction can mean the inter prediction indicator. For example, the inter prediction indicator can indicate one of uni-directional and bi-directional. Alternatively, the inter prediction indicator can have a first value of '0' for an inter mode that uses only reference pictures in the L0 reference picture list, a second value of '1' for an inter mode that uses only reference pictures in the L1 reference picture list, and a third value of '2' for an inter mode that uses at least two of the reference pictures in the L0 reference picture list and the reference pictures in the L1 reference picture list.

[0175] Prediction List Utilization Flag: The prediction list utilization flag for a specific reference image list may indicate whether at least one reference image within the specific reference image list is used to generate a prediction block of the target block. For example, a value of the prediction list utilization flag for a specific reference image list of '0' may indicate that a prediction block is not generated using a reference image within the specific reference image list. A value of the prediction list utilization flag for a specific reference image list of '1' may indicate that a prediction block is generated using a reference image within the specific reference image list.

[0176] - An inter prediction indicator can be derived using a prediction list utilization flag. Conversely, a prediction list utilization flag can be derived using an inter prediction indicator. For example, an inter prediction indicator can be derived using prediction list utilization flags for a plurality of reference image lists. If an inter prediction indicator indicates that specific reference lists among a plurality of reference image lists are used, the prediction list utilization flags of the specific reference lists indicated by the inter prediction indicator among the prediction list utilization flags of the plurality of reference image lists can be set to '1', and the prediction list utilization flags of the remaining reference image lists not indicated by the inter prediction indicator can be set to '0'.

[0177] Reference Direction: The reference direction may point to a list of reference images used for prediction of the target block. For example, the reference direction may point to one or more of the reference image list L0 and the reference image list L1.

[0178] - The reference direction may not indicate that the directions of the reference images in the reference image list are limited to the forward direction or the backward direction, but only indicates the reference image list used for prediction of the target block. That is, each of the reference image list L0 and the reference image list L1 may include forward images and backward images. Here, the forward direction may indicate the direction from the target image to the previous image of the target image. Forward inter prediction may be inter prediction that uses the previous image of the target image as a reference image. Backward direction may indicate the direction from the target image to the subsequent image of the target image. Backward inter prediction may be inter prediction that uses the subsequent image of the target image as a reference image.

[0179] - A unidirectional reference direction may mean that one reference image list is used. A bidirectional reference direction may mean that two reference image lists are used. For example, the reference direction may indicate that only the reference image list L0 is used, that only the reference image list L1 is used, or that two reference image lists are used. Additionally, the reference direction may be indicated by an inter prediction indicator.

[0180] Picture Order Count (POC): The POC of a picture can indicate the display order or output order of the picture.

[0181] Motion information: Motion information may be information used to specify a reference block. The motion information may include information used for inter prediction, such as a motion vector (MV), a reference picture index, a reference picture, an inter prediction indicator, a prediction list utilization flag, and the like. In addition, the motion information may include information used in a specific inter prediction mode, such as an MV candidate, an MV candidate index, a merge candidate, and a merge index. In addition, the motion information may include information related to a block vector described below. The information related to a block vector may mean information including at least one of a block vector, a block vector candidate, and a block vector candidate index.

[0182] - Multiple motion information for multiple reference image lists can be used for inter prediction for a target block, respectively. Motion information for a specific reference image list can be used for prediction using the specific reference image list. Multiple (intermediate) prediction blocks can be derived from the multiple motion information. A (final) prediction block for the target block can be generated using statistical values ​​for the multiple (intermediate) prediction blocks.

[0183] MV: MV can be a two-dimensional vector used in inter prediction. It can represent the offset between a target block and a reference block. Alternatively, it can represent the difference between the locations of a target block and a reference block.

[0184] - For example, MV is (mv x , mv y ) can be expressed in the form of mv x can represent the horizontal component, and mv y can represent vertical components.

[0185] -The zero vector can be (0, 0) MV.

[0186] Block Vector (BV): A BV can be a two-dimensional vector used in intra-block copy prediction. A BV can represent the offset between a target block within a target image and a reference block within the target image. In other words, a BV can represent the displacement between the target block and the reference block within the target image.

[0187] - For example, BV is similar to MV (bv x , bv y ) can be expressed in the form of bv x can represent the horizontal component, bvy can represent vertical components.

[0188] -The zero vector can be (0, 0) BV.

[0189] Motion Information Candidate: In a specific prediction, the motion information of the target block can be selected from among motion information candidates determined by a specific method. The motion information candidate may refer to the motion information of a reference block, or it may refer to the reference block itself containing the motion information. Here, the reference block may be a block determined by a specific method for selecting a motion information candidate.

[0190] Candidate List: A candidate list may be a list containing one or more candidates. For example, the candidate list may include a motion information candidate list, a merge candidate list, an MV candidate list, an MPM list, etc. The candidate list may be generated in the same manner by the encoding device and the decoding device. In other words, the candidate list used by the encoding device and the candidate list used by the decoding device may be the same, and the same candidate list may be shared by the encoding device and the decoding device. The encoding device may select a candidate to be used for processing the target block from among the candidates in the candidate list. An indicator indicating the selected candidate may be signaled from the encoding device to the decoding device. The decoding device may use the indicator to specify a candidate to be used for processing the target block from among the candidates in the candidate list. Alternatively, the encoding device and the decoding device may specify a candidate to be used for processing the target block from among the candidates in the candidate list according to the same rule.

[0191] Motion information candidate list: The motion information candidate list may mean a list constructed using one or more motion information candidates.

[0192] Motion information candidate index: The motion information candidate index may be an identifier or indicator that indicates a motion information candidate used for prediction of a target block among the motion information candidates in the motion information candidate list.

[0193] - In a specific inter prediction mode, motion information of other reconstructed blocks may be used to derive motion information of the target block. The other blocks may include neighboring blocks. In this specific inter prediction mode, the motion information for the target block itself is not individually signaled, but other information used to derive motion information of the target block based on the motion information of other reconstructed blocks may be signaled. In this case, the other information may include information indicating which of the other reconstructed blocks' motion information is used to derive motion information of the target block, such as a motion information candidate index.

[0194] - For example, these inter prediction modes may include AMVP mode, merge mode, and skip mode. The motion information candidate index may be a merge index or an MV candidate index.

[0195] - In embodiments, MV may be part of motion information. In embodiments, information about motion information, such as motion information candidates, motion information candidate lists, and motion information candidate indices, may be replaced with information about MVs, such as MV candidates, MV candidate lists, and MV candidate indices, and the description of motion information may also be applied to MVs.

[0196] Merge: Merge can refer to the merging of motion information across multiple blocks, or it can refer to applying motion information from another block to the target block. In other words, merge mode can refer to a mode in which the motion information of the target block is derived from the motion information of neighboring blocks.

[0197] Merge Candidate: A merge candidate may refer to a specific (restored) block used for merging the target block, or may refer to motion information for the specific block. Alternatively, the merge candidate may include motion information for the specific block.

[0198] - Merge candidates for the target block may include spatial merge candidates, temporal merge candidates, history-based candidates, average candidates based on the average of two merge candidates, and zero merge candidates.

[0199] Merge Candidate List: A merge candidate list may be a list constructed using one or more merge candidates.

[0200] Merge Index: A merge index may be an indicator that points to a merge candidate among the merge candidates in the merge candidate list, which is used for prediction of the target block. The motion information of the merge candidate indicated by the merge index among the merge candidates in the merge candidate list may be used as motion information of the target block.

[0201] Neighboring block: A neighboring block can refer to a block adjacent to the target block. Neighboring blocks can include spatial and temporal neighboring blocks. Neighboring blocks can also refer to reconstructed neighboring blocks within the reference image. Neighboring blocks do not necessarily have to be adjacent to the target block.

[0202] Spatial neighboring blocks: Spatial neighboring blocks can be blocks that are spatially adjacent to the target block.

[0203] - The target block and spatial neighboring blocks can be included within the target image.

[0204] - A spatial neighboring block may include a block whose boundary is at least partially adjacent to a boundary of the target block. Alternatively, a spatial neighboring block may include a block whose distance from the target block is less than or equal to a specific value.

[0205] - A spatial neighboring block may include a block diagonally adjacent to a vertex of the target block.

[0206] - Spatial neighboring blocks may include an upper left block adjacent to the upper left of the target block, an upper block adjacent to the upper right of the target block, an upper right block entered at the upper right of the target block, a left block adjacent to the left of the target block, a right block adjacent to the right of the target block, a lower left block adjacent to the lower left of the target block, a lower block adjacent to the lower bottom of the target block, and a lower right block adjacent to the lower right of the target block.

[0207] Temporal neighboring blocks: Temporal neighboring blocks can be blocks that are temporally adjacent to the target block.

[0208] - A temporal neighboring block may include a collocated block (COL block). A collocated block may be a block within a reconstructed image within a reference image buffer. A collocated picture (col picture) may refer to an image that includes a collocated block. A collocated picture may be an image included in a reference image list.

[0209] - Call blocks can be determined based on the location of the target block within the target image. Two blocks being "temporally adjacent" can mean that the locations of the two blocks satisfy certain conditions.

[0210] - The position of a call block within a call image may be the same as the position of a target block within a target image. Alternatively, the position of a call block within a call image may correspond to the position of a target block within a target image. Here, the correspondence of the positions of blocks may mean that the areas of the blocks are identical, that an area of ​​one block is included in an area of ​​another block, or that one block occupies a specific position of another block.

[0211] - For example, the location of a call block within a call image may be identical to the location of a target block within the target image. Alternatively, a call block may be a block containing a call sample within a call image. A call sample may be a sample having coordinates identical to the coordinates of a specific sample in the target block.

[0212] - A temporal neighboring block may be a block that is temporally adjacent to a spatial neighboring block of the target block.

[0213] Neighbor sample: A neighbor sample may refer to a sample within a neighboring block. Neighbor samples may include predicted samples, reconstructed samples, residual samples, and decoded samples.

[0214] Search range: The search range can refer to a two-dimensional region where MVs are searched during inter prediction. For example, when the optimal MV must be derived for processing a target block, the optimal MV can be selected from among the MVs pointing within the search range.

[0215] Transform coefficient: The transform coefficient may be a coefficient generated by performing a transformation on the residual block. Alternatively, the transform coefficient may be a coefficient value generated by performing dequantization on a quantized level.

[0216] Quantized level: A quantized level can be an integer quantity used as input to dequantization.

[0217] Quantization: Quantization can be the process of generating quantized levels for transform coefficients. Quantized levels can be generated by applying quantization to transform coefficients. The transform can also be considered part of quantization.

[0218] Dequantization: Dequantization can be the process of multiplying a quantized level by a factor. By applying dequantization to a quantized level, (restored) transform coefficients can be generated.

[0219] Quantization Parameter (QP): QP can refer to an argument used to generate quantized levels for transform coefficients in quantization. QP can also refer to an argument used to generate (restored) transform coefficients for quantized levels in dequantization. Alternatively, QP can be a value mapped to the quantization step size.

[0220] Delta QP: Delta QP can be the difference between the QP predicted by a specific process and the QP of the target block. In other words, the QP of the target block can be the sum of the predicted QP and the delta QP.

[0221] Quantization matrix: A quantization matrix can be a matrix used in quantization or inverse quantization to improve the subjective or objective quality of an image.

[0222] Quantization matrix coefficients: Quantization matrix coefficients can be each element within a quantization matrix.

[0223] Scan: A scan can refer to the arrangement of values ​​within a block or matrix. The values ​​can be coefficients. For example, a scan can refer to arranging values ​​arranged in a two-dimensional form into a one-dimensional form, or it can refer to rearranging values ​​arranged in a one-dimensional form into a two-dimensional form. An inverse scan can be the opposite arrangement (or rearrangement) of the arrangement performed in a scan.

[0224] Non-zero transform coefficient: A non-zero transform coefficient can mean a transform coefficient with a non-zero value or a quantized level with a non-zero value.

[0225] Bitstream: A bitstream may refer to a sequence of bits containing encoded information generated by encoding an image. A bitstream may contain information according to specific syntax elements. For example, information may contain syntax elements. An encoding device may generate a bitstream containing information according to specific syntax elements. A decoding device may obtain information from the bitstream according to specific syntax elements.

[0226] Signaling: Signaling information may indicate that information is transmitted from an encoding device to a decoding device via a bitstream. For example, the information may include a syntax element. Alternatively, signaling may mean that the encoding device includes information in a bitstream. Information signaled by the encoding device may be used by the decoding device. In signaling, the bitstream may be transmitted via a network and may be included in a recording medium. In embodiments, description of information being signaled may include: 1) for signaling information, the encoding device determines and generates information; 2) the encoding device encodes the information to generate encoded information; 3) (encoded) information is transmitted from the encoding device to the decoding device via a bitstream; 4) the decoding device decodes the encoded information to obtain information; and 5) for signaling information, the decoding device determines and generates information via signaling.

[0227] - An encoding device can perform encoding on information to generate encoded information. The encoded information can be signaled via a bitstream. A decoding device can obtain information by decoding the encoded information.

[0228] - When information is signaled for a specific target, it can mean that the information is used for each specific target, and the processing indicated by the information is applied to each specific target. For example, when information is signaled at a specific unit level, it can mean that the information is used / processed for each specific unit.

[0229] - The information being signaled may include one or more sub-information. Signaling a specific piece of information may mean that each piece of information within one or more sub-information pieces contained within the specific information is signaled.

[0230] Selective Signaling: Signaling of information may be performed selectively. Selective signaling of information may mean that the encoding device selectively includes information in the bitstream (under certain conditions). Selective signaling of information may mean that the decoding device selectively obtains information from the bitstream (under certain conditions).

[0231] Omission of signaling: Signaling for information may be omitted. Omission of signaling for information may mean that the encoding device (under certain conditions) does not include the information in the bitstream. Omission of signaling for information may mean that the decoding device (under certain conditions) does not obtain the information from the bitstream. The decoding device may derive the information for which signaling is omitted using other information of the embodiments.

[0232] Symbol: may mean at least one piece of information of a target unit, such as a syntax element of a target unit or target block, a coding parameter, a quantized level, and a transform coefficient. In addition, a symbol may mean a target of entropy encoding or a result of entropy decoding.

[0233] Entropy encoding: Entropy encoding can allocate fewer bits to symbols with a high probability of occurrence, and more bits to symbols with a low probability of occurrence. This allocation reduces the size of the bitstream representing the symbols as they are represented.

[0234] - Entropy coding can use methods such as Variable Length Coding (VLC) and Context-Adaptive Binary Arithmetic Coding (CABAC). For example, in variable length coding, entropy coding can be performed using a variable length table. For example, in CABAC, a binarization method for symbols and a probability model of symbols / bins can be derived for entropy coding, and arithmetic coding using context can be performed.

[0235] Entropy decoding: Entropy decoding can reverse the processes performed in entropy encoding. Symbols can be generated by entropy decoding a bitstream.

[0236] Parsing: Parsing can mean determining the values ​​of syntactic elements by performing entropy decoding on the encoded information in the bitstream. Alternatively, parsing can mean entropy decoding itself.

[0237] Statistical Value: The values ​​of information related to specific entities described in the embodiments may be used as inputs to specific operations. The statistical value may be a value derived by a specific operation on the values ​​related to these specific entities. For example, the statistical value for specific information may be one or more of an average value, a weighted average value, a weighted sum value, a minimum value, a maximum value, a mode, a median value, an interpolated value, a sum of products, and a product of sums of values ​​of the specific information. Additionally, information of the embodiments having specific values ​​determined by operations, such as constants, variables, and coding parameters, may have specific statistical values ​​according to the embodiments.

[0238]

[0239] Coding parameters

[0240] In embodiments, coding parameters may be information required for coding. The coding parameters may include information signaled from an encoding device to a decoding device, information calculated / derived during the coding process described in the embodiments, and information used for the coding process described in the embodiments.

[0241] In embodiments, the coding parameters include a size of a CTU, a size of a unit, a form of a unit, a shape of a unit, a depth of a unit, a minimum unit size, a maximum unit size, a maximum unit depth, a minimum unit depth, a partition information of a unit, QT partition information, BT partition information, a partition direction of a BT partition, a partition shape of a BT partition, TT partition information, a partition direction of a TT partition, a partition shape of a TT partition, MTT partition information, a combination of MTT partitions, a partition direction of an MTT partition, a partition shape of an MTT partition, a prediction mode, an intra prediction mode, a luma intra prediction mode, a chroma intra prediction mode, an intra partition information, an inter partition information, a coding block partition information, a prediction block partition information, a transform block partition information, a reference sample line index, a reference sample filtering method, a reference sample filter tap, a reference sample filter coefficient, a prediction block filter method, a prediction block filter tap, a prediction block filter coefficient, a prediction block boundary filtering method, a prediction block boundary filter tap, a prediction block boundary filter coefficient, an inter prediction mode, motion information, MV, a motion vector difference (MV). Difference (MVD), MVD resolution, MV size, MV representation accuracy, reference picture list, reference picture, reference picture index, inter prediction direction, inter prediction indicator, prediction list utilization flag, POC, MV candidate, MV candidate index, MV candidate list, AMVP mode usage information, merge candidate, merge index, merge candidate list, merge mode usage information, motion information compensation information, skip mode usage information, intra block copy mode usage information, BV (Block Vector), Block Vector Difference (BVD), BVD resolution, BV size, BV representation accuracy, BV candidate, BV candidate index, BV candidate list, filter tap of interpolation filter, filter coefficient of interpolation filter, transformation type, transformation size, transformation selection information, primary transformation usage information,Secondary transform usage information, primary transform selection information, secondary transform selection information, residual block presence information, coded block pattern, coded block flag, QP, delta QP, quantization matrix, deblocking filter usage information, coefficients of the deblocking filter, filter taps of the deblocking filter, strength of the deblocking filter, shape / shape of the deblocking filter, adaptive sample offset usage information, adaptive sample offset value, adaptive sample offset category, adaptive sample offset type, adaptive loop filter usage information, coefficients of the adaptive loop filter, filter taps of the adaptive loop filter, shape / shape of the adaptive loop filter, binarization / debinarization method, context model, context model determination method, context model update method, regular mode usage information, bypass mode usage information, significant coefficient flag, last significant coefficient flag, coefficient group unit coding flag, last significant coefficient position, flag indicating whether the coefficient value is greater than 1, whether the coefficient value is greater than 2 A flag indicating whether the coefficient value is greater than 3, a flag indicating whether the coefficient value is greater than 3, remaining coefficient value information, sign information, context bin, bypass bin, reconstructed sample, reconstructed luma sample, reconstructed chroma sample, residual sample, residual luma sample, residual chroma sample, transform coefficient, luma transform coefficient, chroma transform coefficient, transform coefficient level, luma transform coefficient level, chroma transform coefficient level, transform coefficient level scanning method, quantized level, luma quantized level, chroma quantized level, size of MV search region on the side of the decoding device, shape of MV search region on the side of the decoding device, number of MV search on the side of the decoding device, picture type, slice identification information, slice type, slice partitioning information, tile group identification information, tile group type, tile group partitioning information, tile identification information, tile type, tile partitioning information, bit depth,It may include one or more of input sample bit depth, reconstructed sample bit depth, residual sample bit depth, transform coefficient bit depth, quantized level bit depth, mapping availability information, information about luma signal, information about chroma signal, color space of target block, color space of residual block, and temporal layer information.

[0242] In addition, the coding parameter may further include 1) a value of information that may be included in the coding parameter, 2) a combination of multiple pieces of information that may be included in the coding parameter, 3) a statistical value for information that may be included in the coding parameter, 4) information related to the coding parameter, 5) information used to calculate / derive the coding parameter, and 6) information calculated / derived using the coding parameter.

[0243] In embodiments, "X usage information" may be "information indicating whether X is used / applied / performed." Alternatively, "X usage information" may be "information indicating whether X is available." For example, "specific mode usage information" may be information indicating whether a specific mode is used. The mode information may indicate a mode used for a target block among the modes described in the embodiments. In embodiments, the specific mode usage information may be replaced with mode information, and the description of the specific mode usage information may also be applied to the mode information. "X usage information" and "X indicator" may be used interchangeably.

[0244] In embodiments, coding parameters and syntax elements may correspond to each other. For example, syntax elements of an embodiment may be used as coding parameters, and coding parameters may be signaled as syntax elements.

[0245] In embodiments, “X presence information” may be considered as “information indicating whether X exists” or “information indicating whether information indicating X exists in the bitstream.”

[0246] In embodiments, the “X selection information” may be information indicating one of the candidates or methods for X. The “X selection information” may be considered an “X index.”

[0247] In embodiments, the splitting form of a particular tree may represent one of symmetric splitting and asymmetric splitting, and may represent one of QT, BT, TT, and non-split. The splitting direction of a particular tree may represent one of horizontal and vertical directions.

[0248] In embodiments, when a coding parameter has one of multiple values, "coding parameter" may be replaced with "whether the coding parameter has a specific value among the multiple values ​​available to the coding parameter."

[0249] In embodiments, when a coding parameter points to one of a plurality of objects, “coding parameter” may be replaced with “whether the coding parameter points to a specific object among the plurality of objects.”

[0250] In embodiments, the coding parameters may include at least one of a type of a target picture and a type of a target slice. The type of the target picture may be one of an I-picture, a B-picture, and a P-picture. The type of the target slice may be one of an I-slice, a B-slice, and a P-slice.

[0251] - If the target image to be encoded is an I-slice, the target image can be encoded using data within the image itself without inter-prediction that refers to other images. For example, an I-slice can be encoded only with intra-prediction.

[0252] - If the target image is a P slice, the target image can be encoded through inter prediction using only reference slices existing in one direction. Here, the one-way direction can be forward or backward.

[0253] - When the target image is a B slice, the target image can be encoded through inter prediction using reference slices existing in both directions or inter prediction using reference slices existing in one of the forward and backward directions. Here, the bidirectional directions can be forward and backward.

[0254] - P slices and B slices encoded and / or decoded using reference slices can be considered as images in which inter prediction is used.

[0255]

[0256] System for video coding

[0257] Figure 1 illustrates a system for video coding according to one embodiment.

[0258] The system (100) may include at least one of an encoding device (110) and a decoding device (150).

[0259] Each of the encoding device (110) and the decoding device (150) may be a computer or an electronic apparatus.

[0260]

[0261] Structure of the encoding device

[0262] The encoding device (110) may include a processor (120), storage (140), and a communicator (149).

[0263] The processor (120), storage (140), and communication device (149) can be connected via a bus.

[0264] The processor (120) may be a semiconductor device that executes instructions or computer-executable codes, such as a central processing unit (CPU). The processor (120) may be at least one hardware processor.

[0265] The processor (120) can perform generation and processing of information input to the encoding device (110), output from the encoding device (110), or used within the encoding device (110) in the embodiments, and can perform comparisons and judgments related to such information.

[0266] The processor (120) may include a plurality of components. The plurality of components may include a partitioner (122), a subtractor (124), a transformer (125), a quantizer (126), an inverse quantizer (127), an inverse transformer (128), an adder (129), a filter (130), and an entropy encoder (139).

[0267] At least some of the aforementioned components may be program modules. The program modules may be included in the encoding device (110) in the form of an operating system, applications, and other program modules. The program modules may be instructions or computer-executable codes stored in the storage (140) and executed by the processor (120).

[0268] The storage (140) may include various types of volatile storage media and non-volatile storage media. For example, the storage (140) may include memory such as ROM and RAM.

[0269] The storage (140) can store instructions and computer-executable codes used for the operation of the encoding device (110), and can store information and bitstreams described in the embodiments. The storage (140) can include a reference picture buffer (141).

[0270] The communication device (149) can perform functions related to the communication of information in the encoding device (110). For example, the communication device (149) can transmit a bitstream to the decoding device (150).

[0271] Among the names of components of the encoding device (110), “-er” or “-or” may be replaced with “-unit”. The storage (140) may also be named a storage unit.

[0272]

[0273] Operation of the encoding device

[0274] The encoding device (110) can sequentially encode one or more images of a video.

[0275] The storage (140) can store the original image. The original image can be used as a target image in the encoding device (110).

[0276] The processor (120) can generate a bitstream including encoded information by performing encoding on the target image, and can store the generated bitstream in the storage (140). The generated bitstream can be stored in a computer-readable recording medium, and can be transmitted to the communication device (189) of the decoding device (150) via a wired and / or wireless transmission medium by the communication device (149).

[0277] The segmenter (122) can determine a target block by performing segmentation on the target image.

[0278] The predictor (123) can determine the prediction mode of the target block. The predictor (123) can generate a prediction block of the target block by performing prediction according to the prediction mode.

[0279] The prediction mode of the target block may be one of the available prediction modes. For example, the available prediction modes may include intra prediction, inter prediction, and IBC prediction.

[0280] For example, if the prediction mode is intra prediction, the predictor (123) can perform intra prediction on the target block to generate a prediction block of the target block.

[0281] For example, if the prediction mode is inter prediction, the predictor (123) can perform inter prediction on the target block to generate a prediction block of the target block.

[0282] For example, when the prediction mode is IBC, the predictor (123) can perform IBC prediction on the target block to generate a prediction block of the target block.

[0283] The subtractor (124) can generate a residual block of the target block. The residual block may be the difference between the original block and the predicted block. The original block may be the area pointed to by the target block in the original image. Alternatively, the residual block may refer to a block generated by applying one or more of transformation and quantization to the difference between the original block and the predicted block.

[0284] The transformer (125) can perform a transformation on the residual block to generate transformation coefficients.

[0285] The converter (125) can perform the conversion using one of a plurality of conversion methods.

[0286] For example, the multiple transform methods may include a Discrete Cosine Transform (DCT), a Discrete Sine Transform (DST), a Karhunen-Loeve Transform (KLT), and transforms based on each transform.

[0287] Transform skip mode may be a mode for generating a reconstructed block using a reconstructed residual block and a prediction block for which transformation and inverse transformation have not been performed. When transform skip mode is applied to a target block, transformation and inverse transformation for the target block may be omitted, and only quantization and inverse quantization for the target block may be performed.

[0288] A quantizer (126) can generate quantized levels by applying quantization using quantization parameters to transform coefficients. In embodiments, the quantized levels may also be referred to as transform coefficients.

[0289] An entropy encoder (139) can generate encoded information by performing entropy encoding based on a probability distribution on information for decoding an image. The bitstream can include encoded information.

[0290] Information for decoding an image may include quantized levels and syntax elements produced by a quantizer (126).

[0291] The probability distribution can be determined based on the quantized levels and coding parameters.

[0292] The entropy encoder (139) can use scanning to change the quantized levels in the form of two-dimensional blocks into the form of one-dimensional vectors in order to perform encoding on the quantized levels. In scanning, which scan among the upper right diagonal scan, vertical scan, and horizontal scan will be used can be determined based on coding parameters such as the block size and the block intra prediction mode.

[0293] When encoding is performed on a target image / block, the predictor (123) uses a reference image / block for prediction. The encoded target image / block can be used as a reference image / block for other images / blocks to be processed later. Accordingly, the processor (120) can perform restoration on the encoded target block, and store a restored image including the restored target block generated by the restoration as a reference image in the reference picture buffer (141). Inverse quantization and inverse transformation can be performed on the encoded target block for restoration.

[0294] The dequantizer (127) can generate dequantized transform coefficients by performing dequantization on the quantized level.

[0295] The inverse transformer (128) can generate inverse quantized and inversely transformed coefficients by performing inverse transformation on the inverse quantized transform coefficients. In embodiments, the inverse quantized and / or inversely transformed coefficients may refer to coefficients to which at least one of inverse quantization and inverse transformation has been applied. The inverse quantized and inversely transformed coefficients may be a restored residual block.

[0296] The adder (129) can generate a restored block by combining a predicted block and a restored residual block.

[0297] The restoration block may pass through a filter (130). The filter (130) may apply one or more of a plurality of filters to the target. Each filter of the plurality of filters may be an in-loop filter. The target may be a restoration sample, a restoration block, or a restoration image.

[0298] The reference picture buffer (141) can store a restored block / image provided from the filter (130). The restored image may be an image including a restored block. Alternatively, the restored image may be an image composed of restored blocks.

[0299] The reference picture buffer (141) can provide the stored restored image as a reference image to the predictor (123). In terms of storing the decoded (i.e., restored) picture, the reference picture buffer (141) may also be referred to as a decoded picture buffer (DPB).

[0300]

[0301] Structure of the decryption device

[0302] The decryption device (150) may include a processor (160), a storage (180), and a communication device (189).

[0303] The description of the processor (120), storage (140), and communication device (149) related to the encoding device (110) can also be applied to the processor (160), storage (180), and communication device (189) related to the decoding device (150). Duplicate descriptions are omitted.

[0304] The processor (160) may include a plurality of components. The plurality of components may include an entropy decoder (161), a divider (162), a predictor (163), an inverse quantizer (167), an inverse transformer (168), an adder (169), and a filter (170).

[0305] The storage (180) may include a reference picture buffer (181).

[0306] The communication device (189) can perform functions related to communication of information in the decryption device (150). For example, the communication device (189) can receive a bitstream from the encoding device (110).

[0307] Among the names of components of the decryption device (150), “-er” or “-or” may be replaced with “-unit”. The storage (180) may also be named a storage unit.

[0308]

[0309] Operation of the decryption device

[0310] The communication device (149) of the encoding device (110) can transmit the bitstream generated by the encoding device (100) to the decoding device (150). Alternatively, a computer-readable recording medium storing the bitstream can transmit the bitstream generated by the encoding device (100) to the decoding device (150).

[0311] The communication device (189) can receive a bitstream from the encoding device (110) via a wired and / or wireless transmission medium. The received bitstream can be stored in the storage (180).

[0312] The processor (160) can obtain a bitstream from a storage (180) or a computer-readable recording medium.

[0313] A bitstream may contain encoded information.

[0314] An entropy decoder (161) can generate information for decoding an image by performing entropy decoding based on a probability distribution on the encoded information of a bitstream.

[0315] Information for decoding an image may include quantized levels and syntax elements.

[0316] The entropy decoder (161) can use scanning to change the quantized levels in the form of a one-dimensional vector into the form of a two-dimensional block to perform decoding on the quantized levels. In scanning, which scan among the upper right diagonal scan, vertical scan, and horizontal scan will be used can be determined based on coding parameters such as the block size and the block intra prediction mode.

[0317] The entropy decoder (161) can provide syntax elements to other components of the processor (160), such as the segmenter (162).

[0318]

[0319] A common description of the relationship between the components of the encoding device and the components of the decoding device.

[0320] The decoding device (150) performs decoding using the bitstream generated by the encoding device (110). The encoding device (110) can perform encoding on the target block using a restored image derived within the decoding device (150), rather than an original image that is not provided to the decoding device (150). Therefore, the encoding device (110) and the decoding device (150) may need to generate restored blocks / images in the same manner. In this respect, the descriptions of the divider (122), predictor (123), inverse quantizer (127), inverse transformer (128), adder (129), filter (130), and reference picture buffer (141) of the encoding device (110) disclosed in the embodiments can also be applied to the divider (162), predictor (163), inverse quantizer (167), inverse transformer (168), adder (169), filter (170), and reference picture buffer (181) of the decoding device (150). Duplicate descriptions are omitted.

[0321] Additionally, each of the divider (122), predictor (123), inverse quantizer (127), inverse transformer (128), adder (129), and filter (130) of the encoding device (110) can generate syntax element information that specifies processing for the target. Each of the divider (162), predictor (163), inverse quantizer (167), inverse transformer (168), adder (169), and filter (170) of the decoding device (150) can perform processing for the target (same as that performed in the encoding device (110)) using the syntax element information.

[0322] As described above, corresponding components of the encoding device (110) and the decoding device (150) may perform the same or corresponding functions. In embodiments, the processor may represent the processor (120) of the encoding device (110) and / or the processor (160) of the decoding device (150). For example, in the function related to prediction, the processor may represent a predictor (123), a subtractor (124), and an adder (129), and may represent a predictor (163) and an adder (169). In the function related to transformation, the processing unit may represent a transformer (125) and an inverse transformer (128), and may represent an inverse transformer (168). In the function related to quantization, the processor may represent a quantizer (126) and an inverse quantizer (127), and may represent an inverse quantizer (167). In the function related to entropy encoding / decoding, the processing unit may represent an entropy encoder (139) and / or an entropy decoder (161). In the function related to filtering, the processing unit may represent a filter (130) and / or a filter (170). The storage may represent a storage (140) of an encoding device (110) and / or a storage (180) of a decoding device (150). The reference picture buffer may represent a reference picture buffer (141) of an encoding device (110) and / or a reference picture buffer (181) of a decoding device (150). The communication unit may represent a communication unit (149) of an encoding device (110) and / or a communication unit (189) of a decoding device (150).

[0323]

[0324] Division of the units that make up the image

[0325] Figure 2 shows a segmentation structure of an image according to one embodiment.

[0326] Figure 2 can schematically represent an example in which one unit is divided into multiple sub-units.

[0327] A CU can be used as a basic unit for encoding and decoding images. In addition, a CU can be a basic unit for prediction, transformation, quantization, inverse quantization, inverse transform entropy encoding, and entropy decoding.

[0328] A CU can be used as a unit to which a prediction mode is applied. In other words, during coding, it can be determined which prediction mode among the available prediction modes will be applied to each CU. For example, available prediction modes may include intra prediction, inter prediction, and intra-block copy prediction (IBC).

[0329] The target image (200) can be sequentially divided into units of CTUs. A division structure can be determined for each CTU. The CTU can be divided into CUs according to the division structure. Alternatively, a single CTU can be used as a CU. The size of the CTU can be the maximum CU size.

[0330] Each CU can have depth information. The depth information can indicate the depth of the CU and the size of the CU. The depth of a CTU can be 0. The depth of a CU generated by splitting a CTU can be 1. When a parent CU is split into child CUs, the depth of the child CU can be 1 greater than the depth of the parent CU. The number of split CUs can be a positive integer greater than or equal to 2, including 2, 4, 8, and 16. At least one of the horizontal size and the vertical size of the child CU generated by splitting the parent CU can be smaller than at least one of the horizontal size and the vertical size of the parent CU, depending on the number of child CUs.

[0331] A partitioned CU can be recursively partitioned in the same manner up to a predefined maximum depth or a predefined minimum size. The depth of the smallest coding unit (SCU) can be the predefined maximum depth, and the size of the SCU can be the predefined minimum size. The size of the SCU can be the size of the minimum CU.

[0332] For example, the depth of a CU can range from 0 to 3. Depending on the depth of the CU, the CU can have a size from 64x64 to 8x8. A CTU with a depth of 0 can be a 64x64 block. 0 can be the minimum depth. An SCU with a depth of 3 can be an 8x8 block. 3 can be the maximum depth. A depth of 0 can represent a CTU that is a 64x64 block. A depth of 1 can represent a CU that is a 32x32 block. A depth of 2 can represent a CU that is a 16x16 block. A depth of 3 can represent an SCU that is an 8x8 block.

[0333] The partition information of a CU can indicate whether the CU is partitioned. The partition information can be a 1-bit flag. All CUs except SCUs can include partition information. For example, the partition information of a CU that is not further partitioned can be the first value, '0', and the partition information of a CU that is being partitioned can be the second value, '1'.

[0334] Quad Tree (QT) partitioning may mean that one CU is partitioned into four CUs. When a parent CU is partitioned into four child CUs, the width and height of each child CU may be half the width and half the height of the parent CU, respectively.

[0335] A binary tree (BT) split may mean that one CU is split into two CUs. For example, if a parent CU is split into two child CUs, the width or height of each child CU may be half the width or half the height of the parent CU.

[0336] A Ternary Tree (TT) partition may mean that a single CU is partitioned into three CUs. For example, if a parent CU is partitioned into three child CUs, the three child CUs can be created by partitioning the width or height of the parent CU in a ratio of 1:2:1. The width or height of the child CUs may be 1 / 4, 1 / 2, and 1 / 4 of the width or height of the parent CU, respectively.

[0337] In Fig. 2, QT type segmentation was applied to the first CTU. QT segmentation, BT segmentation, and TT segmentation were applied to the second CTU.

[0338] To partition a CTU, at least one of different types of partitions, such as QT partitioning, BT partitioning, and TT partitioning, may be applied to the CTU. Different types of partitions may be applied based on specific priorities.

[0339] For example, QT partitioning may be preferentially applied to a CTU. A CU to which QT partitioning can no longer be applied may correspond to a leaf node of QT. A CU that is a leaf node of QT may be a root node of BT and / or TT. A CU that is a leaf node of QT may be partitioned into a BT or TT form, or may not be partitioned any further. In this case, QT partitioning may not be applied again to a CU that is created by applying a BT or TT partition to a CU that is a leaf node of QT.

[0340] The partitioning of a CU corresponding to each node of QT can be signaled using QT partitioning information. The QT partitioning information can be a flag. The QT partitioning information of a unit can be information indicating whether the unit is partitioned in a QT form. A first value of the QT partitioning information, '0', can indicate that the CU is not partitioned in a QT form. The QT partitioning information having a first value can indicate a multi-type tree (MTT) partitioning. The MTT partitioning can include a BT partitioning and a TT partitioning. A second value of the QT partitioning information, '1', can indicate that the CU is partitioned in a QT form.

[0341] There may be no priority between BT and TT splits. That is, a CU corresponding to a leaf node of QT may be split into either BT or TT forms. Furthermore, a CU generated by BT or TT splits may be split again into BT or TT forms, or may not be split any further.

[0342] A CU corresponding to a leaf node of QT can become the root node of MTT. For each CU corresponding to an MTT node, the CU may further include MTT-type split direction information and split type information.

[0343] Split direction information can indicate the split direction of MTT splitting. The first value of the split direction information, '0', can indicate that the CU is split horizontally. The second value of the split direction information, '1', can indicate that the CU is split vertically.

[0344] The partition type information can indicate the partition type used for multi-type tree partitioning. The first value of the partition type information, '0', can indicate that the CU is partitioned in the TT form. The second value of the partition type information, '1', can indicate that the CU is partitioned in the BT form.

[0345] Here, each of the aforementioned split direction information and split type information may be a flag having a specific length (e.g., 1 bit).

[0346] The CU's partition information may also include QT partition information, partition direction information, and partition shape information.

[0347] CUs that are no longer split by QT splitting, BT splitting, and TT splitting can be used as units for specific processing, such as prediction, transformation, quantization, inverse quantization, inverse transform, entropy encoding, and entropy decoding. That is, for specific processing, CUs may no longer be split. Therefore, splitting information for splitting such CUs into PUs and / or TUs, etc., may not exist in the bitstream.

[0348] On the other hand, if the size of a CU is larger than the maximum TU size, the CU can be recursively split until the size of the CU becomes smaller than or equal to the maximum TU size. For example, if the size of a CU is 64x64 and the maximum TU size is 32x32, the CU can be split into four 32x32 TUs for transformation. For example, if the size of a CU is 32x64 and the maximum TU size is 32x32, the CU can be split into two 32x32 TUs for transformation.

[0349] In such cases, information regarding whether a CU is split for transformation may not be separately signaled. Whether a CU is split may be determined by comparing the size of the CU (width / height) with the maximum TU size (width / height), without signaling. For example, if the width of the CU is greater than the width of the maximum TU size, the CU may be split into two vertically. Additionally, if the height of the CU is greater than the height of the maximum TU size, the CU may be split into two horizontally.

[0350] For example, the minimum size of a CU may be 4x4. For example, the maximum size of a transform block may be 64x64. For example, the minimum size of a transform block may be 4x4. The QT minimum size may be the minimum size of a CU corresponding to a leaf node of the QT. The MTT maximum depth may be the maximum depth of the path from the root node to the leaf node of the MTT.

[0351] The BT maximum size may represent the maximum size of a CU corresponding to each node of the BT, and the TT maximum size may represent the maximum size of a CU corresponding to each node of the TT. The BT minimum size and / or the TT minimum size may be set to the minimum size of a CU.

[0352] If the depth within the MTT of a CU corresponding to a node of the MTT is equal to the maximum depth of the MTT, the CU may not be split into BT shape and / or TT shape.

[0353] Based on the various sizes and depths of the CUs described above, each piece of information described in the embodiments may or may not be present in the bitstream.

[0354] Information about the maximum or minimum size described in the embodiments may be signaled at a higher level of the CU. In the embodiments, the higher level of the CU may include a video level, a sequence level, a picture level, a subpicture level, a tile group level, a tile level, and a slice level.

[0355] The information described in the embodiments may be signaled separately for different types of slices. The different types of slices may include intra-slices and inter-slices.

[0356]

[0357] Processing blocks according to their properties

[0358] Whether a specific process described in the embodiments is applied / performed may be determined based on the properties of a block related to the specific process. Whether a specific process described in the embodiments is applied / performed may be determined based on whether the properties of a block related to the specific process satisfy a specific condition. For example, a block may include a target block, a neighboring block, and a reference block. A block may include other blocks described in the embodiments. A block may be one of the blocks and units described in the embodiments.

[0359] The blocks to which the specific processing described in the examples is applied may have a square shape or a non-square shape.

[0360] In one embodiment, the block's attributes may include the block's size. Certain processing described in the embodiments may be applied / performed when certain conditions regarding the block's size are met.

[0361] In one embodiment, the specific conditions may include a minimum block size condition and a maximum block size condition. The blocks to which the minimum block size condition applies and the blocks to which the maximum block size condition applies may be different.

[0362] In one embodiment, a minimum block size and / or a maximum block size for a particular process may be predefined.

[0363] In one embodiment, the processing of the embodiment may be applied / performed when the size of the block is greater than or equal to the minimum block size and / or when the size of the block is less than or equal to the maximum block size. Alternatively, in one embodiment, the processing of the embodiment may be applied / performed when the size of the block is greater than the minimum block size and / or when the size of the block is less than the maximum block size.

[0364] In one embodiment, the processing of the embodiment may be applied / performed only when the block size is greater than or equal to the minimum block size and less than or equal to the maximum block size. Alternatively, the processing of the embodiment may be applied / performed only when the block size is greater than or equal to the minimum block size and less than or equal to the maximum block size. Alternatively, the processing of the embodiment may be applied / performed only when the block size is greater than or equal to the minimum block size and less than or equal to the maximum block size. The processing of the embodiment may be applied / performed only when the block size is greater than or equal to the minimum block size and less than or equal to the maximum block size.

[0365] In one embodiment, the processing of the embodiment may be applied / performed only when the block size is a predefined block size.

[0366] In embodiments, the size of a block may be determined in various ways. For example, the size of a block may refer to the width or height of the block. The size of a block may refer to both the width and height of the block. The size of a block may refer to the area of ​​the block. The size of a block may refer to 1) a result value of a known formula using the width and height of the block, 2) a result value of a formula of the embodiment, or 3) a statistical value.

[0367] Additionally, for the first size, the processing of the first embodiment among the embodiments may be applied / performed, and for the second size, the processing of the second embodiment among the embodiments may be applied / performed.

[0368] In embodiments, the block size may be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128, etc. Alternatively, in embodiments, the block size may be (2*SIZE X )x(2*SIZE Y ) etc. SIZE X can be one of the integers greater than or equal to 1. SIZE Y can be one of the integers greater than or equal to 1.

[0369]

[0370] Predictive information for prediction

[0371] Prediction information can be used to generate a prediction block for the target block.

[0372] The encoding device (110) can generate prediction information required for prediction and can generate a bitstream including the prediction information. The prediction information can be signaled from the encoding device (110) to the decoding device (150) via the bitstream. The decoding device (150) can obtain the prediction information from the bitstream and perform prediction on the target block using the prediction information, thereby generating a prediction block.

[0373] Prediction information may include intra-prediction information, inter-prediction information, and IBC prediction information. In embodiments, prediction information may be replaced with intra-prediction information, inter-prediction information, and / or IBC information. Intra-prediction information may include information used for intra-prediction as described in embodiments. Inter-prediction information may include information used for inter-prediction as described in embodiments. IBC information may include information used for IBC prediction as described in embodiments.

[0374]

[0375] Intra prediction

[0376] Figure 3 illustrates the structure of intra prediction according to one embodiment.

[0377] Intra prediction can be performed using reference samples and coding parameters of the target block. The reference sample can be a (restored) sample within the (restored) reference block. Alternatively, an intermediate prediction sample can be generated using a sample described in the embodiment, such as a reconstructed sample, and a reference sample can be generated again using the intermediate prediction sample. Processing described in the embodiment, such as filtering, can be applied when generating the reference sample.

[0378] A reference block may be a (spatial) neighboring block of the target block. The coding parameters may be coding parameters for the target block and / or coding parameters for the reference block. In intra prediction, a reference sample may mean a neighboring sample.

[0379] A prediction block can be generated by performing intra prediction on a target block according to an intra prediction mode based on reference samples within a target image and information related to the reference samples. The size of the target block and the size of the prediction block can be the same.

[0380] In embodiments, the prediction block may be a PU. Alternatively, the prediction block may correspond to a CU or TU described in the embodiments. The prediction block may have a square or rectangular shape.

[0381] An intra prediction mode can be expressed by at least one of a mode number, a mode value, a mode angle, and a mode direction. The prediction directions of a plurality of intra prediction modes for a target block are illustrated in the lower right corner of Fig. 3. Among the plurality of intra prediction modes, the remaining intra prediction modes excluding the DC and planar modes may be directional modes. A directional mode may be an intra prediction mode having a specific direction or a specific angle. The intra prediction mode for the target block may be selected from among directional modes and non-directional modes.

[0382] In the lower right rectangle representing the target block, the number '0' may represent the planar mode, which is a non-directional intra prediction mode. The number '1' may represent the DC mode, which is a non-directional intra prediction mode. In the lower right rectangle representing the target block, the arrows from the center to the periphery of the rectangle may represent the prediction directions of the directional intra prediction modes. In addition, the number indicated close to the arrow may represent an example of the mode value assigned to the intra prediction mode or the prediction direction of the intra prediction mode.

[0383] Intra prediction can be performed based on an intra prediction mode for the target block. One of the available intra prediction modes for the target block can be used as the intra prediction mode for the target block.

[0384] The number of intra prediction modes available to a target block may be a predefined value. Alternatively, the number of intra prediction modes available to a target block may be determined based on the properties of the prediction block. For example, the properties of the prediction block may include coding parameters such as shape, size, and color components.

[0385] For example, in Figure 3, the directional modes depicted by the dotted lines (i.e., the directional modes numbered between -14 and -1, or between 67 and 80) can only be applied to predictions for non-square blocks. Therefore, the number of intra prediction modes available for predictions for square blocks can be 67 (planar mode, DC mode, and 65 directional modes).

[0386] For example, the number of available intra prediction modes may vary depending on whether the color component of the block is a luma signal or a chroma signal. The number of available intra prediction modes for a block containing a luma component may be greater than the number of available intra prediction modes for a block containing a chroma component.

[0387] Intra prediction modes may include horizontal-below mode, horizontal mode, vertical mode, and vertical-right mode. The horizontal-below mode may be an intra prediction mode located below the horizontal mode. The vertical-right mode may be a mode located to the right of the vertical mode. For example, in FIG. 3, the mode value of the horizontal mode may be 18. The mode value of the vertical mode may be 50. Intra prediction modes whose mode values ​​are one of 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, and 66 may be vertical-right modes. Intra prediction modes whose mode value is one of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, and 17 can be horizontal bottom modes.

[0388] The number of intra prediction modes and the mode number of each intra prediction mode described above may be merely exemplary. The number of intra prediction modes and the mode number of each intra prediction mode described above may be defined differently depending on the embodiment, implementation, and / or needs.

[0389] When the intra prediction mode is the planar mode, when generating a prediction block of a target block, a sample value of the prediction sample can be generated using a weighted sum (weighted sum) of an upper reference sample of the target sample, a left reference sample of the target sample, an upper right reference sample of the target block, and a lower left reference sample of the target block, depending on the position of the prediction sample within the prediction block.

[0390] When the intra prediction mode is DC mode, a prediction block may be generated based on an average of sample values ​​of a plurality of reference samples. The plurality of reference samples may include upper reference samples and left reference samples of the target block. The value of the prediction sample of the prediction block may be determined based on an average of the sample values ​​of the plurality of reference samples. In addition, filtering using the values ​​of the reference samples may be performed for specific rows and / or specific columns within the target block. The specific rows may be one or more upper rows adjacent to the upper reference samples. The specific columns may be one or more left columns adjacent to the left reference samples.

[0391] When the intra prediction mode is a directional mode, a prediction block can be generated using the top reference sample, the left reference sample, the top right reference sample, and / or the bottom left reference sample of the target block.

[0392] The intra prediction mode of the target block may be determined based on the intra prediction mode of a neighboring block of the target block. Information for determining the intra prediction mode of the target block may be signaled.

[0393] For example, if the intra prediction modes of the target block and the neighboring block are the same, an indicator indicating that the intra prediction modes of the target block and the neighboring block are the same can be signaled.

[0394] For example, an indicator may be signaled that indicates an intra prediction mode that is the same as the intra prediction mode of the target block among the intra prediction modes of multiple neighboring blocks.

[0395] For example, if the intra prediction modes of the target block and neighboring blocks are different, an indicator indicating the intra prediction mode of the target block may be signaled. Alternatively, information used to derive the intra prediction mode of the target block based on the intra prediction mode of the neighboring block may be signaled.

[0396] Reference samples used for intra prediction for the target block may include lower left reference samples, left reference samples, upper left reference samples, upper reference samples, and upper right reference samples.

[0397] For example, the left reference samples may be reconstructed reference samples adjacent to the left side of the target block. The top reference samples may be reconstructed reference samples adjacent to the top side of the target block. The top left reference sample may be a reconstructed reference sample diagonally adjacent to the top left side of the target block. The bottom left reference samples may be reference samples located below the left reference samples among samples located on the same line as the left sample line composed of the left reference samples. The top right reference samples may be reference samples located on the right side of the top reference samples among samples located on the same line as the top sample line composed of the top reference samples.

[0398] Reference samples used for intra prediction for a target block can be determined based on the intra prediction mode of the target block. One or more reference samples can be used to determine the sample value of a prediction sample of a prediction block. In FIG. 3, the direction of the intra prediction mode indicated by the arrow can represent the direction from the prediction sample to the reference sample. The direction of the intra prediction mode can represent the dependency relationship between the reference samples and the prediction samples. For example, depending on the intra prediction mode, the sample value of a specific reference sample can be used as the sample value of at least one sample of the prediction block. Here, the specific reference sample and the at least one sample of the prediction block can be samples designated by a straight line in the direction of the intra prediction mode. In other words, the sample value of the specific reference sample can be copied as the sample value of the prediction sample located in the reverse direction of the direction of the intra prediction mode. Alternatively, the sample value of the prediction sample of the prediction block can be the sample value of the reference sample located in the direction of the intra prediction mode based on the position of the prediction sample.

[0399] The reference samples used for intra prediction may not be limited to samples immediately adjacent to the target block. As illustrated in FIG. 3, at least one of reference sample lines 0 to 3 may be used for intra prediction of the target block.

[0400] Each reference sample line of FIG. 3 may include one or more reference samples. A smaller number of a reference sample line may be a line of reference samples closer to the target block. Reference sample line 0 may be a line of reference samples immediately adjacent to the target block. When the upper left coordinates of the target block are (X, Y), the horizontal length is W, and the vertical length is H, the reference samples of reference sample line 0 may be samples whose x-coordinate is X-1 or whose y-coordinate is Y-1. Here, the y-coordinates of the reference samples whose x-coordinate is X-1 may be Y-1 to Y+2H. The x-coordinates of the reference samples whose y-coordinate is Y-1 may be X-1 to X+2W. The reference samples of reference sample line A may be samples whose x-coordinate is XA-1 or whose y-coordinate is YA-1. Here, the y-coordinates of the reference samples whose x-coordinate is XA-1 may be YA-1 to Y+2H+A. The x-coordinates of reference samples whose y-coordinate is YA-1 can be XA-1 to X+2W+A. A can be 1, 2, or 3.

[0401] Instead of obtaining samples from the reconstructed neighboring blocks, samples of segment A and segment F can be derived using padding using the nearest samples from segment B and segment E, respectively.

[0402] A reference sample line index may indicate a reference sample line among multiple reference sample lines used for intra prediction of a target block. For example, the reference sample line index may have a value of one of 0 to 3. The reference sample line index may be signaled.

[0403] When inter-color component intra prediction is used for a target block, a prediction block of a second color component can be generated based on a reconstructed block of a first color component for the target block. For example, the first color component can be a luma component, and the second color component can be a chroma component.

[0404] For intra prediction between color components, parameters between the first color component and the second color component can be derived based on a template. For example, the parameters can be parameters of a linear model.

[0405] For example, the template may include a top reference sample and / or a left reference sample of the target block, and may include a top reference sample and / or a left reference sample of the restoration block of the first color component corresponding to these reference samples.

[0406] Once the parameters are derived, a prediction block of the second color component for the target block can be generated by applying the reconstructed block of the first color component to a linear model. Depending on the image format or the type of intra prediction between color components, subsampling / downsampling can be performed on the surrounding samples of the reconstructed block of the first color component and the reconstructed block of the first color component. When subsampling is performed, the derivation of the parameters and intra prediction between color components can be performed using the corresponding samples derived by the subsampling.

[0407] Intra Sub-Partitions (ISP) prediction may refer to sequential intra prediction for multiple sub-blocks generated by dividing a target block. In ISP prediction, a target block may be divided into two or four sub-blocks in the horizontal and / or vertical directions. The divided sub-blocks may be sequentially reconstructed. As intra prediction is performed on a sub-block, a sub-prediction block for the sub-block may be generated. Additionally, as inverse quantization and / or inverse transformation is performed on the sub-block, a sub-residual block for the sub-block may be generated. A reconstructed sub-block may be generated by adding the sub-prediction block to the sub-residual block. The reconstructed sub-block may be used as a reference sample for intra predictions for other sub-blocks to be processed subsequently.

[0408] In performing prediction on a target block, it can be determined whether samples included in a reconstructed neighboring block can be used as reference samples of the target block. If there is an unavailable sample among the samples of the neighboring block that cannot be used as a reference sample of the target block, a value generated by copying and / or interpolating using the sample value of at least one sample among the samples included in the reconstructed neighboring block can replace the sample value of the unavailable sample. If the value generated by copying and / or interpolating is replaced with the sample value of the sample, the sample can be used as a reference sample of the target block.

[0409] In intra prediction, the sample value of a prediction sample of a prediction block can be determined by the sample value of a reference sample. The position of the reference sample can be specified by the position of the prediction sample and the direction of the intra prediction mode. If the position specified by the position of the prediction sample and the direction of the intra prediction mode is an integer position, the sample value of one reference sample pointed to by the integer position can be used to determine the sample value of the prediction sample of the prediction block. If the position specified by the position of the prediction sample and the direction of the intra prediction mode is not an integer position, an interpolated reference sample can be generated based on two reference samples closest to the specified position. The sample value of the interpolated reference sample can be used to determine the sample value of the prediction sample. That is, when the position specified by the position of the prediction sample and the direction of the intra prediction mode represents a space between two reference samples, an interpolated sample value can be generated based on the sample values ​​of the two samples.

[0410]

[0411] Inter prediction

[0412] Figure 4 shows the structure of inter prediction to explain the inter prediction process according to one embodiment.

[0413] The rectangle illustrated in Fig. 4 can represent an image. Additionally, the arrow in Fig. 4 can represent a prediction direction.

[0414] Each picture composing a video can be classified into an I picture (i.e., an intra picture), a P picture (i.e., a uni-prediction picture), and a B picture (i.e., a bi-prediction picture) according to its coding type. Coding can be performed for each picture according to its coding type.

[0415] If the target picture is an I-picture, coding for the target picture can be performed using information within the target picture without inter prediction referring to other images. For example, coding for the I-picture can be performed using intra prediction and / or IBC prediction.

[0416] Coding for P pictures and B pictures can be performed by at least one of intra prediction, IBC prediction, and inter prediction using a reference picture.

[0417] If the target picture is a P picture, coding for the target picture can be performed using unidirectional inter prediction using one reference picture list.

[0418] When the target picture is a B picture, coding for the target picture can be performed using unidirectional inter prediction or bidirectional inter prediction using two reference picture lists.

[0419] Below, inter prediction for a target block in inter mode according to an embodiment is specifically described.

[0420] When the prediction mode of the target block is inter mode, inter prediction can be performed on the target block. The target block can be a prediction block or a split prediction block.

[0421] Inter prediction can be performed using reference images and motion information. In inter prediction, a reference image can be selected using a reference image index, and a reference block corresponding to a target block within the reference image can be determined using motion information. A prediction block for the target block can be generated using the determined reference block.

[0422] Motion information can be derived using coding parameters, etc. For example, motion information can be derived using motion information of a reconstructed neighboring block, motion information of a call block, and / or motion information of a block adjacent to a call block.

[0423] In embodiments, a candidate list may be used for inter prediction. The candidate list may include multiple candidates. An index indicating a candidate used for inter prediction for a target block among the candidates in the candidate list may be signaled. The candidate list may be derived in the same manner based on the same information in the encoding device (110) and the decoding device (150). Here, the same information may include a restored image and a restored block. Furthermore, in order to specify a candidate by index, the order of the candidates within the candidate list may need to be consistent.

[0424] In one embodiment, prediction of a target block can be performed by using motion information of a spatial candidate or a temporal candidate as motion information of the target block. The motion information of the spatial candidate may be referred to as spatial motion information. The motion information of the temporal candidate may be referred to as temporal motion information.

[0425] A spatial candidate may be a restored spatial neighboring block that is spatially adjacent to the target block.

[0426] A spatial candidate may be a block that 1) exists within the target image, 2) has already been restored through decryption, and 3) is adjacent to the target block.

[0427] Spatial candidates may include the left block, the top block, the bottom left block, the top right block, and the top left block of the target block.

[0428] A temporal candidate may be a restored temporal neighboring block corresponding to a target block in a restored COL image.

[0429] In embodiments, the motion information of a spatial candidate may be motion information of a block containing the spatial candidate. The motion information of a temporal candidate may be motion information of a block containing the temporal candidate.

[0430] In inter prediction, a call (COL) block for a target block can be identified. The area of ​​the target block within the target image and the area of ​​the call block within the call image may be identical. In other words, a call block may be a block occupying a specific area within the call image. The specific area may correspond to the area of ​​the target block within the call image.

[0431] A temporal candidate may be a location inside and / or outside a call block within a call image.

[0432] For example, a call block may include a first call block and a second call block. When the upper left coordinates of a call block are (xP, yP) and the size of the call block is (nPSW, nPSH), the first call block may be a block occupying coordinates (xP + nPSW, yP + nPSH). The second call block may be a block occupying coordinates (xP + (nPSW >> 1), yP + (nPSH >> 1)). The second call block may be optionally used as a call block when the first call block is unavailable.

[0433] The MV of the target block can be determined based on the MV of the call block. Scaling can be performed on the MV of the call block. The scaled MV of the call block can be used as the MV of the target block or as the predicted MV. Alternatively, the MV of the temporal candidate stored in the candidate list associated with inter prediction can be a scaled MV.

[0434] The ratio of the scaled MV and the MV of the call block may be equal to the ratio of the first temporal distance and the second temporal distance. The first temporal distance may be the distance between the reference image and the target image of the target block. The second temporal distance may be the distance between the reference image and the call image of the call block.

[0435] The method by which motion information is derived can be determined by the inter prediction mode of the target block. For example, as the inter prediction mode, AMVP mode, merge mode, skip mode, merge mode with MVD, subblock merge mode, GPM, Combined Inter Intra Prediction (CIIP) mode, and affine inter mode can be used. In the embodiments below, each of the inter prediction modes is described.

[0436]

[0437] AMVP mode

[0438] When the AMVP mode is used as a prediction mode, an MV candidate list including one or more MV candidates can be generated using the MV of the spatial candidate, the MV of the temporal candidate, the history-based MV candidate, and the zero vector. At least one of the MV of the spatial candidate, the MV of the temporal candidate, and the zero vector can be determined and used as an MV candidate.

[0439] A spatial candidate may include a reconstructed spatial neighboring block. The MV of the reconstructed spatial neighboring block may be referred to as a spatial MV candidate (spatial motion vector candidate). A temporal candidate may include a called block and a block adjacent to the called block. The MV of the called block or the MV of a block adjacent to the called block may be referred to as a temporal MV candidate (temporal motion vector candidate). A history-based MV candidate may be an MV in a list including MVs of other blocks that were previously encoded / decoded prior to encoding / decoding of the target block.

[0440] The encoding device (110) can use the MV candidate list to determine an MV to be used for encoding the target block within the search range. The maximum number of MV candidates in the MV candidate list can be predefined. N can represent a predefined maximum number. For example, N can be 2. Alternatively, the maximum number of candidates can be signaled from the encoding device to the decoding device or derived from the decoding device. The encoding device (110) can determine an MV candidate to be used as a prediction MV of the target block among the MV candidates in the MV candidate list. The MV to be used for encoding the target block can be an MV that can be encoded at the minimum cost. The encoding device (110) can determine whether to use the AMVP mode in encoding the target block, and can generate AMVP mode usage information indicating whether the AMVP mode is used.

[0441] Inter prediction information may include 1) AMVP mode usage information, 2) MV candidate index, 3) MVD, 4) MVD resolution information, 5) reference direction, and 6) reference image index, and may include a residual block. The inter prediction information may be signaled from the encoding device (110) to the decoding device (150) in the form of a bitstream.

[0442] The decoding device (150) can obtain AMVP mode usage information from the bitstream. If the AMVP mode usage information indicates that the AMVP mode is used, the decoding device (150) can obtain an MV candidate index, an MVD, MVD resolution information, a reference direction, and a reference image index from the bitstream. Among the MV candidates included in the MV candidate list, an MV candidate indicated by the MV candidate index can be selected as the prediction MV of the target block.

[0443] The MVD may represent the difference between the MV to be actually used for inter prediction of the target block and the predicted MV. The encoding device (110) may derive a predicted MV that is close to the MV to be actually used for inter prediction of the target block in order to use an MVD with the smallest possible size. The decoding device (150) may derive the MV of the target block by combining the MVD and the predicted MV. In other words, the MV of the target block derived by the decoding device (150) may be the sum of the MVD and the predicted MV candidate.

[0444] Additionally, the encoding device (110) can generate MVD resolution information. The MVD resolution information may be information used to adjust the resolution of the MVD. The decoding device (150) can adjust the resolution of the MVD using the MVD resolution information.

[0445] Meanwhile, the encoding device (110) can calculate the MVD based on the affine model. The affine control point MV of the target block can be derived based on the sum of the affine control point MV candidate and the MVD. The MV of each subblock within the target block can be derived using the affine control point MV.

[0446]

[0447] Merge mode

[0448] When merge mode is used, a merge candidate list including multiple merge candidates can be generated using motion information of spatial candidates and motion information of temporal candidates. The motion information can include 1) MV, 2) reference image index, and 3) reference direction. The merge candidate can be motion information.

[0449] Merge candidates may include 1) a spatial merge candidate generated based on a spatial candidate, 2) a temporal merge candidate generated based on a temporal candidate, 3) a history-based merge candidate, 4) an average merge candidate, and 5) a zero merge candidate.

[0450] A history-based merge candidate may be motion information within a list that includes motion information of other blocks that were previously encoded / decoded prior to encoding / decoding of the target block.

[0451] An average merge candidate may be a merge candidate generated based on the average of two merge candidates in the merge candidate list.

[0452] A zero merge candidate may be zero vector motion information. Zero vector motion information may be motion information whose MV is a zero vector.

[0453] Merge candidates can be added to the merge candidate list according to a predefined method and a predefined order so that the merge candidate list has a set number of merge candidates. The same merge candidate list can be constructed in the encoding device (110) and the decoding device (150) through the predefined method and the predefined order.

[0454] The encoding device (110) can select a merge candidate to be used for encoding a target block from among the merge candidates in the merge candidate list. The encoding device (110) can determine whether to use a merge mode in encoding the target block, and can generate merge mode usage information indicating whether the merge mode is used.

[0455] Inter prediction information may include 1) merge mode usage information, 2) merge index, and 3) correction information, and may include a residual block. Inter prediction information may be signaled from an encoding device (110) in bitstream form to a decoding device (150) in bitstream form.

[0456] The decoding device (150) can obtain merge mode usage information from the bitstream. If the merge mode usage information indicates that the merge mode is being used, the decoding device (150) can obtain information related to the merge mode, such as a merge index, from the bitstream.

[0457] The encoding device (110) can select an optimal merge candidate from among the merge candidates included in the merge candidate list, and set the value of the merge index to point to the selected merge candidate.

[0458] Correction information may be information used for correcting an MV. The encoding device (110) may generate the correction information. The decoding device (150) may perform correction on the MV of the merge candidate selected by the merge index based on the correction information, thereby deriving a corrected MV. The corrected MV may be used as the MV of the target block.

[0459] In one embodiment, the correction information may include an MVD. The correction information may include one or more of correction usage information, correction direction information, and correction size information. The correction usage information may indicate whether correction is used for the MV. A merge mode that performs correction for the MV based on the correction information may be referred to as a merge mode with an MVD.

[0460] In merge mode, prediction for a target block can be performed using a merge candidate pointed to by a merge index among the merge candidates included in the merge candidate list.

[0461] Motion information of the target block can be derived from 1) MV, 2) reference image index, and 3) reference direction of the merge candidate pointed to by the merge index.

[0462] In one embodiment, the merge candidates in the merge candidate list may be specific modes that derive inter-prediction information. The merge candidate may be information indicating a specific mode that derives inter-prediction information. Inter-prediction information of the target block may be derived according to the specific mode indicated by the merge candidate. From this perspective, a specific mode may be considered a specific inter-prediction information derivation mode or a specific motion information derivation mode. A specific mode may include a series of processes that derive inter-prediction information.

[0463] Inter prediction information of a target block can be derived based on a specific mode indicated by a merge candidate selected by a merge index among the merge candidates in the merge candidate list. For example, the specific modes may include a subblock-level motion information derivation mode and an affine motion information derivation mode, and may include other modes for deriving motion information described in the embodiments.

[0464] Skip mode may be a mode that does not use residual blocks. That is, when skip mode is used, the reconstructed block may be identical to the predicted block. The description of merge mode in the embodiments may also apply to skip mode. The difference between merge mode and skip mode may be whether or not residual blocks are signaled and used. That is, skip mode may be similar to merge mode except that residual blocks are not transmitted / used, and the description of merge mode may also apply to skip mode.

[0465] The subblock merge mode may be a mode in which motion information of a target subblock is derived for a target subblock within a target block. When the subblock merge mode is applied, a list of subblock merge candidates may be generated using affine control point motion vector merge candidates and / or subblock-based temporal merge candidates. The subblock-based temporal merge candidates may be motion information of a call subblock of the target subblock.

[0466] In GPM, a first prediction block and a second prediction block can be generated using two pieces of motion information for a target block. For each coordinate of the target block, a final prediction sample of a final prediction block can be generated using a weighted sum of the first prediction sample of the first prediction block and the second prediction sample of the second prediction block.

[0467] Here, the first weight for the weighted consensus first prediction sample and the second weight for the weighted consensus second prediction sample can be determined based on the boundary of the GPM. The boundary can represent a dividing line that divides the target block. Based on the boundary, the target block can be divided into a first divided region and a second divided region.

[0468] If the distance between the final prediction sample and the boundary is less than or equal to a reference value, the value of the final prediction sample of the final prediction block may be determined using a weighted sum of the first prediction sample of the first prediction block and the second prediction sample of the second prediction block. If the distance between the final prediction sample and the boundary is greater than the reference value, one of the first weight and the second weight may be 1, and the other may be 0.

[0469] Combined Inter-Intra Prediction (CIIP) mode may be a mode that derives a prediction sample of a target block using a weighted sum of prediction samples generated by inter prediction and prediction samples generated by intra prediction.

[0470] In the aforementioned modes, self-improvement of the derived motion information can be performed, and the improved motion information can be used as motion information for the target block. For example, blocks within a specific region determined based on the derived motion information can be searched, and the motion information of the block with the smallest sum of absolute differences (SAD) value among the searched blocks can be used as the improved motion information for the target block. The specific region can be a square region within a reference image specified by the motion information. The point indicated by the motion information can be the center of the specific region.

[0471] In the aforementioned modes, compensation for prediction samples derived through inter prediction can be performed using optical flow.

[0472]

[0473] Figure 5 shows the order in which spatial candidates are added to the candidate list according to one embodiment.

[0474] In Fig. 5, the locations of spatial candidates are shown.

[0475] The large block in the center can represent the target block. The five smaller blocks adjacent to the target block can represent spatial candidates.

[0476] The coordinates of the target block can be (xP, yP), and the size of the target block can be (nPSW, nPSH).

[0477] A spatial candidate A0 may be a block adjacent to the lower left of the target block. A0 may be a block that occupies samples at coordinates (xP - 1, yP + nPSH).

[0478] A spatial candidate A1 may be a block adjacent to the left of the target block. A1 may be the bottommost block among the blocks adjacent to the left of the target block. Alternatively, A1 may be a block adjacent to the top of A0. A1 may be a block that occupies a sample at coordinates (xP - 1, yP + nPSH - 1).

[0479] A spatial candidate B0 may be a block adjacent to the upper right of the target block. B0 may be a block that occupies a sample at coordinates (xP + nPSW, yP - 1).

[0480] A spatial candidate B1 may be a block adjacent to the top of the target block. B1 may be the rightmost block among the blocks adjacent to the top of the target block. Alternatively, B1 may be a block adjacent to the left of B0. B1 may be a block that occupies a sample at coordinates (xP + nPSW - 1, yP - 1).

[0481] A spatial candidate B2 may be a block adjacent to the upper left of the target block. B2 may be a block that occupies a sample at coordinates (xP - 1, yP - 1).

[0482] As shown in Figure 5, when adding spatial candidates to the candidate list, B1, A1, The order of B0, A0 and B2 can be used, i.e. B1, A1, Available spatial candidates can be added to the candidate list in the order of B0, A0, and B2. The order in which the spatial candidates are added to the merge candidate list illustrated in Fig. 5 may be merely an example.

[0483] The above candidate list may include a motion information candidate list, a merge candidate list, an MV candidate list, a BV candidate list, and an MPM list.

[0484] To include a spatial or temporal candidate in the candidate list, its availability can be determined. If the candidate block is outside the boundaries of an image, slice, or tile, the candidate block's availability can be set to false. The phrase "availability is set to false" can mean "it is set to non-availability."

[0485] The maximum number of candidates in a candidate list can be set. N can represent the set maximum number. The set maximum number can be signaled through a parameter set or header, etc. For example, the maximum number of candidates in the candidate list for a target block within a slice can be set by the slice header. For example, the default value of N can be 5.

[0486]

[0487] IBC mode

[0488] IBC mode may be an intra-block copy prediction mode that generates prediction blocks for target blocks by referencing already-restored regions within the target image. In this respect, IBC mode may also be referred to as a current image reference mode. A block vector (BV) may be used to specify the already-restored region.

[0489] Whether the target block is encoded / decoded in IBC mode can be determined using IBC mode usage information. The encoding device (110) can determine whether to use IBC mode in encoding the target block and can generate IBC mode usage information indicating whether IBC mode is used. The decoding device (150) can obtain IBC mode usage information from the bitstream.

[0490] In IBC mode, a prediction block of a target block can be generated based on a block vector (BV). The BV can specify a reference block. The BV can indicate displacement between the target block and the reference block. The reference block can be a block within the target image. The description of the MV in the embodiments can also be applied to the BV.

[0491] The IBC mode may include skip mode, merge mode, and AMVP mode. The description of the AMVP mode, merge mode, and skip mode of the embodiments may also be similarly applied to the AMVP mode, merge mode, and skip mode of the IBC mode.

[0492] In skip mode or merge mode, a merge candidate list can be constructed, and a merge index can specify one merge candidate among the merge candidates in the merge candidate list. The BV of the specified merge candidate can be used as the BV of the target block.

[0493] In AMVP mode, BVD can be used. The description of MVD in the embodiments can also be applied to BVD.

[0494] The reference block in IBC mode may be limited to a block within an already reconstructed region of the target image. Alternatively, the reference block may be contained within at least one of the target CTU or the left CTUs. For example, the value of BV may be limited so that the reference block is located within a specific region. The specific region may be an area of ​​three blocks of a specific size that are encoded / decoded before the block of a specific size that contains the target block. The specific size may be 64x64.

[0495]

[0496] Transformation and quantization

[0497] A quantized level can be generated by performing transformation and / or quantization on a residual block. The residual block can represent the difference between the original block and the predicted block. A reconstructed residual block can be generated by performing inverse quantization and / or inverse transformation on the quantized level. The reconstructed residual block can represent the difference between the reconstructed block and the predicted block.

[0498] When a transformation or inverse transformation is performed, a separable transformation or a 2-dimensional (2D) non-separable transformation can be performed on the residual block. A separable transformation can be a transformation that performs 1-dimensional (1D) transformations on the residual block in each of the horizontal and vertical directions.

[0499] The transform kernels used for the transformation may include various DCT kernels such as DCT type 2 (DCT-II), 1) DST kernels, and 3) kernels induced by training. For 1D transform, the DCT type and DST type may include DCT-V, DCT-VIII, DST-I, and DST-VII in addition to DCT-II.

[0500] A set of transforms may be used to determine the DCT type, DST type, or learning-induced kernel to be used for the transformation. Each transform set may include multiple transform candidates. Each transform candidate may be a DCT type, a DST type, or a learning-induced kernel.

[0501] The encoding device (110) can perform transformation and inverse transformation using transformation candidates included in the transformation set. The decoding device (150) can perform inverse transformation using transformation candidates included in the transformation set. Transform selection information indicating which transformation candidate among a plurality of transformation candidates included in the transformation set applied to the residual block is used can be signaled. The transformation selection information can include vertical transformation selection information and horizontal transformation selection information. The vertical transformation selection information can indicate which transformation among the transformations included in the transformation set is used for vertical transformation. The horizontal transformation selection information can indicate which transformation among the transformations included in the transformation set is used for horizontal transformation.

[0502] The transform may include at least one of a primary transform and a secondary transform. A primary transform coefficient may be generated by performing a primary transform on a residual block, and a secondary transform coefficient may be generated by performing a secondary transform on the transform coefficient. Here, the transform coefficient may include a primary transform coefficient and a secondary transform coefficient.

[0503] The primary transformation may mean Multiple Transform Selection (MTS), which applies different transformations for each of the 1D directions (i.e., vertical and horizontal directions).

[0504] A secondary transform may be a transform for improving the energy concentration of the transform coefficients generated by the primary transform. The secondary transform may be 1) a separable transform like the primary transform, or 2) a 2D non-separable transform. The 2D non-separable transform may refer to a low frequency non-separable transform (LFNST) or a non-separable primary transform (NSPT).

[0505] NSPT can be applied to specific block sizes such as 4x4, 4x8, 8x4, 4x16, 16x4, 8x8, 8x16, and 16x8 for intra coding.

[0506] The primary transform can be performed using at least one of a plurality of predefined transform methods. For example, the plurality of predefined transform methods can include DCT, DST, and KLT. In addition, the primary transform can be a transform having various transform types according to a transform kernel function defining DCT and DST. For example, the primary transform can include a plurality of transforms such as DCT-2, DCT-4, DCT-5, DCT-7, DCT-8, DST-1, DST-2, DST-4, DST-7, and DST-8 according to a plurality of transform kernels.

[0507] In one embodiment, the transform type may be determined based on coding parameters associated with the target block. For example, the transform type may be determined based on one or more of: 1) a prediction mode of the target block (e.g., one of intra prediction and inter prediction), 2) a size of the target block, 3) a shape of the target block, 4) an intra prediction mode of the target block, 5) a component of the target block (e.g., one of a luma component and a chroma component), and 6) a split type applied to the target block (e.g., one of QT, BT, TT, and non-split).

[0508] As with the first-order transformation, a set of transformations can also be defined for the second-order transformation. The methods for deriving and / or determining the set of transformations of the embodiments can be applied to both the first-order transformation and the second-order transformation.

[0509] In one embodiment, a primary transformation and / or a secondary transformation may be determined for a specific target. The transformation selection information may include transformation target information. The transformation target information may indicate the target to which the primary transformation and / or the secondary transformation is applied.

[0510] For example, a first-order transform and / or a second-order transform may be applied to one or more of the signal components, including the luma component and the chroma component.

[0511] In one embodiment, the transform selection information may include primary transform usage information and secondary transform usage information. The primary transform usage information may indicate whether the primary transform is applied to the residual block of the target block. The secondary transform usage information may indicate whether the secondary transform is applied to the residual block of the target block.

[0512] In one embodiment, whether a primary transform and / or a secondary transform is applied may be determined based on coding parameters for the target / neighboring blocks, such as the size and shape of the target / neighboring blocks.

[0513] In one embodiment, the transform selection information may include primary transform selection information and secondary transform selection information. The primary transform selection information may indicate a transform method to be applied to a residual block among a plurality of transform methods that may be used in the primary transform. The primary transform selection information may be a primary transform index. The secondary transform selection information may indicate a transform method to be applied to a transform coefficient among a plurality of transform methods that may be used in the secondary transform. The secondary transform selection information may be a secondary transform index.

[0514] In one embodiment, the transformation methods of the first and second transformations may each be derived based on specific information such as coding parameters. For example, the coding parameters may include coding parameters for target / neighboring blocks.

[0515] In embodiments, information related to transformation, such as transformation selection information, and sub-information of the transformation selection information may be signaled for a specific target. For example, the specific target may be a CU.

[0516] Information related to transformation, such as transformation selection information, and sub-information of transformation selection information can be derived for a specific target. For example, the specific target may be a CU.

[0517] Quantized levels can be generated by performing quantization on the result or residual block generated by performing the first transform and / or the second transform.

[0518] The description of the transformation described above can also be applied to the inverse transformation. In such an application, the reverse processing of the processing described for the transformation can be performed in the inverse transformation. The term "transformation" in the name related to the transformation can be changed to "inverse transformation." Furthermore, the input of the transformation can be considered the output of the inverse transformation. The output of the transformation can be considered the input of the inverse transformation. The decoding device (150) can obtain information related to the transformation, such as transformation selection information, and can perform the reverse processing of the processing related to the transformation indicated by the information related to the transformation using the information related to the transformation.

[0519] A target block may include multiple subblocks. Each subblock may be defined according to a minimum block size or a minimum block shape. The target block may be divided into multiple subblocks, and each subblock may include coefficients of sizes such as 4x4, 2x8, and 8x2. The target block may be a transform block. Transform coefficients or quantized levels may be expressed in the form of a block. The transform coefficients may be quantized transform coefficients.

[0520] The transform coefficients or quantized levels can be scanned according to at least one of the scanning types, such as diagonal scanning, vertical scanning, and horizontal scanning. The diagonal scanning can be right-upper diagonal scanning or left-lower diagonal scanning.

[0521] For example, coefficients can be transformed or arranged into a one-dimensional vector by scanning the coefficients of a block using diagonal scanning. Vertical scanning can scan coefficients in the form of two-dimensional blocks in the column direction. Horizontal scanning can scan coefficients in the form of two-dimensional blocks in the row direction.

[0522] The scanning type for coefficients can be determined based on coding parameters such as the intra prediction mode, block size, and block shape. For example, whether diagonal scanning, vertical scanning, or horizontal scanning is used can be determined based on coding parameters such as the intra prediction mode, block size, and block shape. A block can be a transform unit.

[0523] Scanning for each scanning type can start at a specific starting point and end at a specific ending point.

[0524] In scanning, a scanning order based on the scanning type may first be applied between subblocks. Next, a scanning order based on the scanning type may be applied to transform coefficients or quantized levels within the subblock.

[0525] The encoding device (110) can perform entropy encoding on transform coefficients or quantized levels to generate a bitstream including entropy-encoded transform coefficients or entropy-encoded quantized levels.

[0526] The decoding device (150) can obtain entropy-encoded transform coefficients or entropy-encoded quantized levels from a bitstream and perform entropy decoding to generate transform coefficients or quantized levels. The coefficients can be arranged in the form of two-dimensional blocks through inverse scanning. The arrangement of the inverse scanning can be a rearrangement opposite to the arrangement of the scanning.

[0527] Inverse scanning of coefficients can generate inversely scanned transform coefficients or inversely scanned quantized levels. At this time, the inverse scanning types of the inverse scanning can include diagonal scanning, vertical scanning, and horizontal scanning, and the inverse scanning type of the inverse transformation corresponding to the scanning type of the transformation can be selected.

[0528] In the decoding device (150), inverse quantization can be performed on (inversely scanned) coefficients. Depending on whether a second inverse transform is performed, a second inverse transform can be performed on the result generated by performing inverse quantization. In addition, depending on whether a first inverse transform is performed, a first inverse transform can be performed on the result generated by performing the second inverse transform. A restored residual block can be generated by selectively performing the second inverse transform and the first inverse transform on the coefficients.

[0529]

[0530] Filtering

[0531] To improve the image quality, filtering may be performed on blocks. The values ​​of target samples may be determined or updated through filtering.

[0532] The target sample may be one of the samples described in the embodiments. For example, the target sample may be one or more of the samples described in the embodiments, such as a prediction sample, a reference sample, a residual sample, a restored sample, and a restored sample with filtering applied.

[0533] The target sample may be a sample within one or more of a target picture, a target slice, a target CTB, a target block, a reference sample line, and a template. The target block may be one of the blocks described in the embodiments. For example, the target block may be one or more of the blocks described in the embodiments, such as a transform block, a prediction block, a reference block, a residual block, and a reconstruction block.

[0534] In embodiments, the filtering process described as being applied to one object may also be applied to other objects. For example, the filtering process described in a specific in-loop filtering may also be applied to transform blocks, prediction blocks, reference blocks, and residual blocks.

[0535] A specific type of filtering may be used for the filters of the embodiments. The type of filtering may include filter taps (or filter tap lengths), filter shapes, filter strengths, filter coefficients (or weights), and offsets.

[0536] The filter tab may indicate the number of input samples used for the filter. The input samples may include the target sample. Alternatively, the input samples may include a specific value determined for the target sample. The input samples may include one or more reference samples. The one or more reference samples may be determined based on an attribute of the target block described in the embodiments. The attribute may include a coding parameter. For example, an attribute of the target sample may include a position of the target sample. One or more reference samples may be specified based on a relative position with respect to the position of the target sample.

[0537] A filter shape can represent the shape formed by input samples. A specific value determined for a target sample can be considered a target sample. In other words, if a specific value determined for a target sample is used as an input sample of a filter, the target sample can also be considered to form a filter shape.

[0538] The number of samples whose values ​​are determined by filtering may be multiple. The filter strength may indicate the range of samples whose values ​​are determined by filtering. The filter strength may be either a strong filtering strength or a weak filtering strength. The number of samples whose values ​​are determined by a strong filtering strength may be greater than the number of samples whose values ​​are determined by a weak filtering strength. Alternatively, the filter strength may indicate the range of values ​​that are changed by filtering. The range of sample values ​​that are changed by a strong filtering strength may be wider than the range of sample values ​​that are changed by a weak filtering strength.

[0539] The filter coefficients can be coefficients or weights of the input samples.

[0540] An offset can be a specific value that is added to the result calculated using the values ​​and coefficients of the input samples, such as a weighted sum.

[0541] Filtering, interpolation, and sampling may have in common that they update the values ​​of samples. Therefore, the description of any one of filtering, interpolation, and sampling in the embodiments may also apply to any other of filtering, interpolation, and sampling. Here, sampling may include at least one of upsampling, downsampling, and subsampling.

[0542] Filtering may include filtering performed by predictor (123) and predictor (163), etc.

[0543] In encoding a target block, a prediction error may exist between the original samples of the original block and the prediction samples of the prediction block. To reduce the prediction error, filtering may be performed on at least one of the prediction samples of the prediction block and the reference samples referenced for prediction.

[0544] For example, in intra prediction, the reference sample may include one or more of the upper left reference sample, the upper reference sample, the upper right reference sample, the left reference sample, and the lower left reference sample. Filtering on the predicted sample may be performed by applying specific weights to the predicted sample, the left reference sample, the upper reference sample, and / or the upper left reference sample, respectively.

[0545] Filtering of at least one of the prediction sample and the reference sample may be performed based on the attributes of the target block and the attributes of the prediction sample. For example, whether filtering is performed, the type of filter, the area to which the filtering is applied, the filtering weights, the reference sample, the range of the reference sample, and the location of the reference sample may each be determined based on the attributes of the target block and the attributes of the prediction sample.

[0546] For example, the properties of the target block may include information related to the target block described in the embodiments, such as 1) size of the target block, 2) prediction mode, 3) intra prediction mode, 4) reference sample line, 5) sample value, and 6) coding parameter.

[0547] For example, the attributes of a prediction sample may include information related to the prediction sample described in the embodiments, such as 1) a sample value of the prediction sample and 2) a location within a target block, and may include coding parameters related to the prediction sample.

[0548] Filtering may include in-loop filtering performed by filter (130) and filter (170), etc.

[0549]

[0550] Figure 6 illustrates multiple in-loop filters according to an example.

[0551] The plurality of in-loop filters of the in-loop filtering may include one or more of Luma Mapping with Chroma Scaling (LMCS), a deblocking filter, a Sample Adaptive Offset (SAO), and an Adaptive Loop Filter (ALF).

[0552] Multiple in-loop filters can be connected sequentially. For example, the multiple in-loop filters can be connected in the order of LMCS, deblocking filter, SAO, and ALF. Furthermore, the multiple in-loop filters can be connected in any order among all available permutations of the multiple in-loop filters. The output from one of the multiple in-loop filters can be used as the input to the next filter.

[0553] As illustrated in FIG. 6, an input image may be input to the first filter. The input image may be a block described in the embodiments. For example, the input image may be a reconstructed block generated by an adder (129) or an adder (169). The output from one filter may be input to the next filter. An output image may be generated by the last filter. The output image may be a filtered block described in the embodiments. For example, the output image may be a filtered reconstructed image generated by a filter (130) or a filter (170).

[0554] The target block can represent an image input to the filter. The filtered target block can represent an image output from the filter.

[0555] LMCS may include luma signal mapping to a luma signal of a target block and chroma signal scaling to a chroma signal of the target block.

[0556] Luma signal mapping can perform codeword redistribution for the luma signal.

[0557] Luma signal mapping can include forward mapping and reverse mapping. In forward mapping, the existing dynamic range can be divided into multiple intervals. The mapped dynamic range can be determined by performing codeword redistribution on the input image using a linear model for each interval. In reverse mapping, reverse mapping is performed from the mapped dynamic range to the existing dynamic range.

[0558] Chroma scaling can correct chroma signals based on the correlation between a luma signal and a corresponding chroma signal.

[0559] Forward mapping can be performed between inter prediction for a luma signal and reconstruction for the luma signal, and between inter prediction for the luma signal and chroma scaling. Backward mapping can be performed between reconstruction for the luma signal and in-loop filtering for the luma signal. Chroma scaling can be performed between inverse transformation and reconstruction for the chroma signal.

[0560] According to this structure, inverse quantizations for luma and chroma signals, inverse transformations for luma and chroma signals, prediction for luma signals, and restoration for luma signals can be performed within the mapped dynamic range. In-loop filterings for luma and chroma signals, inter predictions for luma and chroma signals, intra prediction for chroma signals, and restoration for chroma signals can be performed within the existing dynamic range.

[0561] A deblocking filter can remove block distortion occurring at boundaries between blocks within a restored image. For example, the blocks may be transform blocks. Furthermore, the blocks may be subblocks of a specific block described in the embodiments. Here, the boundaries between blocks may refer to samples adjacent to the boundaries between blocks.

[0562] Deblocking filters can be applied to vertical and horizontal boundaries between blocks. After filtering the vertical boundaries of blocks, filtering can be performed again on the horizontal boundaries of the filtered blocks.

[0563] A deblocking filter may be applied selectively. Whether to apply a deblocking filter to a target block may be determined based on at least one of the sample(s) contained within a specific number of columns or rows within the target block and the sample(s) contained within a specific number of columns or rows within a neighboring block adjacent to a specific boundary.

[0564] When a deblocking filter is applied to a target block, the filter to be applied may be determined based on the strength of the required deblocking filtering. In other words, among multiple other filters, a filter determined based on the strength of the deblocking filtering may be applied to the target block. The multiple filters may include one of a long-tap filter, a strong filter, a weak filter, and a Gaussian filter.

[0565] The maximum length of the deblocking filter can be determined based on the properties of the target block, such as the size of the target block, the components of the target block, and the coding parameters.

[0566] SAO can compensate for distortion between the original and reconstructed images on a sample-by-sample basis. To compensate, SAO can apply an appropriate offset to the sample values ​​of each sample. That is, the offset can be added to the sample values.

[0567] An offset can be determined for the target block. For example, an offset can be determined for each component of the CTB. The determined offset can be applied to samples within a specific component of the CTB.

[0568] SAO may include SAO using Edge Offset (EO) and SAO using Band Offset (BO). Depending on the characteristics of samples within a specific block, such as a CTU, whether SAO using EO or SAO using BO may be performed may be determined.

[0569] In SAO using EO, distortion correction of samples can be performed based on the direction of the edge within the target block. Pattern classes of EO can include horizontal patterns, vertical patterns, 135 degree diagonal patterns, and 45 degree diagonal patterns. For a target block, information indicating a pattern class applied to the target block and multiple offsets of the pattern class can be signaled. There can be four offsets. For a target sample within the target block, adjacent samples of the target sample can be determined based on the direction of the pattern class. An offset to be applied to the target sample can be determined based on the pattern of the adjacent samples.

[0570] In an offset using BO, distortion of a sample can be corrected by classifying the brightness values ​​of samples within a target block into specific bands. The bit depth of an input image can be divided into m sections. For example, m can be 32. The specific bands can be n consecutive sections among the m sections. For example, n can be 4. N offsets for the n sections can be signaled. Additionally, information indicating a first section selected as one of the n sections among the m sections can be signaled. The offset of the section to which the target sample corresponds can be added to the sample value of the target sample of the target unit.

[0571] ALF can compensate for distortion between the restored image and the original image.

[0572] The filter coefficients of ALF can be signaled via the bitstream.

[0573] The filter shape of ALF can be determined by the components of the target block. For example, a 7x7 diamond-shaped filter can be used for the luma component. A 5x5 diamond-shaped filter can be used for the chroma component.

[0574] In ALF, the characteristics of a specific block can be determined for a specific block, and the class of the specific block can be determined based on the characteristics. In other words, the determination of characteristics and class of ALF can be performed in units of 4x4 blocks. Filter coefficients can be calculated based on the class. A specific block can be a 4x4 block.

[0575] One of 25 classes can be determined as the class of a specific block based on the direction and activity determined using the gradient of the specific block. Rotation, vertical symmetry, and / or diagonal symmetry transformations can be applied to the filter based on the gradient of the specific block.

[0576] Information regarding whether ALF applies can be signaled for specific units, such as CTB.

[0577] An index indicating a filter to be applied to a specific unit among available filters may be signaled. Here, the available filters may include fixed filters and filters configured using a parameter set. For example, the parameter set may be an adaptive parameter set (APS). The fixed filters may be identically predefined in the encoding device (110) and the decoding device (150). The filter coefficients of the filters configured using the parameter set may be determined based on coding parameters.

[0578]

[0579] Entropy encoding and entropy decoding

[0580] Figure 7 illustrates entropy encoding and entropy decoding according to an example.

[0581] The processes of entropy encoding by the entropy encoder (139) are illustrated at the top of Fig. 7.

[0582] The entropy encoder (139) may include a context modeler, a binarization unit, and an entropy encoder. The context modeler may include a context selection unit and a context memory.

[0583] The binarization unit can generate bins for syntactic elements by performing binarization on the syntactic elements of the target block. Binarization may be a process of converting syntactic elements into the form of bins.

[0584] Information about syntactic elements and bins can be provided from the binarization unit to the context selection unit.

[0585] A context modeler can perform context updates.

[0586] Context can mean occurrence probability information for each bin for syntactic elements that have already been encoded.

[0587] The context modeler can update the context to apply current probability information to the entropy encoding of the bins of the syntactic elements of the target block. The updated context can be stored in the context memory. At this time, the updated context corresponding to the syntactic elements of the target block (or bins within the syntactic elements of the target block) can be derived by the context modeler.

[0588] The context selector can select a context corresponding to a bin of a syntactic element of a target block. The selected context can be loaded from the context memory and used as an updated context for entropy encoding of the bins of the syntactic element of the target block.

[0589] The updated context can be used for entropy encoding of syntactic elements of the target block.

[0590] The entropy encoding unit can generate encoded information about syntactic elements of a target block by performing entropy encoding using the generated bins and the updated context, and can generate a bitstream including the encoded information. The entropy encoding unit can use at least one of an arithmetic encoding method and a bypass encoding method.

[0591] The processes of entropy decryption by the entropy decoder (161) are shown at the bottom of Fig. 7.

[0592] The entropy decoder (161) may include a context modeler, an entropy decoder, and an inverse binarizer. The context modeler may include a context selection unit and a context memory.

[0593] A context modeler can perform context updates.

[0594] Context can mean the occurrence probability information of each bin for syntactic elements that have already been decoded.

[0595] The context modeler can update the context to apply the currently decoded probability information to entropy decoding for the bins of the syntactic elements of the target block. The updated context can be stored in the context memory. At this time, the updated context corresponding to the syntactic elements of the target block (or the bins within the syntactic elements of the target block) can be derived by the context modeler.

[0596] The context selector can select a context corresponding to a blank of a syntactic element of a target block. The selected context can be loaded from the context memory and used as an updated context for entropy decoding of the syntactic element of the target block.

[0597] The updated context can be used for entropy decoding of syntactic elements of the target block.

[0598] The entropy decoding unit can generate bins for the delimiting elements of the target block by performing entropy decoding on the encoded information of the bitstream based on the updated context. The entropy decoding unit can use at least one of an arithmetic decoding method and a bypass decoding method.

[0599] The debinarization unit can obtain a syntactic element of the target block by performing debinarization on at least one of the generated bins. The debinarization may be a process of converting at least one of the bins into a form of a syntactic element.

[0600] Information about syntactic elements and bins can be provided from the de-binarization unit to the context selection unit.

[0601] A syntax element may be one of the coding parameters described in the embodiments.

[0602]

[0603] Methods for binarization, debinarization, entropy encoding, and entropy decoding

[0604] In embodiments, one or more of the binarization methods, inverse binarization methods, entropy encoding methods and entropy decoding methods listed below may be used to perform signaling for specific information.

[0605] - Signed 0-th order Exponential Golomb binarization / debinarization method (abbreviated as se(v))

[0606] - k-order exponential-Golomb binarization / inverse binarization method with sign (abbreviated as sek(v))

[0607] - 0-order exponent-Golomb binarization / inverse binarization method for unsigned positive integers (abbreviated as ue(v))

[0608] - k-order exponential-Golomb binarization / inverse binarization method for unsigned positive integers (abbreviated as uek(v))

[0609] - Fixed-length binarization / debinarization method (abbreviated as f(n))

[0610] - Truncated Rice binarization / debinarization method or truncated unary binarization / debinarization method (abbreviated as tu(v))

[0611] - Truncated binary binarization / debinarization method (abbreviated as tb(v))

[0612] - Context-adaptive arithmetic encoding / decoding method (abbreviated as ae(v))

[0613] - bit string in bytes (abbreviated as b(8))

[0614] - Signed integer binarization / debinarization method (abbreviated as i(n))

[0615] - Unsigned positive integer binarization / debinarization method (abbreviated as u(n)) ('u(n)' can also mean fixed-length binarization / debinarization method.)

[0616] - Unary binarization / inverse binarization method

[0617]

[0618] Adaptive execution of the processes of the embodiments

[0619] The processing of the embodiments can be performed in the same and / or corresponding manner in the encoding device (110) and the decoding device (150). In addition, a combination of one or more of the above embodiments can be used in encoding and / or decoding of an image.

[0620] The order in which the embodiments are applied may be different in the encoding device (110) and the decoding device (150). Alternatively, the order in which the embodiments are applied may be (at least partially) the same in the encoding device (110) and the decoding device (150).

[0621] The processing of the embodiments may be performed for each specific object. The processing of the embodiments may be performed identically for specific objects. For example, a specific object may include a luma signal and a chroma signal.

[0622] The processing of the embodiments can be selectively applied / performed based on specific conditions or specific targets.

[0623] In one embodiment, the processing of the embodiment may be selectively applied / performed according to a temporal layer. Temporal layer information for a specific processing may be information indicating a temporal layer to which the processing may be applied / performed. Temporal layer information may be signaled for a specific processing. The temporal layer information may indicate the lowest layer and / or highest layer to which the specific processing may be applied, and may indicate a specific layer to which the specific processing is applied / performed. Alternatively, a fixed temporal layer to which the processing of the embodiment is applied / performed may be defined.

[0624] In one embodiment, a type to which processing of the embodiments is applied / performed may be defined, and whether processing of the embodiment is applied / performed may be determined based on the defined type. The type may include a picture type, a slice type, a tile group type, etc.

[0625] According to the description of the embodiments, when applying / performing a specific process to a specific object, a specific condition may be required, and the specific process may be processed under a specific decision. If it is determined whether a specific condition is met based on a specific coding parameter, or a specific decision is made based on a specific coding parameter, it can be interpreted that such a specific coding parameter can be replaced with another coding parameter. In other words, the coding parameters that affect a specific condition or a specific decision described in the embodiments can be considered merely exemplary, and in addition to the specified coding parameter, one or more other coding parameters, or a combination of one or more other coding parameters, can be understood to perform the role of the specified coding parameter.

[0626] The processing of the embodiments may be applied / performed based on the size of at least one of the blocks described in the embodiments. For example, the blocks may include a coding block, a prediction block, a transform block, a reference block, a current block, and a target block. Alternatively, the blocks may include adjacent blocks of the embodiments. Here, the size may be defined as a minimum size and / or a maximum size for the processing of the embodiments, or may be defined as a fixed size for the processing of the embodiments. In addition, for the processing of the embodiments, the first embodiment may be applied to a first size, and the second embodiment may be applied to a second size. In other words, the processing of the embodiments may be applied in a complex manner depending on the size. In addition, the processing of the embodiments may be applied only when the size of the block is greater than or equal to the minimum size and less than or equal to the maximum size. In other words, the processing of the embodiments may be applied only when the size of the block is within a specific range.

[0627] The aforementioned encoding and decoding devices can predict a target block using bi-prediction. Bi-prediction may refer to a method of generating multiple reference blocks, i.e., intermediate prediction blocks, from at least one reference picture, and using the two intermediate prediction blocks to generate a final prediction block for the target block. The final prediction block may be performed by averaging or weighting the intermediate prediction blocks.

[0628] In the disclosure below, the terms 'intermediate prediction block' and 'reference block' may be used interchangeably.

[0629] In the bidirectional prediction of the present disclosure, pictures reconstructed before the target picture including the target block can be used as reference pictures. The reference pictures can be pictures that precede or follow the target picture in output order, i.e., display order.

[0630] Additionally or alternatively, the target picture may be used as a reference picture. In this case, a region within the target picture that was reconstructed before the target block may be used to predict the target block.

[0631] In the bidirectional prediction of the present disclosure, two or more reference pictures may be used. The two or more reference pictures may be different or the same.

[0632] In some embodiments, one or more new reference pictures may be used, which are generated by processing at least one reference picture into various forms.

[0633] A new reference picture can be generated based on a neural network. For example, a new reference picture can be generated by inputting at least one reference picture that has been encoded or decoded prior to the target picture into a neural network.

[0634] Reference pictures to be input into a neural network can be determined based on the picture order count (POC). That is, reference pictures to be input into a neural network can be selected from a reference picture list based on the difference in POC between the reference pictures in the reference picture list and the target picture.

[0635] For example, multiple reference pictures can be selected in the order of smallest POC difference from the target picture among the reference pictures in the reference picture list.

[0636] As another example, the reference picture with the smallest POC difference from the target picture may be selected from each of reference picture list 1 and reference picture list 2.

[0637] As input data for a neural network for generating a new reference picture, a reference picture and a quantization parameter map associated with the reference picture may be used. The quantization parameter map may be a two-dimensional array of slice-wise quantization parameters of the reference picture or a two-dimensional array of block-wise quantization parameters.

[0638] To generate new reference pictures, optical flow computed between reference pictures can be additionally applied. After applying optical flow to feature maps for the reference pictures and synthesizing them, the synthesized feature maps can be input into a residual network consisting of at least one residual block to generate new reference pictures.

[0639] When computing optical flow between reference pictures, the reference pictures can be downsampled to reduce computational complexity. The optical flow can then be computed using the downsampled reference pictures and then upsampled.

[0640] Hereinafter, for convenience of explanation, bi-prediction using two intermediate prediction blocks is described, but the present invention is not limited thereto, and three or more intermediate prediction blocks may be used for bi-prediction.

[0641] In a bi-prediction for a target block, one final prediction block can be generated using two intermediate prediction blocks.

[0642] The quantity-prediction method may include average-based quantity-prediction, weighted-average-based quantity-prediction, optical-flow-based quantity-prediction, and neural network-based quantity-prediction.

[0643] Average-based quantity prediction is a method of predicting a target block by calculating the average of two intermediate prediction blocks.

[0644] Weighted average-based prediction is a method of predicting a target block by taking a weighted average of two intermediate prediction blocks.

[0645] Equation 1 shows an example of generating a prediction block in a weighted average-based quantity-prediction.

[0646] [Formula 1]

[0647] P bi-pred = ((8 - w) * P0+ w * P1+ 4) >> 3

[0648] P0 may be a first intermediate prediction block. P1 may be a second intermediate prediction block.

[0649] w may be weight information indicating weights for intermediate prediction blocks. Alternatively, w may represent a weight mode that uses a specific weight within a weight set. In other words, a weighted average prediction mode may be defined based on w.

[0650] Inter prediction information may include weight information. Weights for intermediate prediction blocks may be determined by w.

[0651] w can be set for a specific unit. For example, the specific unit can be a unit described above in the embodiment. For example, the specific unit can be a coding unit (CU). Alternatively, the specific unit can be a CTU, a slice, a tile, or a picture.

[0652] Information about w may be encoded or decoded in the above-described specific unit. Information about w may be an index indicating one of the available weights. This index may be referred to as a weighted average prediction mode index.

[0653] In weighted average-based quantity prediction, one of the available weights can be selected as the weight. The available weights can be, for example, {-2, 3, 4, 5, 10}.

[0654] If the target picture is a low-delay picture, for example, all five available weights can be candidates for exploration. For example, one of {-2, 3, 4, 5, 10} can be selected as a weight.

[0655] If the target picture is not a low-latency picture, for example, three of the five available weights can be candidates for exploration. For example, one of {3, 4, 5} can be selected as a weight.

[0656] A weighted average prediction mode in which the weights applied to two intermediate prediction blocks are equal can be referred to as an equal weighted average prediction mode. For example, in Equation 1, w=4 indicates an equal weighted average prediction mode. The equal weighted average prediction mode can refer to the above-mentioned average-based positive prediction.

[0657] The first and second intermediate prediction blocks may be generated through inter-prediction, i.e., motion compensation. However, the scope of the present invention is not limited thereto, and at least one of the first intermediate prediction block or the second intermediate prediction block may be generated in IBC mode or intra-prediction.

[0658] The quantity prediction of the present disclosure may include optical-flow based quantity prediction.

[0659] Optical flow-based volume prediction is a method of calculating motion offset (i.e., optical flow) on a pixel-by-pixel or sub-block-by-subblock basis through a gradient operation on samples within two intermediate prediction blocks generated through motion compensation, and estimating pixel changes within the intermediate prediction blocks based on the optical flow to correct sample values. In some embodiments, optical flow-based volume prediction may be constrained to be executed only when mean-based volume prediction is applied, i.e., when the weights applied to the two intermediate prediction blocks are the same.

[0660] The quantity-prediction of the present disclosure may include neural network-based quantity-prediction.

[0661] Neural network-based quantile prediction may refer to a prediction method that constructs input data including two intermediate prediction blocks or reference blocks, and inputs the input data into a neural network to generate a prediction block for a target block.

[0662] Neural network-based quantity prediction may be performed based on at least one of the attributes and / or coding parameters of the target block. Alternatively, whether neural network-based quantity prediction is performed may be determined based on the attributes and / or coding parameters of the target block.

[0663] Here, the attribute of the target block may be at least one of the target block's size or its components. That is, whether to perform neural network-based quantity prediction may be determined based on the target block's size. Alternatively, whether to perform neural network-based quantity prediction may be determined based on whether the target block is a luma component block or a chroma component block.

[0664] The coding parameters may be weights set for the target block. For example, whether neural network-based quantity prediction is performed may be determined based on the weights set for quantity prediction of the target block.

[0665] In embodiments, the method of applying the neural network technique in the positive prediction may include 1) a first method of applying the neural network technique depending on a weighted average prediction mode index, 2) a second method of applying the neural network technique depending on an equally weighted average prediction mode index, 3) a third method of applying the neural network technique based on an independent coding structure, and 4) a fourth method of adaptively applying the neural network technique based on an independent coding structure.

[0666] The first method applies neural network-based quantity prediction based on whether the weights to be used for quantity prediction of the target block belong to a predefined set of multiple weights. For example, the predefined set of multiple weights may be {-2, 3, 4, 5, 10} as described above. Alternatively, the set of multiple weights may be a subset of {-2, 3, 4, 5, 10}.

[0667] The second method is to apply neural network-based quantity prediction based on whether the weights to be applied to the target block are uniformly weighted.

[0668] Additionally, the first and second methods may be applied based on the block size.

[0669] The third method is a method for determining whether to apply neural network-based quantity prediction without relying on the weights of the target block.

[0670] In some embodiments, neural network-based quantity prediction can be applied without relying on the weights of the target block, in which case the weighted average prediction modes can be replaced with the neural network-based quantity prediction.

[0671] In some other embodiments, whether to apply neural network-based quantization prediction may be determined based on whether the target block size falls within a predefined set of block sizes. If the target block size falls within the predefined set of block sizes, neural network-based quantization prediction may be applied to the target block, and existing weighted average prediction modes may be replaced with the neural network-based quantization prediction. If the target block size does not fall within the predefined set of block sizes, quantization prediction may be performed according to the existing weighted average prediction mode.

[0672] In the fourth method, a single prediction method can be adaptively selected through competition between neural network-based prediction and weighted average prediction modes. In the fourth method, existing prediction methods (e.g., average-based prediction, weighted average-based prediction, optical flow-based prediction) and neural network-based prediction can compete with each other under certain conditions. The certain conditions can be predefined in the encoding and decoding devices, or can be signaled from the encoding device to the decoding device.

[0673] In some embodiments, prediction blocks may be generated according to neural network-based positive prediction and weighted average prediction modes, and encoding costs (SAT, SSE, SATD, rate-distortion cost) may be compared to determine an optimal positive prediction method. The determined optimal positive prediction method may then be signaled from the encoder to the decoder.

[0674] In another embodiment, the quantile prediction method can be determined using surrounding quantile / decoding information. For example, if at least one adjacent block is determined using a neural network-based quantile prediction method, or if the number of adjacent blocks determined using a neural network-based quantile prediction method is greater than a certain number or greater than the number of adjacent blocks determined using a weight-based quantile prediction method, the quantile prediction method of the target block can be determined using a neural network-based method.

[0675] As another embodiment, when the target block is encoded in merge mode, the neural network-based bi-prediction mode can be performed on the target block if the candidate surrounding blocks for the merge mode are encoded in the neural network-based bi-prediction mode.

[0676] An encoding device or a decoding device may select a quantity prediction method to be applied to a target block from among a plurality of quantity prediction methods (average-based quantity prediction, weighted average-based quantity prediction, optical flow-based quantity prediction, and neural network-based quantity prediction) according to any one of the first to fourth methods of applying a neural network technique described above. The target block may be predicted according to the selected quantity prediction method.

[0677] The encoding device or decoding device may select one of the methods including at least two of the first to fourth methods.

[0678] FIG. 8 is an exemplary diagram illustrating a method for selecting one of the first to fourth methods for applying a neural network technique according to one embodiment of the present disclosure.

[0679] The encoding device can determine whether to apply a neural network technique to a target block based on a weighted average prediction mode index (S810). If it is determined to apply a neural network technique based on a weighted average prediction mode index, a quantity-prediction method to be applied to the target block is selected from among multiple quantity-prediction methods according to the first method (S815).

[0680] If the neural network technique is not applied depending on the weighted average prediction mode index, the encoding device determines whether to apply the neural network technique depending on the equal weighted average prediction index (S820). If the neural network technique is applied depending on the equal weighted average prediction index, a quantity-prediction method to be applied to the target block is selected from among multiple quantity-prediction methods according to the second method (S825).

[0681] If the neural network technique is not applied based on the equal weighted average prediction index, the encoding device determines whether to apply a neural network technique based on an independent encoding structure (S830). If the neural network technique is applied based on an independent encoding structure, a quantity-prediction method to be applied to the target block is selected from among multiple quantity-prediction methods according to the third method (S835).

[0682] When not applying a neural network technique based on an independent encoding structure, the encoding device selects a quantity-prediction method to be applied to the target block according to the fourth method of adaptively applying a neural network technique based on an independent encoding structure (S845).

[0683] The encoding device can encode information indicating which of the first to fourth methods was used and transmit it to the decoding device. The decoding device can select one of the first to fourth methods based on the information received from the encoding device.

[0684]

[0685] After a method for applying a neural network technique is determined, the encoding device or the decoding device can select a quantity-prediction method to be applied to a target block from among a plurality of quantity-prediction methods including a neural network-based quantity-prediction method based on one or more search methods.

[0686] In some embodiments, the search scheme may be selected from among 1) a first search scheme that searches all weighted average prediction modes for all block sizes, 2) a second search scheme that searches only equal weighted average prediction modes for all block sizes, 3) a third search scheme that searches all weighted average prediction modes for a specific block size, 4) a fourth search scheme that searches only equal weighted average prediction modes for a specific block size, 5) a fifth search scheme that does not search all weighted average prediction modes for a specific block size, and 6) a sixth search scheme that does not search all weighted average prediction modes for all block sizes.

[0687] The first search method may be a method in which a quantity-prediction based on a neural network technique is performed for all block sizes of the target block and all weights within the weight set.

[0688] A second search method may involve performing a neural network-based quantitative prediction for all block sizes and equal weights of the target block. Equal weights may mean that the weights for the two intermediate prediction blocks are equal. For example, referring to Equation 1, equal weights may mean that w is 4.

[0689] A third search method may be one in which quantity prediction based on a neural network technique is performed for all weights within a weight set when the target block has a specific block size.

[0690] A fourth search method may be a method in which volume prediction based on a neural network technique is performed for equal weights when the target block has a specific block size.

[0691] The fifth search method may be a method in which quantity prediction based on a neural network technique is performed on a target block when the target block has a specific block size.

[0692] The sixth search method may be a method in which quantity prediction based on a neural network technique is performed on the target block.

[0693] The selection of the search method can be performed depending on the first to fourth methods for applying the neural network technique.

[0694] Additionally, information about the selection of the search method can be signaled from the encoding device to the decoding device.

[0695] Below, the first to fourth methods for applying the neural network technique are described in detail with reference to FIGS. 9 to 12.

[0696] The first method for applying neural network techniques

[0697] FIG. 9 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the first method.

[0698] As described above, in the first method, neural network-based quantity prediction can be performed dependently on the weighted average prediction mode index.

[0699] In S910, it can be determined whether w exists within a weight set. For example, the weight set may be {-2, 3, 4, 5, 10}. Existence of w within the weight set may mean that w is equal to one of the values ​​within the weight set.

[0700] S920 can be performed if w exists in the weight set.

[0701] In S920, it can be determined whether a first search method is performed to search all weighted average prediction modes for all block sizes.

[0702] When the first search method is performed according to the above judgment, a neural network-based quantity prediction is performed for the target block in S930.

[0703] If the first search method is not performed according to the above judgment, S940 may be performed to perform a third search method that searches all weighted average prediction modes for a specific block size.

[0704] In step (S940), it can be determined whether the block size of the target block is within the block sizes to which the neural network technique is applied.

[0705] In the embodiments, cnn_size represents block sizes to which the neural network technique is applied. For example, cnn_size may be {128Х128, 64Х64, 32Х32}. Alternatively, cnn_size may be other block sizes described in the embodiments.

[0706] In the embodiments, b_size represents the block size of the target block.

[0707] In embodiments, the block size of the target block being within the block sizes to which the neural network technique is applied may mean that the block size of the target block is equal to one of the block sizes to which the neural network technique is applied.

[0708] If the block size of the target block is within the block sizes to which the neural network technique is applied, the process proceeds to S950, whereby quantity prediction based on the neural network technique can be performed for the target block.

[0709] If the block size of the target block does not exist within the block sizes to which the neural network technique is applied, the process proceeds to S960 to determine whether w is 4.

[0710] In embodiments, w being 4 may mean that the weights for the two intermediate prediction blocks are equal.

[0711] In embodiments, 4 may be replaced with another value meaning that the weights for the two intermediate prediction blocks are equal.

[0712] In an embodiment, w being 4 may mean a weight mode in which the same weight is used for the two intermediate prediction blocks.

[0713] In an embodiment, w being 4 may mean that an equal weighted average prediction mode is used.

[0714] If w is 4, the process proceeds to S970, where average-quantity prediction or optical-flow-based-quantity prediction can be performed on the target block.

[0715] If w is not equal to 4, the process proceeds to S980, where a weighted average-based quantity prediction can be performed for the target block.

[0716] In the first method of applying the neural network technique to depend on the weighted average prediction mode index, either a first search method that searches all weighted average prediction modes for all block sizes or a third search method that searches all weighted average prediction modes for a specific block size can be selectively used.

[0717] For example, if the first method is selected for a target block, and a third search method is selected that searches all weighted average prediction modes for three block sizes (e.g., block sizes {128X128, 64X64, 32X32}), then the neural network-based quantity prediction can replace the conventional quantity prediction (i.e., average quantity prediction, weighted average-based quantity prediction, or optical-flow-based quantity prediction) for the three block sizes. Accordingly, if the size of the target block is equal to one of the three block sizes, the target block can be predicted by the neural network-based quantity prediction. On the other hand, if the size of the target block is not equal to the three block sizes, the target block can be predicted using the conventional quantity prediction determined by the weights.

[0718]

[0719] A second approach to applying neural network techniques

[0720] Figure 10 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the second method.

[0721] As described above, in the second method, neural network-based quantity prediction can be performed dependently on the equally weighted average prediction mode index.

[0722] In S1010, it can be determined whether w exists in the weight set.

[0723] If w exists in the weight set, it can be determined whether the process proceeds to S1015 and a second search method is performed that searches only the equal weighted average prediction mode for all block sizes.

[0724] If the second search method is performed according to the above judgment, it can be determined whether w is 4 in S1020. If w is 4, then quantity prediction based on a neural network technique can be performed on the target block in S1025.

[0725] If it is determined that the second search method is not performed at S1015, the process proceeds to S1030, where a fourth search method is performed to search all weighted average prediction modes for a specific block size. Accordingly, it can be determined whether the block size of the target block is within the block sizes to which the neural network technique is applied.

[0726] If the block size of the target block is within the block sizes to which the neural network technique is applied, it can be determined in S1035 whether w is 4. If w is 4, then quantity prediction based on the neural network technique can be performed for the target block in S1040.

[0727] If the block size of the target block does not fall within the block sizes to which the neural network technique is applied, the process proceeds to S1045 to determine whether w is 4. If w is 4, the process proceeds to S1050 to perform average-quantity prediction or optical-flow-based quantity prediction on the target block. On the other hand, if w is not 4, the process proceeds to S1050 to perform weighted-average-based quantity prediction on the target block.

[0728] Meanwhile, in FIG. 10, it is illustrated that the process is terminated if w is determined to be not 4 in S1020 and S1035. However, in another embodiment, if w is determined to be not 4 in S1020 and S1035, a weighted average prediction mode corresponding to w may be applied to perform weighted average-based quantization prediction for the target block.

[0729] In the second method of applying the neural network technique to depend on the equal weight average prediction mode index, either the second search method that searches only the equal weight average prediction mode for all block sizes or the fourth search method that searches only the equal weight average prediction mode for a specific block size can be selected.

[0730] For example, if a second method is selected that applies a neural network technique to depend on the equal weighted average prediction mode index, and a second search method is selected that searches only the equal weighted average prediction mode for all block sizes, the weighted average prediction mode with w equal to 4 can be replaced by a neural network-based quantile prediction regardless of the block size.

[0731]

[0732] A third approach to applying neural network techniques

[0733] Figure 11 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the third method.

[0734] As mentioned above, the third method may refer to a method of applying a neural network technique based on an independent encoding structure.

[0735] In S1110, it can be determined whether a fifth search method is performed that does not search all weighted average prediction modes for a specific block size.

[0736] If the fifth search method is performed based on the above judgment, S1120 may be performed. If the fifth search method is not performed based on the above judgment, a sixth search method that does not search all weighted average prediction modes for all block sizes may be applied, and accordingly, neural network-based quantity prediction may be performed for the target block at S1180.

[0737] In S1120, it can be determined whether the block size of the target block is within the block sizes to which the neural network technique is applied.

[0738] If the block size of the target block is within the block sizes to which the neural network technique is applied, the process proceeds to S1130, where neural network-based quantity prediction can be performed for the target block.

[0739] If the block size of the target block does not exist within the block sizes to which the neural network technique is applied, S1140 may be performed.

[0740] In S1140, it can be determined whether w exists in the weight set.

[0741] If w exists in the weight set, it can be determined in S1150 whether w is 4.

[0742] If w is 4, S1160 is performed, and accordingly, average quantity-prediction or optical-flow based quantity-prediction can be performed for the target block.

[0743] If w is not 4, then a weighted average-based quantity prediction can be performed for the target block in S1170.

[0744] A third method of applying a neural network technique based on an independent encoding structure may optionally use a fifth search method that does not search all weighted average prediction modes for a specific block size or a sixth search method that does not search all weighted average prediction modes for all block sizes.

[0745] For example, if the third method is selected and the sixth search method, which does not search all weighted average prediction modes for all block sizes, is selected, the neural network-based bi-prediction mode can be used as an independent encoding / decoding mode for encoding / decoding the target block. In this case, the existing bi-prediction modes may not be performed.

[0746]

[0747] A Fourth Approach to Applying Neural Network Techniques

[0748] Figure 12 is an exemplary flowchart illustrating a method for determining a quantity-prediction method to be applied to a target block according to the fourth method.

[0749] As mentioned above, the fourth method may refer to a method of adaptively applying neural network techniques based on an independent encoding structure.

[0750] In S1210, it can be determined whether a fifth search method is performed that does not search all weighted average prediction modes for a specific block size.

[0751] If the fifth search method is performed according to the above judgment, S1215 may be performed. If the fifth search method is not performed according to the above judgment, S1265 may be performed, which performs the sixth search method that does not search all weighted average prediction modes for all block sizes.

[0752] In S1215, it can be determined whether the block size of the target block is within the block sizes to which the neural network technique is applied.

[0753] If the block size of the target block is within the block sizes to which the neural network technique is applied, S1220 may be performed. On the other hand, if the block size of the target block is not within the block sizes to which the neural network technique is applied, S1245 may be performed.

[0754] In S1220, it can be determined whether w exists in the weight set.

[0755] If w exists in the weight set, S1225 is performed, otherwise, a neural network technique-based quantity prediction can be performed for the target block in S1240.

[0756] In S1225, it can be determined whether w is 4.

[0757] If w is 4, then averaged-quantity prediction or optical-flow-based quantization can be performed on the target block in S1230, otherwise, weighted-average-based quantization can be performed on the target block in S1235.

[0758] Meanwhile, in S1245, it can be determined whether w exists in the weight set.

[0759] If w exists in the weight set, it can be determined in S1250 whether w is 4.

[0760] If w is 4, then averaged-average or optical-flow based-quantity prediction can be performed in S1255, otherwise, weighted-average based-quantity prediction can be performed for the target block in S1260.

[0761] Meanwhile, in S1265, which performs the sixth search method that does not search all weighted average prediction modes for all block sizes, it can be determined whether w exists in the weight set.

[0762] If w exists in the weight set, S1270 may be performed, otherwise, neural network-based quantity-prediction may be performed for the target block in S1285.

[0763] In S1270, it can be determined whether w is 4 or not.

[0764] If w is 4, then averaged-quantity prediction or optical-flow-based quantization can be performed on the target block in S1275, otherwise, weighted-average-based quantization can be performed on the target block in S1280.

[0765] A fourth method of adaptively applying a neural network technique based on an independent encoding structure may be a fifth search method that does not search all weighted average prediction modes for a specific block size, or a sixth search method that does not search all weighted average prediction modes for all block sizes.

[0766] For example, if the fourth method is selected and the fifth search method that does not search all weighted average prediction modes for a specific block size is selected, then if the block size of the target block corresponds to the specific block size, the neural network-based bi-prediction mode can be used for encoding / decoding the target block as an independent encoding / decoding mode.

[0767] In Fig. 12, it is illustrated that existing quantity-prediction methods (average quantity-prediction, optical-flow-based quantity-prediction, and weighted average-based quantity-prediction) and neural network-based quantity-prediction are selectively applied depending on whether w is included in a predefined weight set. For example, depending on the judgment result of S1220, the existing quantity-prediction method is selected (S1230 to S1235) or the neural network-based quantity-prediction (S1240). In addition, depending on the judgment result of S1265, the existing quantity-prediction method is selected (S1275 to S1280) or the neural network-based quantity-prediction (S1285) is selected.

[0768] In another embodiment, conventional quantity prediction and neural network-based quantity prediction can be determined through competition. For example, conventional quantity prediction and neural network-based quantity prediction can be adaptively selected regardless of whether w is included in a predefined weight set, or even if w is included in a predefined weight set.

[0769] For example, a positive-prediction mode can be selected for a target block by comparing the encoding costs of a neural network-based positive-prediction mode with those of existing positive-prediction modes. For example, a positive-prediction method with a lower encoding cost can be selected.

[0770] As another example, the quantile prediction method can be determined using surrounding quantile / decryption information. For example, if at least one adjacent block is determined using a neural network-based quantile prediction method, or if the number of adjacent blocks determined using a neural network-based quantile prediction method is greater than a certain number or greater than the number of adjacent blocks determined using a weight-based quantile prediction method, the quantile prediction method for the target block can be determined using a neural network-based method.

[0771] As another example, when the merge mode is used for a target block, if the candidate for the merge mode is encoded / decoded in a neural network-based bi-prediction mode, the neural network-based bi-prediction mode may be used for the target block. Here, the candidate may include a spatial candidate and / or a temporal candidate.

[0772] Meanwhile, if a fifth search method that does not search all weighted average prediction modes for a specific block size is selected, the neural network-based bi-prediction mode may not be used for blocks of sizes different from the predefined specific block sizes, and only the existing bi-prediction modes may be used. Alternatively, even if the block size of the target block does not correspond to the specific block size, if the merge mode is used for the target block and the merge candidate selected for the target block is encoded / decoded with the neural network-based bi-prediction mode, the neural network-based bi-prediction mode may be used for the target block.

[0773] Below, we describe the prediction process using neural network-based quantity prediction.

[0774] The unit of input and output in a neural network can be defined as a processing unit. In embodiments, the processing unit may be a block. Alternatively, in embodiments, the processing unit may be a pixel, a patch, a tile, a slice, or a picture.

[0775] In neural network-based quantization prediction, the neural network can receive input data configured based on pixel values ​​corresponding to a processing unit and output predicted pixels for that processing unit.

[0776] The input data may consist of pixels within a reference picture or pixels derived from those pixels. For example, the input data may be pixels / blocks / patches / slices / tiles within the reference picture. Alternatively, the input data may be the reference picture itself.

[0777] As described above, in the embodiments of the present disclosure, the reference picture may be a picture encoded or decoded before the target picture to be currently encoded or decoded, i.e., a neighboring picture.

[0778] When composing input data from neighboring pictures, the neighboring pictures can be used as input data without scaling.

[0779] Alternatively, scaling can be applied to neighboring pictures based on various conditions. Pixels within the scaled neighboring pictures can be used as input data. For example, scaling using picture order count (POC) can be applied.

[0780] In embodiments of the present disclosure, a target picture may be used as a reference picture. For example, pixels within the target picture that have already been encoded or decoded may be used to construct input data.

[0781] When configuring input data for a target block to be encoded or decoded within a target picture, blocks adjacent to the target block may be used to configure the input data. For example, neighboring blocks spatially adjacent to the target block may be used to configure the input data. Alternatively, non-adjacent neighboring blocks adjacent to the target block may be used to configure the input data. Here, the fact that a neighboring block is adjacent to the target block may mean that the neighboring block is within an area determined based on the position and size of the target block.

[0782] When constructing input data from a target picture, scaling may not be applied to neighboring blocks.

[0783] Alternatively, scaling can be applied to neighboring blocks based on various conditions. For example, the various conditions can include 1) the distance between the target block and the neighboring blocks, and 2) the prediction mode for the target block.

[0784] In embodiments of the present disclosure, a new picture generated by processing at least one reference picture may be used as a reference picture. For example, the new picture may be generated based on a neural network as described above.

[0785] Below, we will describe an example where the processing unit of a neural network is a block. In the following description, input data, which is constructed based on a reference picture and input to the neural network, may include a reference block or an intermediate prediction block.

[0786] In embodiments, when a neighboring picture encoded or decoded before the target picture is used as a reference picture, an intermediate prediction block may be generated through inter prediction or motion compensation.

[0787] In embodiments, when the target picture is used as a reference picture, the intermediate prediction block may be generated based on intra prediction or IBC mode.

[0788] Neural networks can have an attention-residual-convolution based structure.

[0789] The output of the neural network can be used to generate prediction blocks for the target block.

[0790] In some embodiments, the neural network can be trained such that its output is a prediction block of the target block.

[0791] In some other embodiments, the average of two intermediate prediction blocks and the output of a neural network may be added to generate a predicted block of the target block. In this case, the neural network may be trained to minimize the difference between the actual pixel values ​​of the target block and the average pixel values ​​of the two intermediate prediction blocks.

[0792] When neural network-based quantile prediction is performed, at least one model among a prediction block model, a coded information combination model, an extended prediction block model, or an iteratively trained model may be used.

[0793] Neural network-based quantity prediction using the aforementioned models is explained with reference to FIGS. 13 to 16.

[0794] The terms used in Figures 13 to 15 are defined as follows.

[0795] - P0 and P1 can represent two intermediate prediction blocks.

[0796] - N rb can represent the number of residual blocks.

[0797] - P f can represent the block finally generated through the neural network model.

[0798] - Info can represent a map that reconstructs one or more pieces of encoded information to be the same size as the prediction block.

[0799] - Ex0 and Ex1 can represent the region composed of the surrounding top and left samples of the two middle prediction blocks.

[0800] FIG. 13 is an exemplary diagram illustrating neural network-based quantity prediction using a prediction block model according to one embodiment of the present disclosure.

[0801] As illustrated in Fig. 13, the prediction block model may be a neural network model that takes two intermediate prediction blocks P0 and P1 as input.

[0802] Referring to FIG. 13, each of the intermediate prediction blocks can be input to a first convolution module including N layers. Here, N can be an integer greater than or equal to 1. For example, N can be 3.

[0803] The first convolution module can output feature maps for each of the two intermediate prediction blocks. A dot product operation and a sigmoid activation function can be performed on the feature maps output from the first convolution module to extract an attention-based map representing the correlation between the feature maps corresponding to P0 and P1.

[0804] The attention-based map is output from the Att module and can be element-wise multiplied with the output of at least one layer of the first convolution module. The two feature maps multiplied by the attention-based map (hereinafter referred to as "attention-based feature maps") are combined in the Concat module to produce the final prediction block P. f can be used to create .

[0805] In some embodiments, as illustrated in FIG. 13, the attention-based feature maps may be generated by component-wise multiplication of the two feature maps output from the first layer of the first convolution module with the attention-based map. Additionally, the outputs of the last layer of the first convolution module may be combined with the attention-based feature maps in the Concat module.

[0806] In some other embodiments, the attention-based feature maps may be generated by element-wise multiplying the attention-based map with each of the two feature maps output from the last layer of the first convolution module. The Concat module may combine the attention-based feature maps with the outputs of the last layer of the first convolution module. In addition to or instead of the outputs of the last layer, the Concat module may combine the outputs of the first layer of the first convolution module.

[0807] To enhance or improve the feature map output from the Concat module, a second convolution module, a residual network (ResNet), and a third convolution module can be used.

[0808] The second and third convolution modules may be composed of one or more convolution layers (Conv2D).

[0809] The residual network is N rb It can be composed of residual blocks. Each residual block can include one convolutional neural network layer, one ReLU (Rectified Linear Unit) activation function, and one convolutional layer.

[0810] The output image from the third convolution module is the final prediction block P f can be used to generate. According to an embodiment, the final prediction block P f can be generated by combining the average of two intermediate prediction blocks with the output image output from the third convolution module.

[0811] In the above, the number of layers constituting the convolution modules, the number of feature maps, the number of activation functions, and the number of residual blocks can be one or more integers such as 1, 2, and 3.

[0812] FIG. 14 is an exemplary diagram illustrating neural network-based quantity prediction using an encoded information combination model according to one embodiment of the present disclosure.

[0813] In this embodiment, additional encoding information other than the two intermediate prediction blocks P0 and P1 can be used to configure input data for input to the neural network. That is, the intermediate prediction blocks P0 and P1 and additional encoding information can be combined and input to the neural network.

[0814] Additional encoding information may include at least one of mode information, such as information indicating a skip mode, information indicating a merge mode, information indicating a subblock mode, and information indicating an affine mode.

[0815] Additionally or alternatively, the encoding information may include information about quantization parameters. The quantization parameter information may be at least one of a slice quantization parameter (Slice QP) and a slice quantization parameter of a reference picture (Reference slice QP).

[0816] The slice quantization parameter (Slice QP) may be a quantization parameter of a slice containing a target block. The slice quantization parameter (Reference slice QP) of a reference picture may refer to a quantization parameter of a slice containing an intermediate prediction block.

[0817] Alternatively, the quantization parameter information may be block-level quantization parameters. For example, the quantization parameters of the target block and / or the quantization parameters of the intermediate prediction block.

[0818] For encoded information having values ​​of 0 and 1, such as information indicating skip mode, merge mode, sub-block mode, and affine mode, a value greater than 0 and less than 1, such as 0.01, may be used instead of 0 for the stability of neural network model training.

[0819] For quantization parameters, they can be converted to normalized values ​​by dividing by the maximum value, 63.

[0820] Each scalar value of the encoded information can be filled into a matrix equal to the size of the prediction block, and then combined with the prediction block to be converted into a feature map through one or more two-dimensional convolution layers.

[0821] Since the operation after configuring the input data is the same as described in Fig. 13, further explanation is omitted.

[0822] FIG. 15 is an exemplary diagram illustrating neural network-based quantile prediction using an extended prediction block model according to one embodiment of the present disclosure.

[0823] As illustrated in FIG. 15, in addition to the pixels within each intermediate prediction block, at least one of the upper pixels and the left pixels surrounding the intermediate prediction block may be used to construct input data. In other words, the intermediate prediction blocks (P0 and P1) and the upper and / or left surrounding pixels may be combined to construct input data for input to the neural network.

[0824] For example, the top one or more rows and the left one or more columns can be set as extended regions Ex0 and Ex1 to be combined with each of the two intermediate prediction blocks.

[0825] If a decoded pixel does not exist in the extended area, the value of the missing pixel can be filled with a specific value derived based on the bit depth. For example, the value of the missing pixel can be filled with a midpoint value in the range of pixel values ​​at the given bit depth. In other words, it can be filled with half the number of pixel values ​​that can be represented at the given bit depth, i.e., 2(bitdepth - 1).

[0826] For example, if the bit depth is 8, the pixel values ​​restored to 128 can fill in the expanded area where there is no existing pixel value. As another example, if the bit depth is 10, the pixel values ​​restored to 512 can fill in the expanded area where there is no existing pixel value.

[0827] The number of extended rows and columns can be an integer greater than or equal to 1. For example, the number of extended rows and columns can be 1, 2, or 3.

[0828] Since the operation after configuring the input data is the same as described in Fig. 13, further explanation is omitted.

[0829] The neural networks used for the above-described quantity prediction may differ from each other based on at least one of the prediction method (intra prediction, inter prediction), quantization value (QP), slice type (I, B, P), or block size.

[0830] Neural networks with the same neural network structure but trained using different training data can mean different neural networks.

[0831] Neural networks trained with the same training data but having different neural network structures can mean different neural networks.

[0832] That is, if at least one of the training data or the neural network structure is different, they can be defined as different neural networks.

[0833]

[0834] The neural networks described above can be trained to receive two intermediate prediction blocks as input and output a block similar to the actual target block.

[0835] In some embodiments, a neural network may be trained using an iterative training technique. The iterative training technique may include multiple training rounds (e.g., two rounds) and a single testing round.

[0836] FIG. 16 is an exemplary diagram illustrating a method for generating a neural network model to be used for quantity prediction through a process of iterative training according to one embodiment of the present disclosure.

[0837] In the first training phase (training phase 1), the first training data set can be constructed using two intermediate prediction blocks generated from positive prediction during encoding and decoding of the training images to be used for training.

[0838] At this time, quantization parameter information may be additionally included in the primary training data set. The quantization parameter information may be slice-level quantization parameter information or block-level quantization parameter information.

[0839] The first version of the neural network model, Model V1, can be built by training using the primary training data.

[0840] In the second training phase (training phase 2), Model V1 built in the first training phase can be merged into the encoder / decoder and utilized in the prediction process.

[0841] After encoding the training images using the combined encoder / decoder of Model V1, a secondary training dataset can be constructed using the two intermediate prediction blocks generated during the quantization process during decoding. As with the primary training process, quantization parameter information can be included in the secondary training dataset.

[0842] The training videos used in the second training session may be different from or identical to the training videos used in the first training session.

[0843] A second version of the neural network model, Model V2, can be built by training using the secondary training data set.

[0844] The neural network model Model V2 can be built as a final model to be used in the verification process and can be used for the aforementioned neural network-based quantity prediction.

[0845]

[0846] Below, another aspect of the present disclosure, a neural network-based super-resolution technique, is described.

[0847] When super-resolution techniques are applied, the encoding device downsamples a high-resolution (HR) image to obtain a low-resolution (LR) image and encodes the low-resolution image. The decoding device restores the low-resolution image and then upsamples it to convert it into a high-resolution image for output.

[0848] The present disclosure relates to a neural network-based super-resolution technique for converting a low-resolution image into a high-resolution image close to the original image.

[0849] Neural network-based super-resolution techniques can be used in conjunction with the aforementioned neural network-based quantity prediction. That is, the following neural network-based super-resolution techniques can be applied to images restored using neural network-based quantity prediction.

[0850] FIG. 17 is a block diagram illustrating an encoding device for applying a super-resolution technique according to one embodiment of the present disclosure.

[0851] The encoding device illustrated in FIG. 17 may additionally include a downsampling unit (1710) and an up / downsampling unit (1720) compared to the encoding device (110) of FIG. 1. Components other than the downsampling unit (1710) and the up / downsampling unit (1720) have substantially the same functions as the components illustrated in FIG. 1, and therefore, their descriptions are omitted to avoid redundant descriptions.

[0852] The downsampling unit (1710) can generate a downsampled picture for a target picture to be encoded.

[0853] The downsampling unit (110) can generate a downsampled picture by applying an interpolation filter defined in the Reference Picture Resampling (RPR) technique. For example, a downsampled picture can be generated through a 12-tap filter as illustrated in FIG. 18.

[0854] The up / downsampling unit (1720) can upsample or downsample a target picture to be encoded or a restored picture for the current picture and store it in a reference picture buffer (1985). The picture stored in the buffer (141) can be used as a reference picture for inter prediction.

[0855] Additionally or alternatively, the up / downsampling unit (1720) can upsample or downsample a reference picture stored in the buffer (141), and the upsampled or downsampled reference picture can be used for inter prediction.

[0856] For example, if the size of the reference picture is larger than that of the target picture, the up / downsampling unit (1720) can generate a reference picture of the same size as the target picture through downsampling of the reference picture. At this time, an interpolation filter defined in the reference picture resampling technique can be used.

[0857] On the other hand, if the size of the reference picture is smaller than that of the target picture, the up / down sampling unit (1720) can generate a reference picture of the same size as the target picture through upsampling of the reference picture.

[0858] Meanwhile, the upsampler of the up / downsampling unit (1720) can perform upsampling using at least one of a reference picture resampling technique or a neural network-based upsampling technique. The upsampling process performed by the up / downsampling unit (1720) will be described later.

[0859] The encoding device can determine whether to perform downsampling and a scale factor for downsampling. For example, if it is determined that downsampling is not to be performed, the scale factor can be determined as 1. If it is determined that downsampling is to be performed, the scale factor can be selected from a plurality of predefined values. The plurality of predefined values ​​can include a plurality of values ​​among, for example, 4, 3, 2, 1.5, and 1.25. Downsampling is performed at a ratio of (1 / scale factor). For example, if the scale factor is 2, the size of the picture is reduced by half in the horizontal and vertical directions.

[0860] The scale factor is determined for each specific unit, and that information can be signaled to the video decoding device.

[0861] For example, it can be determined at the video sequence or GOP (Group of Pictures) level.

[0862] As another example, the scale factor can be determined from a random access picture (RAP), such as an Instantaneous Decoding Refresh (IDR) picture. The scale factor determined from the current RAP can be applied to pictures prior to the next RAP.

[0863] As another example, the scale factor can be determined at the picture, subpicture, slice, or tile level, or at the block level, such as CTU or CU.

[0864] The scale factor can be determined based on the peak signal-to-noise ratio (PSNR). Additionally, information about the quantization parameters can be utilized.

[0865] As an exemplary embodiment, when determining a scale factor in units of GOP, the encoding device (3100) may determine the scale factor by comparing a case where at least one picture among the pictures in the GOP is encoded at the original resolution with a case where it is encoded by downsampling to a low resolution.

[0866] For example, the encoding device can downsample the first picture in the display order among the pictures in the GOP (i.e., the picture with the smallest POC) by applying a specific scale factor, and then upsample it back to the original resolution (picture size). Here, the downsampling and upsampling can be performed according to a reference picture resampling technique.

[0867] Hereinafter, a picture that is upsampled after downsampling is referred to as a re-scaling picture. The encoding device can calculate the PSNR between the re-scaled picture for the first picture and the first picture, and compare the PSNR with a predefined threshold to check self-similarity, which indicates whether the re-scaled picture is sufficiently similar to the first picture.

[0868] The encoding device can select one of the plurality of scale factors by performing a self-similarity test on at least some of the predefined plurality of scale factors, and can apply the selected scale factor to all pictures within the GOP.

[0869] The threshold for self-similarity judgment can be determined based on the quantization parameters applied to the first picture.

[0870] The threshold value can be a fixed value regardless of the scale factor, or it can be determined differently for each scale factor. For example, the threshold value can be higher for larger scale factors.

[0871] Additionally, the threshold value can be set differently depending on whether the coding structure is AI (All Intra), RA (Random Access), or LD (Low Delay).

[0872] FIG. 19 is a block diagram illustrating a decoding device for applying a super-resolution technique according to one embodiment of the present disclosure.

[0873] The decoding device illustrated in FIG. 19 may additionally include an up / down sampling unit (1920) compared to the decoding device (150) of FIG. 1. Since the functions of other components other than the up / down sampling unit (1920) are substantially the same as those of the decoding device (150) illustrated in FIG. 1, their descriptions are omitted to avoid redundant descriptions.

[0874] The up / down sampling unit (1920) may have the same structure and function as the up / down sampling unit (1720) of FIG. 17.

[0875] The up / downsampling unit (1920) can upsample or downsample a restored picture for a target picture and store it in a reference picture buffer (181). The picture stored in the buffer (181) can be used as a reference picture for inter prediction.

[0876] Additionally or alternatively, the up / downsampling unit (1920) can upsample or downsample a reference picture stored in the buffer (181), and the upsampled or downsampled reference picture can be used for inter prediction.

[0877] Below, the upsampling process performed by the up / downsampling unit (11720, 1920) of the encoding device and decoding device is described in detail.

[0878] 1. Upsampling method

[0879] Upsampling can be performed using a reference picture resampling technique or a neural network-based super-resolution technique. A neural network-based super-resolution technique can include a single-image neural network-based super-resolution technique or a multi-image neural network-based super-resolution technique.

[0880] An encoding device or a decoding device can perform upsampling by selecting one of a plurality of techniques.

[0881] FIG. 20 is an exemplary flowchart illustrating a method for selecting an upsampling method according to one embodiment of the present disclosure.

[0882] In S2010, it can be determined whether to apply a neural network-based upsampling technique.

[0883] If it is determined that a neural network-based upsampling technique is not applicable, a reference picture resampling method may be selected in S2015.

[0884] On the other hand, if a neural network-based upsampling technique is applied, it can be determined whether a multi-image neural network-based super-resolution technique is applied in S2020.

[0885] If it is determined that the multi-image neural network-based super-resolution technique is not applicable, the single-image neural network-based super-resolution technique may be selected in S2025.

[0886] On the other hand, if it is determined that a multi-image neural network-based super-resolution technique is applicable, the multi-image neural network-based super-resolution technique may be selected in S2035.

[0887] Information about the selected upsampling method can be encoded by the encoding device and transmitted to the decoding device.

[0888]

[0889] 2. Configuring input data for upsampling

[0890] The method of configuring input data may include a single image input configuration based on the current picture, a multiple image input configuration based on the current picture duplication, a multiple image input configuration based on the current picture and a list of reference pictures, etc.

[0891] The current picture-based single-image input configuration method can mean a method that uses the current picture (target picture) as a neural network input.

[0892] The current picture duplication-based multi-image input configuration method may mean duplicating and using the current picture when multiple neural network inputs are required.

[0893] The current picture and reference picture list-based multi-image input configuration method means a method of selecting at least one of the reference pictures included in the reference picture list and the current picture and using them as neural network inputs.

[0894] The way the input data is structured may depend on the upsampling method.

[0895] For example, in upsampling using a reference picture resampling technique or a single-image neural network-based super-resolution technique, a current picture-based single-image input configuration method can be used.

[0896] For example, in upsampling based on a multi-image neural network-based super-resolution technique, a current picture duplication-based multi-image input configuration method or a current picture and reference picture list-based multi-image input configuration method can be used.

[0897] FIG. 21 is an exemplary flowchart illustrating a method for selecting an input data configuration method according to one embodiment of the present disclosure.

[0898] It can be determined whether a neural network-based super-resolution technique is applied as an upsampling method in S2110.

[0899] If a neural network-based super-resolution technique is not applied (i.e., a reference picture resampling technique is selected), the current picture-based single image input configuration method in S2120 can be selected as the input data configuration method.

[0900] When a neural network-based super-resolution technique is applied, it can be determined whether a multi-image neural network-based super-resolution technique is applied in S2130.

[0901] If a multi-image neural network-based super-resolution technique is not applied (i.e., a single-image neural network-based super-resolution technique is applied), a current picture-based single-image input configuration method may be selected in S2120.

[0902] When a multi-image neural network-based super-resolution technique is applied, it can be determined whether a reference picture list is utilized in S2140.

[0903] If the reference picture list is not utilized, the current picture duplication-based multi-image input configuration method in S2150 can be selected as the input data configuration method.

[0904] For example, if the number of input pictures of a multi-image neural network is defined as 5, the current picture can be duplicated to configure 5 pictures as input. Here, the number of input pictures can mean one of the numbers 2 or greater, such as 2, 3, 4, and 5.

[0905] When a reference picture list is utilized, a multi-image input configuration method based on the current picture and reference picture list in S2160 can be selected as the input data configuration method.

[0906] For example, if the number of input pictures of a multi-image neural network is defined as 5, the input data of the neural network can be configured by sequentially combining two L0 reference pictures referenced by the current picture within reference picture list 1 (L0 reference picture list), the current picture, and two L1 reference pictures referenced by the current picture within reference picture list 2 (L1 reference picture list).

[0907] As another example, one or more reference pictures among two L0 reference pictures referenced by the current picture may be extracted that have a POC smaller than the POC of the current picture. Similarly, one or more reference pictures among two L1 reference pictures referenced by the current picture may be extracted that have a POC larger than the POC of the current picture. The input data may be composed by aligning the extracted reference pictures and the current picture based on the POC.

[0908] Here, if the number of configured pictures is less than the defined input picture number of 5, the current picture can be duplicated by the insufficient number to configure 5 pictures as input data.

[0909] As another example, among the two L0 reference pictures and two L1 reference pictures referenced by the current picture, only those reference pictures with a temporal ID lower than or equal to the current picture can be extracted for composing the input data. The input data can be composed by combining the extracted reference pictures and the current picture.

[0910] Here, if the number of configured pictures is less than the defined number of input pictures, which is 5, the current picture can be duplicated by the insufficient number to configure 5 pictures as input data.

[0911] In the examples above, the number of input pictures is described as 5, but the number of input pictures can be one of 2 or more numbers, such as 2, 3, 4, and 5.

[0912]

[0913] 3. Component-based neural network model

[0914] When upsampling using a neural network-based super-resolution technique is applied, a neural network model can be selected based on components.

[0915] A neural network model can be classified based on a luminance component and two chrominance components (e.g., a blue chrominance component U and a red chrominance component V).

[0916] For example, neural network models trained individually for each component can be used. For example, a neural network model for the luminance component (Y), a neural network model for the first chrominance component (e.g., the blue chrominance component U), and a neural network model for the second chrominance component (e.g., the red chrominance component V) can be trained and used individually.

[0917] As another example, a chrominance integration model trained with input data that integrates two chrominance components can be used.

[0918] As another example, a luminance-chrominance integrated model trained with input data that integrates both luminance and two chrominance components can be used.

[0919] FIG. 22 is an exemplary flowchart illustrating a method for selecting a neural network model according to one embodiment of the present disclosure.

[0920] According to the example of Fig. 22, it is determined whether a neural network-based upsampling technique is applied to both the luminance component and the chrominance component in S2210.

[0921] When a neural network-based upsampling technique is applied to only one of the luminance component and the chrominance component, it is determined in S2220 whether the neural network-based upsampling technique is applied to the luminance component.

[0922] When a neural network-based upsampling technique is applied to the luminance component, a luminance model may be selected in S2230. In this case, upsampling for the chrominance component may be performed using an interpolation filter defined in the reference picture resampling technique. For example, the interpolation filter illustrated in FIG. 24 may be used.

[0923] If neural network-based upsampling is determined to be applied to the chrominance component in S2220, a chrominance integration model may be selected in S2240. In this case, upsampling for the luminance component may be performed using an interpolation filter defined in the reference picture resampling technique. For example, the interpolation filter illustrated in FIG. 23 may be used.

[0924] Meanwhile, if it is determined that a neural network-based upsampling technique is applied to both the luminance component and the chrominance component in S2210, it can be determined whether to apply separate neural network models to the luminance component and the chrominance component in S2250.

[0925] If separate neural network models are not applied to the luminance component and chrominance component, a single luminance-chrominance integrated model can be selected to upsample the luminance component and chrominance component in S2260.

[0926] When separate neural network models are used for the luminance component and the chrominance component, it can be determined in S2270 whether to apply separate neural network models to the two chrominance components.

[0927] When individual neural network models are applied to two chrominance components, a luminance model, a first chrominance model, and a second chrominance model may be selected for performing component-by-component upsampling in S2290. Otherwise, a luminance model for upsampling the luminance component and a chrominance integrated model for upsampling the two chrominance components may be selected in S2280.

[0928] The encoding device can signal information to the decoding device to indicate the selected neural network model.

[0929]

[0930] 4. Perform upsampling

[0931] Upsampling using a reference picture resampling technique can be performed using a predefined interpolation filter. For example, upsampling for the luminance component can be performed using the filter illustrated in FIG. 23, and upsampling for the chrominance component can be performed using the filter illustrated in FIG. 24.

[0932] Neural network-based upsampling can be performed using a trained neural network. Neural network-based upsampling is described below with reference to FIGS. 25 to 34.

[0933] The definitions of the terms indicated in Figures 25 to 34 are as follows.

[0934] - LR rec can represent a reconstructed (restored) low-resolution picture.

[0935] - HR rec can represent a reconstructed high-resolution picture.

[0936] - N rb can represent the number of residual blocks that constitute the residual network.

[0937] - QP can represent a feature map based on quantization parameter information. QP can be a block-based quantization parameter map that constitutes a reconstructed low-resolution picture, or a slice-based quantization parameter map.

[0938] - Slice can represent a feature map based on slice information of a picture.

[0939] - Partition can represent a feature map based on the block division information of a picture.

[0940] - LR pred can represent a predicted low-resolution picture based on prediction blocks.

[0941] - HR rpr can represent a reconstructed high-resolution picture after performing upsampling based on reference picture resampling.

[0942] - LR ref,0 , LR ref,1 , LR ref,2 , LR ref,3 The reconstructed low-resolution picture may represent a reference picture referenced by the reconstructed low-resolution picture and / or a picture copied from the reconstructed low-resolution picture.

[0943] - Optical Flow Estimation can represent an optical-flow prediction module.

[0944] - Feature Propagation Block can represent a feature propagation block.

[0945] FIGS. 25 to 30 are exemplary diagrams showing neural network structures that can be used for upsampling using a single image neural network-based super-resolution technique according to one embodiment of the present disclosure.

[0946] In some embodiments, as illustrated in FIGS. 25 to 28, a residual convolutional neural network may be used as a neural network for upsampling, replacing reference picture resampling-based upsampling. The structure of the residual convolutional neural network may be modified depending on the input type and the presence or absence of residual connections (skip connections).

[0947] For example, as illustrated in FIG. 25, a reconstructed low-resolution picture can be used as input and a reconstructed high-resolution picture can be output from the neural network.

[0948] In neural network-based upsampling, the reconstructed low-resolution picture can be input to a single convolutional neural network layer.

[0949] The feature map for the reconstructed low-resolution picture output through the convolutional neural network layer is N rb A residual network consisting of residual blocks can be input. Each residual block can include, for example, one convolutional neural network layer, one ReLU (Rectified Linear Unit) activation function, and one convolutional neural network layer.

[0950] The output of the residual network is sequentially passed through one convolutional neural network layer, one pixel shuffle module, and one convolutional neural network layer, and a reconstructed high-resolution picture can be output accordingly.

[0951] In the neural network structure illustrated in Figure 25, the number of convolutional neural network layers, the number of feature maps, the number of activation functions, the number of residual blocks, and the number of pixel shuffle modules can be configured in various ways as numbers greater than or equal to 1, such as 1, 2, and 3.

[0952] As another example, as illustrated in FIG. 26, a global residual connection can be additionally applied to the neural network structure illustrated in FIG. 25. After applying bicubic interpolation to the reconstructed low-resolution picture to generate a high-resolution picture, the output of the convolutional neural network can be added to output the final reconstructed high-resolution picture.

[0953] Here, in addition to the bicubic interpolation method, image interpolation methods such as nearest neighbor and bilinear interpolation can be used.

[0954] As another example, as illustrated in FIG. 27, a feature map based on additional encoding information may be combined with the input of the neural network illustrated in FIG. 25. Here, the additional encoding information may include one or more of quantization parameters, slice information, block partition information, and predicted low-resolution pictures.

[0955] As another example, as illustrated in FIG. 28, a feature map based on additional encoding information can be combined with the input of the neural network illustrated in FIG. 26.

[0956] In some other embodiments, a residual convolutional neural network combining upsampling based on reference picture resampling may be used as a neural network for upsampling, as illustrated in FIGS. 29 and 30 . The residual convolutional neural network's structure can be modified depending on the input format and the presence or absence of residual connections (skip connections).

[0957] For example, as shown in Fig. 29, a reconstructed high-resolution picture (HR) is obtained after performing upsampling based on reference picture resampling. rpr ) is a reconstructed low-resolution picture (LR) rec ) can be combined with and used as input to a convolutional neural network.

[0958] Here, the reconstructed low-resolution picture (LR rec ) to match the size of the reference picture, and then perform upsampling based on resampling the reference picture to reconstruct the high-resolution picture (HR rpr ) can be applied to the PixelUnShuffle module.

[0959] Alternatively, a reconstructed high-resolution picture (HR) based on reference picture resampling rpr ) is transformed through wavelet decomposition to generate images corresponding to multiple subbands, and the images corresponding to each subband are reconstructed as low-resolution pictures (LR). rec ) can be input into the neural network. For example, HRrpr can be transformed into four subband images with the size reduced by half in the horizontal and vertical directions by the first wavelet transform. The four subband images are HR rpr This is an image in which the frequency is decomposed into low-frequency components (L) and high-frequency components (H) in the horizontal and vertical directions, respectively, and may be an image corresponding to LL, LH, HL, and HH, respectively.

[0960] As another example, as shown in Fig. 30, a reconstructed high-resolution picture (HR) is obtained after performing upsampling based on reference picture resampling. rpr ) can be connected to the global residual. LR rec The output of the neural network and HR rpr This can be added and the final reconstructed high-resolution picture can be output.

[0961] As another example, a neural network combining the structures of Figs. 29 and 30 can be used. In this case, as described above, the reconstructed high-resolution picture (HR rpr ) from which multiple sub-band images (e.g., LL, LH, HL, HH) are wavelet-transformed to reconstruct a low-resolution picture (LR) rec ) can be combined and used as input to a neural network.

[0962] For residual convolutional neural networks combining upsampling based on reference picture resampling, feature maps based on additional encoding information can also be combined with the input.

[0963] FIGS. 31 to 34 are exemplary diagrams showing neural network structures that can be used for upsampling using a multi-image neural network-based super-resolution technique according to one embodiment of the present disclosure.

[0964] The input data to the multi-image neural network is the reconstructed low-resolution picture LR. recIn addition, at least one (LR) of the reference pictures and / or duplicated reconstructed low-resolution pictures referenced by the reconstructed low-resolution pictures ref,0 , LR ref,1 , LR ref,2 , LR ref,3 ) may be included.

[0965] Here, the number of input pictures can mean one of two or more numbers, such as 2, 3, 4, and 5.

[0966] A multi-image neural network may have a structure in which a feature map is generated for each input picture, and the feature map generated for one input picture is propagated to generate a feature map for another input picture.

[0967] The propagation can be performed by at least one of forward propagation, backward propagation, and arbitrary access based propagation.

[0968] The multiple input pictures input to the multi-image neural network can be arranged in a predefined order as described above. For example, they can be arranged based on the POC.

[0969] Forward propagation can mean a propagation method in which the feature map generated for the previous input picture is passed on to generate the feature map for the next input picture in a predefined order.

[0970] Backpropagation can mean a propagation method in which the feature map generated for the previous input picture is passed on to generate the feature map for the next input picture in the reverse order of a predefined order.

[0971] Random access-based propagation can mean a propagation method in which the generated feature map for one input picture is propagated to generate the feature map for another input picture, regardless of the predefined order.

[0972] In embodiments, the multi-image neural network can perform upsampling by hierarchically configuring at least one feature propagation process among forward propagation, backpropagation, and random access-based propagation.

[0973] In some embodiments, as illustrated in FIG. 31, the multi-image neural network may include a forward propagation process.

[0974] The predefined order for input pictures is LR ref,0 , LR ref,1 , LR ref,2 , LR ref,3 , and LR rec It is assumed that the order is . For convenience of explanation, Fig. 31 shows LR which are reference pictures and / or duplicated reconstructed low-resolution pictures. ref,0 , LR ref,1 , LR ref,2 , LR ref,3 LR, a low-resolution picture that has been reconstructed rec are arranged in a predefined order. When a multi-image neural network is constructed using forward propagation, the reconstructed low-resolution picture LR is output to the reconstructed high-resolution picture. rec can be placed in the last order.

[0975] Once the order of multiple input pictures is defined, the optical flow between adjacent pictures can be predicted by the optical flow estimation module. Optical flow estimation can be performed using methods such as the Lucas-Kanade method or a convolutional neural network-based method.

[0976] Each input picture can be fed into one convolutional neural network layer. The number of convolutional neural network layers and the number of feature maps can mean one or more numbers, such as 1, 2, and 3.

[0977] The feature map for each input picture output from the convolutional neural network layer can be processed by the feature propagation block.

[0978] FIG. 34 is a diagram illustrating the structure of a feature propagation block according to one embodiment of the present disclosure.

[0979] The feature propagation block may include an alignment module and a residual network. The residual network may be N, each of which includes one convolutional neural network layer, one Rectified Linear Unit (ReLU) activation function, one convolutional neural network layer, and a skip connection. rb It may contain residual blocks of dogs.

[0980] Here, the number of convolutional neural network layers, the number of feature maps, the number of activation functions, and the number of residual blocks can mean one or more numbers such as 1, 2, and 3.

[0981] The alignment module can perform a warping operation using the optical flow information inferred during the optical flow prediction process.

[0982] The input of the alignment module can consist of a feature map that has passed through the convolutional neural network layer at the current point in time and a feature map for an adjacent picture that has previously passed through the feature propagation block.

[0983] Features that pass through the feature propagation block can be forward propagated to the feature propagation block for adjacent pictures in a predefined order.

[0984] When the last input picture passes through the feature propagation block, the features are sequentially passed through one convolutional neural network layer, one pixel shuffle module, and one convolutional neural network layer, and the reconstructed high-resolution picture HR is generated accordingly. rec This can be printed.

[0985] Here, the number of convolutional neural network layers, the number of feature maps, and the number of pixel shuffle modules can mean one or more numbers such as 1, 2, and 3.

[0986] In some other embodiments, as illustrated in FIG. 32, the multi-image neural network may include backpropagation and forward propagation processes.

[0987] The predefined order for input pictures is LR ref,0 , LR ref,1 , LR ref,2 , LR ref,3 , and LR rec It is assumed that the order is . If the multi-image neural network includes backpropagation and forward propagation processes, the reconstructed low-resolution picture LR that outputs the reconstructed high-resolution picture rec can be placed in the last order.

[0988] In the backpropagation process, the reconstructed low-resolution picture LR is located at the end. rec The output of the feature propagation block for LR rec It is input as a feature propagation block for another input picture adjacent to it. Feature propagation (backpropagation) can be sequentially performed between feature propagation blocks of adjacent pictures in the reverse order of the predefined order.

[0989] Once the feature propagation block for the first input picture is reached in a predefined order, forward propagation can be performed again. That is, the output of the feature propagation block for the first input picture is propagated to the feature propagation blocks for the pictures adjacent to the first input picture in a predefined order. Feature propagation (forward propagation) can be performed sequentially between feature propagation blocks in a predefined order.

[0990] In the forward propagation process, when the last input picture passes through the feature propagation block, the features sequentially pass through one convolutional neural network layer, one pixel shuffle module, and one convolutional neural network layer, and the reconstructed high-resolution picture HR is generated accordingly. rec This can be printed.

[0991] In some other embodiments, the multi-image neural network may include a random access-based propagation process. Additionally, the multi-image neural network may further include at least one of forward propagation and backpropagation.

[0992] Figure 33 shows the structure of a multi-image neural network including forward propagation, backpropagation, and random access-based propagation processes.

[0993] For convenience of explanation, the predefined order is LR ref,0 , LR ref,1 , LR rec , LR ref,2 , and LR ref,3 It is assumed that the order is LR, i.e., the reconstructed low-resolution picture that will output the reconstructed high-resolution picture. rec can be positioned in the middle of a predefined order.

[0994] In a random access-based propagation process, the feature map for a specific input picture can be propagated to the feature propagation block of another input picture, regardless of the predefined order.

[0995] For example, as shown in Figure 33, the first input picture LR ref,0 The output of the feature propagation block for the last input picture LR ref,3 The feature propagation block is propagated to LR. And, ref,3 The output of the feature propagation block for the second input picture LR ref,1 The feature propagation block is propagated to LR. ref,1 The output of the feature propagation block for the fourth input picture LRref,2 The feature propagation block is propagated to .

[0996] Through this process, the input picture LR located in the middle rec When the processing of a specific propagation block is finished, the feature map output from the feature propagation block sequentially passes through one convolutional neural network layer, one pixel shuffle module, and one convolutional neural network layer, and thus the reconstructed high-resolution picture HR rec This could be it.

[0997] FIG. 35 and FIG. 36 are diagrams illustrating the configuration of an input picture input to a neural network according to one embodiment of the present disclosure.

[0998] Y in Fig. 35 and Fig. 36 LR , U LR , V LR can represent the reconstructed low-resolution picture for the luminance component, blue chrominance component, and red chrominance component, respectively. U IHR and V IHR Each of them can represent an intermediate high-resolution picture of blue and red chrominance components, upsampled to be the same size as the luminance component picture. U IHR and V IHR can be generated, for example, based on a pre-trained neural network.

[0999] In the embodiments, the input picture fed to the neural network, e.g., LR rec , can be configured in various ways according to the component-based neural network model described with reference to FIG. 22.

[1000] For example, when a luminance model is selected in FIG. 22 (S2230), one luminance component (Y) picture can be used as input, as in (A) of FIG. 35.

[1001] As another example, when a color difference integration model is selected (S2240), as illustrated in (B) of FIG. 35, a first color difference component (U) picture and a second color difference component (V) picture can be combined and used as input.

[1002] As another example, when the luminance-chrominance integrated model is selected (S2260), the luminance component, the first and second chrominance component pictures can be combined and used as input, as shown in (C) of FIG. 35.

[1003] Here, a pixel unshuffle module can be applied to make the luminance component picture the same size as the chrominance component picture.

[1004] Meanwhile, when the luminance component, first and second chrominance component pictures are combined, the order of input of the chrominance components can be changed.

[1005] For example, information indicating whether the order of chrominance components is adjusted can be signaled. The information can be signaled, for example, at the SPS or picture level. When chrominance order adjustment is enabled, pictures of luminance components and chrominance components can be input and output in the order of Y, V, and U, and when chrominance order adjustment is disabled, pictures of luminance components and chrominance components can be input and output in the order of Y, U, and V.

[1006] Additionally, even if the input order of the neural network is configured as Y, U, and V during the learning process, the input order of the neural network can be different, such as Y, V, and U, during the inference process.

[1007] As another example, when a luminance model and a chrominance integration model are selected (S2280), as illustrated in (D) of FIG. 35, one luminance component picture can be used as input to the luminance model, and a picture in which the first and second chrominance component pictures are combined can be used as input to the chrominance integration model.

[1008] As another example, when a luminance model, a first chrominance model, and a second chrominance model are each selected (S2290), as illustrated in (E) of FIG. 35, the luminance component picture, the first chrominance component picture, and the second chrominance component picture can be used as inputs to the luminance model, the first chrominance model, and the second chrominance model, respectively.

[1009] As another example, as illustrated in FIG. 36, a neural network model (e.g., a first chrominance model, a second chrominance model, a chrominance integration model) used for upsampling of chrominance components may use a luminance component as input data as additional encoding information.

[1010] For example, as illustrated in (A) of FIG. 36, a pixel unshuffle module can be applied to a luminance component picture to make it the same size as a chrominance component picture.

[1011] As another example, as illustrated in (B) of FIG. 36, a neural network-based super-resolution model for chrominance components learned in advance can be applied to chrominance component pictures to make them the same size as the luminance component pictures.

[1012]

[1013] In the embodiments, the methods may be described based on a flowchart comprising a series of steps or units. The methods of the embodiments are not limited to the described order of the steps, and some steps may be performed in a different order than the described order or may be performed concurrently with other steps. Furthermore, the steps described by the flowchart or the like may not be exclusive. Other steps may be included between the steps described by the flowchart or the like. One or more steps described by the flowchart or the like may be deleted or omitted.

[1014] The embodiments may include examples of various aspects. While not all possible combinations to illustrate various aspects can be described, those skilled in the art will recognize that other combinations are possible in addition to those explicitly described. Accordingly, the present invention encompasses all other alterations, modifications, and variations that fall within the scope of the following claims.

[1015] The embodiments may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium.

[1016] The computer-readable recording medium may include a non-transitory computer-readable recording medium. The computer-readable recording medium may include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically configured to store and execute program instructions such as ROMs, RAMs, flash memories, and the like. The hardware devices may be configured to operate as one or more software modules to perform processes according to the embodiments, and vice versa.

[1017] A computer-readable recording medium may contain program commands, data files, data structures, etc., either singly or in combination. The program commands recorded on the computer-readable recording medium may be specially designed and configured for the embodiments, or may be known and available to those skilled in the art of computer software.

[1018] Program instructions may include machine language code, such as that generated by a compiler, or may include high-level language code that can be executed by a computer using an interpreter or the like. Program instructions may also be referred to as computer-executable code or a program. In embodiments, program instructions, computer-executable code, and a program may be interchangeable.

[1019] A computer-readable recording medium may contain information used in the embodiments. For example, the computer-readable recording medium may contain a bitstream, and the bitstream may contain information described in the embodiments. The information described in the embodiments may include syntax elements. The information described in the embodiments, such as syntax elements, may be understood as computer-executable code in that it drives an encoding device and a decoding device to perform specific processing.

[1020] A bitstream may contain computer-executable code. The computer-executable code may include information described in the embodiments, such as syntax elements. In other words, information described in the embodiments, such as syntax elements, may be considered computer-executable code within the bitstream or a part of the computer-executable code.

[1021] Although the present invention has been described above with specific details such as specific components and limited embodiments and drawings, such descriptions are provided only to help a more general understanding of the present invention, and the present invention is not limited to the described embodiments, and those with ordinary knowledge in the technical field to which the present invention pertains can make various modifications and variations from this description.

[1022] Therefore, the spirit of the present invention should not be limited to the described embodiments, and all modifications equivalent to or equivalent to the claims described below, as well as the claims, are considered to fall within the scope of the spirit of the present invention.

[1023] CROSS-REFERENCE TO RELATED APPLICATION

[1024] This patent application claims priority to Korean patent application numbers 10-2024-0003844 and 10-2024-0003850, filed on January 9, 2024, the entire contents of which are incorporated herein by reference.

Claims

1. In the video decryption method, A step of determining at least one reference picture for a target block; A step of generating a plurality of reference blocks from at least one reference picture; A step of configuring input data using the above plurality of reference blocks; and A step of generating a prediction block for a target block from the input data using a trained neural network. An image decoding method characterized by including a.

2. In paragraph 1, The step of determining at least one reference picture is: A step of selecting a plurality of reference pictures within at least one reference picture list; and A step of inputting the above multiple reference pictures into a neural network to generate at least one new reference picture, An image decoding method, characterized in that at least one reference picture for the target block includes the new reference picture.

3. In paragraph 1, A method for decoding an image, characterized in that at least one reference picture includes a target picture to which a target block belongs.

4. In paragraph 1, The above input data is, An image decoding method, characterized in that it further includes at least one of quantization parameter information and encoding mode information of the target block.

5. In paragraph 4, The above encoding mode information is, A video decoding method, characterized in that it includes at least one of information indicating a skip mode, information indicating a merge mode, information indicating a subblock mode, or information indicating an affine mode.

6. In paragraph 4, The above quantization parameter information is, An image decoding method, characterized in that it includes at least one of a quantization parameter of a slice including the target block or a quantization parameter of slices including the reference blocks.

7. In paragraph 1, The steps for configuring the above input data are: For each of the plurality of reference blocks, a step of generating extended reference blocks by combining at least one restored sample among one or more left columns or one or more upper rows surrounding the reference block; and A step of configuring the input data using the above extended reference blocks. An image decoding method characterized by including a.

8. In paragraph 1, The step of generating a prediction block for the above target block is: A step of deriving feature maps corresponding to a plurality of reference blocks constituting the input data using a module including one or more convolutional layers; A step of deriving an attention-based map representing the correlation between the above feature maps; A step of applying the above attention-based map to the above feature maps to derive attention-based feature maps; and A step of generating a prediction block for the target block using the above attention-based feature maps. An image decoding method characterized by including a.

9. In the video encoding method, A step of determining at least one reference picture for a target block; A step of generating a plurality of reference blocks from at least one reference picture; A step of configuring input data using the above plurality of reference blocks; and A step of generating a prediction block for a target block from the input data using a trained neural network. An image encoding method characterized by including a.

10. A method for providing image data to an image decoding method, A step of generating a bitstream by encoding a target block; and Comprising a step of transmitting the bitstream to the image decoding device, The steps of generating the above bitstream are: A step of determining at least one reference picture for the target block; A step of generating a plurality of reference blocks from at least one reference picture; A step of configuring input data using the above plurality of reference blocks; and A step of generating a prediction block for a target block from the input data using a trained neural network. A method characterized by comprising:

Images