WO2022108361A1 - 신경망 특징맵 양자화 방법 및 장치 - Google Patents
신경망 특징맵 양자화 방법 및 장치 Download PDFInfo
- Publication number
- WO2022108361A1 WO2022108361A1 PCT/KR2021/016998 KR2021016998W WO2022108361A1 WO 2022108361 A1 WO2022108361 A1 WO 2022108361A1 KR 2021016998 W KR2021016998 W KR 2021016998W WO 2022108361 A1 WO2022108361 A1 WO 2022108361A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- feature map
- neural network
- quantization
- layer
- present
- Prior art date
Links
- 238000013139 quantization Methods 0.000 title claims abstract description 293
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 266
- 238000000034 method Methods 0.000 title claims description 92
- 238000003672 processing method Methods 0.000 claims abstract description 28
- 238000009826 distribution Methods 0.000 claims description 111
- 238000010606 normalization Methods 0.000 claims description 43
- 238000012545 processing Methods 0.000 claims description 6
- 238000010586 diagram Methods 0.000 description 62
- 230000008569 process Effects 0.000 description 31
- 238000009827 uniform distribution Methods 0.000 description 31
- 230000006870 function Effects 0.000 description 29
- 238000000605 extraction Methods 0.000 description 21
- 238000005070 sampling Methods 0.000 description 11
- 238000004891 communication Methods 0.000 description 9
- 238000013507 mapping Methods 0.000 description 7
- 238000005192 partition Methods 0.000 description 5
- 238000011176 pooling Methods 0.000 description 5
- 230000006835 compression Effects 0.000 description 4
- 238000007906 compression Methods 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000004590 computer program Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 238000013515 script Methods 0.000 description 3
- 238000000638 solvent extraction Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000012952 Resampling Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0495—Quantised networks; Sparse networks; Compressed networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/771—Feature selection, e.g. selecting representative features from a multi-dimensional feature space
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/13—Adaptive entropy coding, e.g. adaptive variable length coding [AVLC] or context adaptive binary arithmetic coding [CABAC]
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
- H04N19/136—Incoming video signal characteristics or properties
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/42—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals characterised by implementation details or hardware specially adapted for video compression or decompression, e.g. dedicated software implementation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/70—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals characterised by syntax aspects related to video coding, e.g. related to compression standards
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/90—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
- H04N19/91—Entropy coding, e.g. variable length coding [VLC] or arithmetic coding
Definitions
- the present invention relates to a method and apparatus for quantizing a neural network feature map. Specifically, the present invention relates to a method and apparatus for quantizing a neural network feature map using a neural network structure. In addition, the present invention relates to a method and apparatus for quantizing a neural network feature map using feature map classification.
- a video image is compression-encoded by removing spatial and temporal redundancy and inter-view redundancy, which may be transmitted through a communication line or stored in a form suitable for a storage medium.
- An object of the present invention is to improve the coding efficiency of a video signal through efficient neural network feature map quantization.
- a neural network-based signal processing method and apparatus can generate a feature map using a multi-neural network including a plurality of neural networks, and perform quantization on the feature map.
- the quantization may be performed based on a structure of the multiple neural network or an attribute of the feature map.
- the property of the feature map may include a distribution type of sample values in the feature map.
- the quantization may be performed by a quantization method mapped to the distribution type.
- the distribution type may include at least one of a normal distribution, a Gaussian distribution, and a Laplace distribution.
- the performing of the quantization may include performing normalization on sample values in the feature map using a normalization method mapped to the distribution type. .
- the structure of the multiple neural network includes whether the multiple neural networks are connected in series, whether the multiple neural networks are connected in parallel, whether the multiple neural networks are connected in series and in parallel, or the feature map. It may include at least one of the types of layers adjacent to the generated current layer.
- the quantization may be performed by a quantization method mapped to a type of the adjacent layer, and the type of the layer is at least one of a batch normalization layer and a summation layer.
- the method may further include classifying the feature map into a plurality of classes, and the attribute of the feature map may include the class of the feature map.
- the feature map may include a plurality of channels.
- the feature map may be classified into the plurality of classes including one or more channels based on the similarity between the plurality of channels.
- the feature map may be spatially classified based on spatial similarity of input images.
- the present invention it is possible to improve the coding efficiency of a video signal.
- FIG. 1 is a diagram illustrating a hierarchical structure of a multi-neural network according to an embodiment of the present invention.
- FIG. 2 is a diagram illustrating an example of a neural network hierarchical structure according to an embodiment of the present invention.
- FIG. 3 is a diagram illustrating a case in which the characteristic of a feature map is a uniform distribution according to an embodiment of the present invention.
- FIG. 4 is a diagram illustrating a case in which a characteristic of a feature map is a Gaussian distribution according to an embodiment of the present invention.
- FIG. 5 is a diagram illustrating a case in which a characteristic of a feature map is a Laplace distribution as an embodiment of the present invention.
- FIG. 6 illustrates a neural network feature map encoder that encodes a feature map of a neural network as an embodiment of the present invention.
- FIG. 7 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- FIG. 8 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- FIG. 9 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- FIG. 10 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- FIG. 11 is a diagram illustrating a uniform distribution quantization process as an embodiment of the present invention.
- FIG. 12 is a diagram illustrating a Gaussian distribution quantization process as an embodiment of the present invention.
- FIG. 13 is a diagram illustrating a Laplace distribution quantization process as an embodiment of the present invention.
- FIG. 14 illustrates a neural network feature map decoding unit that decodes a neural network feature map as an embodiment of the present invention.
- 15 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- 16 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- 17 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- FIG. 18 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- FIG. 19 is a diagram illustrating a uniform distribution inverse quantization process as an embodiment of the present invention.
- 20 is a diagram illustrating a Gaussian distribution inverse quantization process as an embodiment of the present invention.
- 21 is a diagram illustrating a Laplace distribution inverse quantization process as an embodiment of the present invention.
- FIG. 22 is a diagram for explaining a neural network structure extracted through a neural network structure feature extraction unit as an embodiment of the present invention.
- FIG. 23 is a diagram conceptually illustrating an example of various types of information output through the neural network structure feature extraction unit when the neural network structure is used as an input of the neural network structure feature extraction unit as an embodiment of the present invention.
- FIG. 24 is a diagram illustrating a neural network feature map encoder that encodes a feature map of a neural network as an embodiment of the present invention.
- 25 is a diagram illustrating a feature map classification unit according to an embodiment of the present invention.
- 26 is a diagram illustrating a feature map classification unit according to an embodiment to which the present invention is applied.
- FIG. 27 is a diagram illustrating a feature map classification unit according to an embodiment to which the present invention is applied.
- FIG. 28 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- 29 is a diagram illustrating a partial quantization flowchart of a feature map according to an embodiment of the present invention.
- FIG. 30 is a diagram illustrating a flowchart of partial quantization of a feature map according to an embodiment of the present invention.
- 31 is a diagram illustrating a flowchart of partial quantization of a feature map according to an embodiment of the present invention.
- 32 is a block diagram of a neural network feature map decoding unit according to an embodiment of the present invention.
- 33 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- a neural network-based signal processing method and apparatus can generate a feature map using a multi-neural network including a plurality of neural networks, and perform quantization on the feature map.
- the quantization may be performed based on a structure of the multiple neural network or an attribute of the feature map.
- the property of the feature map may include a distribution type of sample values in the feature map.
- the quantization may be performed by a quantization method mapped to the distribution type.
- the distribution type may include at least one of a normal distribution, a Gaussian distribution, and a Laplace distribution.
- the performing of the quantization may include performing normalization on sample values in the feature map using a normalization method mapped to the distribution type. .
- the structure of the multiple neural network includes whether the multiple neural networks are connected in series, whether the multiple neural networks are connected in parallel, whether the multiple neural networks are connected in series and in parallel, or the feature map. It may include at least one of the types of layers adjacent to the generated current layer.
- the quantization may be performed by a quantization method mapped to a type of the adjacent layer, and the type of the layer is at least one of a batch normalization layer and a summation layer.
- the method may further include classifying the feature map into a plurality of classes, and the attribute of the feature map may include the class of the feature map.
- the feature map may include a plurality of channels.
- the feature map may be classified into the plurality of classes including one or more channels based on the similarity between the plurality of channels.
- the feature map may be spatially classified based on spatial similarity of input images.
- a part when a part is 'connected' to another part, it includes not only a case in which it is directly connected, but also a case in which it is electrically connected with another element interposed therebetween.
- a part of the configuration of the apparatus or a part of the steps of the method may be omitted.
- the order of some of the components of the apparatus or some of the steps of a method may be changed.
- other components or other steps may be inserted into some of the components of the device or some of the steps of the method.
- each component shown in the embodiment of the present invention is shown independently to represent different characteristic functions, and it does not mean that each component is made of separate hardware or a single software component. That is, each component is listed as each component for convenience of description, and at least two components of each component are combined to form one component, or one component can be divided into a plurality of components to perform a function. Integrated embodiments and separate embodiments of each of these components are also included in the scope of the present invention without departing from the essence of the present invention.
- a decoding device (Video Decoding Apparatus), which will be described later, is a civilian security camera, a civilian security system, a military security camera, a military security system, a personal computer (PC), a notebook computer, a portable multimedia player (PMP, Portable MultimediaPlayer), It may be a device included in a server terminal such as a wireless communication terminal, a smart phone, a TV application server, and a service server, and communication for performing communication with the terminal, wired and wireless communication network, etc.
- Various devices including a communication device such as a modem, a memory for storing various programs and data for decoding or decoding an image, inter- or intra-screen prediction for decoding an image, and a microprocessor for operating and controlling the program by executing the program. can mean
- the video encoded as a bitstream by the encoder is transmitted in real time or in non-real time through a wired/wireless communication network such as the Internet, a local area network, a wireless LAN network, a WiBro network, or a mobile communication network, or through a cable, a universal serial bus (USB , Universal Serial Bus) can be transmitted to an image decoding device through various communication interfaces, such as, decoded, restored as an image, and reproduced.
- the bitstream generated by the encoder may be stored in a memory.
- the memory may include both a volatile memory and a non-volatile memory. In this specification, a memory may be expressed as a recording medium storing a bitstream.
- a moving picture may be composed of a series of pictures, and each picture may be divided into coding units such as blocks.
- picture described below can be used in place of other terms having the same meaning, such as image and frame. There will be.
- coding unit may be substituted for other terms having the same meaning, such as a unit block, a block, and the like.
- a method and apparatus for compressing a feature map that is a result (or intermediate result) of a neural network in more detail, information of a neural network structure is used in compressing the feature map.
- a method and apparatus are proposed.
- An embodiment of the present invention provides a method and apparatus for using a plurality of different quantizers in compressing a feature map using information of a neural network structure.
- a structural feature of a neural network is analyzed, and a different quantizer or inverse quantizer is selectively used according to the analyzed feature to improve compression performance. and a decryption method and apparatus.
- the characteristics of the feature map can be considered.
- FIG. 1 is a diagram illustrating a hierarchical structure of a multi-neural network according to an embodiment of the present invention.
- the neural network according to the present embodiment may have a neural network structure composed of multiple neural networks (ie, multiple neural networks).
- each neural network in the multiple neural network may include multiple neural network layers.
- Data expressed in various forms may be transmitted.
- the data may be transmitted in the form of a tensor that is 3D data between adjacent neural networks.
- each neural network in the multi-neural network may be composed of a plurality of layers for performing a function of the neural network.
- the neural network may refer to the entire multi-neural network including a plurality of neural networks, may refer to one neural network among the multiple neural networks, and may refer to all or a part of a neural network layer included in the neural network.
- FIG. 2 is a diagram illustrating an example of a neural network hierarchical structure according to an embodiment of the present invention.
- one neural network layer may include at least one of a filtering layer, an offset summation layer, a first sampling layer, a batch normalization layer, a nonlinear mapping layer, a summation layer, and a second sampling layer.
- the neural network layer shown in FIG. 2 is an example, and the order of the layers may be different from the figure.
- a convolution operation may be performed in the filtering layer.
- the filter used for convolution may be a filter having various dimensions such as one-dimensional, two-dimensional, and three-dimensional.
- a predetermined offset value may be summed in the offset summation layer. In this case, the same number of offset values as the number of filters used in the filtering layer may be summed.
- data may be transmitted to the next priority layer without the offset summation layer.
- sampling may be performed at a predetermined sampling rate on the offset-added data in the first sampling layer.
- Sampling may be performed on all data to which the convolution and offset are summed.
- the position of the data to which the convolution and offset are summed is first sampled, and the convolution and offset summation may be performed only at the corresponding position. have.
- batch normalization may be performed in the batch normalization layer. Batch normalization may normalize the feature map using mean and/or variance values. In this case, the average and/or variance value may be a value learned in the learning process.
- the feature map may be mapped by a nonlinear mapping function.
- various nonlinear functions such as Rectified Linear Unit (ReLU), Leaky ReLU (Leak ReLU), sigmoid, and tanh (Hyperbolic Tangent) may be used as the nonlinear mapping function.
- ReLU Rectified Linear Unit
- Leaky ReLU Leak ReLU
- sigmoid sigmoid
- tanh Hyperbolic Tangent
- the feature map generated in the current or previous layer may be summed with another predetermined feature map.
- the other feature map may be one of previously generated feature maps.
- the summation may mean addition.
- the summation may refer to a combination in which data is connected in a specific dimension.
- the current feature map may be spatially downsampled.
- a downsampling method max pooling sampling with the largest value within a specific range, average pooling sampling with an average value, median pooling sampling with an intermediate value, and DCT using DCT Various downsampling methods such as pooling may be used.
- FIG. 3 is a diagram illustrating a case in which the characteristic of a feature map is a uniform distribution according to an embodiment of the present invention.
- quantization may be performed on the feature map in consideration of the characteristics of the feature map.
- the feature map may be an output of a neural network or a neural network layer.
- the characteristic of the feature map may be a distribution characteristic of values of the feature map.
- the value may be a pixel, sample, or coefficient value of the feature map.
- a quantization method corresponding to the characteristic of the feature map may be predefined.
- a distribution of values of a feature map output from any one layer (referred to as an n-th layer) among layers of a multi-neural network may be a uniform distribution.
- the function of the current neural network may be a case in which a function of generating or predicting noise of a uniform distribution is performed.
- quantization suitable for the uniform distribution when the characteristic of the feature map is a uniform distribution, quantization suitable for the uniform distribution may be applied.
- inverse quantization suitable for uniform distribution when quantization suitable for uniform distribution is applied to the corresponding feature map, inverse quantization suitable for uniform distribution may also be applied to the inverse quantization.
- FIG. 4 is a diagram illustrating a case in which a characteristic of a feature map is a Gaussian distribution according to an embodiment of the present invention.
- a distribution of values of a feature map output from an n-th layer among layers of a multi-neural network may be a Gaussian distribution. It is a graph showing a Gaussian distribution with mean ⁇ and variance ⁇ as an example. In general, when learning using a large amount of data, the distribution of the feature map and the distribution of data are similar, and most general data may follow a Gaussian distribution.
- quantization suitable for the Gaussian distribution when the characteristic of the feature map is a Gaussian distribution, quantization suitable for the Gaussian distribution may be applied.
- quantization suitable for the Gaussian distribution when quantization suitable for the Gaussian distribution is applied to the corresponding feature map, inverse quantization suitable for the Gaussian distribution may also be applied to the inverse quantization.
- FIG. 5 is a diagram illustrating a case in which a characteristic of a feature map is a Laplace distribution as an embodiment of the present invention.
- a distribution of values of a feature map output from an n-th layer among layers of a multi-neural network may be a Laplace distribution.
- the distribution of the feature map may be a Laplace distribution.
- the feature map generated in the current layer may be at least one of a high-frequency component signal, a difference signal, or a detail signal of the feature map generated in the previous layer.
- the signal in general, a Laplace distribution with an average of 0 may be exhibited. That is, when the next layer is the summation layer, the currently generated feature map may generally be a Laplace distribution.
- quantization suitable for the Laplace distribution when the characteristic of the feature map is a Laplace distribution, quantization suitable for the Laplace distribution may be applied.
- quantization suitable for the Laplace distribution when quantization suitable for the Laplace distribution is applied to the feature map, inverse quantization suitable for the Laplace distribution may also be applied to the inverse quantization.
- FIG. 6 illustrates a neural network feature map encoder that encodes a feature map of a neural network as an embodiment of the present invention.
- the neural network feature map encoder may encode a feature map of a neural network generated from multiple neural networks.
- the neural network feature map encoder may include a quantizer, a transform quantizer (or a transform unit), an entropy encoder, a neural network structure feature extractor, and a neural network structure encoder.
- the configuration of the neural network feature map encoder shown in FIG. 6 is an example, and some configurations may be omitted or may be implemented to further include other configurations.
- the multi-neural network may be composed of a plurality of neural networks, and each neural network may be connected in series or in parallel. Alternatively, with respect to one data, some of the multiple neural network structures may be connected in series and others may be connected in parallel.
- a feature map which is an intermediate result (or output), may be generated in the continuous neural network connection.
- one feature map When neural networks are connected in series, one feature map may be generated. And, when the neural networks are connected in parallel, one or more plurality of feature maps may be generated. Each of the plurality of feature maps may have the same size or may be different from each other.
- one or more feature maps that are results (or intermediate results) of multiple neural networks may be compressed through a neural network feature encoder and transmitted to a decoder or stored in a storage device.
- the quantization unit may quantize the input feature map.
- the feature map (or pixel values in the feature map) may be a value expressed as a floating point number. In this case, it may be converted into an integer that can be expressed with a bit depth supported by the encoder. If the values of the feature map are integer types, the values of the feature map may be mapped to a range that can be expressed by the bit depth supported by the encoder.
- the structural features of the neural network in which the feature map is generated are received from the neural network structural feature extractor, and different quantization methods can be selectively or adaptively used according to the features.
- the quantized feature map may be input to the transform quantization unit.
- the transform quantization unit may be referred to as a transform unit.
- the neural network structure feature extraction unit may analyze the structure of the multiple neural network, extract the features, and transmit the extracted features to the quantization unit.
- the feature may be a type of a neural network layer and a next neural network layer in which a feature map to be currently encoded is generated.
- the characteristic may be a position of a layer, such as the number of the neural network layer in which the current neural network is generated in the entire multi-neural network.
- the characteristic when the neural networks are connected in parallel, the characteristic may be a location of a number of parallel connections and index information of the parallel connection.
- the transform quantization unit may transform and quantize the received feature map for encoding and transmit it to the entropy encoding unit.
- spatial transformation of spatially converting high-dimensional data into low-dimensional data may be performed.
- quantization in transform quantization may mean quantization for rate control.
- the feature map is three-dimensional data, and the length along the axis of each dimension may be expressed by horizontal, vertical, depth, or channels.
- the feature map may be converted into two-dimensional data such as an image by connecting all channels of the feature map to one channel.
- the transformed 2D data may be transformed and quantized through an existing image or video encoding method.
- frequency transformation such as DCT and DST may be performed on the feature map, and quantization according to frequency may be performed in the frequency domain.
- the neural network structure encoder receives information on the entire neural network or a partial neural network structure from the multi-neural network and performs symbolization to encode, and the symbolized neural network structure may be transmitted to the entropy encoder.
- the entropy encoding unit may receive the received transform quantized feature map and the neural network structure and entropy-encode it to generate a bitstream.
- FIG. 7 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 7 may be an example of a process performed by a quantizer.
- this embodiment may be performed in the quantization unit of the neural network feature map encoder described above with reference to FIG. 6 .
- the quantization unit may extract (or obtain) a histogram (or a characteristic) by using all values of the feature map.
- the quantizer may determine whether the distribution of the extracted histogram is a Gaussian distribution.
- information on whether a Gaussian distribution is present may be transmitted to a decoder through an entropy encoder.
- the similarity with the Gaussian function obtained through the average and variance of the feature map can be measured and judged through the similarity. If the Gaussian distribution is followed, Gaussian distribution quantization may be performed. Otherwise, the quantizer may check whether the Laplace distribution is followed.
- information on whether the Laplace distribution is followed may be transmitted to the decoder through the entropy encoder.
- the similarity between the Laplace function created using the mean and variance of the feature map and the distribution of the feature map can be measured and judged based on the similarity. If it is determined that the Laplace distribution is followed, Laplace distribution quantization may be performed, and in the opposite case, uniform distribution quantization may be performed.
- FIG. 8 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 8 may be an example of a process performed by a quantizer.
- this embodiment may be performed in the quantization unit of the neural network feature map encoder described above with reference to FIG. 6 .
- the quantizer may determine whether the next layer is the summation layer. In the case of a summation layer, Laplace distribution quantization may be performed. Conversely, in the case of a layer other than the summation layer, Gaussian distribution quantization may be performed.
- FIG. 9 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 9 may be an example of a process performed by a quantizer.
- this embodiment may be performed in the quantization unit of the neural network feature map encoder described above with reference to FIG. 6 .
- the quantizer may check whether the previous layer is a batch normalization layer through the neural network feature received from the neural network structure feature extractor.
- the previous layer may mean a layer in which a feature map to be currently encoded is generated.
- Gaussian distribution quantization may be performed.
- the quantizer may directly perform Gaussian distribution quantization.
- the previous layer is not a batch normalization layer, it can be checked whether the next layer is a summation layer.
- Laplace distribution quantization may be performed.
- Gaussian distribution quantization may be performed.
- FIG. 10 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 10 may be an example of a process performed by a quantizer.
- this embodiment may be performed in the quantization unit of the neural network feature map encoder described above with reference to FIG. 6 .
- the quantizer may first determine whether to use uniform distribution quantization. In this case, whether to use uniform distribution quantization may be determined by a user input or an agreement between the encoder and the decoder. Alternatively, it may be determined according to a specific layer index. The determined uniform distribution quantization usage information may be transmitted to the decoder through the entropy encoder.
- the quantizer may check whether the previous layer is a batch normalization layer through the neural network feature received from the neural network structure feature extractor. When uniform distribution quantization is used, uniform distribution quantization may be performed. Otherwise, the quantizer may check whether the previous layer is a batch normalization layer.
- Gaussian distribution quantization may be performed.
- the quantizer may check whether the next layer is the summation layer. If the next layer is a summation layer, Laplace distribution quantization may be performed. When the next layer is a layer other than the summation layer, Gaussian distribution quantization may be performed.
- FIG. 11 is a diagram illustrating a uniform distribution quantization process as an embodiment of the present invention.
- a quantizer when uniform distribution quantization is performed, a quantizer (or an encoding apparatus, or an encoder) may perform uniform distribution normalization, uniform distribution quantization, and bit depth clipping.
- the order of the steps shown in FIG. 11 may be changed, and some steps may be omitted or other steps may be added.
- uniform distribution normalization may be performed as in Equation 1 below.
- f, f min , f max , and f norm may represent a feature map value, a minimum value among feature maps, a maximum value among feature maps, and a normalized feature map value, respectively. That is, when the current feature map follows a uniform distribution, the quantization unit may perform linear normalization by mapping the minimum value of the feature map to 0 and the maximum value to (1 ⁇ bitdepth) - 1 of the feature map.
- the normalized feature map may be uniformly distributed quantized through Equation 2 below.
- Q step and level may represent a quantization size and a quantized feature map value, respectively.
- floor(_) may represent a discard operation (or a function).
- the offset U may be an offset for rounding.
- the above-described variables may be information on a boundary of a quantization interval that fits a distribution. For example, when f norm is quantized to 3, Q step is 1, and offset is 0.5, the quantization interval may be [2.5, 3.5), and 2.5 and 3.5 may be the quantization interval boundaries. That is, the quantization interval may be determined as [f norm - offset, f norm + offset -1].
- bit depth clipping may be performed through Equation 3 below.
- the Clip3(min, max, value) function receives the minimum, maximum, and input values for clipping as input and outputs the input value, or sets the minimum value if the input value is less than the minimum value, and the maximum value if the input value is greater than the maximum value. Represents a function that outputs.
- FIG. 12 is a diagram illustrating a Gaussian distribution quantization process as an embodiment of the present invention.
- the quantizer when Gaussian distribution quantization is performed, the quantizer (or the encoding apparatus, the encoder) may perform Gaussian distribution normalization, Gaussian distribution quantization, and bit depth clipping.
- the order of the steps shown in FIG. 12 may be changed, and some steps may be omitted or other steps may be added.
- Gaussian distribution normalization may be performed as in Equation 4 below.
- f, ⁇ , ⁇ , and f norm may mean a feature map value, a feature map average, a feature map variance, and a normalized feature map value, respectively.
- the normalized feature map may be quantized with a Gaussian distribution through Equation 5 below.
- Q step , offset G , and level may mean a quantization size, an offset for rounding, and a quantized feature map value, respectively.
- floor(_) may represent a discard operation (or a function).
- the offset G may be an offset for rounding.
- the above-described variables may be information on a boundary of a quantization interval that fits a distribution.
- bit depth clipping may be performed.
- bit depth clipping may be performed by Equation 3 described above.
- FIG. 13 is a diagram illustrating a Laplace distribution quantization process as an embodiment of the present invention.
- the quantizer when Laplace distribution quantization is performed, the quantizer (or the encoding apparatus, the encoder) may perform Laplace distribution normalization, Laplace distribution quantization, and bit depth clipping.
- the order of the steps shown in FIG. 13 may be changed, and some steps may be omitted or other steps may be added.
- normalization of the Laplace distribution may be performed as in Equation 6 below.
- f, scale, and f norm may mean a feature map value, a feature map scale value, and a normalized feature map value, respectively.
- the normalized feature map may be quantized by Laplace distribution through Equation 7 below.
- Q step and level may mean a quantization size and a quantized feature map value, respectively.
- floor(_) may mean a discard function (or operation).
- offset G may be an offset for rounding.
- the above-described variables may be information on a boundary of a quantization interval that fits a distribution.
- bit depth clipping may be performed.
- bit depth clipping may be performed by Equation 3 described above.
- FIG. 14 illustrates a neural network feature map decoding unit that decodes a neural network feature map as an embodiment of the present invention.
- the neural network feature map decoding unit may decode the neural network feature map.
- the neural network feature map decoding unit may include an entropy decoding unit, an inverse transform quantization unit (or an inverse transform unit), an inverse quantization unit, a neural network structure decoding unit, and a neural network structure feature extracting unit.
- the configuration of the neural network feature map decoding unit shown in FIG. 14 is an example, and some components may be omitted or may be implemented to further include other components.
- the neural network feature map decoder may decode the bitstream received from the encoder to restore the feature map and/or the neural network.
- the reconstructed neural network may be an entire multi-neural network or a partial neural network.
- the transmitted feature map may be the entire neural network after the generated layer.
- the entropy decoding unit may decode the received bitstream to generate a transform-quantized feature map, and may transmit it to the inverse transform quantization unit.
- the symbolized neural network structure can be restored and transmitted to the neural network structure decoder.
- the inverse transform quantization unit may inverse quantize and inverse transform the transform-quantized feature map to transmit the inverse quantization unit.
- the neural network structure decoder may restore the neural network structure by decoding the symbolized neural network structure received from the entropy decoder.
- the restored neural network structure may be transferred to the neural network structure feature extraction step, and the neural network structure feature extraction step may be the same as the step included in the neural network feature map encoding unit.
- the neural network structure feature extraction unit may extract various information such as the entire neural network structure, the order of layers, the index of the layers, and types before or after the current feature map from the reconstructed neural network structure and deliver it to the inverse quantization unit.
- the inverse quantizer may adaptively or selectively perform inverse quantization through the received inverse transform quantized (or inverse transformed) feature map and the features of the neural network structure.
- the inverse quantization step may convert the data format of the feature map according to the format of data used in the subsequent neural network.
- the inverse quantization unit restores the integer type, and in the case of a floating-point-based neural network, it can be restored on a floating-point basis.
- the restored feature map can be an input to the neural network.
- 15 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 15 may be an example of a process performed by the inverse quantizer.
- the present embodiment may be performed in the inverse quantization unit of the neural network feature map decoding unit described above with reference to FIG. 14 .
- the inverse quantizer may receive information on whether the distribution of the current feature map follows a Gaussian distribution or a Laplace distribution from the entropy decoder. When a Gaussian distribution is followed, Gaussian distribution inverse quantization may be performed. Otherwise, the inverse quantizer may additionally check whether the Laplace distribution is followed.
- Laplace distribution quantization may be performed, otherwise, uniform distribution quantization may be performed.
- 16 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 16 may be an example of a process performed by the inverse quantizer.
- the present embodiment may be performed in the inverse quantization unit of the neural network feature map decoding unit described above with reference to FIG. 14 .
- the inverse quantizer may determine whether the next layer is the summation layer through the neural network feature received from the neural network structure feature extractor. When the next layer is the summation layer, Laplace distribution inverse quantization may be performed. If it is a layer other than the summation layer, Gaussian distribution inverse quantization may be performed.
- 17 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 17 may be an example of a process performed by the inverse quantizer.
- the present embodiment may be performed in the inverse quantization unit of the neural network feature map decoding unit described above with reference to FIG. 14 .
- the inverse quantizer may check whether the previous layer is a batch normalization layer through the neural network feature received from the neural network structure feature extractor.
- the previous layer may mean a layer in which a feature map to be currently encoded is generated.
- Gaussian distribution inverse quantization may be performed.
- the inverse quantizer may check whether the next layer is the summation layer. If it is a summation layer, Laplace distribution inverse quantization may be performed. In the case of a layer other than the summation layer, Gaussian distribution inverse quantization may be performed.
- FIG. 18 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 18 may be an example of a process performed by the inverse quantizer.
- the present embodiment may be performed in the inverse quantization unit of the neural network feature map decoding unit described above with reference to FIG. 14 .
- the inverse quantizer may first determine whether to use uniform distribution inverse quantization. In this case, whether to use uniform distribution quantization may be transmitted through the entropy decoder. Alternatively, it may be determined by an agreement between the encoder and the decoder. Alternatively, it may be determined according to a specific layer index.
- the inverse quantizer may check whether the previous layer is a batch normalization layer through the neural network feature received from the neural network structure feature extractor. When uniform distribution inverse quantization is used, uniform distribution inverse quantization may be performed. Otherwise, the inverse quantizer may check whether the previous layer is a batch normalization layer.
- Gaussian distribution inverse quantization may be performed.
- the inverse quantizer may check whether the next layer is the summation layer. In the case of a summation layer, Laplace distribution inverse quantization may be performed. In the case of a layer other than the summation layer, Gaussian distribution inverse quantization may be performed.
- FIG. 19 is a diagram illustrating a uniform distribution inverse quantization process as an embodiment of the present invention.
- the inverse quantizer may perform quantization size scaling, bit depth clipping, and uniform distribution denormalization.
- the order of the steps shown in FIG. 19 may be changed, and some steps may be omitted or other steps may be added.
- quantization magnitude scaling may be performed through Equation 8 below.
- level, Q step , and f dq may mean a quantized feature value, a quantized size, and an inverse quantized feature value, respectively.
- Bit depth clipping may be applied to the inverse quantized feature value through Equation 9 below.
- f, f min , and f max may mean a restored feature map value, a minimum value among the feature maps, and a maximum value among the feature maps, respectively.
- f min and f max may be received from the encoder through the entropy decoder.
- 20 is a diagram illustrating a Gaussian distribution inverse quantization process as an embodiment of the present invention.
- the inverse quantizer (or, a decoding apparatus or a decoder) may perform quantization size scaling, bit depth clipping, and Gaussian distribution denormalization.
- the order of the steps shown in FIG. 20 may be changed, and some steps may be omitted or other steps may be added.
- the quantization magnitude scaling may be performed through Equation 11 below.
- level, Q step , and f dq may mean a quantized feature value, a quantized size, and an inverse quantized feature value, respectively.
- Bit depth clipping may be applied to the inverse quantized feature value through Equation 12 below.
- the clipped feature value may then be uniformly distributed denormalization through Equation 13 below.
- f, ⁇ , and ⁇ may mean a reconstructed feature map value, a feature map average, and a feature map variance value, respectively.
- ⁇ and ⁇ may be received from the encoder through the entropy decoder.
- 21 is a diagram illustrating a Laplace distribution inverse quantization process as an embodiment of the present invention.
- the inverse quantizer (or, a decoding apparatus or a decoder) may perform quantization size scaling, bit depth clipping, and Laplace distribution denormalization processes.
- the order of the steps shown in FIG. 21 may be changed, and some steps may be omitted or other steps may be added.
- the quantization magnitude scaling may be performed through Equation 14 below.
- level, Q step , and f dq may mean a quantized feature value, a quantized size, and an inverse quantized feature value, respectively.
- bit depth clipping may be applied to the inverse quantized feature value through Equation 15 below.
- the clipped feature value may be uniformly distributed denormalization through Equation 16 below.
- f and scale may mean a reconstructed feature map value and a feature map scale value, respectively.
- the scale may be received from the encoder through the entropy decoder.
- FIG. 22 is a diagram for explaining a neural network structure extracted through a neural network structure feature extraction unit as an embodiment of the present invention.
- the neural network structure feature extraction unit may be the neural network structure feature extraction unit described with reference to FIGS. 6 and 14 .
- the neural network structure may be extracted through the neural network structure feature extraction unit, and the extracted features may be used in the feature map encoding process.
- the neural network may have N layers, and when data is input, a feature map may be generated in all layers except for the last layer. In this case, the number of generated feature maps may be N-1 or less.
- Each layer may be of a different type.
- layers of a specific type pattern may be continuously connected. For example, a convolutional layer, an offset summing layer, and an activation layer may be iteratively connected.
- the layer type may be at least one of a convolutional layer, an offset summing layer, a sampling layer, a batch normalization layer, an activation layer, a summing layer, and a pooling layer.
- FIG. 23 is a diagram conceptually illustrating an example of various types of information output through the neural network structure feature extraction unit when the neural network structure is used as an input of the neural network structure feature extraction unit as an embodiment of the present invention.
- the neural network structure may be the neural network structure described above with reference to FIG. 22 .
- the neural network structure feature extraction unit may be the neural network structure feature extraction unit described with reference to FIGS. 6 and 14 .
- the neural network structure extractor may extract a type of an n-th layer to which a current feature map is output and an n+1 layer to which a current feature map is input, from the received neural network structure.
- the neural network structure extractor may check whether the type of the n-th layer is a batch normalization layer.
- the encoder may transmit a parameter used for batch normalization to the quantization layer.
- the decoder may transmit the batch normalization parameters to the inverse quantizer.
- the neural network structure extractor may check whether the type of the n+1th layer is the summation layer.
- the summation layer information on whether the summation layer is present may be transmitted to the quantizer or the inverse quantizer.
- the layer index for which layer and the summation is to be performed and the summation type may be transmitted to the quantizer.
- the summing layer may be at least one of an addition layer and a concatenation layer.
- An embodiment of the present invention proposes a method and apparatus for adaptively quantizing a feature map through feature map analysis for efficient compression of a feature map of a neural network. More specifically, for efficient feature map compression, a quantization method and apparatus for determining importance among feature map channels and applying different quantization sizes according to importance are proposed.
- the feature map which is an intermediate result of the neural network, may mean different features for each channel.
- each characteristic may have different influences on the final prediction predicted through the subsequent neural network. Therefore, among the channels of the feature map, by classifying the channels that have a large influence on the final prediction and the channels that have little influence on the final prediction, the quantization size is adaptively adjusted for each channel using the classified result to improve the encoding efficiency of the feature map compression. do.
- FIG. 24 is a diagram illustrating a neural network feature map encoder that encodes a feature map of a neural network as an embodiment of the present invention.
- the neural network feature map encoder may encode a feature map of a neural network generated from multiple neural networks.
- the neural network feature map encoder may include a feature map classifier, a quantizer, a transform quantizer (or a transformer), an entropy encoder, a neural network structure feature extractor, and a neural network structure encoder.
- the configuration of the neural network feature map encoder shown in FIG. 24 is an example, and some configurations may be omitted or may be implemented to further include other configurations. In the description of the present embodiment, a portion overlapping with the description of FIG. 6 will be omitted.
- the multi-neural network may be composed of a plurality of neural networks, and each neural network may be connected in series or in parallel. Alternatively, with respect to one data, some of the multiple neural network structures may be connected in series and others may be connected in parallel.
- a feature map which is an intermediate result (or output), may be generated in the continuous neural network connection.
- one feature map When neural networks are connected in series, one feature map may be generated. And, when the neural networks are connected in parallel, one or more plurality of feature maps may be generated. Each of the plurality of feature maps may have the same size or may be different from each other.
- one or more feature maps that are results (or intermediate results) of multiple neural networks may be compressed through a neural network feature map encoder and transmitted to a decoder or stored in a storage device.
- the feature map classifier may classify the input feature map and deliver the classified feature map to the quantizer.
- the classification information generated by the feature map classifier may be transmitted to the decoder through the entropy encoder.
- the classification information may be a classification index according to a channel of the feature map, a classification index according to a spatial location, a classification index for a spatial mask, and the like.
- the feature maps classified by the feature map classifier may be transmitted to the quantizer.
- the quantization unit may generate a quantized feature map by individually quantizing the received classified feature map according to a classification index.
- the generated quantized feature map may be transmitted to the transform quantization unit.
- the transform quantization unit may perform transform quantization (or transform) for encoding the received quantized feature map.
- quantization in transform quantization may mean quantization for rate control.
- the transform quantization unit may reconstruct the feature map for each classification index and transform it into 2D data or transform the shape into 1D data. Alternatively, a frequency domain transformation used for general image and video encoding may be applied. After transforming to the frequency domain, the quantized coefficients may be transmitted to the entropy encoder by quantization for rate control.
- 25 is a diagram illustrating a feature map classification unit according to an embodiment of the present invention.
- the feature map may have a size of (width H, height W, channel C). Such a feature map may be classified into N classes by the feature map classification unit.
- the feature map classification unit may have the configuration described above with reference to FIG. 24 .
- the feature map classification unit may store and manage the channel index of the original tensor for each class in the form of a list.
- the channel index information may be transmitted to the decoder through the entropy encoder for quantization and dequantization.
- the feature map classification unit may classify the feature map for each channel by using the degree of similarity between channels. Since most feature maps extract feature values while maintaining the spatial and structural features of the input image, a set of channels with high similarity between channels may exist although the size of the feature values for each channel is different.
- the feature map classifier may classify the channels using a machine learning-based classification method such as a k-means algorithm using the similarity between channels or a deep learning-based classification method.
- a machine learning-based classification method such as a k-means algorithm using the similarity between channels or a deep learning-based classification method.
- the number of classifications may be the same according to the algorithm used.
- each class may have a different number, and in this case, information on the number of each list may be transmitted to the decoder.
- 26 is a diagram illustrating a feature map classification unit according to an embodiment to which the present invention is applied.
- a feature map having a size of (width, height, channel) may be classified into a class by classifying a specific spatial region.
- the classified location information may be transmitted to a decoder through entropy encoding in the form of a class index map.
- the feature map classification unit may classify the input image and video, and apply the classification result to the same location by reflecting the classified result to the feature.
- the feature map classification unit may change the resolution to the same resolution through resampling to reflect the resolution difference.
- the feature map may be classified into blocks having a specific size.
- partition information, partition depth, class index, and the like of the block may be transmitted to the decoder through the entropy encoder.
- FIG. 27 is a diagram illustrating a feature map classification unit according to an embodiment to which the present invention is applied.
- a feature map having a size of (width, height, channel) may be classified into a class by classifying a specific spatial region.
- the classified location information may be transmitted to a decoder through entropy encoding in the form of a class index map.
- the feature map classification unit may classify the input image and video, and apply the classification result to the same location by reflecting the classified result to the feature.
- the feature map classification unit may spatially divide the feature map into blocks and classify the feature map by mapping a class index for each block.
- the block division may be divided into 4 divisions, 2 divisions, and 3 divisions. Also, it may be a diagonal division, and the result of the diagonal division may also be said to be one block.
- the feature map may be divided in various directions such as 8 divisions or 4 divisions of the 3D data itself.
- a block partition can be divided into several layers. Block partition information, partition depth, class index, and the like may be transmitted to the decoder through the entropy encoder.
- FIG. 28 is a diagram illustrating a flowchart of quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 28 may be an example of a process performed by a quantizer.
- this embodiment may be performed in the quantization unit of the neural network feature map encoder described above with reference to FIG. 24 .
- the quantization unit may determine whether to perform partial quantization, and perform full quantization or partial quantization according to the determination result.
- the determination of whether to perform partial quantization may be input from a user.
- the quantizer may determine whether to perform partial quantization according to the function of the neural network. For example, when the function of the neural network has a function of segmenting an image or a video, the quantization unit may perform partial quantization. Alternatively, when the function of the neural network has a function of predicting the position of an object in an image or video, partial quantization may be performed.
- partial quantization may be performed.
- Information on whether partial quantization is performed may be transmitted to the decoder through the entropy encoder.
- the encoder and the decoder may determine it through the neural network function information, respectively.
- the quantization method may be predefined in the encoder and the decoder according to the function of the neural network.
- 29 is a diagram illustrating a partial quantization flowchart of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 29 may be an example of a process performed by the quantizer.
- this embodiment may be performed by the quantization unit of the neural network feature map encoder described above with reference to FIG. 24 .
- the quantizer may check whether channel quantization is performed.
- Information on whether to perform may be input from the encoder user. Alternatively, information on whether to perform may be determined by information determined in a higher stage.
- information on whether to perform channel quantization may be transmitted to a decoder through an entropy encoder.
- channel feature extraction may be performed.
- the quantizer may extract the inter-channel feature of the feature map.
- the inter-channel feature may be an average value of values included in the channel.
- the inter-channel feature may be a variance value.
- the inter-channel feature may be a feature extracted through another neural network.
- the inter-channel feature may be structural complexity.
- the extracted features may be transferred to the feature map channel classification step.
- channels may be classified using features received in the feature map channel classification step.
- the number of classified classes may be less than or equal to the number of channels of the existing feature map.
- the feature map channel classification step the feature map channel may be classified based on the similarity of features for each channel.
- One or more features extracted for each channel may be combined into a one-dimensional vector, and may be classified into K classes through a classification algorithm such as a k-means algorithm based on the similarity between the channel feature vectors. Then, the classified channel information may be transferred to the channel merging and splitting step.
- channel merging and splitting step since encoding efficiency may be reduced when the number of channels included in one class is too small, one or more classes may be merged with other classes to form one class. Alternatively, if too many channels are included in one class, channel division may be performed.
- the finally configured channel information for each class may be delivered to the channel quantization step.
- feature map quantization may be performed by applying different quantization methods according to the class classified in the above step.
- different quantization methods may indicate different quantization steps.
- different quantization methods may indicate that quantized values have different bit depths.
- different quantization methods may indicate that different nonlinear mapping functions are used for quantization.
- Information on whether to perform domain quantization may be input from a user of the encoder. Alternatively, information on whether to perform domain quantization may be determined by information determined in a higher stage. Here, information on whether to perform channel quantization may be transmitted to a decoder through an entropy encoder.
- a domain feature extraction step may be performed.
- a block feature extraction step may be performed for block quantization.
- the quantization unit may extract features from the received feature map for each spatial location and deliver the extracted features to the region feature classification step.
- the quantizer may classify the region based on the received spatial features. In this case, the classified area may be shared and used by all channels of the feature map. Then, the classified regions may be transferred to a region merging and division step.
- the final region may be determined by performing merging and dividing based on the number of classes, the size, width, height, number of pixels, etc. of the divided regions.
- the finally determined regions may be transferred to a region quantization step.
- different quantization methods may be applied according to the received classified domain.
- features may be extracted from the received feature map in units of a specific block size.
- a block may mean a cube that is data in a three-dimensional space, and may also be referred to as a tensor. That is, the block may be unit data obtained by dividing the feature map into smaller units.
- the extracted features may be transferred to a block feature classification step.
- the quantizer may classify the blocks using the received unit block features. Merge and division may be performed into blocks having different widths, heights, and depths using the classified classes.
- the split information of the merged and split blocks may be transmitted to the decoder through the entropy encoder.
- the division information may exist hierarchically, and may be divided into tree structures of various structures, such as 8 division, 4 division, and 2 division.
- class information of the merged and split blocks may be transmitted to the decoder through the entropy encoder in the form of a classification map.
- the values of the classification map may mean indexes of classes.
- the quantization unit may encode the index difference between the prediction value and the current block using neighboring blocks of the current block, and may transmit the encoded difference to the entropy encoding unit.
- different quantization methods may be applied using class information in blocks of various sizes.
- different quantization methods may indicate different quantization sizes.
- various methods such as an offset value, a scale value, vector quantization, and scalar quantization may be applied for each class.
- Information related to the quantization method for each class may be transmitted to the decoder through the entropy encoder.
- FIG. 30 is a diagram illustrating a flowchart of partial quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 30 may be an example of a process performed by the quantizer.
- this embodiment may be performed by the quantization unit of the neural network feature map encoder described above with reference to FIG. 24 .
- the method described above with reference to FIGS. 28 and 29 may be applied to the present embodiment, and a related overlapping description will be omitted.
- the quantization unit may classify the input feature map and extract one or more information on the quantization method by using the classified feature map.
- information on whether partial quantization is included may be included in the extracted information. If partial quantization is not used, full quantization may be performed. Otherwise, partial quantization may be performed.
- the partial quantization related information may be transmitted to the decoder through the entropy encoder.
- 31 is a diagram illustrating a flowchart of partial quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 31 may be an example of a process performed by a quantizer.
- this embodiment may be performed by the quantization unit of the neural network feature map encoder described above with reference to FIG. 24 .
- the method described above with reference to FIGS. 28 to 30 may be applied to the present embodiment, and a related overlapping description will be omitted.
- one or more information on a quantization method may be extracted by classifying an input feature map and using the classified feature map.
- information on whether the channel is quantized may be included in the extracted information. If information on channel quantization is included, characteristics for each channel may be additionally included.
- information on whether to quantize a region may be included in the extracted information. If domain quantization is performed, information on features of a spatial domain may be included.
- the quantizer may perform region merging and division by using the region characteristics.
- block unit features may be included, and block merging and division may be performed using them.
- Information on block division may be transmitted to a decoder through an entropy encoder.
- 32 is a block diagram of a neural network feature map decoding unit according to an embodiment of the present invention.
- the neural network feature map decoding unit may decode the neural network feature map.
- the neural network feature map decoding unit may include an entropy decoding unit, an inverse transform quantization unit (or an inverse transform unit), a feature map dividing unit, and an inverse quantization unit.
- the configuration of the neural network feature map decoding unit shown in FIG. 32 is an example, and some components may be omitted or may be implemented to further include other components.
- the neural network feature map decoder may generate a restored feature map by decoding the bitstream received from the encoder.
- the neural network may be a neural network used by an encoder and a decoder agreement.
- the neural network may be all or a part of the neural network restored through the decoder/encoder.
- the entropy decoding unit may decode the received bitstream to restore the transform-quantized feature map and transmit it to the inverse transform quantization unit. Also, the entropy decoder may restore the feature map classification information and transmit it to the inverse transform quantizer and the inverse quantizer. The inverse transform quantization unit may inverse quantize and inverse transform the transform quantized feature map, and transmit it to the inverse quantizer.
- the inverse quantization unit may perform individual inverse quantization according to classification through the classification information received from the entropy decoder.
- the inverse quantized feature map may be finally delivered to the neural network.
- 33 is a diagram illustrating a flow chart of inverse quantization of a feature map according to an embodiment of the present invention.
- the present embodiment described with reference to FIG. 33 may be an example of a process performed by the inverse quantizer.
- the present embodiment may be performed by the inverse quantization unit of the neural network feature map decoding unit described above with reference to FIG. 32 .
- the inverse quantization unit may receive the received feature map and information related to the inverse quantization method from the entropy decoder.
- the inverse quantization-related information may include information on whether full inverse quantization is performed.
- the inverse quantization related information may include information on whether channel inverse quantization is performed.
- the inverse quantization-related information may additionally include information related to the feature map classification for each channel. can do.
- the inverse quantization-related information may include information on whether to perform domain inverse quantization.
- information used for domain division may be additionally included. For example, in the form of a classification map, it may be transmitted from the encoder through the entropy decoder.
- information related to a quantization step used for inverse quantization, an offset, and the like may be included.
- the inverse quantization unit may check whether the entire inverse quantization is performed. When it is determined to perform full inverse quantization, full inverse quantization may be performed. For full inverse quantization, the same inverse quantization method may be performed on the entire feature map. In this case, a quantization step, an offset, a scaling value, etc. necessary for inverse quantization may be received from the entropy decoder. Alternatively, a predetermined set may be used according to a value received from the entropy decoder.
- inverse quantization related information may check whether channel inverse quantization is performed.
- the feature map may be transmitted to the feature map channel division step.
- the inverse quantizer may divide the channel through the received channel partitioning information to deliver the divided feature map in the channel inverse quantization step.
- the inverse quantizer may check whether domain inverse quantization is performed.
- the received feature map may be transferred to the feature map region division step.
- the inverse quantizer may divide the feature map into a plurality of regions using the received feature map and region segmentation information received from the entropy encoder.
- the divided feature map may be transferred to a region inverse quantization step.
- the inverse quantizer may perform different quantization methods for each domain.
- a feature map block partitioning step may be performed.
- the inverse quantization unit divides the feature map into various sizes and shapes through block division information received from the entropy decoder, receives class information for each block, and applies a quantization method according to the class.
- Embodiments according to the present invention may be implemented by various means, for example, hardware, firmware, software, or a combination thereof.
- an embodiment of the present invention is one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), FPGAs (field programmable gate arrays), a processor, a controller, a microcontroller, may be implemented by a microprocessor.
- ASICs application specific integrated circuits
- DSPs digital signal processors
- DSPDs digital signal processing devices
- PLDs programmable logic devices
- FPGAs field programmable gate arrays
- an embodiment of the present invention is implemented in the form of a module, procedure, or function that performs the functions or operations described above, and a recording medium readable through various computer means.
- the recording medium may include a program command, a data file, a data structure, etc. alone or in combination.
- the program instructions recorded on the recording medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software.
- the recording medium includes a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a compact disk read only memory (CD-ROM), a digital video disk (DVD), and a floppy disk.
- Magneto-Optical Media such as a disk
- hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like.
- program instructions may include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those generated by a compiler.
- Such hardware devices may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.
- the device or terminal according to the present invention may be driven by a command to cause one or more processors to perform the functions and processes described above.
- such instructions may include interpreted instructions, such as script instructions, such as JavaScript or ECMAScript instructions, or executable code or other instructions stored on a computer-readable medium.
- the device according to the present invention may be implemented in a distributed manner across a network, such as a server farm, or may be implemented in a single computer device.
- a computer program (also known as a program, software, software application, script or code) mounted on the device according to the invention and executing the method according to the invention includes compiled or interpreted language or a priori or procedural language. It can be written in any form of programming language, and can be deployed in any form, including stand-alone programs, modules, components, subroutines, or other units suitable for use in a computer environment.
- a computer program does not necessarily correspond to a file in a file system.
- a program may be in a single file provided to the requested program, or in multiple interacting files (eg, files that store one or more modules, subprograms, or portions of code), or portions of files that hold other programs or data. (eg, one or more scripts stored within a markup language document).
- the computer program may be deployed to be executed on a single computer or multiple computers located at one site or distributed over a plurality of sites and interconnected by a communication network.
- the present invention can be applied to a neural network-based quantization method and apparatus.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Signal Processing (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biodiversity & Conservation Biology (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
Description
Claims (10)
- 신경망 기반의 신호 처리 방법에 있어서,복수의 신경망을 포함하는 다중 신경망을 이용하여 특징맵(feature map)을 생성하는 단계; 및상기 특징맵에 대하여 양자화를 수행하는 단계를 포함하되,상기 양자화는 상기 다중 신경망의 구조 또는 상기 특징맵의 속성(attribute)에 기초하여 수행되는, 신경망 기반 신호 처리 방법.
- 제1항에 있어서,상기 특징맵의 속성은 상기 특징맵 내 샘플 값들의 분포 유형을 포함하고, 그리고,상기 양자화는 상기 분포 유형에 맵핑되는 양자화 방법에 의해 수행되는, 신경망 기반 신호 처리 방법.
- 제2항에 있어서,상기 분포 유형은 정규 분포, 가우시안 분포 또는 라플라스 분포 중 적어도 하나를 포함하는, 신경망 기반 신호 처리 방법.
- 제2항에 있어서,상기 양자화를 수행하는 단계는,상기 분포 유형에 맵핑되는 정규화 방법으로 상기 특징맵 내 샘플 값들에 대한 정규화를 수행하는 단계를 포함하는, 신경망 기반 신호 처리 방법.
- 제1항에 있어서,상기 다중 신경망의 구조는 상기 다중 신경망의 직렬 연결 여부, 상기 다중 신경망의 병렬 연결 여부, 상기 다중 신경망의 직렬 및 병렬 연결 여부 또는 상기 특징맵이 생성된 현재 계층에 인접하는 계층의 유형 중 적어도 하나를 포함하는, 신경망 기반 신호 처리 방법.
- 제5항에 있어서,상기 양자화는 상기 인접하는 계층의 유형에 맵핑되는 양자화 방법에 의해 수행되고, 그리고,상기 계층의 유형은 배치 정규화 계층 또는 합산 계층 중 적어도 하나를 포함하는, 신경망 기반 신호 처리 방법.
- 제1항에 있어서,상기 특징맵을 복수의 클래스로 분류하는 단계를 더 포함하고, 그리고,상기 특징맵의 속성은 상기 특징맵의 클래스를 포함하는, 신경망 기반 신호 처리 방법.
- 제7항에 있어서,상기 특징맵은 복수의 채널을 포함하고,상기 특징맵은 상기 복수의 채널간 유사성에 기초하여 하나 이상의 채널을 포함하는 상기 복수의 클래스로 분류되는, 신경망 기반 신호 처리 방법.
- 제7항에 있어서,상기 특징맵은 입력 이미지의 공간적 유사성에 기초하여 공간적으로 분류되는, 신경망 기반 신호 처리 방법.
- 신경망 기반의 신호 처리 장치에 있어서,상기 신호 처리 장치를 제어하는 프로세서; 및상기 프로세서와 결합되고, 데이터를 저장하는 메모리를 포함하되,상기 프로세서는,복수의 신경망을 포함하는 다중 신경망을 이용하여 특징맵(feature map)을 생성하고, 그리고,상기 특징맵에 대하여 양자화를 수행하되,상기 양자화는 상기 다중 신경망의 구조 또는 상기 특징맵의 속성(attribute)에 기초하여 수행되는, 신경망 기반 신호 처리 장치.
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202180084852.1A CN116746149A (zh) | 2020-11-18 | 2021-11-18 | 神经网络特征图量化方法和装置 |
US18/253,381 US20230421764A1 (en) | 2020-11-18 | 2021-11-18 | Neural network feature map quantization method and device |
KR1020237020605A KR20230107869A (ko) | 2020-11-18 | 2021-11-18 | 신경망 특징맵 양자화 방법 및 장치 |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR10-2020-0154765 | 2020-11-18 | ||
KR10-2020-0154766 | 2020-11-18 | ||
KR20200154766 | 2020-11-18 | ||
KR20200154765 | 2020-11-18 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022108361A1 true WO2022108361A1 (ko) | 2022-05-27 |
Family
ID=81709506
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2021/016998 WO2022108361A1 (ko) | 2020-11-18 | 2021-11-18 | 신경망 특징맵 양자화 방법 및 장치 |
Country Status (3)
Country | Link |
---|---|
US (1) | US20230421764A1 (ko) |
KR (1) | KR20230107869A (ko) |
WO (1) | WO2022108361A1 (ko) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2024025386A1 (ko) * | 2022-07-28 | 2024-02-01 | 한국전자기술연구원 | 표준 정규분포 기반 양자화 기법을 이용한 딥러닝 네트워크 부호화/복호화 방법 및 장치 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20190014900A (ko) * | 2017-08-04 | 2019-02-13 | 삼성전자주식회사 | 뉴럴 네트워크의 파라미터들을 양자화하는 방법 및 장치 |
KR20190054454A (ko) * | 2017-11-13 | 2019-05-22 | 삼성전자주식회사 | 인공 신경망의 양자화 방법 및 장치 |
KR20190125141A (ko) * | 2018-04-27 | 2019-11-06 | 삼성전자주식회사 | 뉴럴 네트워크의 파라미터들을 양자화하는 방법 및 장치 |
US20200143169A1 (en) * | 2018-11-01 | 2020-05-07 | Microsoft Technology Licensing, Llc | Video recognition using multiple modalities |
KR20200110613A (ko) * | 2019-03-15 | 2020-09-24 | 삼성전자주식회사 | 정리 및 양자화 공동 실행 심층 신경망 |
-
2021
- 2021-11-18 US US18/253,381 patent/US20230421764A1/en active Pending
- 2021-11-18 KR KR1020237020605A patent/KR20230107869A/ko unknown
- 2021-11-18 WO PCT/KR2021/016998 patent/WO2022108361A1/ko active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20190014900A (ko) * | 2017-08-04 | 2019-02-13 | 삼성전자주식회사 | 뉴럴 네트워크의 파라미터들을 양자화하는 방법 및 장치 |
KR20190054454A (ko) * | 2017-11-13 | 2019-05-22 | 삼성전자주식회사 | 인공 신경망의 양자화 방법 및 장치 |
KR20190125141A (ko) * | 2018-04-27 | 2019-11-06 | 삼성전자주식회사 | 뉴럴 네트워크의 파라미터들을 양자화하는 방법 및 장치 |
US20200143169A1 (en) * | 2018-11-01 | 2020-05-07 | Microsoft Technology Licensing, Llc | Video recognition using multiple modalities |
KR20200110613A (ko) * | 2019-03-15 | 2020-09-24 | 삼성전자주식회사 | 정리 및 양자화 공동 실행 심층 신경망 |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2024025386A1 (ko) * | 2022-07-28 | 2024-02-01 | 한국전자기술연구원 | 표준 정규분포 기반 양자화 기법을 이용한 딥러닝 네트워크 부호화/복호화 방법 및 장치 |
Also Published As
Publication number | Publication date |
---|---|
US20230421764A1 (en) | 2023-12-28 |
KR20230107869A (ko) | 2023-07-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2019009489A1 (ko) | 영상을 부호화/복호화 하는 방법 및 그 장치 | |
WO2016064185A1 (ko) | 최적화 함수를 이용하여 그래프 기반 예측을 수행하는 방법 및 장치 | |
WO2019009490A1 (ko) | 영상을 부호화/복호화 하는 방법 및 그 장치 | |
WO2017065525A2 (ko) | 영상을 부호화 또는 복호화하는 방법 및 장치 | |
WO2020009357A1 (ko) | Cclm에 기반한 인트라 예측 방법 및 그 장치 | |
WO2019009488A1 (ko) | 영상을 부호화 또는 복호화하는 방법 및 장치 | |
WO2017014585A1 (ko) | 그래프 기반 변환을 이용하여 비디오 신호를 처리하는 방법 및 장치 | |
WO2017010850A1 (ko) | 분리 가능한 그래프 기반 변환을 이용하여 비디오 신호를 처리하는 방법 및 장치 | |
WO2016129872A1 (ko) | 그래프 기반 변환을 이용하여 비디오 신호를 처리하는 방법 및 장치 | |
WO2018038554A1 (ko) | 이차 변환을 이용한 비디오 신호의 인코딩/디코딩 방법 및 장치 | |
WO2017155334A1 (ko) | 비디오 코딩 방법 및 장치 | |
WO2012044076A2 (ko) | 비디오의 부호화 방법 및 장치, 복호화 방법 및 장치 | |
WO2011071325A2 (en) | Method and apparatus for encoding and decoding image by using rotational transform | |
WO2016190690A1 (ko) | 적응적인 분리가능한 그래프 기반 변환을 이용하여 비디오 신호를 처리하는 방법 및 장치 | |
WO2021101243A1 (en) | Apparatus and method for using ai metadata related to image quality | |
WO2019143026A1 (ko) | 특징맵 압축을 이용한 이미지 처리 방법 및 장치 | |
WO2022071695A1 (ko) | 영상을 처리하는 디바이스 및 그 동작 방법 | |
WO2016195455A1 (ko) | 그래프 기반 변환을 이용하여 비디오 신호를 처리하는 방법 및 장치 | |
WO2020032609A1 (ko) | 영상 코딩 시스템에서 어파인 머지 후보 리스트를 사용하는 어파인 움직임 예측에 기반한 영상 디코딩 방법 및 장치 | |
WO2019009447A1 (ko) | 영상을 부호화/복호화 하는 방법 및 그 장치 | |
WO2022108361A1 (ko) | 신경망 특징맵 양자화 방법 및 장치 | |
WO2019143024A1 (ko) | 라인 단위 연산을 이용한 초해상화 방법 및 장치 | |
WO2020076142A1 (ko) | 교차성분 선형 모델을 이용한 비디오 신호 처리 방법 및 장치 | |
WO2019143027A1 (ko) | 이미지 파이프라인 처리 방법 및 장치 | |
WO2019088657A1 (ko) | 은닉 변수를 이용하는 영상 및 신경망 압축을 위한 방법 및 장치 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21895129 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18253381 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202180084852.1 Country of ref document: CN |
|
ENP | Entry into the national phase |
Ref document number: 20237020605 Country of ref document: KR Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21895129 Country of ref document: EP Kind code of ref document: A1 |