Feature encoding / decoding method and apparatus based on prediction model training, and recording medium having bit stream stored therein

Abstract

Provided are a feature encoding / decoding method and apparatus, and a computer-readable recording medium generated by the feature encoding method. According to the present disclosure, a decoding method performed by a feature decoding device comprises the steps of: training a prediction model based on a reconstructed neighbor sample; predicting the current block by using the trained prediction model; and reconstructing the current block based on residual information of the current block and the predicted current block.

Description

Technical Field

[0001] The present disclosure relates to a feature encoding / decoding method and apparatus, and in particular, to a feature encoding / decoding method and apparatus based on prediction model training, and a recording medium storing a bit stream generated by the feature encoding method / apparatus of the present disclosure. Background Art

[0002] With the development of machine learning technology, the demand for AI services based on image processing is increasing. In order to effectively process the large amount of image data required for AI services within limited resources, image compression technology that optimizes machine task performance is essential. However, existing image compression technology has been developed with the goal of high-resolution, high-quality image processing for human vision, and there is a problem that it is not suitable for AI services. Accordingly, the research and development of new machine-oriented image compression technology suitable for AI services is actively underway. Summary of the invention

[0003] Technical issues

[0004] An object of the present disclosure is to provide a feature encoding / decoding method and apparatus with improved encoding / decoding efficiency.

[0005] Another object of the present disclosure is to provide a feature encoding / decoding method and device that trains a prediction model and uses the prediction model to make predictions.

[0006] Another object of the present disclosure is to provide a feature encoding / decoding method and device that configures a list of fully trained prediction models and uses one of the prediction models in the list for prediction or as an initial value for training.

[0007] Another object of the present disclosure is to provide a method for transmitting a bit stream generated by the feature encoding method or device according to the present disclosure.

[0008] Another object of the present disclosure is to provide a recording medium storing a bit stream generated by the feature encoding method or device according to the present disclosure.

[0009] Another object of the present disclosure is to provide a recording medium storing a bit stream of a feature received and decoded by a feature decoding apparatus according to the present disclosure and used for reconstruction.

[0010] The technical problems solved by the present disclosure are not limited to the above-mentioned technical problems, and other technical problems not described herein will become apparent to those skilled in the art from the following description.

[0011] Technical Solution

[0012] According to an aspect of the present disclosure, a feature decoding method performed by a feature decoding device may include: training a prediction model based on reconstructed neighbor samples, predicting a current block by using the trained prediction model, and reconstructing the current block based on residual information of the current block and the predicted current block.

[0013] According to another aspect of the present disclosure, a feature encoding method performed by a feature encoding device may include: training a prediction model based on reconstructed neighbor samples, predicting a current block using the trained prediction model, and encoding residual information derived based on the current block and the predicted current block.

[0014] A recording medium according to another aspect of the present disclosure may store a bit stream generated by the feature encoding method or the feature encoding device of the present disclosure.

[0015] A bitstream transmission method according to another aspect of the present disclosure may send a bitstream generated by the feature encoding method or the feature encoding device of the present disclosure to a feature decoding device.

[0016] The features briefly summarized above with respect to the present disclosure are merely exemplary aspects of the disclosure described in detail below and do not limit the scope of the disclosure.

[0017] Beneficial Effects

[0018] According to the present disclosure, it is possible to provide a feature encoding / decoding method and apparatus with improved encoding / decoding efficiency.

[0019] Furthermore, according to the present disclosure, since the number of prediction models available for prediction can be actually increased through training, a prediction model suitable for the characteristics of the current block can be applied.

[0020] Furthermore, according to the present disclosure, the amount of bits in a bitstream can be reduced compared to a variable prediction model.

[0021] Those skilled in the art will appreciate that the effects that can be achieved by the present disclosure are not limited to the above specific description, and other advantages of the present disclosure will be more clearly understood from the detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a diagram schematically showing a VCM system to which an embodiment of the present disclosure is applicable.

[0023] Figure 2 Schematically shows a VCM pipeline structure to which an embodiment of the present disclosure can be applied.

[0024] Figure 3 FIG. 4 is a diagram schematically showing an image / video encoder to which an embodiment of the present disclosure can be applied.

[0025] Figure 4 FIG. 1 is a diagram schematically showing an image / video decoder to which an embodiment of the present disclosure can be applied.

[0026] Figure 5 is a flow chart schematically illustrating a feature / feature map encoding process to which embodiments of the present disclosure may be applied.

[0027] Figure 6 is a flow chart schematically illustrating a feature / feature map decoding process to which embodiments of the present disclosure may be applied.

[0028] Figure 7 is a diagram showing an example of reconstructed neighbor samples and current encoding / decoding target samples that can be used to train a prediction model.

[0029] Figure 8 is a flowchart illustrating a feature encoding method according to an embodiment of the present disclosure.

[0030] Fig. 9 is a flow chart illustrating a feature decoding method according to an embodiment of the present disclosure.

[0031] Fig.10 is a flowchart illustrating a feature encoding method according to another embodiment of the present disclosure.

[0032] Fig.11 is a flowchart illustrating a feature decoding method according to another embodiment of the present disclosure.

[0033] Fig.12a and Figure 12b is a flowchart illustrating a feature encoding method according to yet another embodiment of the present disclosure.

[0034] Fig.13a and Fig.13b is a flowchart illustrating a feature decoding method according to yet another embodiment of the present disclosure.

[0035] Fig.14 : is a diagram showing an example of a content streaming system to which an embodiment of the present disclosure is applicable.

[0036] Fig.15 is a view showing another example of a content streaming system to which an embodiment of the present disclosure is applicable. DETAILED DESCRIPTION

[0037] Hereinafter, the embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily implement them. However, the present disclosure can be implemented in various forms and is not limited to the embodiments described herein.

[0038] When describing the present disclosure, if it is determined that the detailed description of related known functions or configurations makes the scope of the present disclosure unnecessarily ambiguous, the detailed description thereof will be omitted. In the drawings, parts irrelevant to the description of the present disclosure are omitted, and like reference numerals are attached to like parts.

[0039] In the present disclosure, when a component is "connected", "coupled" or "linked" to another component, it may include not only a direct connection relationship but also an indirect connection relationship with intermediate components. In addition, when a component "includes" or "has" other components, unless otherwise specified, it means that other components may be further included rather than excluded.

[0040] In the present disclosure, the terms first, second, etc. may be used only for the purpose of distinguishing one component from other components, and do not limit the order or importance of the components unless otherwise specified. Therefore, within the scope of the present disclosure, a first component in one embodiment may be referred to as a second component in another embodiment, and similarly, a second component in one embodiment may be referred to as a first component in another embodiment.

[0041] In the present disclosure, components that are distinguished from each other are intended to clearly describe each feature and do not mean that these components are necessarily separated. That is, multiple components can be integrated and implemented in one hardware or software unit, or one component can be distributed and implemented in multiple hardware or software units. Therefore, even if not otherwise specified, implementations in which these components are integrated or components are distributed are also included in the scope of the present disclosure.

[0042] In the present disclosure, the components described in various embodiments do not necessarily mean essential components, and some components may be optional components. Therefore, embodiments consisting of a subset of the components described in the embodiments are also included in the scope of the present disclosure. In addition, in addition to the components described in various embodiments, embodiments including other components are also included in the scope of the present disclosure.

[0043] The present disclosure relates to encoding and decoding of images, and terms used in the present disclosure may have general meanings commonly used in the technical field to which the present disclosure belongs unless newly defined in the present disclosure.

[0044] The present disclosure may be applied to methods disclosed in the Versatile Video Coding (VVC) standard and / or the Machine Video Coding (VCM) standard. In addition, the present disclosure may be applied to methods disclosed in the Basic Video Coding (EVC) standard, the AOMedia Video 1 (AV1) standard, the second generation audio video coding standard (AVS2), or the next generation video / image coding standard (e.g., H.267 or H.268, etc.).

[0045] The present disclosure provides various embodiments related to video / image coding, and unless otherwise specified, these embodiments may be performed in combination with each other. In the present disclosure, "video" refers to a collection of a series of images according to the passage of time. "Image" may be information generated by artificial intelligence (AI). Input information used in the process of performing a series of tasks by AI, information generated in the process of information processing, and output information may be used as an image. In the present disclosure, "picture" generally refers to a unit representing an image within a specific time period, and a slice / tile is a coding unit that constitutes a part of a picture when encoding. A picture may be composed of one or more slices / tiles. In addition, a slice / tile may include one or more coding tree units (CTUs). A CTU may be divided into one or more CUs. A tile is a rectangular area present in a specific tile row and a specific tile column in a picture, and may be composed of multiple CTUs. A tile column may be defined as a rectangular area of ​​a CTU, may have the same height as a picture, and may have a width specified by a syntax element signaled from a bitstream portion such as a picture parameter set. A tile row may be defined as a rectangular area of ​​a CTU, may have the same width as a picture, and may have a height specified by a syntax element signaled from a bitstream portion such as a picture parameter set. Tile scanning is a certain method of sequentially ordering CTUs that partition a picture. Here, CTUs may be sequentially ordered according to a CTU raster scan within a tile, and tiles in a picture may be sequentially ordered according to a raster scan order of tiles of a picture. A slice may contain an integer number of complete tiles, or may contain a continuous integer number of complete CTU rows within a tile of a picture. A slice may be contained exclusively in a single NAL unit. A picture may consist of one or more tile groups. A tile group may include one or more tiles. A tile may indicate a rectangular area of ​​a CTU row within a tile in a picture. A tile may include one or more tiles. A tile may refer to a rectangular area of ​​a CTU row in a tile. A tile may be split into multiple tiles, and each tile may include one or more CTU rows belonging to a tile. Tiles that are not split into multiple tiles may also be considered tiles.

[0046] In the present disclosure, "pixel" or "picture element" may mean the smallest unit constituting a picture (or image). In addition, "sample" may be used as a term corresponding to a pixel. A sample may generally represent a pixel or a value of a pixel, and may represent only a pixel / pixel value of a luminance component or only a pixel / pixel value of a chrominance component.

[0047] In an embodiment, particularly when applied to VCM, when there is a picture composed of a set of components with different characteristics and meanings, the pixel / pixel value may represent the pixel / pixel value of the component generated by the combination, synthesis and analysis of independent information or individual components. For example, in an RGB input, only the pixel / pixel value of R may be represented, only the pixel / pixel value of G may be represented, or only the pixel / pixel value of B may be represented. For example, only the pixel / pixel value of the brightness component synthesized using the R, G and B components may be represented. For example, only the pixel / pixel value of the image and information extracted by analyzing the R, G and B components from the component may be represented.

[0048] In the present disclosure, a "unit" may represent a basic unit of image processing. The unit may include at least one of a specific area of ​​a picture and information related to the area. A unit may include a luminance block and two chrominance (e.g., Cb and Cr) blocks. In some cases, the unit may be used interchangeably with terms such as "sample array", "block" or "area". In general, an M×N block may include M columns and N rows of samples (or sample arrays) or a set (or array) of transform coefficients. In an embodiment, in particular, particularly when applied to VCM, the unit may represent a basic unit containing information for performing a specific task.

[0049] In the present disclosure, "current block" may mean one of "current coding block", "current coding unit", "coding target block", "decoding target block" or "processing target block". When prediction is performed, "current block" may mean "current prediction block" or "prediction target block". When transform (inverse transform) / quantization (dequantization) is performed, "current block" may mean "current transform block" or "transform target block". When filtering is performed, "current block" may mean "filtering target block".

[0050] In addition, in the present disclosure, "current block" may mean "luminance block of the current block", unless explicitly stated as a chrominance block. "Chrominance block of the current block" may be expressed by including an explicit description of a chrominance block such as "chrominance block" or "current chrominance block".

[0051] In the present disclosure, the terms " / " and "," should be interpreted as indicating "and / or". For example, the expressions "A / B" and "A, B" may mean "A and / or B". In addition, "A / B / C" and "A, B, C" may mean "at least one of A, B, and / or C".

[0052] In the present disclosure, the term "or" should be interpreted as indicating "and / or". For example, the expression "A or B" may include 1) only "A", 2) only "B", and / or 3) both "A and B". In other words, in the present disclosure, the term "or" should be interpreted as indicating "additionally or alternatively".

[0053] The present disclosure relates to machine video / image coding (VCM).

[0054] VCM refers to compression techniques for encoding / decoding parts of source images / videos or information obtained from source images / videos for machine vision purposes. In VCM, the encoding / decoding target may be referred to as a feature. The feature may refer to information extracted from the source image / video based on the mission purpose, requirements, surrounding environment, etc. The feature may have a different information form from the source image / video, and accordingly, the compression method and expression format of the feature may also be different from that of the video source.

[0055] VCM can be applied to a variety of application areas. For example, in a monitoring system that identifies and tracks objects or people, VCM can be used to store or send object identification information. In addition, in intelligent transportation or smart traffic systems, VCM can be used to send vehicle location information collected from GPS, sensing information collected from LIDAR, radar, etc., and various vehicle control information to other vehicles or infrastructure. In addition, in the field of smart cities, VCM can be used to perform individual tasks of interconnected sensor nodes or devices.

[0056] The present disclosure provides various embodiments of feature / feature map compilation. Unless otherwise specified, the embodiments of the present disclosure may be implemented individually or in combination of two or more.

[0057] Overview of VCM System

[0058] Figure 1 Schematically shows a VCM system to which an embodiment of the present disclosure can be applied.

[0059] Reference Figure 1 , the VCM system may include an encoding device 10 and a decoding device 20.

[0060] The encoding device 10 can generate a bitstream by compressing / encoding features / feature maps extracted from the source image / video, and transmit the generated bitstream to the decoding device 20 through a storage medium or a network. The encoding device 10 may also be referred to as a feature encoding device. In a VCM system, features / feature maps may be generated at each hidden layer of a neural network. The size and number of channels of the generated feature map may vary depending on the type of neural network or the position of the hidden layer. In the present disclosure, a feature map may be referred to as a feature set, and a feature or feature map may be referred to as "feature information".

[0061] The encoding device 10 may include a feature acquisition unit 11 , an encoding unit 12 , and a transmission unit 13 .

[0062] The feature acquisition unit 11 can acquire features / feature maps for the source image / video. Depending on the implementation, the feature acquisition unit 11 can acquire features / feature maps from an external device, such as a feature extraction network. In this case, the feature acquisition unit 11 performs a feature receiving interface function. Alternatively, the feature acquisition unit 11 can acquire features / feature maps by executing a neural network (e.g., CNN, DNN, etc.) using the source image / video as input. In this case, the feature acquisition unit 11 performs a feature extraction network function.

[0063] According to an embodiment, the encoding device 10 may further include a source image generator (not shown) for obtaining a source image / video. The source image generator may be implemented using an image sensor, a camera module, etc., and the source image / video may be obtained through an image / video capture, synthesis, or generation process. In this case, the generated source image / video may be sent to a feature extraction network and used as input data for extracting features / feature maps.

[0064] The encoding unit 12 may encode the features / feature maps acquired by the feature acquisition unit 11. The encoding unit 12 may perform a series of processes such as prediction, transformation, and quantization to increase encoding efficiency. The encoded data (encoded feature / feature map information) may be output in the form of a bitstream. The bitstream containing the encoded feature / feature map information may be referred to as a VCM bitstream.

[0065] The transmission unit 13 may obtain feature / feature graph information or data output in the form of a bit stream, and forward the feature / feature graph information or data to the decoding device 20 or another external object in the form of a file or stream via a digital storage medium or a network. Here, the digital storage medium may include various storage media, such as USB, SD, CD, DVD, Blu-ray, HDD, and SSD. The transmission unit 13 may include an element for generating a media file having a predetermined file format or an element for transmitting data via a broadcast / communication network. The transmission unit 13 may be provided as a transmission device separate from the encoding unit 12, and in this case, the transmission device may include at least one processor for obtaining feature / feature graph information or data output in the form of a bit stream and a transmission unit for forwarding feature / feature graph information or data in the form of a file or stream.

[0066] The decoding device 20 may obtain feature / feature map information from the encoding device 10 and reconstruct the feature / feature map based on the obtained information.

[0067] The decoding device 20 may include a receiving unit 21 and a decoding unit 22 .

[0068] The receiving unit 21 may receive a bit stream from the encoding device 10 , obtain feature / feature map information from the received bit stream, and send it to the decoding unit 22 .

[0069] The decoding unit 22 may decode the feature / feature map based on the acquired feature / feature map information. The decoding unit 22 may perform a series of processes corresponding to the operations of the encoding unit 12, such as dequantization, inverse transformation, and prediction, to increase decoding efficiency.

[0070] Depending on the implementation, the decoding device 20 may further include a task analysis / rendering unit 23 .

[0071] The task analysis / rendering unit 23 can perform task analysis based on the decoded features / feature graphs. In addition, the task analysis / rendering unit 23 can render the decoded features / feature graphs into a form suitable for task execution. Various machine (oriented) tasks can be performed based on the task analysis results and the rendered features / feature graphs.

[0072] As described above, the VCM system can encode / decode features extracted from source images / videos according to user and / or machine requests, task objectives, and surrounding environments, and perform various machine (oriented) tasks based on the decoded features. The VCM system can be implemented by extending / redesigning the video / image compilation system, and can perform various encoding / decoding methods defined in the VCM standard.

[0073] VCM pipeline

[0074] Figure 2 Schematically shows a VCM pipeline structure to which an embodiment of the present disclosure can be applied.

[0075] Reference Figure 2 , the VCM pipeline 200 may include a first pipeline 210 for encoding / decoding an image / video and a second pipeline 220 for encoding / decoding a feature / feature map. In the present disclosure, the first pipeline 210 may be referred to as a video codec pipeline, and the second pipeline 220 may be referred to as a feature codec pipeline.

[0076] The first pipeline 210 may include a first stage 211 for encoding an input image / video and a second stage 212 for decoding the encoded image / video to generate a reconstructed image / video. The reconstructed image / video may be used for human viewing, ie, human vision.

[0077] The second pipeline 220 may include a third stage 221 for extracting features / feature maps from an input image / video, a fourth stage 222 for encoding the extracted features / feature maps, and a fifth stage 223 for decoding the encoded features / feature maps to generate reconstructed features / feature maps. The reconstructed features / feature maps can be used for machine (vision) tasks. Here, machine (vision) tasks may refer to tasks in which images / videos are consumed by machines. Machine (vision) tasks can be applied to service scenarios such as, for example, monitoring, smart transportation, smart cities, smart industries, smart content, etc. Depending on the implementation, the reconstructed features / feature maps can be used for human vision.

[0078] Depending on the implementation, the features / feature maps encoded in the fourth stage 222 may be transmitted to the first stage 221 and used to encode the image / video. In this case, an additional bitstream may be generated based on the encoded features / feature maps, and the generated additional bitstream may be transmitted to the second stage 222 and used to decode the image / video.

[0079] Depending on the implementation, the features / feature maps decoded in the fifth stage 223 may be passed to the second stage 222 and used to decode the image / video.

[0080] Figure 2 The case where the VCM pipeline 200 includes the first pipeline 210 and the second pipeline 220 is shown, but this is only an example and the embodiments of the present disclosure are not limited thereto. For example, the VCM pipeline 200 may include only the second pipeline 220, or the second pipeline 220 may be expanded to multiple feature codec pipelines.

[0081] In addition, in the first pipeline 210, the first stage 211 may be performed by an image / video encoder, and the second stage 212 may be performed by an image / video decoder. In addition, in the second pipeline 220, the third stage 221 may be performed by a VCM encoder (or a feature / feature map encoder), and the fourth stage 222 may be performed by a VCM decoder (or a feature / feature map encoder). Hereinafter, the encoder / decoder structure will be described in detail.

[0082] Encoder

[0083] Figure 3 FIG. 4 is a diagram schematically showing an image / video encoder to which an embodiment of the present disclosure can be applied.

[0084] Reference Figure 3, the image / video encoder 300 may further include an image segmenter 310, a predictor 320, a residual processor 330, an entropy encoder 340, and an adder 350, a filter 360, and a memory 370. The predictor 320 may include an inter-frame predictor 321 and an intra-frame predictor 322. The residual processor 330 may include a transformer 332, a quantizer 333, a dequantizer 334, and an inverse transformer 335. The residual processor 330 may further include a subtractor 331. The adder 350 may be referred to as a reconstructor or a reconstructed block generator. Depending on the implementation, the image segmenter 310, the predictor 320, the residual processor 330, the entropy encoder 340, the adder 350, and the filter 360 may be configured by one or more hardware components (e.g., an encoder chipset or processor). In addition, the memory 370 may include a decoded picture buffer (DPB) and may be configured by a digital storage medium. The above-mentioned hardware components may further include a memory 370 as an internal / external component.

[0085] The image splitter 310 may split the input image (or picture, frame) input to the image / video encoder 300 into one or more processing units. As an example, a processing unit may be referred to as a coding unit (CU). The coding unit may be recursively split from a coding tree unit (CTU) or a maximum coding unit (LCU) according to a quadtree binary tree ternary tree (QTBTTT) structure. For example, a coding unit may be split into a plurality of coding units of a deeper depth based on a quadtree structure, a binary tree structure, and / or a ternary structure. In this case, for example, a quadtree structure may be applied first, and a binary tree structure and / or a ternary tree structure may be applied later. Alternatively, a binary tree structure may be applied first. The image / video coding process according to the present disclosure may be performed based on a final coding unit that is no longer split. In this case, the maximum coding unit may be used as the final coding unit based on coding efficiency according to image characteristics, or, if necessary, the coding unit may be recursively split into coding units of a deeper depth to use a coding unit with an optimal size as the final coding unit. Here, the coding process may include processes such as prediction, transformation, and reconstruction, which will be described later. As another example, the processing unit may further include a prediction unit (PU) or a transform unit (TU). In this case, both the prediction unit and the transform unit may be divided or partitioned from the above-mentioned final coding unit. The prediction unit may be a unit for sample prediction, and the transform unit may be a unit for deriving a transform coefficient and / or a unit for deriving a residual signal from a transform coefficient.

[0086] In some cases, the unit may be used interchangeably with terms such as block or region. In general, an M×N block may represent a set of samples or transform coefficients consisting of M columns and N rows. A sample may generally represent a pixel or a pixel value, and may represent only a pixel / pixel value of a luminance component, or only a pixel / pixel value of a chrominance component. A sample may be used as a term corresponding to a pixel or a pixel element.

[0087] The image / video encoder 300 can generate a residual signal (residual block, residual sample array) by subtracting a prediction signal (prediction block, prediction sample array) output from the inter predictor 321 or the intra predictor 322 from the input image signal (original block, original sample array) and transmit the generated residual signal to the transformer 332. In this case, as shown, the unit that subtracts the prediction signal (prediction block, prediction sample array) from the input image signal (original block, prediction sample array) within the image / video encoder 300 can be referred to as a subtractor 331. The predictor can perform prediction on a processing target block (hereinafter referred to as a current block) and generate a prediction block including prediction samples for the current block. The predictor can determine whether intra prediction or inter prediction is applied in the current block or CU unit. The predictor can generate various information related to prediction such as prediction mode information and transmit it to the entropy encoder 340. Information about prediction can be encoded in the entropy encoder 340 and output in the form of a bitstream.

[0088] The intra-frame predictor 322 can predict the current block by referring to the samples in the current picture. At this time, depending on the prediction mode, the referenced sample can be located in the neighbor of the current block or can be located far away from the current block. In intra-frame prediction, the prediction mode may include multiple non-directional modes and multiple directional modes. The non-directional mode may include, for example, a DC mode and a plane mode. Depending on the level of detail of the prediction direction, the directional mode may include, for example, 33 directional prediction modes or 65 directional prediction modes. However, this is only an example, and more or less directional prediction modes may be used depending on the setting. The intra-frame predictor 322 may determine the prediction mode applied to the current block by using the prediction mode applied to the neighboring blocks.

[0089] The inter-frame predictor 321 may derive a prediction block for the current block based on a reference block (reference sample array) specified by a motion vector on a reference picture. In this case, in order to reduce the amount of motion information sent in the inter-frame prediction mode, the motion information may be predicted in units of blocks, sub-blocks, or samples based on the correlation of the motion information between the neighboring blocks and the current block. The motion information may include a motion vector and a reference picture index. The motion information may further include inter-frame prediction direction (L0 prediction, L1 prediction, Bi prediction, etc.) information. In the case of inter-frame prediction, the neighboring blocks may include spatial neighboring blocks present in the current picture and temporal neighboring blocks present in the reference picture. The reference picture including the reference block and the reference picture including the temporal neighboring block may be the same or different. The temporal neighboring block may be referred to as a collocated reference block, a collocated CU (colCU), etc., and the reference picture including the temporal neighboring block may be referred to as a collocated picture (colPic). For example, the inter-frame predictor 321 may construct a motion information candidate list based on the neighboring blocks, and generate information indicating which candidate is used to derive the motion vector and / or reference picture index of the current block. Inter prediction may be performed based on various prediction modes, and for example, in the case of skip mode and merge mode, the inter predictor 321 may use motion information of a neighboring block as motion information of a current block. In the case of skip mode, unlike merge mode, a residual signal may not be transmitted. In the case of motion vector prediction (MVP) mode, a motion vector of a neighboring block may be used as a motion vector predictor, and a motion vector difference may be signaled to indicate a motion vector of the current block.

[0090] The predictor 320 can generate a prediction signal based on various prediction methods. For example, for the prediction of a block, the predictor can apply not only intra prediction or inter prediction, but also both intra prediction and inter prediction at the same time. This can be called combined inter and intra prediction (CIIP). In addition, the predictor can be based on an intra block copy (IBC) prediction mode or a palette mode for the prediction of the block. The IBC prediction mode or the palette mode can be used for content image / video coding of games, etc., such as screen content coding (SCC). IBC basically performs prediction within the current picture, but can be performed similarly to inter prediction because a reference block is derived within the current picture. That is, IBC can use at least one of the inter prediction techniques described in the present disclosure. The palette mode can be regarded as an example of intra coding or intra prediction. When the palette mode is applied, the sample value within the picture can be sent with a signal based on information about the palette table and the palette index.

[0091] The prediction signal generated by the predictor 320 can be used to generate a reconstruction signal or to generate a residual signal. The transformer 332 can generate a transform coefficient by applying a transform technique to the residual signal. For example, the transform technique may include at least one of a discrete cosine transform (DCT), a discrete sine transform (DST), a karhunen-loève transform (KLT), a graph-based transform (GBT), or a conditional nonlinear transform (CNT). Here, GBT refers to a transform obtained from a graph when the relationship information between pixels is represented by a graph. CNT refers to a transform obtained based on a prediction signal generated using all previously reconstructed pixels. In addition, the transform process can be applied to square pixel blocks of the same size, or can be applied to non-square blocks of variable size.

[0092] The quantizer 130 may quantize the transform coefficients and send them to the entropy encoder 190. The entropy encoder 190 may encode the quantized signal (information about the quantized transform coefficients) and output a bitstream. Information about the quantized transform coefficients may be referred to as residual information. The quantizer 130 may reorder the quantized transform coefficients in the form of blocks into a one-dimensional vector form based on the coefficient scanning order, and generate information about the quantized transform coefficients based on the quantized transform coefficients in the form of a one-dimensional vector. The entropy encoder 340 may perform various encoding methods such as, for example, exponential Golomb, context adaptive variable length coding (CAVLC), context adaptive binary arithmetic coding (CABAC), etc. The entropy encoder 340 may encode information necessary for video / image reconstruction other than the quantized transform coefficients (e.g., the values ​​of syntax elements, etc.) together or separately. The encoded information (e.g., encoded video / image information) can be transmitted or stored in units of a network abstraction layer (NAL) in the form of a bitstream. The video / image information may further include information about various parameter sets such as an adaptive parameter set (APS), a picture parameter set (PPS), a sequence parameter set (SPS), or a video parameter set (VPS). In addition, the video / image information may further include general constraint information. In addition, the video / image information may further include methods, purposes, etc. for generating and using the encoded information. In the present disclosure, information and / or syntax elements transmitted / signaled from the image / video encoder to the image / video decoder may be included in the image / video information. The image / video information may be encoded and included in the bitstream through the above-mentioned encoding process. The bitstream may be sent through a network or may be stored in a digital storage medium. The network may include a broadcast network and / or a communication network, and the digital storage medium may include various storage media such as USB, SD, CD, DVD, Blu-ray, HDD, SSD, etc. A transmitter (not shown) that transmits a signal output from the entropy encoder 340 and / or a storage unit (not shown) that stores the signal may be configured as an internal / external element of the image / video encoder 300, or the transmitter may be included in the entropy encoder 340.

[0093] The quantized transform coefficients output from the quantizer 130 can be used to generate a prediction signal. For example, the residual signal (residual block or residual sample) can be reconstructed by applying dequantization and inverse transformation to the quantized transform coefficients by the dequantizer 334 and the inverse transformer 335. The adder 350 adds the reconstructed residual signal to the prediction signal output from the inter-frame predictor 321 or the intra-frame predictor 322 to generate a reconstructed signal (reconstructed picture, reconstructed block, reconstructed sample array). In the case where there is no residual for the processing target block, such as the case where the skip mode is applied, the prediction block can be used as a reconstructed block. The adder 350 can be referred to as a reconstructor or a reconstructed block generator. The generated reconstructed signal can be used for intra-frame prediction of the next processing target block in the current picture, and can be used for inter-frame prediction of the next picture by filtering as described below.

[0094] Additionally, luminance mapping with chroma scaling (LMCS) is applied during picture encoding and / or reconstruction.

[0095] The filter 360 can improve the subjective / objective image quality by applying filtering to the reconstructed signal. For example, the filter 360 can generate a modified reconstructed picture by applying various filtering methods to the reconstructed picture, and store the modified reconstructed picture in the memory 370, specifically, in the DPB of the memory 370. Various filtering methods may include, for example, deblocking filtering, sample adaptive offset, adaptive loop filter, bilateral filter, etc. The filter 360 can generate various information related to filtering and send the generated information to the entropy encoder 190. The information related to filtering can be encoded by the entropy encoder 340 and output in the form of a bit stream.

[0096] The modified reconstructed picture transmitted to the memory 370 may be used as a reference picture in the inter predictor 321. Through this, prediction mismatch between the encoder and the decoder may be avoided and coding efficiency may be improved.

[0097] The DPB of the memory 370 may store the modified reconstructed picture for use as a reference picture in the inter-frame predictor 321. The memory 370 may store the motion information of the block from which the motion information in the current picture is derived (or encoded) and / or the motion information of the block in the reconstructed picture. The stored motion information may be transmitted to the inter-frame predictor 321 to be used as the motion information of the spatial neighboring block or the motion information of the temporal neighboring block. The memory 370 may store the reconstructed samples of the reconstructed block in the current picture, and may transmit the stored reconstructed samples to the intra-frame predictor 322.

[0098] In addition, the VCM encoder (or feature / feature map encoder) basically performs a series of processes such as prediction, transformation, and quantization to encode the feature / feature map, and thus may basically have the same Figure 3 The VCM encoder has the same / similar structure as the image / video encoder 300 described above. However, the VCM encoder is different from the image / video encoder 300 in that the feature / feature map is the encoding target, and therefore the name of each unit (or component) (e.g., image splitter 310, etc.) and its specific operation content may be different from the image / video encoder 300. The specific operation of the VCM encoder will be described in detail later.

[0099] Decoder

[0100] Figure 4 FIG. 1 is a diagram schematically showing an image / video decoder to which an embodiment of the present disclosure can be applied.

[0101] Reference Figure 4 , the image / video decoder 400 may include an entropy decoder 410, a residual processor 420, a predictor 430, an adder 440, a filter 450, and a memory 460. The predictor 430 may include an inter-frame predictor 431 and an intra-frame predictor 432. The residual processor 420 may include a dequantizer 421 and an inverse transformer 422. Depending on the implementation, the entropy decoder 410, the residual processor 420, the predictor 430, the adder 440, and the filter 450 may be configured by one hardware component (e.g., a decoder). Chipset or processor) configuration. In addition, the memory 460 may include a decoded picture buffer (DPB) and may be configured by a digital storage medium. The hardware component may further include a memory 460 as an internal / external component.

[0102] When a bit stream containing video / image information is input, the image / video decoder 400 can be used with Figure 3 The image / video can be reconstructed correspondingly to the process of processing image / video information in the image / video encoder 300. For example, the image / video decoder 400 can derive the unit / block based on the block segmentation related information obtained from the bitstream. The image / video decoder 400 can perform decoding using the processing unit applied in the image / video encoder. Accordingly, the decoding processing unit can be, for example, a coding unit, and the coding unit can be segmented from the coding tree unit or the maximum coding unit according to a quadtree structure, a binary tree structure and / or a ternary tree structure. One or more transform units can be derived from the coding unit. In addition, the reconstructed image signal decoded and output by the image / video decoder 400 can be played by a playback device.

[0103] The image / video decoder 400 may receive the data in the form of a bit stream from Figure 3The signal output by the encoder of the video decoder 400 is decoded by the entropy decoder 410. For example, the entropy decoder 410 can parse the bitstream to derive the information (e.g., image / video information) necessary for image reconstruction (or picture reconstruction). The image / video information may further include information about various parameter sets, such as an adaptive parameter set (APS), a picture parameter set (PPS), a sequence parameter set (SPS), or a video parameter set (VPS). In addition, the image / video information may further include general constraint information. In addition, the image / video information may include methods, purposes, etc. for generating and using decoding information. The image / video decoder 400 may further decode the picture based on the information about the parameter set and / or the general constraint information. The information and / or syntax elements sent / received by the signal may be decoded and obtained from the bitstream through the decoding process. For example, the entropy decoder 410 may decode the information in the bitstream based on a coding method such as exponential Golomb coding, CAVLC, or CABAC, and output the values ​​of the syntax elements necessary for image reconstruction and the quantized values ​​of the transform coefficients related to the residual. More specifically, in the CABAC entropy decoding method, a bin corresponding to each syntax element can be received in a bitstream, the decoding target syntax element information and the decoding information of the neighboring blocks and the decoding target blocks or the information about the symbol / bin decoded in the previous step can be used to determine the context model, and the bin can be arithmetically decoded to generate a symbol corresponding to the value of each syntax element. At this time, the CABAC entropy decoding method can use the information about the decoded symbol / bin for the context model of the next symbol / bin after determining the context model to update the context model. The information about prediction among the information decoded in the entropy decoder 410 is provided to the predictor (inter-frame predictor 432 and intra-frame predictor 431), and the residual value obtained by performing entropy decoding in the entropy decoder 410, that is, the quantized transform coefficient and related parameter information can be input to the residual processor 420. The residual processor 420 can derive a residual signal (residual block, residual sample, residual sample array). In addition, the information about filtering among the information decoded by the entropy decoder 410 can be provided to the filter 450. Meanwhile, a receiver (not shown) receiving a signal output from the image / video encoder may be further configured as an internal / external element of the image / video decoder 400, or the receiver may be a component of the entropy decoder 410. Meanwhile, the image / video decoder according to the present disclosure may be referred to as an image / video decoding device, and the image / video decoder may be divided into an information decoder (image / video information decoder) and a sample decoder (image / video sample decoder).In this case, the information decoder may include an entropy decoder 410 , and the sample decoder may include at least one of a dequantizer 321 , an inverse transformer 322 , an adder 440 , a filter 450 and a memory 460 , an inter-frame predictor 432 or an intra-frame predictor 431 .

[0104] The dequantizer 421 may dequantize the quantized transform coefficients and output the transform coefficients. The dequantizer 421 may rearrange the quantized transform coefficients into a two-dimensional block form. In this case, the rearrangement may be performed based on the coefficient scanning order performed in the image / video encoder. The dequantizer 321 may dequantize the quantized transform coefficients using quantization parameters (e.g., quantization step size information) and obtain the transform coefficients.

[0105] The inverse transformer 422 inversely transforms the transform coefficients to obtain a residual signal (residual block, residual sample array).

[0106] The predictor 430 may perform prediction on the current block and generate a prediction block including prediction samples for the current block. The predictor may determine whether to apply intra prediction or inter prediction to the current block based on information about the prediction output from the entropy decoder 410, and may determine a specific intra / inter prediction mode.

[0107] The predictor 420 can generate a prediction signal based on various prediction methods. For example, the predictor can not only apply intra prediction or inter prediction to the prediction of a block, but also apply intra prediction and inter prediction at the same time. This can be called combined inter and intra prediction (CIIP). In addition, the predictor can be based on an intra block copy (IBC) prediction mode or a palette mode for the prediction of the block. The IBC prediction mode or the palette mode can be used, for example, for image / video coding of content such as games, such as screen content coding (SCC). In IBC, the prediction is basically performed within the current picture, but can be performed similarly to inter prediction because a reference block is derived within the current picture. That is, IBC can use at least one of the inter prediction techniques described in this document. The palette mode can be regarded as an example of intra coding or intra prediction. When the palette mode is applied, information about the palette table and the palette index can be included in the image / video information and sent with a signal.

[0108] The intra-frame predictor 431 can predict the current block by referring to samples in the current picture. Depending on the prediction mode, the referenced samples can be located in the neighborhood of the current block, or can be located far away from the current block. In intra-frame prediction, the prediction mode can include multiple non-directional modes and multiple directional modes. The intra-frame predictor 431 can use the prediction mode applied to the neighboring blocks to determine the prediction mode applied to the current block.

[0109] The inter-frame predictor 432 may derive a prediction block for the current block based on a reference block (reference sample array) specified by a motion vector in a reference picture. At this time, in order to reduce the amount of motion information sent in the inter-frame prediction mode, the motion information may be predicted in units of blocks, sub-blocks, or samples based on the correlation of the motion information between the neighboring blocks and the current block. The motion information may include a motion vector and a reference picture index. The motion information may further include inter-frame prediction direction (L0 prediction, L1 prediction, Bi prediction, etc.) information. In the case of inter-frame prediction, the neighboring blocks may include spatial neighboring blocks existing in the current picture and temporal neighboring blocks existing in the reference picture. For example, the inter-frame predictor 432 may construct a motion information candidate list based on the neighboring blocks, and derive the motion vector and / or reference picture index of the current block based on the received candidate selection information. Inter-frame prediction may be performed based on various prediction modes, and information about the prediction may include information indicating a mode of inter-frame prediction for the current block.

[0110] The adder 440 can generate a reconstructed signal (reconstructed picture, reconstructed block, reconstructed sample array) by adding the obtained residual signal to the prediction signal (prediction block, prediction sample array) output from the predictor (including the inter-frame predictor 432 and / or the intra-frame predictor 431). If there is no residual for the processing target block, such as when the skip mode is applied, the prediction block can be used as the reconstructed block.

[0111] The adder 440 may be referred to as a reconstructor or a reconstructed block generator. The generated reconstructed signal may be used for intra prediction of the next processing target block in the current picture, may be output after filtering as described later, or may be used for inter prediction of the next picture.

[0112] In addition, luma mapping with chroma scaling (LMCS) can be applied during picture decoding.

[0113] The filter 450 can improve the subjective / objective image quality by applying filtering to the reconstructed signal. For example, the filter 450 can generate a modified reconstructed picture by applying various filtering methods to the reconstructed picture, and send the modified reconstructed picture to the memory 460, specifically to the DPB of the memory 460. The various filtering methods may include, for example, deblocking filtering, sample adaptive offset, adaptive loop filter, bilateral filter, etc.

[0114] The (modified) reconstructed picture stored in the DPB of the memory 460 may be used as a reference picture in the inter-frame predictor 432. The memory 460 may store the motion information of the block from which the motion information in the current picture is derived and / or the motion information of the block in the reconstructed picture. The stored motion information may be transmitted to the inter-frame predictor 432 to be used as the motion information of the spatial neighboring block or the motion information of the temporal neighboring block. The memory 460 may store the reconstructed samples of the reconstructed block in the current picture and transmit them to the intra-frame predictor 431.

[0115] In addition, the VCM decoder (or feature / feature map decoder) performs a series of processes such as prediction, inverse transformation, and dequantization to decode the feature / feature map, and may have substantially the same functions as those described above with reference to Figure 4 The VCM decoder has the same / similar structure as the image / video decoder 400 described above. However, the VCM decoder is different from the image / video decoder 400 in that the feature / feature map is a decoding target, and the name of each unit (or component) (e.g., DPB, etc.) and its specific operation may be different from the image / video decoder 400. The operation of the VCM decoder may correspond to the operation of the VCM encoder, and the specific operation will be described in detail later.

[0116] Feature / Feature Map Encoding Process

[0117] Figure 5 is a flow chart schematically illustrating a feature / feature map encoding process to which embodiments of the present disclosure may be applied.

[0118] Reference Figure 5 The feature / feature map encoding process may include a prediction process (S510), a residual processing process (S520), and an information encoding process (S530).

[0119] The prediction process (S510) can be referred to above. Figure 3 The predictor 320 described herein performs.

[0120] Specifically, the intra-frame predictor 322 can predict the current block (i.e., the set of current encoding target feature elements) by referring to the feature elements in the current feature / feature map. Intra-frame prediction can be performed based on the spatial similarity of the feature elements constituting the feature / feature map. For example, the feature elements included in the same region of interest (RoI) in the image / video can be estimated to have similar data distribution characteristics. Therefore, the intra-frame predictor 322 can predict the current block by referring to the feature elements that have been reconstructed in the region of interest including the current block. At this time, depending on the prediction mode, the referenced feature elements can be located at a position adjacent to the current block or can be located at a position away from the current block. The intra-frame prediction mode for feature / feature map encoding may include multiple non-directional prediction modes and multiple directional prediction modes. The non-directional prediction mode may include, for example, a prediction mode corresponding to a DC mode and a plane mode of an image / video encoding process. In addition, the directional mode may include a prediction mode corresponding to, for example, 33 directional modes or 65 directional modes of an image / video encoding process. However, this is an example, and the type and number of intra-frame prediction modes may be set / changed in various ways depending on the implementation.

[0121] The inter-frame predictor 321 may predict the current block based on a reference block (i.e., a set of reference feature elements) specified by motion information about a reference feature / feature map. Inter-frame prediction may be performed based on the temporal similarity of feature elements constituting a feature / feature map. For example, temporally continuous features may have similar data distribution characteristics. Therefore, the inter-frame predictor 321 may predict the current block by referring to the reconstructed feature elements of features that are temporally adjacent to the current feature. At this time, the motion information for specifying the reference feature element may include a motion vector and a reference feature / feature map index. The motion information may further include information about the inter-frame prediction direction (e.g., L0 prediction, L1 prediction, Bi prediction, etc.). In the case of inter-frame prediction, the neighboring blocks may include spatial neighboring blocks present in the current feature / feature map and temporal neighboring blocks present in the reference feature / feature map. The reference feature / feature map including the reference block and the reference feature / feature map including the temporal neighboring block may be the same or different. The temporal neighboring block may be referred to as a collocated reference block, etc., and the reference feature / feature map including the temporal neighboring block may be referred to as a collocated feature / feature map. The inter-frame predictor 321 can construct a motion information candidate list based on neighboring blocks and generate information indicating which candidate is used to derive the motion vector and / or reference feature / feature map index of the current block. Inter-frame prediction can be performed based on various prediction modes. For example, in the case of skip mode and merge mode, the inter-frame predictor 321 can use the motion information of the neighboring blocks as the motion information of the current block. In the case of skip mode, unlike the merge mode, the residual signal may not be sent. In the case of motion vector prediction (MVP) mode, the motion vector of the neighboring block is used as a motion vector predictor, and the motion vector of the current block can be indicated by sending a motion vector difference with a signal. In addition to the above-mentioned intra-frame prediction and inter-frame prediction, the predictor 320 can also generate a prediction signal based on various prediction methods.

[0122] The prediction signal generated by the predictor 320 may be used to generate a residual signal (residual block, residual feature element) (S520). The residual processing process (S520) may be referred to above. Figure 3 The residual processor 330 described above is performed. In addition, a (quantized) transform coefficient may be generated by a transform and / or quantization process for the residual signal, and the entropy encoder 340 may encode information about the (quantized) transform coefficient in the bitstream as residual information (S530). In addition, in addition to the residual information, the entropy encoder 340 may also encode information necessary for feature / feature map reconstruction in the bitstream, such as prediction information (e.g., prediction mode information, motion information, etc.).

[0123] In addition, the feature / feature map encoding process may further include not only a process for encoding information used for feature / feature map reconstruction (for example, prediction information, residual information, segmentation information, etc.) and outputting it in the form of a bit stream (S530), a process for generating a reconstructed feature / feature map for the current feature / feature map, and a process for applying loop filtering to the reconstructed feature / feature map (optional).

[0124] The VCM encoder can derive (modified) residual features from the quantized transform coefficients by dequantization and inverse transformation, and generate reconstructed features / feature maps based on the predicted features and (modified) residual features of the output of step S510. The reconstructed features / feature maps generated in this way can be the same as the reconstructed features / feature maps generated in the VCM decoder. When the loop filtering process is performed on the reconstructed features / feature maps, the modified reconstructed features / feature maps can be generated by performing a loop filtering process on the reconstructed features / feature maps. The modified reconstructed features / feature maps can be stored in a decoded feature buffer (DFB) or a memory and used as reference features / feature maps in a later feature / feature map prediction process. In addition, (loop) filtering related information (parameters) can be encoded and output in the form of a bitstream. Through the loop filtering process, noise that may occur during feature / feature map coding can be removed, and the task performance based on the feature / feature map can be improved. In addition, by performing the loop filtering process at both the encoder stage and the decoder stage, the identity of the prediction results can be guaranteed, the reliability of the feature / feature map coding can be improved, and the amount of data transmission for feature / feature map coding can be reduced.

[0125] Feature / Feature Map Decoding Process

[0126] Figure 6 is a flow chart schematically illustrating a feature / feature map decoding process to which embodiments of the present disclosure may be applied.

[0127] Reference Figure 6, the feature / feature map decoding process may include an image / video information acquisition process (S610), a feature / feature map reconstruction process (S620 to S640), and a loop filtering process (S650) for reconstructing the feature / feature map. The feature / feature map reconstruction process may be performed on the prediction signal and the residual signal obtained by the inter / intra prediction (S620) and residual processing (S630) described in the present disclosure, and the dequantization and inverse transformation process for the quantized transform coefficients described in the present disclosure. The modified reconstructed feature / feature map may be generated by a loop filtering process for the reconstructed feature / feature map, and the modified reconstructed feature / feature map may be output as the decoded feature / feature map. The decoded feature / feature map may be stored in a decoded feature buffer (DFB) or a memory, and used as a reference feature / feature map in the inter prediction process when decoding the feature / feature map. In some cases, the above-mentioned loop filtering process may be omitted. In this case, the reconstructed features / feature maps can be output as decoded features / feature maps without change and stored in a decoded feature buffer (DFB) or memory, and then used as reference features / feature maps in the inter-frame prediction process when decoding the features / feature maps.

[0128] Implementation

[0129] In the case where a compressed bit stream of an image is generated to perform a machine task, in the prior art, any one predetermined prediction model is selected to perform intra-frame prediction.

[0130] Since the number of selectable prediction models is limited, the prior art may not be sufficient to apply a prediction model suitable for the characteristics of the current block, thereby possibly increasing the bit amount of the bitstream.

[0131] When a variable prediction model is applied to perform prediction of features suitable for the current block, the same parameters need to be used on the feature encoding device 10 and the feature decoding device 20. Therefore, the parameters of the prediction model used in the feature encoding device 10 need to be sent to the feature decoding device 20, and the increase in the variability of the prediction model may also increase the number of corresponding parameters, thereby possibly increasing the bit amount of the bitstream.

[0132] The present disclosure proposes a method for performing intra-frame prediction on a current block, which method not only uses a prediction model that has been fully trained and thus has predetermined parameters, but also uses a prediction model with newly learned parameters, which are learned by using reconstructed neighbor reference samples (reconstructed neighbor samples).

[0133] In addition, the present disclosure proposes a method for storing fully learned parameters in a list and then using one of the parameters in the list as a parameter for current block prediction. In addition, the present disclosure also proposes a method for using one of the parameters in the list as an initial value for training to predict the current block.

[0134] Therefore, through the present disclosure, if a compressed bitstream of an image is generated to perform a machine task, the bitstream having a smaller size (fewer bits) can enable the machine task to be performed with a similar degree of accuracy compared to existing methods.

[0135] The embodiments described in this disclosure may operate alone or in combination thereof.

[0136] Figure 8 is a flowchart illustrating a feature encoding method according to an embodiment of the present disclosure.

[0137] Reference Figure 8 , the feature encoding device 10 may train a prediction model based on the reconstructed neighbor samples (S810). "Prediction model" may be "parameters of the prediction model", and "training of the prediction model" may be "training of the parameters of the prediction model". The training of the prediction model may be performed n times or more (n is an integer equal to or greater than 0).

[0138] The reconstructed neighbor samples may be samples decoded by intra prediction based on training, or samples decoded by using an existing intra prediction method. In addition, the reconstructed neighbor samples may be samples decoded by using other prediction methods including inter prediction.

[0139] exist Figure 7 An example of a reconstructed neighbor sample and a current encoding / decoding target sample (current sample or current block) that can be used to train a prediction model is shown in FIG. Figure 7 , the samples in the area marked with vector x represent the current encoding / decoding target samples, and the samples in the areas marked with r, r0, r1, r2, r3, x0, x1, x2 and x3 represent the reconstructed neighbor samples adjacent to the upper left of the current block.

[0140] exist Figure 7 The position distribution, shape, and number of neighbor samples used as training data in are only an example for the convenience of description, and the actual neighbor samples can have the same Figure 7 For example, the reconstructed neighbor samples used as training data to ensure the independence of encoding and decoding of intra prediction modes can be restricted to be used only in intra encoding and decoding modes, or can be restricted to be used only in predefined same encoding and decoding units (e.g., tiles, slices, etc.).

[0141] In the case of using the value of the reconstructed neighbor sample r to predict the value of the sample (current sample) x of the current block, the value of the sample (current sample) x of the current block is given by the parameter θ t The prediction model f predicts the value of the prediction sample x pred It can be expressed as shown in Equation 1.

[0142] [Equation 1]

[0143] x pred =f(r;θ t )

[0144] In Equation 1, θ t It represents the parameter value of the prediction model after t training processes according to the predetermined training method.

[0145] Training θ t The process can be described as repeated search so that the loss function L tr Minimize θ t The optimization process, loss function L tr It can be represented by the sum of the L2 norm, as shown in Equation 2.

[0146] [Equation 2]

[0147] L tr =∑ n ||x n -f(r n θ t )||2

[0148] As shown in Equation 2, training θ t The process can start from the initial value and use N pairs of reconstructed reference samples {r o ,x0},{r1,x1},{r2,x2},{r3,x3},…,{r N ,x N} as training data, and reconstructed x N As input. The loss function in equation 2 is just an example for convenience of description, and different equations can be used.

[0149] The feature encoding device 10 may predict the current block using the trained prediction model (S820). The feature encoding device 10 may generate residual information for the current block based on the predicted current block (prediction block) and the current block, and encode the generated residual information (S830).

[0150] Fig. 9 is a flow chart illustrating a feature decoding method according to an embodiment of the present disclosure.

[0151] Reference Fig. 9, the feature decoding device 20 can train a prediction model based on the reconstructed neighbor samples (S910).

[0152] The prediction model may be a parameter of the prediction model, and the training of the prediction model may be the training of the parameter of the prediction model. The training of the prediction model may be performed n times or more (n is an integer greater than or equal to 0).

[0153] The reconstructed neighbor samples may be samples decoded by intra prediction based on training, or samples decoded by using an existing intra prediction method. In addition, the reconstructed neighbor samples may be samples decoded by using other prediction methods including inter prediction. Examples of the reconstructed neighbor samples and the current encoding / decoding target samples used for training the prediction model may be the same as Figure 7 The same example.

[0154] The feature decoding device 20 may predict the current block by using the trained prediction model (S920). The feature decoding device 20 may reconstruct the current block based on the prediction block and the residual block for the current block (S930). The residual block may be derived based on the residual information signaled from the feature encoding device 10.

[0155] Implementation Method 1

[0156] Embodiment 1 corresponds to a method for training a prediction model by using reconstructed neighbor samples and then performing prediction by the trained prediction model.

[0157] Fig.10 is a flowchart showing a feature encoding method according to embodiment 1. Fig.10 The feature encoding method shows that the parameters θ of the training prediction model t , the process of predicting the current sample and encoding its difference.

[0158] Reference Fig.10 , a variety of variables that can be used to train the prediction model can be configured (S1002), and the prediction model can be trained based on the reconstructed neighbor samples (S1004). The parameters θ of the prediction model t The training (S1004) can start from the initial value θ0 and continue to reduce the loss function L tr , as exemplified by Equation 2.

[0159] The prediction model with the initial value θ0 may be a “first prediction model”. According to an embodiment, the first prediction model may be a prediction with the initial value θ0, or a prediction model derived as a result of previous training when the prediction model is repeatedly trained. Therefore, the training of the prediction model may include a process of predicting the current block based on the reconstructed neighbor samples and the first prediction model.

[0160] As shown in Equation 2, the loss function L tr The value derived based on the difference between the predicted sample and the reconstructed neighbor sample may correspond to the value derived based on the difference between the predicted sample and the reconstructed neighbor sample. Therefore, the training of the prediction model may include a process of training the first model so that the value derived based on the difference between the predicted sample and the reconstructed neighbor sample decreases.

[0161] The training results can be stored as θ t+1 , and such a process can be repeatedly performed (t=t+1) (S1006). The above process can be repeated for a predefined number of times T max (S1012), and can be obtained by using the neighbor sample r to which the reconstruction is applied and the training result θ t The prediction model is used to derive the predicted value x of Equation 1 pred (S1006). In addition, as shown in Equation 3, the loss function L(x, x) with respect to the value x of the current sample can be derived in the form of L2 norm. pred )(S1008).

[0162] [Equation 3]

[0163] L(x,x pred )=||xx pred ||2

[0164] The loss function of Equation 3 is just an example for convenience of description, and a different equation may be used.

[0165] The loss function L(x,x pred ) can be L min , the corresponding time value θ t It can be set to θ opt , and the corresponding time t can be set to t_opt (information about the number of times) (S1016). Alternatively, in the loss function L(x,x pred ) is less than the predefined threshold L thr (S1008) If the training cannot be performed up to the maximum number of times T max (S1012), then the value θ corresponding to the time can be t Set to θ opt , the corresponding time t can be set to t_opt (S1014).

[0166] The intra-frame prediction result x for encoding can be derived as shown in Equation 4 pred , and the value x of the current sample and the predicted value x can be derived as shown in Equation 5 pred The difference between diff .

[0167] [Equation 4]

[0168] x pred =f(r;θ opt )

[0169] [Equation 5]

[0170] x diff =xx pred

[0171] Can be used to compare the difference x diff Encoding is performed (S1024), and the feature encoding device 10 is notified by a signal to the feature decoding device 20. In addition, if the training process is repeated the same number of times as the encoding process during decoding, the same training result θ as the encoding process can be derived. t , and since this can be applied to prediction, t_opt can also be encoded (S1020) and can be notified from the feature encoding device 10 to the feature decoding device 20 by a signal.

[0172] If t_opt==0 (S1018), f_skip (a first flag) may be encoded (S1022) and may be signaled from the feature encoding device 10 to the feature decoding device 20, the first flag indicating that the training process of the prediction model does not need to be performed. In this case, encoding of t_opt may be skipped.

[0173] According to an embodiment, a default value of t_opt may be predefined in the feature encoding device 10 and the feature decoding device 20. In this case, t_opt may not be encoded. Fig.10 A case where only one prediction model is defined is described, but a plurality of prediction models may be defined, and this case may have an index encoded to select any one prediction model from the plurality of prediction models.

[0174] Fig.11 : is a flowchart showing the feature decoding method according to embodiment 1. Fig.11 The feature decoding method shows that the parameters θ of the training prediction model t , the process of predicting the current sample and reconstructing the current sample by using the decoded difference.

[0175] Reference Fig.11 , a variety of variables that can be used to train the prediction model can be configured (S1100), and the prediction model can be trained based on the reconstructed neighbor samples (S1030). Similar to the feature encoding method, in the training prediction model parameter θ t When θ0 is θ0, the following process can be repeatedly performed (S1140): starting from the initial value θ0, continuing to reduce the loss function L illustrated in equation 2 tr, and store the training results as θ t+1 .

[0176] The prediction model with the initial value θ0 may be a “first prediction model”. According to an embodiment, the first prediction model may be a prediction with the initial value θ0 or a prediction model derived as a result of previous training when the training of the prediction model is repeated. Therefore, the training of the prediction model may include a process of predicting the current block based on the reconstructed neighbor samples and the first prediction model.

[0177] The loss function Ltr illustrated in Equation 2 may correspond to a value derived based on the difference between the predicted sample and the reconstructed neighbor sample. Therefore, the training of the prediction model may include a process of training the first model so that the value derived based on the difference between the predicted sample and the reconstructed neighbor sample is reduced.

[0178] When the training process is repeated the same number of times as indicated by the decoded t_opt (S1120) (S1150), the feature decoding device 20 can thus obtain the same θ as that used by the feature encoding device 10 for prediction. t_opt The same prediction model parameters. As the training process is repeated the same number of times as indicated by t_opt, and the results are used for prediction, the same prediction value x as in the feature encoding device 10 can be derived as shown in Equation 6 pred (S1160).

[0179] [Equation 6]

[0180] x pred =f(r;θ t_opt )

[0181] Alternatively, when the value of f_skip (first flag) as a flag indicating that a separate prediction model parameter training process is not required is 1 (S1110), decoding of t_opt (S1120) and θ t_opt The training process (S1130, S1140, S1150) uses the initial default value of the prediction model to perform prediction, so that the prediction value x can be derived pred (S1160).

[0182] As shown in Equation 7, by using the predicted value x pred The difference between decoding The reconstructed value of the current sample can be exported .

[0183] [Equation 7]

[0184]

[0185] According to an embodiment, a default value of t_opt may be predefined in the feature encoding device 10 and the feature decoding device 20. In this case, t_opt may not be decoded. Fig.11 A case where only one prediction model is defined is described, but a plurality of prediction models may be defined, and this case may have an index that is decoded to select any one prediction model from the plurality of prediction models.

[0186] Tables 1 and 2 show examples of syntax elements that are signaled to perform a feature encoding method and a feature decoding method according to implementation 1. Specifically, Table 1 shows syntax elements indicating whether training-based intra prediction is used within a video sequence. Table 1 shows syntax elements at the SPS level indicating whether training-based intra prediction is used and the size of the available blocks, but this is only an example, and syntax elements may be defined at other levels such as PPS, picture header, slice header, etc. Table 2 shows syntax elements indicating whether the intra prediction mode of an individual CU is training-based intra prediction, the index of the prediction model used, whether additional prediction model parameters are trained, and the number of training times.

[0187] [Table 1]

[0188]

[0189] [Table 2]

[0190]

[0191] sps_lip_enabled_flag indicates whether training-based intra prediction is used in the video sequence. sps_lip_enabled_greater_flag indicates whether training-based intra prediction can be used in blocks larger than a predefined size in the video sequence. Table 2 describes that the predefined size is 4 and the number of sps_lip_enabled_greater_flag is 1, but this is just an example for the convenience of description. That is, the predefined size can be defined as another value, and there can be multiple sps_lip_enabled_greater_flags for various block sizes.

[0192] intra_lip_flag (intra-frame prediction mode information) indicates whether the prediction mode of the corresponding CU is intra-frame prediction based on training. intra_lip_model_idx (prediction model information) indicates the index of the predefined prediction model to be used when applying training-based intra-frame prediction to the corresponding CU. intra_lip_model_idx can indicate the prediction model to be used as the first prediction model among at least one or more candidate prediction models. intra_lip_skip_learning_flag (first flag) indicates whether an additional training process for the prediction model parameters is required when decoding the corresponding CU, which is a syntax element corresponding to f_skip. When intra_lip_skip_learning_flag has a value of 0, it indicates that additional training should be performed, and when intra_lip_skip_learning_flag has a value of 1, it indicates that prediction is performed by using the initial default value of the corresponding prediction model parameter without further training. intra_lip_num_itr indicates the number of times (number information) training is performed using the CU samples reconstructed when decoding the CU and the determined prediction model. intra_lip_num_itr is a syntax element corresponding to t_opt. intra_lip_num_itr can be encoded in various ways, for example, by encoding the value of t_opt as it is or showing the difference from a default value (default number), such as t_opt = default value + intra_lip_num_itr.

[0193] Implementation Method 2

[0194] Implementation 2 proposes a method for storing trained parameters (or trained prediction models) in a list, a method for performing training by using any one of the prediction models stored in the list as an initial value for training, and a method for performing prediction by using any one of the prediction models stored in the list without performing a training process.

[0195] The list in which trained prediction models or untrained prediction models are stored may be referred to as a “prediction model candidate list” or a “parameter list T”, and the prediction model selected from the prediction models in the prediction model candidate list may be referred to as a “first prediction model”.

[0196] Fig.12a and Figure 12b is a flowchart showing a feature encoding method according to Embodiment 2. The feature encoding method of FIG. 12 shows the parameters θ of the prediction model selected and trained. t , the process of predicting the current sample and encoding its difference.

[0197] Reference Fig.12a and Figure 12b , the prediction model used to predict the reconstructed block can be stored in a prediction model candidate list (or parameter list T) (S1228). That is, the prediction model candidate list may include at least one prediction model candidate for predicting at least one adjacent block. For example, if block n is fully encoded and parameter t is used when encoding the block To make a prediction, when encoding block n, the prediction model candidate list can store For ease of description, Fig.12a and Figure 12b Describes each are stored in the prediction model candidate list, but it is possible to store only or old items can be deleted according to predetermined rules.

[0198] As in Equation 2, the parameters θ of the prediction model can be performed t training to reduce the loss function L tr (S1214), and the training result can be stored as θ t+1 (S1216). In addition, these processes may be repeated. In this case, the initial value θ0 of the training may start from the default initial value as in Embodiment 1, or from the n_idxth parameter among the parameters stored in the prediction model candidate list. That is, the first prediction model can be determined based on one of the prediction model candidate list and the predetermined prediction model. Specifically, the first prediction model corresponding to the trained initial value can be determined as the predetermined prediction model (default initial value), or can be determined as the n_idexth prediction model among the prediction models stored in the prediction model candidate list.

[0199] The corresponding process can be repeated a predefined number of times T max (S1222), and can be obtained by using the neighbor sample r to which the reconstruction is applied and the training result θ t The prediction model of derives the predicted value x as in Equation 1 pred (S1216). In addition, as shown in Equation 3, the loss function L(x, x) with respect to the value x of the current sample can be derived. pred )(S1218).

[0200] The loss function L(x,x pred ) can be L min , the corresponding time value θ t It can be set to θ opt , the corresponding time t can be set to t_opt (information about the number of times) (S1226). Alternatively, in the loss function L(x,xpred ) is less than the predefined threshold L thr In the case of (S1218), the training cannot be performed up to the maximum number of times T max (S1222), then the value θ corresponding to the time can be t Set to θ opt , and the corresponding time t can be set to t_opt S1224.

[0201] According to an embodiment, without performing a training process of a prediction model, prediction may be performed using a predefined default initial value as is, or the n_idx-th prediction model in the prediction model candidate list may be used. Prediction is performed (S1204).

[0202] As shown in Equation 4 and Equation 5, the prediction result x can be derived pred and the values ​​of the current sample x and x pred The difference between diff . You can use the difference x diff Encoding is performed (S1232), and a signal is sent from the feature encoding device 10 to the feature decoding device 20. In addition, if the training process is repeated the same number of times as in encoding during decoding, the same training result θ as in encoding can be derived. t , and since this can be applied to prediction, t_opt can also be encoded (S1230), and can be signaled from the feature encoding device 10 to the feature decoding device 20. If t_opt == 0 (S1208), f_skip (first flag) as a flag indicating that the training process of the prediction model does not need to be performed can be encoded (S1212), and signaled from the feature encoding device 10 to the feature decoding device 20. In this case, the encoding of t_opt can be skipped.

[0203] like Fig.12a and Figure 12b As shown, before the training process, it can be determined whether the initial value θ0 of the prediction model training starts from the default initial value, or starts from one of the prediction models in the prediction model candidate list, or whether to use the prediction model in the prediction model candidate list as it is when the training process is skipped (S1202 and S1203). However, this is only an example, and the encoding process (or training process) can be partially or completely performed, and then a more efficient method is selected and performed. In such an example, in the θ used for prediction opt If θ is the result value after additional training process, opt Added to the list of prediction model candidates.

[0204] Although in Fig.12a and Figure 12b There is no expression in θ, but there may be a opt In addition, there may be a condition for deleting a model stored in the prediction model candidate list from the prediction model candidate list. Some of the conditions.

[0205] The value f_list indicating whether to perform training and prediction using the prediction model stored in the prediction model candidate list or to perform prediction using the default initial value may be encoded (S1203 and S1204) and signaled to the feature decoding device 20 from the feature encoding device 10. In this case, the index n_idx (selection information) indicating the prediction model used for prediction may be encoded (S1204) and signaled to the feature decoding device 20 from the feature encoding device 10.

[0206] According to an embodiment, a default value of t_opt may be predefined in the feature encoding device 10 and the feature decoding device 20. In this case, t_opt may not be encoded. Fig.12a and Figure 12b A case where only one prediction model is defined is described, but a plurality of prediction models may be defined, and this case may have an index encoded to select any one prediction model from the plurality of prediction models. In addition, a plurality of prediction model candidate lists may exist.

[0207] Fig.13a and Fig.13b This is a flowchart showing a feature decoding method according to embodiment 2. Fig.13a and Fig.13b The feature decoding method shows how to select and train the parameters θ of the prediction model t , the process of predicting the current sample and reconstructing the current sample by using the decoded difference.

[0208] Reference Fig.13a and Fig.13b , we can diff Decoding is performed, and t=0 can be set (S1302). According to the value of the flag f_list (S1304) indicating whether to use the prediction model stored in the prediction model candidate list to perform training and prediction, the initial value θ0 of the prediction model parameter can be determined. Specifically, when f_list==1, the index n_idx (selection information) can be decoded, and the initial value θ0 of the prediction model parameter can be determined as the prediction model indicated by n_idx in the prediction model candidate list. When f_list==0, the initial value θ0 of the prediction model parameter may be determined as a predefined default value ( S1308 ).

[0209] Similar to the feature encoding method, as illustrated in Equation 3, the parameters θ of the prediction model can be repeatedly performed t training (S1314) to reduce the loss function L tr , and store the training results as θ t+1 (S1316). When the training process is repeated as many times as indicated by the number of decoded t_opt (S1312) (S1318), the feature decoding device 20 can thus obtain the same θ as that used by the feature encoding device 10 for prediction. opt Alternatively, when f_skip (the first flag) as a flag indicating that a separate training process is not required is decoded to a value of 1 (S1310), decoding of t_opt and training of prediction model parameters (S1304) may not be performed according to the value of f_list, and a default value of the prediction model or the n_idx-th prediction model in the prediction model candidate list may be used. As θ opt .

[0210] Since the training process is repeated as many times as indicated by t_opt and prediction is performed using the result, the same prediction value x as in the feature encoding device 10 can be derived as shown in Equation 6. pred (S1322). In addition, as shown in Equation 7, by using the predicted value x pred The difference between decoding The reconstructed value of the current sample can be exported

[0211] According to an embodiment, a default value of t_opt may be predefined in the feature encoding device 10 and the feature decoding device 20. In this case, t_opt may not be decoded. Fig.13a and Fig.13b The case where only one prediction model is defined is described, but multiple prediction models may be defined, and this case may have an index that is decoded to select any one prediction model from the multiple prediction models. In addition, there may be multiple prediction model candidate lists.

[0212] When f_skip==0 (S1310), the θ used for prediction can be opt Add to the prediction model candidate list (S1320). Fig.13a and Fig.13b There is no expression in θ, but there may be one that is used to convert θ opt In addition, there may be a condition for deleting a model stored in the prediction model candidate list from the prediction model candidate list. Some of the conditions.

[0213] Tables 3 and 4 show examples of syntax elements that are signaled to perform a feature encoding method and a feature decoding method according to implementation mode 2. Specifically, Table 3 shows syntax elements indicating whether training-based intra-frame prediction is used within a video sequence and the size of a block to which the training-based intra-frame prediction is applicable. Table 3 shows syntax elements at the SPS level, but this is only an example, and syntax elements can be defined at other levels such as PPS, picture header, slice header, etc. Table 4 shows syntax elements indicating whether the intra-frame prediction mode of an individual CU is based on training intra-frame prediction, the index of the prediction model used, whether to use a prediction model stored in a prediction model candidate list, whether to train additional prediction model parameters, and the number of training times.

[0214] [Table 3]

[0215]

[0216] [Table 4]

[0217]

[0218] sps_lip_enabled_flag indicates whether training-based intra prediction is used in the video sequence. sps_lip_enabled_greater_flag indicates whether training-based intra prediction can be used in blocks larger than a predefined size in the video sequence. Table 3 describes that the predefined size is 4 and the number of sps_lip_enabled_greater_flag is 1, but this is just an example for the convenience of description. That is, the predefined size can be defined as another value, and there can be multiple sps_lip_enabled_greater_flags for various block sizes. sps_lip_parm_list_flag (first flag) indicates whether the prediction model stored in the prediction model candidate list is used in the video sequence.

[0219] intra_lip_flag (intra-frame prediction mode information) indicates whether the prediction mode of the corresponding CU is intra-frame prediction based on training. intra_lip_model_idx (prediction model information) indicates the index of the predefined prediction model to be used when applying training-based intra-frame prediction to the corresponding CU. intra_lip_model_idx can indicate a prediction model to be used as the first prediction model among at least one or more candidate prediction models. intra_lip_parm_from_list_flag indicates whether the value stored in the prediction model candidate list is used as the initial value of the prediction model parameter when decoding the corresponding CU. intra_lip_parm_from_list_flag is a syntax element corresponding to f_lst. When intra_lip_parm_from_list_flag==0, the default initial value is used, and when intra_lip_parm_from_list_flag==1, the index indicating the value stored in the prediction model candidate list is additionally decoded. intra_lip_parm_idx is a syntax element indicating a prediction model among the prediction models stored in the prediction model candidate list to be used as the initial value for the current CU to be decoded. intra_lip_parm_idx is a syntax element corresponding to n_idx (selection information). intra_lip_skip_learning_flag (first flag) indicates whether an additional training process for the prediction model parameters is required when decoding the corresponding CU, which is a syntax element corresponding to f_skip. When intra_lip_skip_learning_flag has a value of 0, it indicates that additional training should be performed, and when intra_lip_skip_learning_flag has a value of 1, it indicates that prediction is performed by using the initial default value of the corresponding prediction model parameter without further training. intra_lip_num_itr indicates the number of times (number information) that training is performed using the sample of the CU reconstructed when the CU is decoded and the determined prediction model. intra_lip_num_itr is a syntax element corresponding to t_opt. intra_lip_num_itr can be encoded in various ways, for example, the value of t_opt is encoded as it is, or the difference from the default value (default number of times) is shown, such as t_opt = default value + intra_lip_num_itr.

[0220] Implementation 3

[0221] The above embodiment has described the encoding of f_skip (first flag), which indicates that if t_opt == 0, there is no need to perform the training process of the prediction model in the decoding process.

[0222] According to an embodiment, t_opt==0 indicates that zero training is performed, and this may indicate that no training is performed. Therefore, when no training is performed, the feature encoding device 10 may encode the value of t_opt as 0, and when the value of t_opt is 0, the feature decoding device 20 may not perform training. In addition, when the value of t_opt is 0, f_skip may not be encoded and decoded.

[0223] Although the exemplary method of the present disclosure described above is represented as a series of operations for the sake of clarity of description, it is not intended to limit the order in which the steps are performed, and the steps may be performed simultaneously or in different orders when necessary. In order to implement the method according to the present disclosure, the steps described may further include other steps, may include remaining steps other than some steps, or may include other additional steps other than some steps.

[0224] In the present disclosure, an image encoding device or an image decoding device that performs a predetermined operation (step) may perform an operation (step) that confirms the execution condition or situation of the corresponding operation (step). For example, if it is described that a predetermined operation is performed when a predetermined condition is met, the image encoding device or the image decoding device may perform the predetermined operation after determining whether the predetermined condition is met.

[0225] The various embodiments of the present disclosure are not a list of all possible combinations, but are intended to describe representative aspects of the present disclosure, and the contents described in the various embodiments may be applied independently or in combination of two or more.

[0226] The embodiments described in the present disclosure may be implemented and executed on a processor, microprocessor, controller or chip. For example, the functional units shown in the figures may be implemented and executed on a computer, processor, microprocessor, controller or chip. In this case, information (e.g., information about instructions) or algorithms for implementation may be stored in a digital storage medium.

[0227] In addition, the decoder (decoding device) and encoder (encoding device) of the embodiment of the present disclosure can be included in multimedia broadcast sending and receiving equipment, mobile communication terminals, home theater video equipment, digital theater video equipment, surveillance cameras, video chat equipment, real-time communication equipment such as video communication, mobile streaming equipment, storage media, cameras, video on demand (VoD) service providing equipment, OTT video (over-the-top video) equipment, Internet streaming service providing equipment, three-dimensional (3D) video equipment, augmented reality (AR) equipment, videophone video equipment, traffic terminals (e.g., vehicle (including autonomous driving vehicle) terminals, robot terminals, aircraft terminals, ship terminals, etc.) and medical video equipment, etc., and can be used to process video signals or data signals. For example, OTT video equipment may include game consoles, Blu-ray players, Internet access TVs, home theater systems, smart phones, tablet PCs, digital video recorders (DVRs), etc.

[0228] In addition, the processing method of the embodiment of the present disclosure can be generated in the form of a program executed by a computer and stored in a computer-readable recording medium. Multimedia data having a data structure according to the embodiment of this document can also be stored in a computer-readable recording medium. Computer-readable recording media include all types of storage devices and distributed storage devices that store computer-readable data. Computer-readable recording media include, for example, Blu-ray discs (BD), universal serial buses (USB), ROMs, PROMs, EPROMs, EEPROMs, RAMs, CD-ROMs, magnetic tapes, floppy disks, and optical data storage devices. In addition, computer-readable recording media include media implemented in the form of carrier waves (e.g., transmitted via the Internet). In addition, the bit stream generated by the encoding method can be stored in a computer-readable recording medium or sent via a wired or wireless communication network.

[0229] In addition, the embodiments of the present disclosure may be implemented as a computer program product through program code, and the program code may be executed on a computer by the embodiments of the present disclosure. The program code may be stored on a computer-readable carrier.

[0230] Fig.14 is a view illustrating an example of a content streaming system to which an embodiment of the present disclosure is applicable.

[0231] Reference Fig.14 The content streaming transmission system to which the embodiments of the present disclosure are applied may mainly include an encoding server, a streaming media server, a network server, a web storage, a user device and a multimedia input device.

[0232] The encoding server compresses the content input from a multimedia input device such as a smart phone, a camera, a camcorder, etc. into digital data to generate a bitstream and transmits the bitstream to the streaming server. As another example, when a multimedia input device such as a smart phone, a camera, a camcorder, etc. directly generates a bitstream, the encoding server may be omitted.

[0233] The bitstream may be generated by applying the image encoding method or the image encoding device according to the embodiment of the present disclosure, and the streaming server may temporarily store the bitstream in the process of sending or receiving the bitstream.

[0234] The streaming media server sends multimedia data to the user device through the network server based on the user's request, and the network server is used as a medium for notifying the user of the service. When the user requests the desired service from the network server, the network server can transmit it to the streaming media server, and the streaming media server can send the multimedia data to the user. In this case, the content streaming system can include a separate control server. In this case, the control server is used to control the command / response between the devices in the content streaming system.

[0235] The streaming server may receive content from a media storage and / or encoding server. For example, when receiving content from an encoding server, the content may be received in real time. In this case, in order to provide a smooth streaming service, the streaming server may store the bitstream for a predetermined time.

[0236] Examples of user devices may include mobile phones, smart phones, laptop computers, digital broadcast terminals, personal digital assistants (PDAs), portable multimedia players (PMPs), navigation, tablet PCs, tablet PCs, ultrabooks, wearable devices (e.g., smart watches, smart glasses, head-mounted displays), digital televisions, desktop computers, digital signage, etc.

[0237] Each server in the content streaming system may operate as a distributed server, in which case data received from each server may be distributed.

[0238] Fig.15 is a diagram showing another example of a content streaming system to which an embodiment of the present disclosure is applicable.

[0239] Reference Fig.15In an embodiment such as VCM, the task may be performed in the user terminal or may be performed in an external device (e.g., a streaming server, an analysis server, etc.) according to the performance of the device, the user's request, the characteristics of the task to be performed, etc. In this way, in order to send the information necessary for performing the task to the external device, the user terminal may generate a bitstream including the information necessary for performing the task (e.g., information such as the task, the neural network, and / or the usage) directly or through the encoding server.

[0240] In an embodiment, the analysis server can perform the task requested by the user terminal after decoding the encoded information received from the user terminal (or from the encoding server). At this time, the analysis server can send the result obtained by executing the task back to the user terminal, or it can send it to another linked service server (e.g., a network server). For example, the analysis server can send the result obtained by executing the task of determining the fire to a fire-related server. In this case, the analysis server may include a separate control server. In this case, the control server can be used as a command / response between each device associated with the analysis server and the server. In addition, the analysis server can request the desired information from the network server based on the task to be performed by the user device and the task information that can be performed. When the analysis server requests the desired service from the network server, the network server sends it to the analysis server, and the analysis server can send the data to the user terminal. In this case, the control server of the content streaming system can be used as a command / response between the devices in the control streaming system. The scope of the present disclosure includes software or machine executable commands (e.g., operating systems, applications, firmware, programs, etc.) for enabling the operation of the methods according to various embodiments to be performed on a device or computer, non-transitory computer-readable media on which such software or commands are stored and can be executed on a device or computer.

[0241] Industrial Applicability

[0242] Embodiments of the present disclosure may be used to encode or decode features / feature maps.

Claims

1. A feature decoding method performed by a feature decoding device, the feature decoding method comprising the following steps: The prediction model is trained based on the reconstructed neighbor samples; Predicting the current block by using the trained prediction model; as well as The current block is reconstructed based on the residual information of the current block and the predicted current block.

2. The feature decoding method according to claim 1, wherein: The training step comprises the following steps: Predicting the current block based on the reconstructed neighbor samples and a first prediction model; and The first prediction model is trained so that a value derived based on a difference between the predicted current block and the reconstructed neighbor samples decreases.

3. The feature decoding method according to claim 2, wherein: The first prediction model is trained as many times as the number of times the information about the number of times is derived, and Therein, the information about the number of times is obtained from a bitstream.

4. The feature decoding method according to claim 3, wherein: The information about the number of times indicates a number obtained by subtracting a predetermined default number of times from the number of times the first prediction model is trained.

5. The feature decoding method according to claim 1, wherein: The training is performed based on prediction mode information obtained from a bitstream and indicating that a prediction mode of the current block is a training-based prediction.

6. The feature decoding method according to claim 2, wherein: The first prediction model is a prediction model indicated by prediction model information obtained from a bitstream among at least one or more candidate prediction models.

7. The feature decoding method according to claim 2, wherein: The first prediction model is determined based on one of a prediction model candidate list or a predetermined prediction model.

8. The feature decoding method according to claim 7, wherein: Based on the value of the first flag obtained from the bitstream being a first value, the first prediction model is determined based on the prediction model candidate list, and Wherein, the value based on the first mark is a second value, and the first prediction model is determined based on the predetermined prediction model.

9. The feature decoding method according to claim 7, in, The prediction model candidate list includes at least one prediction model candidate for predicting at least one neighbor block, and The first prediction model is determined as a prediction model candidate among the at least one prediction model candidate indicated by selection information obtained from the bitstream.

10. A feature encoding method performed by a feature encoding device, the feature encoding method comprising the following steps: The prediction model is trained based on the reconstructed neighbor samples; Predicting a current block by using the trained prediction model; as well as Residual information derived based on the current block and the predicted current block is encoded.

11. A computer-readable recording medium for storing a bit stream generated by the feature encoding method according to claim 10.

12. A method for transmitting a bit stream generated by a feature encoding method, in, The feature encoding method comprises the following steps: The prediction model is trained based on the reconstructed neighbor samples; predicting the current block by using the trained prediction model; and Residual information derived based on the current block and the predicted current block is encoded.

Images