Image encoding / decoding method and device, and recording medium

Abstract

An image decoding method is disclosed. The image decoding method comprises the steps of: by decoding position information from a bitstream, identifying, in a data block including data related to a target block, a partial region to be decoded first, and the remaining region excluding the partial region; reconstructing the partial region identified by the position information; reconstructing the remaining region by using the reconstructed partial region; and generating a reconstructed block for the data block by using the reconstructed partial region and the reconstructed remaining region.

Description

Method, device and recording medium for video encoding / decoding

[0001] The present invention relates to a method, device, and recording medium for video encoding / decoding. In one aspect, the present invention relates to a method of first encoding / decoding a portion of information of a target block to be encoded, and then encoding / decoding the remaining information of the target block based on that portion of information.

[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] The present disclosure provides a method for improving encoding efficiency by first encoding / decoding some information of a target block to be encoded and then encoding / decoding the remaining information of the target block based on that some information.

[0006] One aspect of the present disclosure provides a video decoding method. The method comprises the steps of: decoding position information from a bitstream, identifying a portion of a data block containing data for a target block, to be decoded first, and a remaining portion excluding the portion; restoring the portion of the data block identified by the position information; restoring the remaining portion of the data block using the restored portion of the data block; and generating a restored block for the data block using the restored portion of the data block and the restored remaining portion of the data block.

[0007] Another aspect of the present disclosure provides a video encoding method. The method comprises the steps of: determining a portion of a data block containing data for a target block to be encoded first and a remaining portion of the data block excluding the portion of the data block; encoding positional information for indicating the positions of the portion of the data block and the remaining portion of the data block; encoding the portion of the data block to generate encoding data for the portion of the data block; and restoring the portion of the data block and encoding the remaining portion of the data block using the restored portion of the data block.

[0008] Another aspect of the present invention provides a method for transmitting a bitstream encoded by the above image encoding method.

[0009] Another aspect of the present invention provides a computer-readable recording medium for storing a bitstream encoded by the image encoding method.

[0010] Figure 1 illustrates a system for video coding according to one embodiment.

[0011] Figure 2 shows a segmentation structure of an image according to one embodiment.

[0012] Figure 3 illustrates the structure of intra prediction according to one embodiment.

[0013] Figure 4 shows the structure of inter prediction to explain the inter prediction process according to one embodiment.

[0014] Figure 5 shows the order in which spatial candidates are added to the candidate list according to one embodiment.

[0015] Figure 6 illustrates multiple in-loop filters according to an example.

[0016] Figure 7 shows the structure of entropy encoding and entropy decoding according to an example.

[0017] Fig. 8 is a diagram illustrating a still image encoding device and decoding device implemented based on a neural network.

[0018] Figure 9 illustrates an example of a latent representation transformation neural network.

[0019] Figure 10 illustrates an example of a latent representation restoration neural network.

[0020] Figure 11 is an example diagram for explaining a neural network-based encoding / decoding structure using a hidden vector probability model dependent on an input image.

[0021] FIG. 12 and FIG. 13 are exemplary diagrams showing a super-prediction information transformation neural network and a super-prediction information restoration neural network, respectively, according to one embodiment of the present disclosure.

[0022] FIG. 14 is a diagram illustrating a neural network-based video encoding device that uses a method for reducing temporal redundancy in image space.

[0023] FIG. 15 is an exemplary block diagram of an encoding device for encoding a target block in an image space according to one embodiment of the present disclosure.

[0024] Figure 16 shows an example of dividing a target block into predetermined intervals and selecting four pieces of information.

[0025] Figure 17 is an example diagram illustrating a method of expressing a binary mask.

[0026] Figure 18 is a diagram illustrating predefined patterns of a binary mask.

[0027] Figure 19 is an example diagram to explain the process of generating a binary mask based on a transformer neural network.

[0028] Figure 20 is an example diagram illustrating a process of classifying some selected information and the remaining unselected information for a target block.

[0029] FIG. 21 is an exemplary diagram of some information prediction units according to one embodiment of the present disclosure.

[0030] FIG. 22 is an exemplary diagram illustrating the operation of a mask autoencoder according to one embodiment of the present disclosure.

[0031] FIG. 23 is a diagram illustrating a remaining information prediction unit according to one embodiment of the present disclosure.

[0032] FIG. 24 is an exemplary diagram illustrating a method for generating the positions of information input to a mask autoencoder within a remaining information prediction unit according to one embodiment of the present disclosure.

[0033] FIG. 25 is an exemplary diagram showing a decryption device that decrypts a target block in an image space according to one embodiment of the present disclosure.

[0034] FIG. 26 is an exemplary diagram showing an encoding device that encodes a target block in a latent representation space according to one embodiment of the present disclosure.

[0035] FIG. 27 is an exemplary diagram illustrating a pre-processing information extraction and encoding unit according to one embodiment of the present disclosure.

[0036] Figure 28 is an example of a neural network for pre-processing information restoration.

[0037] FIG. 29 is an exemplary diagram illustrating some information encoding units according to one embodiment of the present disclosure.

[0038] FIG. 30 is an exemplary diagram showing a remaining information encoding unit according to one embodiment of the present disclosure.

[0039] FIG. 31 is an exemplary diagram showing a decoding device that decodes a target block in a latent representation space according to one embodiment of the present disclosure.

[0040] FIG. 32 is an exemplary diagram showing some information decryption units according to one embodiment of the present disclosure.

[0041] FIG. 33 is an exemplary diagram showing a remaining information decryption unit according to one embodiment of the present disclosure.

[0042] FIG. 34 is a flowchart for explaining an image encoding method performed by an encoding device according to one embodiment of the present disclosure.

[0043] FIG. 35 is a flowchart illustrating an image decoding method performed by a decoding device according to one embodiment of the present disclosure.

[0044] The present invention is capable of various modifications. Furthermore, the present invention may have various embodiments. Specific embodiments are described in the accompanying drawings and detailed description.

[0045] 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 as embodiments.

[0046] 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.

[0047] 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.

[0048] 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.

[0049] 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.

[0050] 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.

[0051] 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.

[0052] 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.

[0053] 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.

[0054] 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.

[0055]

[0056] Interchange between terms in the examples

[0057] 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.

[0058] - 'one or more', 'at least one'

[0059] - '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'.)

[0060] - 'Information', 'Signal'

[0061] - 'value', 'predefined value', 'specific value', 'threshold', 'threshold value', 'baseline value', 'reference value'

[0062] - 'statistical value', 'statistics value'

[0063] - 'indicator', 'index', 'index', 'flag', 'information'

[0064] - 'encoder', 'encoding apparatus'

[0065] - 'decoder', 'decoding apparatus'

[0066] - 'Entropy encoding', 'encoding', 'encoding'

[0067] - 'Entropy decryption', 'decoding', 'decoding'

[0068] - 'coding', 'encoding and / or decoding'

[0069] - 'video', 'moving picture', 'image', 'picture', 'frame', 'screen'

[0070] - 'Reference picture', 'Reference video'

[0071] - 'Reference Picture List (RPL),' 'Reference Image List'

[0072] - 'original', 'input', 'source'

[0073] - 'Block', 'Unit', 'Signal'

[0074] - 'square', 'square shape'

[0075] - 'pixel', 'pixels', 'samples', 'pels'

[0076] - 'region', 'area', 'part', 'segment'

[0077] - 'partition', 'split', 'divide'

[0078] - 'quad', 'quarternary'

[0079] - 'luma component', 'luma', 'luminance component', 'luminance', 'Y'

[0080] - 'chroma component', 'chroma', 'chrominance', 'chrominance component', 'Cb and Cr', 'Cb or Cr', 'Cb', 'Cr', 'U and V', 'U or V', 'U', 'V'

[0081] - 'target', 'current' (e.g. target block and current block, or target image and current image)

[0082] - 'neighbor', 'neighboring', 'adjacent', 'neighbor / neighboring' (e.g., neighboring block, adjacent block, and surrounding block)

[0083] - 'collocated', 'collected'

[0084] - 'reconstruction', 'reconstruction', 'decoding'

[0085] - 'reconstructed', 'reconstructed', 'decoded'

[0086] - 'difference', 'difference', 'difference', 'error', 'residual', 'residual'

[0087] - Largest Coding Unit (LCU), Coding Tree Unit (CTU)

[0088] - 'inter', 'inter-screen'

[0089] - 'Inter prediction', 'inter prediction', 'motion compensation'

[0090] - 'Inter mode', 'Inter prediction mode', 'Inter-screen mode', 'Inter-screen prediction mode'

[0091] - 'Motion vector', 'Predicted motion vector', 'Advanced Motion Vector Prediction (AMVP)'

[0092] - 'list', 'candidate list'

[0093] - 'Spatial candidate', 'Spatial merge candidate'

[0094] - 'Temporal candidate', 'Temporal merge candidate'

[0095] - 'Prediction motion vector candidate', 'motion vector predictor'

[0096] - 'Prediction method', 'Prediction mode'

[0097] - 'Intra', 'Intra'

[0098] - 'Intra prediction', 'Intra prediction'

[0099] - 'Intra mode', 'Intra prediction mode'

[0100] - 'Dequantization', 'scaling'

[0101] - 'Quantization matrix', 'Scaling list'

[0102] - 'Quantization matrix coefficients', 'matrix coefficients'

[0103] - 'Transform coefficient level', 'quantized level', 'quantized coefficient', 'quantized transform coefficient', 'quantized transform coefficient level'

[0104] - 'Dequantized coefficient', 'dequantized transform coefficient'

[0105] - 'Scanning type', 'Scanning direction'

[0106] - 'Directional mode', 'Angle mode', 'Angular mode', 'Intra prediction mode'

[0107] - '(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'

[0108] - 'Merge Mode', 'Movement Merge Mode'

[0109] - 'Geometric Partitioning Mode (GPM)', 'Triangle Partitioning Mode'

[0110] 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.

[0111]

[0112] The range of information and values ​​of information described in the examples

[0113] 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'.

[0114] Information can have one of multiple values. 'n-th value' can mean the nth value among multiple values.

[0115] 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.

[0116] 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.

[0117] 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.

[0118]

[0119] Coding related concepts

[0120] Below, concepts related to coding are described. The descriptions disclosed below can be applied to embodiments.

[0121] 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'.

[0122] - 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.

[0123] - 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.

[0124] - The above description of the predefined values ​​can also be applied to predefined information. In the above descriptions, 'value' can be replaced with 'information'.

[0125] 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."

[0126] 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.

[0127] - 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.

[0128] Coding: Coding can mean encoding and / or decoding of images.

[0129] 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.

[0130] 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."

[0131] - A picture can mean the entire picture, or it can mean a part of a picture, such as a block.

[0132] 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.

[0133] Subpicture: A picture can be divided into one or more subpictures.

[0134] - A subpicture may be a square or rectangular area within a picture. A subpicture may contain one or more CTUs.

[0135] - 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.

[0136] - 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.

[0137] 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.

[0138] 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.

[0139] CTU: An image can be divided into multiple coding tree units (CTUs).

[0140] - 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.

[0141] - 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.

[0142] CTB: CTB can refer to one of Y CTB, Cb CTB, and Cr CTB.

[0143] 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.

[0144] - 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.

[0145] - 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.

[0146] - A unit can contain syntax elements. In other words, a block and its syntax elements can be combined to form a unit.

[0147] - 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.

[0148] - 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.

[0149] - 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.

[0150] - 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.

[0151] - Unit information may include unit type, unit size, unit depth, unit encoding order, and unit decoding order.

[0152] 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.

[0153] 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.

[0154] - 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.

[0155] - 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.

[0156] - 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.

[0157] 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.

[0158] 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.

[0159] - 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.

[0160] 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.

[0161] - 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.

[0162] Parameter set: A parameter set may correspond to header information among the structures within a bitstream.

[0163] - 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).

[0164] - 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.

[0165] - 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.

[0166] MPM (Most Probable Mode): MPM can indicate the intra prediction mode that is likely to be used for intra prediction for the target block.

[0167] - 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.

[0168] - 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.

[0169] MPM List: An MPM list may contain one or more MPMs. The number of MPMs in an MPM list may be predefined.

[0170] 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.

[0171] MPM Usage Directive: The MPM usage directive can indicate whether the MPM list is used for prediction on the target block.

[0172] 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.

[0173] Reference image list: The reference image list may be a list containing one or more reference images used for prediction for the target block.

[0174] - There may be multiple reference image lists. Multiple reference image lists may include List 0 (L0), List 1 (L1), etc.

[0175] - 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.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] Reference sample: A reference sample may be a sample that is referenced for encoding / decoding of a target block, such as prediction and filtering.

[0180] 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.

[0181] 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.

[0182] - 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'.

[0183] 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.

[0184] - 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.

[0185] - 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.

[0186] Picture Order Count (POC): The POC of a picture can indicate the display order or output order of the picture.

[0187] 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.

[0188] - 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.

[0189] 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.

[0190] - 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.

[0191] -The zero vector can be (0, 0) MV.

[0192] 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.

[0193] - 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.

[0194] -The zero vector can be (0, 0) BV.

[0195] 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.

[0196] 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.

[0197] Motion information candidate list: The motion information candidate list may mean a list constructed using one or more motion information candidates.

[0198] 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.

[0199] - 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.

[0200] - 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.

[0201] - 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.

[0202] 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.

[0203] 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.

[0204] - 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.

[0205] Merge Candidate List: A merge candidate list may be a list constructed using one or more merge candidates.

[0206] 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.

[0207] 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.

[0208] Spatial neighboring blocks: Spatial neighboring blocks can be blocks that are spatially adjacent to the target block.

[0209] - The target block and spatial neighboring blocks can be included within the target image.

[0210] - 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.

[0211] - A spatial neighboring block may include a block diagonally adjacent to a vertex of the target block.

[0212] - 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.

[0213] Temporal neighboring blocks: Temporal neighboring blocks can be blocks that are temporally adjacent to the target block.

[0214] - 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.

[0215] - 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.

[0216] - 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.

[0217] - 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.

[0218] - A temporal neighboring block may be a block that is temporally adjacent to a spatial neighboring block of the target block.

[0219] 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.

[0220] 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.

[0221] 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.

[0222] Quantized level: A quantized level can be an integer quantity used as input to dequantization.

[0223] 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.

[0224] 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.

[0225] 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.

[0226] 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.

[0227] 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.

[0228] Quantization matrix coefficients: Quantization matrix coefficients can be each element within a quantization matrix.

[0229] 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.

[0230] 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.

[0231] 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.

[0232] 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.

[0233] - 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.

[0234] - 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.

[0235] - 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.

[0236] 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).

[0237] 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.

[0238] 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.

[0239] 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.

[0240] - 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.

[0241] Entropy decoding: Entropy decoding can reverse the processes performed in entropy encoding. Symbols can be generated by entropy decoding a bitstream.

[0242] 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.

[0243] 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.

[0244]

[0245] Coding parameters

[0246] 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.

[0247] 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.

[0248] 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.

[0249] 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.

[0250] 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.

[0251] 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.”

[0252] 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.”

[0253] 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.

[0254] 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.”

[0255] 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.”

[0256] 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.

[0257] - 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.

[0258] - 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.

[0259] - 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.

[0260] - P slices and B slices encoded and / or decoded using reference slices can be considered as images in which inter prediction is used.

[0261]

[0262] System for video coding

[0263] Figure 1 illustrates a system for video coding according to one embodiment.

[0264] The system (100) may include at least one of an encoding device (110) and a decoding device (150).

[0265] Each of the encoding device (110) and the decoding device (150) may be a computer or an electronic apparatus.

[0266]

[0267] Structure of the encoding device

[0268] The encoding device (110) may include a processor (120), storage (140), and a communicator (149).

[0269] The processor (120), storage (140), and communication device (149) can be connected via a bus.

[0270] 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.

[0271] 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.

[0272] 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).

[0273] 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).

[0274] 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.

[0275] 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).

[0276] 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).

[0277] 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.

[0278]

[0279] Operation of the encoding device

[0280] The encoding device (110) can sequentially encode one or more images of a video.

[0281] The storage (140) can store the original image. The original image can be used as a target image in the encoding device (110).

[0282] 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).

[0283] The segmenter (122) can determine a target block by performing segmentation on the target image.

[0284] 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.

[0285] 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.

[0286] 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.

[0287] 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.

[0288] 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.

[0289] 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.

[0290] The transformer (125) can perform a transformation on the residual block to generate transformation coefficients.

[0291] The converter (125) can perform the conversion using one of a plurality of conversion methods.

[0292] 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.

[0293] 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.

[0294] 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.

[0295] 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.

[0296] Information for decoding an image may include quantized levels and syntax elements produced by a quantizer (126).

[0297] The probability distribution can be determined based on the quantized levels and coding parameters.

[0298] 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.

[0299] 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.

[0300] The dequantizer (127) can generate dequantized transform coefficients by performing dequantization on the quantized level.

[0301] 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.

[0302] The adder (129) can generate a restored block by combining a predicted block and a restored residual block.

[0303] 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.

[0304] 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.

[0305] 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).

[0306]

[0307] Structure of the decryption device

[0308] The decryption device (150) may include a processor (160), a storage (180), and a communication device (189).

[0309] 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.

[0310] 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).

[0311] The storage (180) may include a reference picture buffer (181).

[0312] 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).

[0313] 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.

[0314]

[0315] Operation of the decryption device

[0316] 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).

[0317] 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).

[0318] The processor (160) can obtain a bitstream from a storage (180) or a computer-readable recording medium.

[0319] A bitstream may contain encoded information.

[0320] 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.

[0321] Information for decoding an image may include quantized levels and syntax elements.

[0322] 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.

[0323] The entropy decoder (161) can provide syntax elements to other components of the processor (160), such as the segmenter (162).

[0324]

[0325] A common description of the relationship between the components of the encoding device and the components of the decoding device.

[0326] 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.

[0327] 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.

[0328] 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).

[0329]

[0330] Division of the units that make up the image

[0331] Figure 2 shows a segmentation structure of an image according to one embodiment.

[0332] Figure 2 can schematically represent an example in which one unit is divided into multiple sub-units.

[0333] 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.

[0334] 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).

[0335] 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.

[0336] 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.

[0337] 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.

[0338] 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.

[0339] 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'.

[0340] 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.

[0341] 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.

[0342] 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.

[0343] 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.

[0344] 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.

[0345] 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.

[0346] 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.

[0347] 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.

[0348] 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.

[0349] 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.

[0350] 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.

[0351] Here, each of the aforementioned split direction information and split type information may be a flag having a specific length (e.g., 1 bit).

[0352] The CU's partition information may also include QT partition information, partition direction information, and partition shape information.

[0353] 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.

[0354] 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.

[0355] 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.

[0356] 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.

[0357] The BT maximum size may represent the maximum size of the CU corresponding to each node of the BT, and the TT maximum size may represent the maximum size of the 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 the CU.

[0358] 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.

[0359] 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.

[0360] 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.

[0361] 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.

[0362]

[0363] Processing blocks according to their properties

[0364] 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.

[0365] The blocks to which the specific processing described in the examples is applied may have a square shape or a non-square shape.

[0366] 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.

[0367] 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.

[0368] In one embodiment, a minimum block size and / or a maximum block size for a particular process may be predefined.

[0369] 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.

[0370] 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.

[0371] In one embodiment, the processing of the embodiment may be applied / performed only when the block size is a predefined block size.

[0372] 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.

[0373] 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.

[0374] 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.

[0375]

[0376] Predictive information for prediction

[0377] Prediction information can be used to generate a prediction block for the target block.

[0378] 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.

[0379] 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.

[0380]

[0381] Intra prediction

[0382] Figure 3 illustrates the structure of intra prediction according to one embodiment.

[0383] 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.

[0384] 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.

[0385] 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.

[0386] 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.

[0387] 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.

[0388] 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.

[0389] 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.

[0390] 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.

[0391] 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).

[0392] 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.

[0393] 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.

[0394] 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.

[0395] 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.

[0396] 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.

[0397] 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.

[0398] 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.

[0399] 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.

[0400] 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.

[0401] 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.

[0402] 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.

[0403] 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.

[0404] 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.

[0405] 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.

[0406] 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.

[0407] Instead of obtaining samples from the reconstructed neighboring blocks, the samples of segment A and segment F can be derived using padding using the nearest samples of segment B and segment E, respectively.

[0408] 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.

[0409] 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.

[0410] 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.

[0411] 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.

[0412] 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.

[0413] 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.

[0414] 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.

[0415] 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.

[0416]

[0417] Inter prediction

[0418] Figure 4 shows the structure of inter prediction to explain the inter prediction process according to one embodiment.

[0419] The rectangle illustrated in Fig. 4 can represent an image. Additionally, the arrow in Fig. 4 can represent a prediction direction.

[0420] 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.

[0421] 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.

[0422] 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.

[0423] 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.

[0424] 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.

[0425] Below, inter prediction for a target block in inter mode according to an embodiment is specifically described.

[0426] 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.

[0427] 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.

[0428] 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.

[0429] 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.

[0430] 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.

[0431] A spatial candidate may be a restored spatial neighboring block that is spatially adjacent to the target block.

[0432] 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.

[0433] 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.

[0434] A temporal candidate may be a restored temporal neighboring block corresponding to a target block in a restored COL image.

[0435] 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.

[0436] 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.

[0437] A temporal candidate may be a location inside and / or outside a call block within a call image.

[0438] 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.

[0439] 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.

[0440] 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.

[0441] 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.

[0442]

[0443] AMVP mode

[0444] 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.

[0445] 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.

[0446] 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.

[0447] 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. Inter prediction information may be signaled from an encoding device (110) to a decoding device (150) in the form of a bitstream.

[0448] 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.

[0449] 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.

[0450] 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.

[0451] 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.

[0452]

[0453] Merge mode

[0454] 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.

[0455] 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.

[0456] 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.

[0457] An average merge candidate may be a merge candidate generated based on the average of two merge candidates in the merge candidate list.

[0458] A zero merge candidate may be zero vector motion information. Zero vector motion information may be motion information whose MV is a zero vector.

[0459] 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.

[0460] 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.

[0461] 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.

[0462] 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.

[0463] 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.

[0464] 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.

[0465] 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.

[0466] 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.

[0467] 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.

[0468] 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.

[0469] 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.

[0470] 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.

[0471] 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.

[0472] 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.

[0473] 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.

[0474] 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.

[0475] 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.

[0476] 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.

[0477] In the aforementioned modes, compensation for prediction samples derived through inter prediction can be performed using optical flow.

[0478]

[0479] Figure 5 shows the order in which spatial candidates are added to the candidate list according to one embodiment.

[0480] In Fig. 5, the locations of spatial candidates are shown.

[0481] The large block in the center can represent the target block. The five smaller blocks adjacent to the target block can represent spatial candidates.

[0482] The coordinates of the target block can be (xP, yP), and the size of the target block can be (nPSW, nPSH).

[0483] 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).

[0484] 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).

[0485] 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).

[0486] 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).

[0487] 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).

[0488] 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.

[0489] 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.

[0490] 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."

[0491] 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.

[0492]

[0493] IBC mode

[0494] 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.

[0495] 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.

[0496] 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.

[0497] 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.

[0498] 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.

[0499] In AMVP mode, BVD can be used. The description of MVD in the embodiments can also be applied to BVD.

[0500] 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.

[0501]

[0502] Transformation and quantization

[0503] 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.

[0504] 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.

[0505] 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.

[0506] 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.

[0507] 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.

[0508] 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.

[0509] 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).

[0510] 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).

[0511] NSPT can be applied to specific block sizes such as 4x4, 4x8, 8x4, 4x16, 16x4, 8x8, 8x16, and 16x8 for intra coding.

[0512] 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.

[0513] 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).

[0514] 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.

[0515] 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.

[0516] 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.

[0517] 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.

[0518] 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.

[0519] 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.

[0520] 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.

[0521] 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.

[0522] 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.

[0523] 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.

[0524] 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.

[0525] 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.

[0526] 변환 계수들 또는 양자화된 레벨들은 대각 스캐닝, 수직 스캐닝 및 수평 스캐닝과 같은 스캐닝 타입들 중 적어도 하나의 스캐닝 타입에 따라서 스캐닝 될 수 있다. 대각 스캐닝은 우-상단 대각 스캐닝 또는 좌-하단 대각 스캐닝일 수 있다.

[0527] 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.

[0528] 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.

[0529] Scanning for each scanning type can start at a specific starting point and end at a specific ending point.

[0530] 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.

[0531] 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.

[0532] 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.

[0533] 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.

[0534] 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.

[0535]

[0536] Filtering

[0537] To improve the image quality, filtering may be performed on blocks. The values ​​of target samples may be determined or updated through filtering.

[0538] 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.

[0539] 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.

[0540] 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.

[0541] 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.

[0542] 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.

[0543] 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.

[0544] 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.

[0545] The filter coefficients can be coefficients or weights of the input samples.

[0546] 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.

[0547] 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.

[0548] Filtering may include filtering performed by predictor (123) and predictor (163), etc.

[0549] 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.

[0550] 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.

[0551] 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.

[0552] 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.

[0553] 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.

[0554] Filtering may include in-loop filtering performed by filter (130) and filter (170), etc.

[0555]

[0556] Figure 6 illustrates multiple in-loop filters according to an example.

[0557] 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).

[0558] 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.

[0559] 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).

[0560] The target block can represent an image input to the filter. The filtered target block can represent an image output from the filter.

[0561] 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.

[0562] Luma signal mapping can perform codeword redistribution for the luma signal.

[0563] 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.

[0564] Chroma scaling can correct chroma signals based on the correlation between a luma signal and a corresponding chroma signal.

[0565] 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.

[0566] 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.

[0567] 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.

[0568] 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.

[0569] 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.

[0570] 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.

[0571] 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.

[0572] 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.

[0573] 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.

[0574] 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.

[0575] 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.

[0576] 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.

[0577] ALF can compensate for distortion between the restored image and the original image.

[0578] The filter coefficients of ALF can be signaled via the bitstream.

[0579] 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.

[0580] 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.

[0581] 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.

[0582] Information regarding whether ALF applies can be signaled for specific units, such as CTB.

[0583] 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.

[0584]

[0585] Entropy encoding and entropy decoding

[0586] Figure 7 illustrates entropy encoding and entropy decoding according to an example.

[0587] The processes of entropy encoding by the entropy encoder (139) are illustrated at the top of Fig. 7.

[0588] 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.

[0589] 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.

[0590] Information about syntactic elements and bins can be provided from the binarization unit to the context selection unit.

[0591] A context modeler can perform context updates.

[0592] Context can mean occurrence probability information for each bin for syntactic elements that have already been encoded.

[0593] 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.

[0594] 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.

[0595] The updated context can be used for entropy encoding of syntactic elements of the target block.

[0596] 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.

[0597] The processes of entropy decryption by the entropy decoder (161) are shown at the bottom of Fig. 7.

[0598] 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.

[0599] A context modeler can perform context updates.

[0600] Context can mean the occurrence probability information of each bin for syntactic elements that have already been decoded.

[0601] 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.

[0602] 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.

[0603] The updated context can be used for entropy decoding of syntactic elements of the target block.

[0604] 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.

[0605] 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.

[0606] Information about syntactic elements and bins can be provided from the de-binarization unit to the context selection unit.

[0607] A syntax element may be one of the coding parameters described in the embodiments.

[0608]

[0609] Methods for binarization, debinarization, entropy encoding, and entropy decoding

[0610] 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.

[0611] - Signed 0-th order Exponential Golomb binarization / debinarization method (abbreviated as se(v))

[0612] - k-order exponential-Golomb binarization / inverse binarization method with sign (abbreviated as sek(v))

[0613] - 0-order exponent-Golomb binarization / inverse binarization method for unsigned positive integers (abbreviated as ue(v))

[0614] - k-order exponential-Golomb binarization / inverse binarization method for unsigned positive integers (abbreviated as uek(v))

[0615] - Fixed-length binarization / debinarization method (abbreviated as f(n))

[0616] - Truncated Rice binarization / debinarization method or truncated unary binarization / debinarization method (abbreviated as tu(v))

[0617] - Truncated binary binarization / debinarization method (abbreviated as tb(v))

[0618] - Context-adaptive arithmetic encoding / decoding method (abbreviated as ae(v))

[0619] - bit string in bytes (abbreviated as b(8))

[0620] - Signed integer binarization / debinarization method (abbreviated as i(n))

[0621] - Unsigned positive integer binarization / debinarization method (abbreviated as u(n)) ('u(n)' can also mean fixed-length binarization / debinarization method.)

[0622] - Unary binarization / inverse binarization method

[0623]

[0624] Adaptive execution of the processes of the embodiments

[0625] 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.

[0626] 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).

[0627] 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.

[0628] The processing of the embodiments can be selectively applied / performed based on specific conditions or specific targets.

[0629] 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.

[0630] 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.

[0631] 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.

[0632] 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.

[0633]

[0634] Meanwhile, the encoding device and decoding device may also be implemented based on neural networks.

[0635] Fig. 8 is a diagram illustrating a still image encoding device and decoding device implemented based on a neural network.

[0636] Referring to Fig. 8, the encoding device converts an input image x to be encoded into a latent representation y through a latent representation transformation neural network. The latent representation may be replaced with terms such as a latent vector or a latent feature map.

[0637] The encoding device generates a latent representation probability model for the latent representation y through a latent representation transformation neural network. A bitstream is generated by performing entropy encoding based on the base and the bitstream is transmitted to the decoding device. Each component y of the latent representation y i may be quantized prior to entropy encoding.

[0638] The decoding device receives the bitstream from the encoding device and restores the latent representation or quantized latent representation through entropy decoding based on the latent representation probability model. The latent representation probability model is shared between the video encoding device and the video decoding device.

[0639] Afterwards, the decryption device inputs the restored latent representation into the latent representation restoration neural network to restore the restored image. Prints out.

[0640] The blocks indicated by dotted lines in Figure 8 may represent modules or neural networks containing trainable parameters. The neural network or the parameters constituting the neural network can be trained using a loss function and the backpropagation algorithm.

[0641] The loss function is the MSE (Mean Square Error) between the input image and the restored image and the latent expression probability model. Entropy calculated through It includes both. MSE represents the quality of the restored image, and entropy can represent the bit quantity. An optimization method that considers both the quality and bit quantity of the restored image is called rate-distortion optimization. Here, E stands for Expectation.

[0642] Below, each component illustrated in Fig. 8 is described in detail.

[0643] Transformation neural networks and restoration neural networks

[0644] Figure 9 illustrates an example of a latent representation transformation neural network, and Figure 10 illustrates an example of a latent representation restoration neural network.

[0645] A latent representation transformation neural network is a neural network that receives an image as input and outputs a latent representation, and a latent representation restoration neural network is a neural network that receives a quantized latent representation as input and outputs a restored image.

[0646] Typically, the input image is a three-dimensional image information consisting of three channels of RGB or YCbCr and height and width, and the latent representation can be a feature map with a larger number of channels (e.g., N=128 or 192, etc.) and smaller height and width compared to the input image.

[0647] Referring to Fig. 9, the transformation neural network (g a ) can be composed of four convolution layers and three GDN (Generalized Divisive Normalization) layers.

[0648] Each convolutional layer consists of 5x5 kernels and can have a stride of 2. The GDN layer performs normalization on the components of the latent representation output from the convolutional layer to approximate a Gaussian distribution. The GDN layer is a nonlinear layer that exhibits high efficiency in shallow neural networks.

[0649] Referring to Fig. 10, a restoration neural network (gs) can be constructed by replacing the convolutional layer in the transformation neural network with a transposed convolutional layer and replacing the nonlinear layer GDN with an inverse generalized divisive normalization (IGDN).

[0650] The transformation neural network and restoration neural network illustrated in FIGS. 9 and 10 can be configured using neural network structures such as a fully-connected layer, a vision transformer, etc., in addition to a convolutional layer.

[0651] The kernel and stride sizes used in the convolutional layers of the transformation neural network and the restoration neural network and the number of convolutional layers and GDN layers can be configured in various ways considering rate-distortion optimization and / or hardware resources.

[0652] Additionally, in a structure using multiple convolutional layers, different sizes of kernels and / or different sizes of strides may be used for each convolutional layer, or at least one convolutional layer may use different sizes of kernels and / or different sizes of strides from other convolutional layers.

[0653] Latent representation quantization

[0654] Quantization is the mapping of values ​​within a specific range to a specific value, and can be used to limit the number of values ​​that can be represented.

[0655] Quantization can generally be categorized into uniform and non-uniform quantization. Uniform quantization can refer to a quantization method that divides the range of values ​​into multiple even intervals and maps the values ​​within each even interval to a specific value corresponding to that interval. Non-uniform quantization can refer to a quantization method that divides the range of values ​​into multiple even intervals and maps the values ​​within each interval to a specific value corresponding to that interval.

[0656] Neural network-based image encoding / decoding may also involve quantizing the values ​​of latent representation components. In this example, since the transform neural network guarantees nonlinear transformation, uniform quantization can be used as a quantization method.

[0657] To enable the use of the backpropagation algorithm, quantization at training time can be replaced by adding uniform noise between -0.5 and 0.5, as shown in Equation 1. Quantization at inference time can be implemented using a rounding function, as shown in Equation 2.

[0658] [Mathematical Formula 1]

[0659]

[0660] [Equation 2]

[0661]

[0662] In equations 1 and 2, is a latent expression with added uniform noise, is a quantized latent representation, u(a, b) is a uniform distribution between a and b, and Round( ) represents a rounding operation.

[0663] Entropy encoding and decoding of latent representations

[0664] Entropy coding is a coding method that uses the probability distribution of random variables, and an example of this is arithmetic coding.

[0665] The encoding device and the decoding device perform entropy encoding and entropy decoding, respectively, on the quantized latent representation.

[0666] Entropy encoding and decoding of latent representation is done using latent representation probability model (i.e. entropy model) It is done through, can be designed as a probability distribution estimated through a module or neural network containing learnable parameters.

[0667] The latent expression probability model is a model that represents each component y of the latent expression. i Assuming that are independent of each other, it can be designed as a product of the probability distributions of each component, as in mathematical equation 3.

[0668] [Equation 3]

[0669]

[0670] The probability model of the latent expression with uniform noise added within (-1 / 2, 1 / 2) is as in Equation 4, with the probability distribution of each component and the uniform distribution It can be expressed as a convolution operation between. The probability model of the quantized latent representation is also expressed as Equation 5, with the probability distribution of each component and the uniform distribution. It can be expressed as a convolution operation between .

[0671] [Equation 4]

[0672]

[0673] [Equation 5]

[0674]

[0675] In mathematical expressions 4 and 5, * denotes a convolution operation.

[0676] Hereinafter, unless otherwise specified, the latent representation probability model means the probability model of the quantized latent representation.

[0677] Latent expression probability model

[0678] Latent expression probability models can be divided into input image-independent probability models and input image-dependent probability models.

[0679] A probabilistic model that is independent of the input image is a probabilistic model that applies equally to all input images. That is, the same probabilistic model is used for all input images.

[0680] An input image-dependent probabilistic model is a probabilistic model estimated differently depending on the input image. In other words, different probabilistic models are used depending on the characteristics of the input image. An input image-dependent probabilistic model can be adaptively estimated based on the input image using an additional neural network structure.

[0681] Latent expression probability models can be designed as parametric or non-parametric models. Parametric models can utilize distributions such as the Gaussian or Laplacian distributions, while non-parametric models can be implemented by appropriately utilizing the results of arithmetic operations on learnable parameters and nonlinear functions.

[0682] Figure 8 can be viewed as an example of a case where the probability model of the latent expression is independent of the input image.

[0683] Mathematical expression 6 is an example of a latent expression probability model, which is a latent expression probability model designed with a Laplace distribution using learnable parameters μ and σ.

[0684] [Equation 6]

[0685]

[0686] In mathematical expression 6, Lap represents the Laplace distribution, and μ and σ represent the mean and scale information, which are parameters of the Laplace distribution, and can have the same dimension as the latent representation. are each refers to each component of

[0687] Figure 11 is an example diagram for explaining a neural network-based encoding / decoding structure using a hidden vector probability model dependent on an input image.

[0688] The structure illustrated in FIG. 11 may additionally include a hyperprior information transformation neural network, a hyperprior information probability model, and a hyperprior information restoration neural network in addition to the structure illustrated in FIG. 8.

[0689] If the latent representation probability model is dependent on the input image, additional information for estimating the latent representation probability model for the current input image must be generated and transmitted to the decoding device. This additional information is called hyper-prior information.

[0690] The super-lexical information transformation neural network takes latent representations as input and super-lexical information It is a neural network that outputs the super-dictionary information, and the super-dictionary information restoration neural network is a quantized super-dictionary information It is a neural network that receives as input the parameters of the latent expression probability model and outputs the parameters of the latent expression probability model. The parameters of the latent expression probability model are, for example, the average and standard deviation It could be.

[0691] The super-lexical information can be quantized and entropy-encoded by the encoding device and transmitted to the decoding device. The encoding device uses a super-lexical information probability model for entropy encoding. The super-lexical information probability model can use the previously described input image-independent latent representation probability model.

[0692] FIG. 12 and FIG. 13 are exemplary diagrams showing a super-prediction information transformation neural network and a super-prediction information restoration neural network, respectively, according to one embodiment of the present disclosure.

[0693] Referring to FIG. 12, the super-precision information transformation neural network may include an absolute value function that performs an absolute value operation on an input hidden vector, one convolutional layer with a 3x3 kernel and a stride of 1, two convolutional layers with a 5x5 kernel and a stride of 2, and a nonlinear layer connecting the convolutional layers, such as a Rectified Linear Unit (ReLU).

[0694] Referring to Fig. 13, the hyper-lexical information restoration neural network may include two pre-convolutional layers with 5x5 kernels and a stride of 2, one pre-convolutional layer with 3x3 kernels and a stride of 1, and a non-linear layer connecting the pre-convolutional layers, for example, ReLU. The hyper-lexical information restoration neural network may be implemented by excluding the absolute value operation from the hyper-lexical information transformation neural network and adding one ReLU layer instead.

[0695] The number of convolutional layers that constitute the super-precision information transformation neural network and the restoration neural network, as well as the kernel size and stride value used in the convolutional layers, can be changed as needed, and the nonlinear layer can also be used by selecting one or more of several nonlinear activation functions.

[0696] The neural network-based encoding or decoding described above utilizes at least one neural network. For example, the at least one neural network may include a latent representation transformation neural network for transforming an image-related target signal into a latent representation, a restoration neural network for restoring the target signal from the latent representation, and the like. Furthermore, the at least one neural network may include a latent representation probability model for entropy encoding or decoding the latent representation, a super-dictionary information transformation neural network, and a super-dictionary information restoration neural network.

[0697] These neural networks can be trained using a backpropagation algorithm, which updates the parameters or weights of the neural network in a way that minimizes a specific loss function calculated from training data.

[0698] In particular, neural networks for image encoding are trained end-to-end to simultaneously minimize the distortion between the input image and the restored image and the bit rate of the bitstream transmitted from the encoding device to the decoding device, and this optimization method can be called rate-distortion optimization.

[0699] In general, distortion can be measured using the Mean Square Error (MSE) or Multi-Scale Structural Similarity Index Measure (MS-SSIM) between the input image and the restored image, and the bit rate can be measured using the entropy (i.e., average information amount) calculated by the latent representation probability model. can be used.

[0700] Additionally, a constant λ is used to determine the ratio between distortion and bit rate, and by using λ as in Equation 7, the desired image quality level can be determined. Generally, the larger λ is, the higher the image quality level.

[0701] [Equation 7]

[0702]

[0703] Here, L is the loss function, is the restored image, and d is a distortion function such as MSE or MS-SSIM.

[0704] In particular, when the probability model of the latent representation is dependent on the input image, an additional entropy term for transmitting the super-predicate information can be included in the loss function, as in Equation 8.

[0705] [Equation 8]

[0706]

[0707] The above-described neural network-based still image encoding / decoding method can be applied to video encoding / decoding together with a method for removing temporal redundancy.

[0708] Temporal redundancy can be removed by using previously restored reference images, and images encoded using these reference images are called P-images or B-images. A P-image is an image encoded using one reference image, and a B-image is an image encoded using two reference images. On the other hand, an image encoded without using a reference image is called an I-image, and the encoding of an I-image can be performed in the same way as the encoding of a still image.

[0709] Neural network-based video encoding methods can be divided into methods that reduce temporal redundancy in image space (i.e., image domain) and methods that reduce temporal redundancy in latent representation space.

[0710] The former follows a similar structure to traditional video encoding methods. The latter is a novel method applicable to neural network-based video encoding, addressing error propagation, a problem in video encoding. Error propagation refers to the accumulation of image quality degradation as a reconstructed image is repeatedly used as a reference image.

[0711] FIG. 14 is a diagram illustrating a neural network-based video encoding device that uses a method for reducing temporal redundancy in image space.

[0712] The neural network-based video encoding device illustrated in FIG. 14 may include a neural network for estimating optical flow, a neural network for encoding and decoding motion vectors, a neural network for motion compensation, and a neural network for encoding and decoding residual images.

[0713] The encoding device is the current video Wow ki restored reference video is input into the optical flow neural network (Optical Flow Net in Fig. 14), and the motion vector is output from the neural network. Obtain .

[0714] movement vector is input to the motion vector encoding neural network (MV Encoder Net of Fig. 14) and the motion vector is converted accordingly. , that is, a latent representation for the motion vector is generated. The latent representation for the motion vector is quantized and then entropy encoded. For this purpose, the latent representation transform neural network, quantization, and latent representation entropy encoding described in the encoding of still images can be used.

[0715] Meanwhile, the encoding device uses a motion vector decoding neural network (MV Decoder Net of Fig. 14) to decode the motion vector from the latent representation for the motion vector. The motion vector decoding neural network can be implemented with the structure of the latent representation restoration neural network described above.

[0716] The encoding device inputs the restored motion vector into the motion compensation neural network (Motion Compensation Net in Fig. 14). The motion compensation neural network uses the restored reference image and the restored motion vector to predict the image. creates .

[0717] The encoding device is the current video and prediction video The residual image, which is the difference between Input the residual image transformation neural network (Residual Encoder Network), and generate a latent representation for the residual image accordingly. Obtained. The latent representation for the residual image is quantized and then entropy encoded. For this purpose, the latent representation transform neural network, quantization, and latent representation probability model described in the encoding of still images can be used.

[0718] Meanwhile, the encoding device can perform the following decoding process.

[0719] As mentioned above, the encoding device derives the motion vector from the latent representation for the motion vector through the motion vector decoding neural network. Restore the motion vector restored by the motion compensation neural network and the predicted image using the restored reference image. can be created.

[0720] Meanwhile, the latent representation for the residual image is derived through the residual decoding neural network (Residual Decoder Net in Figure 14). is restored. The residual decoding neural network can be implemented with the structure of the aforementioned latent representation restoration neural network.

[0721] The encoding device restores the current image by adding the restored residual image and the predicted image. That is, the restored image is expressed by the formula can be ultimately generated by.

[0722] Meanwhile, the decoding device decode the quantized latent representation from the bitstream from the encoding device through entropy decoding. and motion vector , and the current image can be restored through the same decoding process as the encoding device.

[0723]

[0724] Hereinafter, an improved image encoding and decoding method implemented using the aforementioned encoding device and decoding device is described.

[0725] Conventional block-based image compression techniques predictively encode the target block based on reconstructed surrounding samples. However, when predicting the block's interior using only surrounding samples is difficult, encoding performance deteriorates due to the large residual difference between the predicted result and the target block.

[0726] To address this issue, a method can be used to reduce the residual between the reconstructed surrounding samples and the partitioned block by recursively partitioning the target block. However, this method suffers from the problem of using additional bits to transmit block partition information.

[0727] The present disclosure provides a method for first encoding and decoding partial information within a target block, and then encoding remaining information within the target block (remaining information) based on the decoded partial information. According to the present disclosure, encoding efficiency can be improved even when it is difficult to predict the interior of a block using only surrounding samples.

[0728] Here, some information within the target block may refer to some samples or some areas within the target block. Encoding some information may refer to encoding some samples or samples within some areas.

[0729] Additionally, the remaining information within the target block may refer to an area within the target block excluding some of the above-mentioned areas. Encoding the remaining information may refer to encoding samples within the remaining area excluding some of the areas within the target block.

[0730] The method according to the present disclosure can be used when encoding is performed directly in the image domain (i.e., image space), and can also be used when encoding is performed after transforming an image into a latent representation space.

[0731] In this specification, these two cases are explained separately for convenience, but their essential feature is the same in that some information is encoded first and then the remaining information of the target block is encoded based on this.

[0732] For convenience of explanation, the following description assumes that encoding and decoding are performed on a block-by-block basis. Here, a block may be a CTU, CU, PU, ​​or TU as described above. However, the present invention is not limited thereto, and the unit of encoding and decoding may also be a tile, slice, or picture.

[0733] The components or "units" included in the devices illustrated in the drawings below may be implemented in hardware, software, or a combination of hardware and software. Furthermore, the functions of each component may be implemented in software, with a microprocessor executing the software functions corresponding to each component. At least some of the components or units may be implemented based on a neural network.

[0734]

[0735] 1. Encoding / decoding in image space

[0736] (1) Encoding in image space

[0737] FIG. 15 is an exemplary block diagram of an encoding device for encoding a target block in an image space according to one embodiment of the present disclosure.

[0738] A partial information determination unit (1510) selects the location of partial information to be encoded first among the information within the target block (X). The location information indicating the location of the selected partial information is encoded and transmitted to a decoding device.

[0739] For example, some information decision units (1510) may output a binary mask (m) that sets positions corresponding to some information as True and positions corresponding to the remaining information as False. The binary mask may be a two-dimensional map that assigns different first and second values ​​to positions corresponding to some information and positions corresponding to the remaining information.

[0740] Additionally, some information determination units (1510) may refer to the restored area to determine some information within the target block (X ref ) can also be used.

[0741] In embodiments of the present disclosure, the restored region may be an area belonging to the current picture to which the target block belongs or a region belonging to a reference picture restored before the current picture. For example, the restored region may be at least one restored block surrounding the target block within the current picture. At least one restored block may be located above and / or to the left of the target block.

[0742] Some information classification unit (1520) receives a target block and a binary mask as input, and selects some information (X) within the target block. key ) and the remaining information (X rest ) and transmits some information and remaining information to some information encoding unit (1530) and remaining information encoding unit (1540), respectively.

[0743] Some information encoding units (1530) perform encoding on some information and output some encoded information. Some information may be encoded on its own without prediction, or may be predictively encoded using reference information.

[0744] Additionally, some information encoding units (1530) restore some information and restore some of the information ( ) is printed.

[0745] The remaining information encoding unit (1540) can encode the remaining information using some of the decoded information. In addition, to encode the remaining information, reference information (X ref ) may be additionally used.

[0746] Information about the binary mask, some encoded information, and the remaining encoded information can be transmitted to the decryption device in the form of a bitstream.

[0747] Below, exemplary embodiments of the operations performed by each component of FIG. 15 are described in detail.

[0748] ■ Some information decision units

[0749] Some information determination units (1510) can determine the location of some information within a target block in the form of a binary mask and encode information representing the binary mask.

[0750] Some of the information may refer to one or more sub-blocks selected from among non-overlapping sub-blocks from the target block to be encoded. The remaining information may refer to unselected sub-blocks.

[0751] For example, a target block can be divided into sub-blocks in a grid shape with predetermined, regular intervals.

[0752] Figure 16 shows an example of dividing a target block having a height H and a width W into predetermined intervals N and selecting four pieces of information.

[0753] Each area divided by the grid becomes a sub-block of the target block.

[0754] Sub-blocks marked in white represent some selected information, and hatched sub-blocks represent the remaining information.

[0755] A binary mask is information that indicates the location of some information, and is information that allows you to distinguish between some information and the rest of the information.

[0756] Figure 17 is an example diagram illustrating a method of expressing a binary mask.

[0757] As shown in (A) of Fig. 17, the binary mask can be expressed as true (e.g., 1) or false (e.g., 0) whether each sample unit within the target block with height H and width W corresponds to some information. Alternatively, as shown in (B) of Fig. 17, the binary mask can be simply expressed by assigning true and false to sub-blocks constituting the target block.

[0758] In some embodiments, some information determination units (1510) may define multiple patterns for a binary mask and select a binary mask to be used for a target block from among the multiple patterns defined. In this case, index information indicating one of the multiple patterns defined as information for the binary mask may be encoded.

[0759] A method of determining a pattern with optimal encoding efficiency among several predetermined patterns and encoding the information can reduce the computational complexity required for binary mask determination and the number of bits of binary mask representation.

[0760] For example, if the encoding target block is divided into 16 sub-blocks as in Fig. 18, the number of possible cases in selecting a binary mask is 2. 16However, several patterns (e.g., three) can be predefined. By determining one pattern among the predefined patterns to be used for the target block and encoding index information indicating the determined pattern among the predefined patterns, the number of bits required to express the binary mask can be reduced.

[0761] Here's an example of how to determine one of the predefined patterns:

[0762] For each binary mask of a predetermined pattern, an encoding and decoding process for a target block as described below is performed using the corresponding binary mask, and accordingly, the amount of bits and distortion (distortion between the target block and the restored block) that occur when using the corresponding binary mask are measured.

[0763] The measured bit rate and distortion are input to the rate-distortion loss function to compute the cost for each pattern's binary mask. Here, the distortion can be calculated using, for example, the Hadamard difference (SATD).

[0764] As a result of the cost measurement based on the rate-distortion loss function, the binary mask of the pattern with the smallest cost can be determined as the binary mask used for encoding the target block.

[0765] In another embodiment, the binary mask for the target block may be determined using a neural network trained to predict a binary mask appropriate for the input block.

[0766] For example, some information decision units (1510) may determine binary masks using a transformer-based neural network.

[0767] Below, we will illustrate the use of a transformer neural network to generate binary masks. However, the present invention is not limited to this method, and other neural networks trained to predict binary masks for input blocks, as described above, may also be used.

[0768] Figure 19 is an example diagram to explain the process of generating a binary mask based on a transformer neural network.

[0769] Referring to Figure 19, each sub-block of the encoding target block is flattened into a one-dimensional space. The flattened vectors are then input into a transformer neural network, which is structured to connect an attention module and a feedforward neural network (FNN). Additionally, vectors containing flattened reference information may be additionally used as input to the transformer neural network.

[0770] The attention module uses a self-attention mechanism to calculate correlations between sub-blocks as weights. The weighted sub-blocks (i.e., 1-dimensional vectors) are input to a forward neural network, which then outputs features for the 1-dimensional vectors.

[0771] Some information decision units (1510) project the output of the neural network for one-dimensional vectors into a one-dimensional space and apply a sigmoid activation function and a rounding operation to the projected values ​​to generate a binary representation for each sub-block.

[0772] The binary representation for the sub-blocks can be determined by a binary mask, which is abbreviated as assigning true (e.g., 1) and false (e.g., 0) to each sub-block.

[0773] The binary mask of the simplified representation can be encoded using lossless encoding such as Huffman coding.

[0774] ■ Some information division units

[0775] Some information distinction unit (1520) inputs a target block and some information locations in the form of a binary mask, and distinguishes some information selected for the target block from the remaining information that is not selected and outputs the result.

[0776] Figure 20 is an example diagram illustrating a process of classifying some selected information and the remaining unselected information for a target block.

[0777] Some information classification units (1520) select sub-blocks corresponding to true in the binary mask (indicated in white in the binary mask of FIG. 20) among the sub-blocks within the target block as some information, and select sub-blocks corresponding to false in the binary mask (indicated with hatching in the binary mask of FIG. 20) as the remaining information.

[0778] Sub-blocks selected with some information and sub-blocks selected with the remaining information can each be output in raster scan order.

[0779] ■ 일부 정보 부호화 유닛

[0780] Some information encoding units (1530) may encode some information selected within the target block.

[0781] Some information can be encoded without prediction itself.

[0782] Alternatively, some information may be predictively encoded using reference information of the target block.

[0783] When predictive encoding is performed, some information encoding units (1530) can generate predictive information for some information using reference information for the target block, and generate and encode residual information, which is the difference between the predictive information and the some information.

[0784] Some of the information itself or residual information can be encoded using the entropy encoder of the conventional encoding device illustrated in Fig. 1. Alternatively, it can be encoded by being converted into a latent representation by the neural network-based encoding device illustrated in Fig. 8 or Fig. 14.

[0785] Predicting for some information may involve a predictive neural network trained to take as input an image with some portions masked and predict the image at the masked location.

[0786] For example, predictions for some information can be made using a predictive neural network that inputs the location and reference information of some information within the target block. In other words, predictions for some information can be generated using a neural network trained to minimize the error between the output and the information, taking the location and reference information of some information as input.

[0787] For this purpose, some information encoding units (1530) may include some information prediction units based on neural networks.

[0788] FIG. 21 is an exemplary diagram of some information prediction units according to one embodiment of the present disclosure.

[0789] Some information prediction units may include a reference information location generation unit (2110), a masked autoencoder (2120), and some information classification units (1520).

[0790] In the embodiments of the present disclosure, a mask autoencoder is used as the predictive neural network, but the present invention is not limited thereto. That is, in addition to the structure of the mask autoencoder, a neural network with a different structure trained to input a partially masked image and predict an image at the masked location may be used.

[0791] The reference information location generation unit (2110) can generate an extension block by expanding a target block of WxH size to include reference information, and generate a binary mask that represents the location of the reference information within the extension block as true (e.g., 1) and represents the location of the target block (i.e., the locations of sub-blocks within the target block) as false (e.g., 0). In Fig. 21, the location of the reference information is indicated in white, and the location corresponding to the target block is indicated with diagonal lines. The binary mask can be used as information for representing the location of the reference information.

[0792] The mask autoencoder (2120) receives the location and reference information as input and outputs a prediction block for the target block. The operation of the mask autoencoder (2120) is described below with reference to FIG. 22.

[0793] Some information classification units (1520) output prediction information for some information by extracting sub-blocks corresponding to some information positions within a prediction block.

[0794] FIG. 22 is an exemplary diagram illustrating the operation of a mask autoencoder according to one embodiment of the present disclosure.

[0795] A mask autoencoder is a neural network based on a transformer that takes as input some selected information and its location from the entire image information and outputs a prediction for the unselected information.

[0796] Referring to FIG. 22, the mask autoencoder may have a structure in which an encoder embedding unit (2210), an encoder neural network (2220), a decoder embedding unit (2230), and a decoder neural network (2240) are sequentially connected.

[0797] The encoder embedding unit (2210) receives selected information from the entire information and the positions of the selected information. Then, it flattens each of the selected pieces of information into a one-dimensional space, adds the positional embedding of the selected information to the flattened one-dimensional vector, and transmits the result to the encoder neural network (2220).

[0798] The encoder neural network (2220) takes as input data a flattened vector in which the position embedding of the selected information is combined, and outputs a latent representation for the input data.

[0799] The decoder embedding unit (2230) receives the latent representation output from the encoder neural network and the positions of selected information. It then adds a mask token to the latent representation by positioning the latent representation at the positions of the selected information and a mask token at the positions of the unselected information. The latent representation with the mask token added is added to the position embedding for the entire information and output to the decoder neural network (2240).

[0800] The decoder neural network (2240) predicts information at unselected locations from the output of the decoder embedding unit (2230).

[0801] Some information prediction units illustrated in Fig. 21 can construct input data using reference information as 'selected information' in Fig. 22 and the location of the reference information as 'location of selected information', and input the input data into a mask autoencoder to generate a prediction block for a target block.

[0802] Meanwhile, some information encoding units (1530) may restore some information. When predictive encoding is performed, some information encoding units may restore residual information for some information and output some restored information using the restored residual information and the predictive information for some information.

[0803] ■ Remaining information encoding unit

[0804] The remaining information encoding unit (1540) can encode the remaining information within the target block using the restored partial information and the positions of the partial information output from the partial information encoding unit (1530) and output the encoded remaining information accordingly. In addition, the remaining information can be decoded for encoding blocks to be encoded later.

[0805] Encoding / decoding of the remaining information can be performed using a conventional encoding / decoding device as illustrated in FIG. 1, or a neural network-based encoding / decoding device as illustrated in FIG. 8 or FIG. 14.

[0806] Additionally, reference information for the target block may be used to encode the remaining information.

[0807] The remaining information encoding unit (1540) can generate prediction information for the remaining information using some information and the location of the remaining information, and encode residual information, which is the difference between the remaining information and the predicted information. The prediction can be performed based on a neural network. To this end, the remaining information encoding unit can include a neural network-based remaining information prediction unit.

[0808] FIG. 23 is a diagram illustrating a remaining information prediction unit according to one embodiment of the present disclosure.

[0809] As illustrated in FIG. 23, the remaining information prediction unit may include a mask autoencoder (2310). The mask autoencoder (2310) of FIG. 23 may have the same structure as the mask autoencoder of FIG. 22.

[0810] In some embodiments, to predict remaining information using a mask autoencoder, some information and the location of the selected information within the target block may be used as the "selected information" and the "location of the selected information" respectively, input to the mask autoencoder of FIG. 22. The mask autoencoder may receive some information and the location of the selected information, and generate prediction information for the remaining information within the target block.

[0811] In another embodiment, additional reference information for the target block may be utilized. In this case, some information and reference information of the target block may be used as "selected information" in FIG. 22, and the location of some information and the location of the reference information may be used as "location of selected information."

[0812] If additional reference information is used, the remaining information prediction unit may include a reference information location addition unit (2320).

[0813] Referring to FIG. 24, the reference information location addition unit (2320) can generate an extension block by expanding a target block to include reference information, and generate a binary mask that expresses the location of the reference information and some information locations within the target block as true (e.g., 1) and the locations of the remaining information as false (e.g., 0) within the extension block.

[0814] The generated binary mask can be used to identify the location of some information and the location of reference information within the target block.

[0815] A mask autoencoder (2310) can receive reference information and some information within a target block, as well as the location of the reference information and the location of some information, and generate prediction information for the remaining information.

[0816] (2) Decryption in image space

[0817] FIG. 25 is an exemplary diagram showing a decryption device that decrypts a target block in an image space according to one embodiment of the present disclosure.

[0818] The location information decryption unit (2510) decodes location information indicating the location of some information to be decrypted first within the target block.

[0819] The location information may be information about a binary mask. The location information decoding unit (2510) outputs a binary mask m that decodes information about the encoded binary mask, thereby making the locations corresponding to some information true and the locations of the remaining information false.

[0820] Some information decryption unit (2520) restores some information by decrypting some encoded information.

[0821] Some of the encoded information may be encoded data for the information itself, or encoded data for residual information for the information. Decoding of some of the encoded information may be performed by an entropy decoder of a conventional decoding device as illustrated in FIG. 1, or a neural network-based decoding device as illustrated in FIG. 8 or 14.

[0822] When some information is predictively encoded using reference information, some of the encoded information may be residual. In this case, the partial information decoding unit (2520) generates prediction information for some information using the reference information and the location of some information output by the location information decoding unit (2510), and adds the prediction information and the residual to restore some information ( ) can be created.

[0823] To predict some information, the information decoding unit (2520) may include some information prediction units based on the same neural network as the encoding device. Since the information prediction units have already been described with reference to FIG. 21, further description is omitted to avoid redundancy.

[0824] The remaining information decryption unit (2530) uses the encoded remaining information to restore the remaining information ( ) can be generated. The encoded remaining information may be information representing the residual for the remaining information.

[0825] For restoration of the remaining information, a conventional decryption device as illustrated in Fig. 1 or a neural network-based decryption device as illustrated in Fig. 8 or 14 may be used.

[0826] The remaining information decoding unit (2530) may include a remaining information prediction unit based on the same neural network as the encoding device for predictive decoding of the remaining information. The remaining information prediction unit may generate prediction information for the remaining information using some information and the locations of the information. The prediction information is added to the residual information to generate restored remaining information.

[0827] Meanwhile, to generate prediction information for the remaining information, the remaining information prediction unit may additionally utilize reference information for the target block.

[0828] As the remaining information prediction units have already been described with reference to Figure 23, further description is omitted.

[0829] The combining unit (2540) combines some of the restored information and the remaining restored information to create a restored block for the target block. ) is generated. In other words, a restoration block is created by arranging some of the restored information and the remaining information in the corresponding locations.

[0830]

[0831] 2. Encoding and decoding in latent representation space

[0832] (1) 잠재표현 공간에서의 부호화

[0833] FIG. 26 is an exemplary diagram showing an encoding device that encodes a target block in a latent representation space according to one embodiment of the present disclosure.

[0834] The image data of the target block may be encoded after being converted into a latent representation. To this end, the encoding device may include a latent representation extraction unit (2610). In addition, the encoding device may include a partial information determination unit (2640), a partial information distinction unit (2650), a partial information encoding unit (2660), and a remaining information encoding unit (2670) to encode the latent representation.

[0835] Additionally, to more accurately estimate the probability model for latent expressions, super-dictionary information may be utilized. When super-dictionary information is utilized, the encoding device may further include a super-dictionary information extraction and encoding unit (2620) and a super-dictionary information feature extraction unit (2630).

[0836] Below, exemplary embodiments of each component of the encoding device illustrated in Fig. 26 are described in detail.

[0837] ■ Latent expression extraction unit

[0838] The latent expression extraction unit (2610) converts the image data of the target block into a latent expression. To this end, the latent expression extraction unit (2610) may be implemented as a latent expression transformation neural network as illustrated in FIG. 9.

[0839] ■ 초사전 정보 추출 및 부호화 유닛

[0840] The super-dictionary information extraction and encoding unit (2620) receives a latent representation for a target block as input, and extracts and encodes super-dictionary information that helps estimate the probability distribution of the latent representation.

[0841] FIG. 27 is an exemplary diagram illustrating a pre-processing information extraction and encoding unit according to one embodiment of the present disclosure.

[0842] The super-dictionary information extraction and encoding unit (2620) may include a super-dictionary information transformation neural network (2710), a quantizer (2720), and an entropy encoder (2730).

[0843] The hyper-lexical information transformation neural network (2710) receives a latent representation for a target block as input and outputs hyper-lexical information for the latent representation. For example, the hyper-lexical information transformation neural network (2710) may output hyper-lexical information whose width and height are each reduced to 1 / 4 of the latent representation.

[0844] The super-precision information transformation neural network (2710) may include three convolutional layers and a ReLU layer positioned between the convolutional layers, as illustrated in FIG. 27. Alternatively, the super-precision information transformation neural network (2710) may be implemented with the structure illustrated in FIG. 12.

[0845] The quantizer (2720) quantizes the hyper-dictionary information output from the hyper-dictionary information transformation neural network (2710). For example, the quantizer (2720) may round the hyper-dictionary information to the nearest decimal place and output the quantized hyper-dictionary information.

[0846] The entropy encoder (2730) can encode quantized super-dictionary information using a probability model for the super-dictionary information. A probability model independent of the input image can be used as the super-dictionary information. For example, range encoding using a probability distribution independent of the input image can be used.

[0847] ■ Pre-processing information feature extraction unit

[0848] The super-dictionary information feature extraction unit (2630) uses the super-dictionary information output from the super-dictionary information transformation neural network (2720) to generate super-dictionary information features, which are parameters of the latent expression probability model.

[0849] The super-prediction information feature extraction unit (2630) can be implemented as a super-prediction information restoration neural network. The super-prediction information restoration neural network can have a structure in which convolutional layers and ReLU activation functions are sequentially connected.

[0850] For example, a hyper-prediction information restoration neural network may include one convolutional layer with a kernel size of 3x3 and a stride of 1, two convolutional layers with a kernel size of 3x3 and a stride of 2, and a ReLU connecting between the convolutional layers, as illustrated in FIG. 28.

[0851] Alternatively, a super-prediction information restoration neural network of the structure illustrated in Fig. 13 may be used.

[0852] ■ 일부 정보 결정 유닛

[0853] A partial information determination unit (2640) selects the location of partial information to be encoded first among the latent representations for the target block. Information indicating the location of the selected partial information is encoded and transmitted to a decoding device.

[0854] For example, some information decision units (2640) may output a binary mask (m) that sets the positions corresponding to some information as True and the positions of the remaining information as False.

[0855] Some information determination units (2640) may perform the same operations as some information determination units (1510) of FIG. 15, in addition to determining the location of some information in the latent representation space.

[0856] For example, some information decision units (2640) may generate binary masks for latent representations of binary target blocks using a neural network trained to predict binary masks for input latent representations.

[0857] In some embodiments, the neural network for generating binary masks may be implemented as a transformer neural network.

[0858] For example, some information determination units (2640) may, as illustrated in FIG. 19, flatten sub-blocks segmented from the latent representation of the target block, and then input the flattened representation (or vector) into a transformer neural network to generate a binary mask. Additionally, a flattened vector for reference information may be additionally input into the transformer neural network.

[0859] As another example, some information decision units (2640) may additionally utilize super-precision information features to generate binary masks. In this case, the super-precision information features output from the super-precision information feature extraction unit (2630) may be input into the transformer neural network of FIG. 19 along with a flattened representation of the latent representation.

[0860] As another example, in addition to the hyper-lexical information features for the latent representation of the target block, hyper-lexical information features for reference information (hereinafter referred to as “reference hyper-lexical information features”) can be additionally input to the transformer neural network.

[0861] Here, the reference information refers to the latent representation of the restored region, and the reference hyper-dictionary information features refer to the features of hyper-dictionary information about the reference information. As previously explained in the encoding in image space, the restored region may be an area belonging to the current picture to which the target block belongs or a region belonging to a reference picture restored before the current picture. For example, the restored region may be at least one restored block surrounding the target block. At least one restored block may be located above and / or to the left of the target block.

[0862] ■ Some information classification units

[0863] Some information distinction units (2650) determine some information (Y) within the latent representation of the target block based on the binary mask output from some information decision units (2640). key ) and the remaining information (Y) excluding some of the information rest) and transmits some information and remaining information to some information encoding unit (2660) and remaining information encoding unit (2670), respectively.

[0864] Some information classification units (2650) can output sub-blocks corresponding to some information among the sub-blocks of the latent expression in raster scan order, and output sub-blocks corresponding to the remaining information among the sub-blocks of the latent expression in raster scan order.

[0865] ■ 일부 정보 부호화 유닛

[0866] A partial information encoding unit (2660) estimates a probability model for encoding some information using the super-dictionary information features corresponding to the latent representation of the target block. Then, the unit entropy encodes some information using the estimated probability model. Prior to entropy encoding, some information may be quantized.

[0867] For more accurate probability model estimation, some information encoding units (2660) may additionally utilize at least one of the prediction information for some information or the super-priority information features of the reference information.

[0868] Additionally, some information encoding units may restore some information and output some of the restored information to the remaining information encoding units (2670).

[0869] FIG. 29 is an exemplary diagram illustrating some information encoding units according to one embodiment of the present disclosure.

[0870] In some embodiments, some information encoding unit (2660) may include an entropy encoder (2930) that encodes some information using super-lexical information features for some information output from some information classification unit (2650).

[0871] The partial information classification unit (2650) receives the super-lexical information features corresponding to the latent representation and a binary mask as input, and extracts the super-lexical information features corresponding to the partial information. The entropy encoder (2930) performs entropy encoding on the partial information using a probability model based on the super-lexical information features of the partial information.

[0872] In another embodiment, some information encoding units (2660) may further include a probability distribution estimation neural network (2920) to increase the accuracy of probability model estimation.

[0873] A probability distribution estimation neural network can estimate a probability model for some information by receiving at least one of the following inputs: hyper-priority information features for some information, predictive information for some information, or hyper-priority information features for reference information. For example, a probability distribution estimation neural network can estimate the parameters of a Gaussian distribution.

[0874] The probability distribution estimation neural network (2920) may be composed of, for example, three convolutional layers with kernel sizes of 1x1 and ReLUs connecting the convolutional layers. To use the probability distribution estimation neural network (2920), the neural network may be trained to estimate a probability model of the latent representation by inputting a prediction signal and super-priority information for the latent representation.

[0875] Meanwhile, to generate prediction information for some information, some information encoding units (2660) may include some information prediction units (2910).

[0876] Some information prediction units (2910) generate prediction information for some information using reference information. The prediction process performed by some information prediction units (2910) may be identical to the prediction process of some information prediction units illustrated in FIG. 21.

[0877] The entropy encoder (2930) performs encoding on some information using the estimated probability model.

[0878] ■ Remaining information encoding unit

[0879] The remaining information encoding unit (2670) estimates a probability model for encoding the remaining information using the super-dictionary information features for the remaining information, excluding some of the super-dictionary information features corresponding to the latent representation of the target block. Then, the remaining information is encoded using the estimated probability model.

[0880] For more accurate probability model estimation, the remaining information encoding unit (2670) may additionally utilize at least one of the prediction information for the remaining information or the super-priority information features of the reference information.

[0881] FIG. 30 is an exemplary diagram showing a remaining information encoding unit according to one embodiment of the present disclosure.

[0882] In some embodiments, the remaining information encoding unit (2670) may include an entropy encoder (3030) that encodes the remaining information using the super-lexical information features for the remaining information output from the partial information classification unit (2650).

[0883] Some information classification units (2650) receive super-lexical information features corresponding to latent expressions and binary masks, and extract super-lexical information features for the remaining information. The entropy encoder (3030) performs entropy encoding on the remaining information using a probability model based on the super-lexical information features for the remaining information.

[0884] In another embodiment, the remaining information encoding unit (2670) may further include a probability distribution estimation neural network (3020) to increase the accuracy of the probability model estimation.

[0885] The probability distribution estimation neural network (3020) can estimate a probability model for the remaining information by receiving at least one of the prediction information for the remaining information, or the hyper-priority information feature of the reference information, along with the hyper-priority information feature for the remaining information. The probability distribution estimation neural network (3020) can estimate the parameters of a Gaussian distribution.

[0886] The probability distribution estimation neural network (3020) may be composed of, for example, three convolutional layers with a kernel size of 1x1 and ReLU connecting the convolutional layers, similar to the probability distribution estimation neural network (2920) of FIG. 29.

[0887] For prediction of remaining information, the remaining information encoding unit (2670) may include a remaining information prediction unit (3010). The remaining information prediction unit (3010) generates prediction information for the remaining information based on some of the restored information. Reference information may additionally be used to generate prediction information for the remaining information. The prediction process of the remaining information prediction unit (3010) may be identical to the prediction process of the remaining information prediction unit illustrated in FIG. 23.

[0888] The entropy encoder (3030) performs encoding on the remaining information using the estimated probability model.

[0889] (2) Decoding in latent expression space

[0890] FIG. 31 is an exemplary diagram showing a decoding device that decodes a target block in a latent representation space according to one embodiment of the present disclosure.

[0891] The decryption device may include a location information decryption unit (3110), a partial information decryption unit (3140), a remaining information decryption unit (3150), a combination unit (3160), and a target block restoration unit (3170).

[0892] The location information decoding unit (3110) decodes location information indicating the location of some information to be decoded first within the latent representation of the target block.

[0893] The location information may be information about a binary mask. The location information decoding unit (2510) outputs a binary mask m that decodes information about the encoded binary mask, thereby making the locations corresponding to some information true and the locations of the remaining information false.

[0894] Some information decryption unit (3140) decrypts some encoded information, thereby restoring some information ( ) is created.

[0895] The remaining information decryption unit (3150) decrypts the encoded remaining information, thereby restoring the remaining information ( ) can be created.

[0896] The combining unit (3160) combines some of the restored information and the remaining restored information to create a latent representation for the target block. ) is generated. In other words, a restored latent representation for the target block is generated by arranging some of the restored information and the remaining information in the corresponding positions.

[0897] The target block restoration unit (3170) restores the restoration block for the target block from the restored latent representation. ) is generated. The target block restoration unit (3170) can be implemented as a latent representation restoration neural network as illustrated in FIG. 10.

[0898] Meanwhile, when super-precise information is used, the decryption device may further include a super-precise information decryption unit (3120) and a super-precise information feature extraction means (3130).

[0899] The super-dictionary information decoding unit (3120) decodes the encoded super-dictionary information using a probability model for the super-dictionary information and outputs the super-dictionary information (Z). A probability model independent of the input image may be used as the probability model for the super-dictionary information. For example, range encoding using a probability distribution independent of the input image may be used.

[0900] The super-dictionary information feature extraction unit (3130) uses the super-dictionary information output from the super-dictionary information decoding unit (3120) to generate super-dictionary information features, which are parameters of a latent representation probability model. The super-dictionary information feature extraction unit (2630) can be implemented as a super-dictionary information restoration neural network as illustrated in FIG. 13 or FIG. 28.

[0901] Below, embodiments for some information decryption units (3140) and remaining information decryption units (3150) are described in detail.

[0902] ■ Some information decryption unit

[0903] The partial information decoding unit (3140) estimates a probability model for decoding some of the encoded information using the super-dictionary information features corresponding to the latent representation of the target block. Then, the unit decodes some of the encoded information using the estimated probability model, thereby restoring some of the information.

[0904] Additionally, for more accurate probability model estimation, at least one of the predictive information for some information or the super-prior information features of the reference information may be utilized.

[0905] Some information decoding units (3140) can estimate a probability model for some information of the latent representation in the same manner as some information encoding units (2660) of the encoding device.

[0906] FIG. 32 is an exemplary diagram showing some information decryption units according to one embodiment of the present disclosure.

[0907] In some embodiments, the information decryption unit (3140) may include an information classification unit (3210) and an entropy decoder (3240).

[0908] Some information classification units (3210) extract super-dictionary information features corresponding to some of the super-dictionary information features for the latent representation of the target block using a binary mask output from the location information decoding unit (3110).

[0909] An entropy decoder (3240) entropy decodes some encoded information and generates some restored information using a probability model based on super-dictionary information features for some information.

[0910] In another embodiment, the partial information decoding unit (3140) may additionally utilize at least one of the prediction information for the partial information or the super-priority information features of the reference information to enhance the accuracy of the probability model estimation. To this end, the partial information decoding unit (3140) may further include the partial information prediction unit (3220) and the probability distribution estimation neural network (3230).

[0911] Some information prediction units (3220) and probability distribution estimation neural networks (3230) are identical to some information prediction units (2910) and probability distribution estimation neural networks (2920) of the encoding device, so further description is omitted to avoid redundant description.

[0912] The entropy decoder (3240) performs entropy decoding on some encoded information using a probability model estimated by a probability distribution estimation neural network (3230).

[0913] ■ 나머지 정보 복호화 유닛

[0914] The remaining information decoding unit (3150) estimates a probability model for decoding the remaining information using the hyper-dictionary information features corresponding to the latent representation of the target block. Then, the remaining information is decoded using the probability model.

[0915] Additionally, for more accurate probability model estimation, at least one of the prediction information for the remaining information or the super-priority information features of the reference information may be utilized.

[0916] The remaining information decoding unit (3150) can estimate a probability model for the remaining information in the same manner as the remaining information encoding unit (2670) of the encoding device.

[0917] FIG. 33 is an exemplary diagram showing a remaining information decryption unit according to one embodiment of the present disclosure.

[0918] In some embodiments, the remaining information decryption unit (3150) may include some information distinction unit (3310) and an entropy decoder (3340).

[0919] Some information classification units (3310) use a binary mask output from a location information decoding unit (3110) to extract super-dictionary information features corresponding to the remaining information among the super-dictionary information features for the latent representation of the target block.

[0920] The entropy decoder (3340) entropy decodes the encoded remaining information using a probability model based on the super-dictionary information features for the remaining information and generates restored remaining information accordingly.

[0921] In another embodiment, the remaining information decoding unit (3150) may additionally utilize at least one of the prediction information for some information or the super-priority information features of the reference information to enhance the accuracy of the probability model estimation. To this end, the remaining information decoding unit (3150) may further include a remaining information prediction unit (3320) and a probability distribution estimation neural network (3330).

[0922] The remaining information prediction unit (3320) and the probability distribution estimation neural network (3330) are identical to the remaining information prediction unit (3010) and the probability distribution estimation neural network (3020) of the encoding device, so further description is omitted to avoid redundant description.

[0923] The entropy decoder (3240) performs entropy decoding on the remaining encoded information using a probability model estimated by a probability distribution estimation neural network (3330).

[0924]

[0925] Hereinafter, as an example, a video encoding method and a video decoding method are disclosed.

[0926] In the embodiments below, the 'partial information' and 'remaining information' of the aforementioned target block or latent representation for the target block are referred to as 'partial region' and 'remaining region', respectively.

[0927] FIG. 34 is a flowchart for explaining an image encoding method performed by an encoding device according to one embodiment of the present disclosure.

[0928] The encoding device first determines a portion of a data block containing data for a target block to be encoded, and generates position information for identifying that portion of the data block and the remaining portion of the data block (S3410). The position information may be encoded and signaled to a decoding device (S3420).

[0929] A data block may be a block composed of samples that constitute a target block in image space. Alternatively, the data block may be a block defining a latent representation of the target block generated through a latent representation transform neural network.

[0930] The encoding device can divide a data block into a plurality of sub-blocks, and determine sub-blocks belonging to some areas of the sub-blocks of the data block and sub-blocks belonging to the remaining areas.

[0931] As mentioned above, location information can be represented as a binary mask that represents some areas and the rest of the areas with different values.

[0932] The encoding device can define patterns of binary masks and determine which binary mask to apply to a data block among the defined patterns. In this case, the position information may be index information indicating one of the defined patterns.

[0933] Alternatively, the encoding device can flatten sub-blocks of a data block to generate a flattened vector, and input this flattened vector into a transformer neural network to generate a binary mask. The generation of a binary mask using a transformer neural network has been described in detail with reference to FIG. 19, and thus further description is omitted.

[0934] The encoding device encodes some areas to generate encoded data of some areas (S3430). Some areas may be quantized prior to encoding.

[0935] Additionally, the encoding device restores some areas (S3440) and encodes the remaining areas using the restored areas (S3450).

[0936] In some embodiments, when the data block is a block in image space, some areas and remaining areas may be encoded using processes performed by the some information encoding unit (1530) and remaining information encoding unit (1540) of FIG. 15, respectively.

[0937] For example, the encoding device may encode some regions themselves without prediction.

[0938] As another example, the encoding device can perform predictive encoding on a portion of the region. As previously described with reference to FIG. 21, the encoding device inputs a restored region around a data block into a mask autoencoder to generate prediction signals for the data block, and extracts prediction signals corresponding to a portion of the region from the prediction signals of the data block based on positional information.

[0939] The encoding device can generate residual signals for some regions based on prediction signals of some regions and encode the residual signals.

[0940] The encoding device can restore the encoded residual signal and restore some areas by adding the restored residual signal and the prediction signal.

[0941] Once some regions are restored, the encoding device can input the restored regions into a mask autoencoder to generate prediction signals for the remaining regions. Then, based on the prediction signals for the remaining regions, residual signals for the remaining regions can be generated and encoded.

[0942] As another embodiment, when the data block is a block in the latent representation space, some areas and remaining areas can be encoded using processes performed by the some information encoding unit (2660) and remaining information encoding unit (2670) of FIG. 26, respectively.

[0943] For example, as described with reference to FIG. 29, the encoding device can estimate a probability model for a portion of a region based on the super-dictionary information features of that portion extracted from the super-dictionary information features of the data block. Then, by encoding the portion of the region using the estimated probability model, encoded data for the portion of the region can be generated. The portion of the region may be quantized prior to encoding.

[0944] To estimate a probability model for a specific region, prediction signals for that region can be used. For example, an encoding device can input the reconstructed region surrounding a data block into a mask autoencoder to generate prediction signals for that region. Then, the generated prediction signals and the hyper-priority information features for that region can be input into a probability distribution estimation neural network to estimate a probability model for that region.

[0945] Additionally, as described with reference to FIG. 30, the encoding device can estimate a probability model for the remaining region based on the super-dictionary information features of the remaining region extracted from the super-dictionary information features of the data block. Then, by encoding the remaining region using the estimated probability model, encoded data for the remaining region can be generated. Some regions may be quantized prior to encoding.

[0946] To estimate the probability model for some domains, the prediction signals for the remaining domains can be used.

[0947] For example, the encoding device can input a portion of the restored region into a mask autoencoder to generate prediction signals for the remaining region. To generate prediction signals for the remaining region, additional reference information can be input into the mask autoencoder.

[0948] The encoding device can estimate a probability model for the remaining region by inputting prediction signals for the remaining region and the super-priority information features of the remaining region into a probability distribution estimation neural network.

[0949]

[0950] FIG. 35 is a flowchart illustrating an image decoding method performed by a decoding device according to one embodiment of the present disclosure.

[0951] The decoding device decodes position information from a bitstream received from the encoding device (S3510), and can identify some information and the remaining information within a data block using the position information (S3520).

[0952] Similar to the encoding device, the decoding device can divide the data block into a plurality of sub-blocks and, using the position information, determine which sub-blocks belong to some areas of the sub-blocks of the data block and which sub-blocks belong to the remaining areas.

[0953] The decryption device restores some areas identified by location information (S3530). Then, the remaining areas are restored using the restored areas (S3540).

[0954] The decryption device can restore the data block using the restored part of the area and the restored remaining area.

[0955] When the data block is a block in the image space, some areas and remaining areas can be restored using processes performed by the part information decoding unit (2520) and remaining information decoding unit (2530) of FIG. 25, respectively.

[0956] If the data block is a block in the latent representation space, some areas and remaining areas can be restored using processes performed by the partial information decoding unit (3140) and remaining information decoding unit (3150) of FIG. 31, respectively.

[0957] The processes performed by some information decryption units (3140) and the remaining information decryption units (3150) have been described with reference to FIGS. 32 and 33, respectively, and therefore further description is omitted.

[0958]

[0959] 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.

[0960] 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.

[0961] 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.

[0962] 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.

[0963] 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.

[0964] 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.

[0965] 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.

[0966] 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.

[0967] 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.

[0968] 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.

[0969] CROSS-REFERENCE TO RELATED APPLICATION

[0970] This patent application claims priority to Korean Patent Application No. 10-2024-0003869, filed on January 9, 2024, the entire contents of which are incorporated herein by reference.

Claims

1. A step of decoding position information from a bitstream, identifying a portion of the data block containing data for a target block to be decoded first and the remaining portion of the data block excluding the portion of the data block; A step of restoring said partial area identified by said location information; A step of restoring the remaining area using the restored portion of the above; and A step of generating a restoration block for the data block using the above-mentioned restored partial area and the above-mentioned restored remaining area. An image decoding method characterized by including a.

2. In paragraph 2, The step of identifying the above-mentioned part of the area and the remaining area is, A step of dividing the above data block into sub-blocks of a predetermined size; and A step of classifying sub-blocks belonging to the part of the area and sub-blocks belonging to the remaining area among the sub-blocks within the data block using the above location information, An image decoding method, characterized in that the above location information is information defining a binary mask in which different values are assigned to sub-blocks belonging to the above-mentioned partial area and sub-blocks belonging to the above-mentioned remaining area.

3. In paragraph 1, An image decoding method, characterized in that the above data block is a block composed of samples in the image space for the target block.

4. In paragraph 3, The steps to restore some of the above areas are: A step of restoring residual signals for said partial region from said bitstream; A step of inputting a restored area around the above data block into a prediction neural network to generate prediction signals for the above partial area, wherein the prediction neural network is a neural network trained to input an image with a portion of the image masked and predict an image at a masked location; and A step of restoring said partial region by using residual signals for said partial region and prediction signals for said partial region. An image decoding method characterized by including a.

5. In paragraph 4, The steps to restore the remaining areas are: A step of restoring residual signals for the remaining area from the bitstream; A step of inputting the above-described restored portion of the region into the above-described prediction neural network to generate prediction signals for the remaining region; and A step of restoring the remaining region based on residual signals for the remaining region and prediction signals for the remaining region. An image decoding method characterized by including a.

6. In paragraph 5, An image decoding method, characterized in that the prediction signals for the remaining area are generated by inputting the restored partial area and the restored area around the data block into the prediction neural network.

7. In paragraph 1, An image decoding method, characterized in that the above data block is a block representing a latent representation of the target block.

8. In paragraph 7, A step of decoding the super-precise information for the data block from the bitstream; and A step of inputting the above-mentioned super-dictionary information into a super-dictionary information restoration neural network to obtain super-dictionary information features for the data block, wherein the super-dictionary information features are information related to probability parameters of a probability model for the data block; An image decoding method characterized by including a.

9. In paragraph 8, The steps to restore some of the above areas are: A step of extracting hyper-precision information features corresponding to the partial region from hyper-precision information features for the latent expression of the target block using the above location information; A step of estimating a probability model for said partial region based on the hyper-prediction information features corresponding to said partial region; and A step of restoring a latent representation for the partial region by decoding encoded data for the partial region included in the bitstream based on a probability model for the partial region. An image decoding method characterized by including a.

10. In paragraph 9, The step of estimating the probability model for the above-mentioned areas is: A step of inputting a restored area around the above data block into a prediction neural network to generate prediction signals for the above partial area, wherein the prediction neural network is a neural network trained to input an image with a portion of the image masked and predict an image at a masked location; and A step of estimating a probability model for the above-mentioned partial region by inputting prediction signals for the above-mentioned partial region and super-priority information features corresponding to the above-mentioned partial region into a probability distribution estimation neural network. An image decoding method characterized by including a.

11. In paragraph 9, The steps to restore the remaining areas are: A step of extracting the super-precision information features corresponding to the remaining area from the super-precision information features for the data block using the above location information; A step of inputting the above-described restored portion of the region into a prediction neural network to generate prediction signals for the remaining region, wherein the prediction neural network is a neural network trained to input an image of which a portion is masked and predict an image of the masked location; A step of estimating a probability model for the remaining region by inputting prediction signals for the remaining region and super-priority information features corresponding to the remaining region into a probability distribution estimation neural network; and A step of restoring the remaining area by decoding the encoded data for the remaining area included in the bitstream based on the probability model for the remaining area. An image decoding method characterized by including a.

12. In paragraph 11, An image decoding method, characterized in that the prediction signals for the remaining area are generated by inputting the restored partial area and the restored area around the data block into the prediction neural network.

13. A step of determining some areas to be encoded first and the remaining areas excluding the some areas within a data block containing data for the target block; A step of encoding location information to indicate the location of the above-mentioned partial area and the above-mentioned remaining area; A step of encoding said partial region to generate encoded data for said partial region; and A step of restoring some areas and encoding the remaining areas using the restored areas. An image encoding method characterized by including a.

14. A method for providing image data to a decryption device, A step of generating a bitstream by encoding a target block; and comprising a step of transmitting the bitstream to the decryption device; The steps for generating the above bitstream are: A step of determining a portion of a data block containing data for the target block to be encoded first and the remaining portion of the data block excluding the portion of the data block; A step of encoding location information to indicate the location of the above-mentioned partial area and the above-mentioned remaining area; A step of encoding said partial region to generate encoded data for said partial region; and A step of restoring some areas and encoding the remaining areas using the restored areas. A method, characterized by including:

Images