WO2022211409A1 - 특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 - Google Patents
특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 Download PDFInfo
- Publication number
- WO2022211409A1 WO2022211409A1 PCT/KR2022/004312 KR2022004312W WO2022211409A1 WO 2022211409 A1 WO2022211409 A1 WO 2022211409A1 KR 2022004312 W KR2022004312 W KR 2022004312W WO 2022211409 A1 WO2022211409 A1 WO 2022211409A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- feature map
- feature
- reduced
- encoding
- generating
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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
Definitions
- the present disclosure relates to a machine vision data coding apparatus and method using feature map reduction.
- MPEG Moving Picture Expert Group
- VCM Video Coding for Machines
- the basic structure of the VCM coding device is as illustrated in FIG. 1 .
- the VCM encoder extracts features as information for machine vision, performs feature conversion as needed, and then encodes them.
- the VCM encoder may refer to the encoded feature when encoding the input image or video.
- the VCM encoder generates a bitstream by encoding features and input images (or residual images) for machine vision.
- the VCM encoder multiplexes and transmits each bitstream generated by encoding a feature and a video.
- the VCM decoder demultiplexes the transmitted bitstream into a feature bitstream and a video bitstream, and then decodes the feature and the video, respectively.
- the VCM decoder may refer to the restored feature when decoding the video. After inverse conversion, the reconstructed features can be used simultaneously for machine vision and human vision.
- an interface for Neural Network may be used to apply a deep learning model to extracting features in a VCM encoder, performing a task for machine vision in a VCM decoder, and the like.
- a representative type of information for machine vision is a feature map extracted by a machine task-specific deep learning model. If the VCM encoder transmits the feature map instead of transmitting the video/video, the computational load in the cloud server including the VCM decoder can be greatly reduced, and there is an advantage in protecting personal information.
- the feature map is a stacked array of multiple channels compared to a general image/video, the size of the feature map is fundamentally large.
- the size of the output feature map of a specific layer of the deep learning model may be smaller than that of the input image/video, there is also a problem that the corresponding layer changes whenever the deep learning model is changed.
- An object of the present invention is to provide a VCM (Video Coding for Machines) coding apparatus and method using a feature map decomposition method based on the feature map decomposition method.
- VCM Video Coding for Machines
- an encoding method for encoding a feature map performed by a machine vision encoding apparatus extracting the feature map from an input image using a deep learning model, wherein the The feature map is generated from the middle layer of the deep learning model; generating a reduced feature map by reducing the size of the feature map; converting the data type of the reduced feature map and rearranging the reduced feature map to generate a converted feature map; and generating a bitstream by encoding the transformed feature map using a video encoder.
- a feature extractor for extracting a feature map from an input image using a deep learning model, wherein the feature map is generated from an intermediate layer of the deep learning model; a feature reducer for generating a reduced feature map by reducing the size of the feature map; a pre-quantizer for quantizing the reduced feature map to transform a data type of the reduced feature map; a repacker that rearranges the quantized feature map in the form of a video sequence to generate a converted feature map; and a feature encoder for generating a bitstream by encoding the transformed feature map using a video encoder.
- a decoding method performed by a machine vision decoding apparatus, the method comprising: decoding a transformed feature map from a bitstream using a video decoder; restoring a reduced feature map by rearranging the transformed feature map and converting a data type of the rearranged transformed feature map; and generating a restored feature map by expanding the size of the reduced feature map, wherein the restored feature map corresponds to a feature map generated from an intermediate layer of a deep learning model in a machine vision encoding apparatus,
- a decryption method is provided.
- a sparsity method for reducing the need for spatial and channel aspects of the feature map and a feature map based on tensor decomposition
- a sparing method that reduces the spatial and channel unnecessary of the feature map As a step of feature transformation, a sparing method that reduces the spatial and channel unnecessary of the feature map, and a tensor decomposition based
- a sparing method that reduces the spatial and channel unnecessary of the feature map As a step of feature transformation, a sparing method that reduces the spatial and channel unnecessary of the feature map, and a tensor decomposition based
- FIG. 1 is a block diagram conceptually illustrating a VCM coding apparatus.
- FIG. 2 is an exemplary diagram illustrating a feature map for each layer of a deep learning model according to an embodiment of the present disclosure.
- FIG. 3 is an exemplary diagram conceptually illustrating operations of a front end and a cloud server.
- FIG. 4 is a block diagram conceptually illustrating a VCM core apparatus for performing intermediate feature compression according to an embodiment of the present disclosure.
- FIG. 5 is a block diagram conceptually illustrating a VCM encoding apparatus using intermediate feature map reduction according to an embodiment of the present disclosure.
- FIG. 6 is a block diagram conceptually illustrating a VCM decoding apparatus using intermediate feature map reconstruction according to an embodiment of the present disclosure.
- FIG. 7 is an exemplary diagram illustrating a feature map reduction method based on feature map sparsity, according to an embodiment of the present disclosure.
- FIG. 8 is an exemplary diagram illustrating a feature map reduction method based on feature map sparsity, according to another embodiment of the present disclosure.
- FIG. 9 is an exemplary diagram illustrating a feature map reduction method using Tucker decomposition according to an embodiment of the present disclosure.
- FIG. 10 is an exemplary diagram illustrating a feature map reduction method using CP decomposition according to an embodiment of the present disclosure.
- FIG. 11 is a flowchart illustrating a VCM encoding method using intermediate feature map reduction according to an embodiment of the present disclosure.
- FIG. 12 is a flowchart illustrating a VCM decoding method using intermediate feature map reconstruction according to an embodiment of the present disclosure.
- This embodiment discloses a machine vision data coding apparatus and method using feature map reduction.
- a sparsification method that reduces the unnecessaryness of the spatial and channel aspects of the feature map, and tensor decomposition
- VCM video coding for machines
- the VCM coding apparatus includes a VCM encoding apparatus and a VCM decoding apparatus.
- the existing codecs for encoding and decoding a video signal so as to be optimized for human vision are referred to as a video encoder and a video decoder.
- a multidimensional array output from each layer of a deep learning model having a hierarchical structure is typically a feature, a feature map, a deep feature, or a deep learning feature map (deep learning). It is called a feature map).
- the feature map for each layer has a form in which two-dimensional arrays with spatial correlation are stacked.
- the height H and the width W of the 2D array are called the height and width of the corresponding feature map
- the number of stacked 2D arrays is called the number C of channels of the corresponding feature map.
- CNN Convolutional Neural Network
- the layer deepens the height and width of the feature map decrease and the size of the channel increases by pooling.
- the short-channel 2D feature map may be regarded as one frame
- the stacked feature maps may be regarded as one continuous video sequence.
- the front-end device acquires information from a user or the physical world, and then transmits the acquired information to a cloud server (ie, a data center), and the cloud server uses the acquired information to perform additional processes and Perform the analysis.
- a cloud server ie, a data center
- a front-end device disposed in a real environment such as a security surveillance camera, a wearable device, etc. may acquire massive data transmitted to a cloud server for visual analysis, as illustrated in FIG. 3 .
- the compression and transmission of video between the front end and the cloud server can rely on the traditional 'compress-then-analyse' method.
- the front end device captures and compresses the visual data, and then sends the coded bitstream to the cloud server. After decoding the bitstream, the cloud server performs shape extraction and visual analysis.
- the cloud server performs shape extraction and visual analysis.
- visual compression at the video signal level can be burdensome and generally unmanageable.
- the computational load of numerous deep learning models that are simultaneously executed for feature extraction may act as a significant difficulty.
- the extracted features are used for task analysis instead of visual signals. It can be a feasible alternative to 'analysis'.
- the top layer ie, the deepest layer
- the low layer ie, the shallow layer
- the top layer usually the top layer ( That is, the characteristics of the deepest layer) may be transmitted to the cloud end.
- lightweight tasks such as comparison between features are performed on the cloud server, and tasks such as feature extraction are distributed to the front end.
- this feature transmission is advantageous for privacy protection. Compared to the transmission of a visual signal in which personal information can be easily exposed, feature transmission can avoid visible information disclosure.
- the deep learning model is designed and trained for a specific task, and the features of the upper layer are very abstract and task-specific, it may be difficult to generalize the features of the upper layer. In other words, it may be difficult for features compressed according to standardization to be applied to other scenarios, and this problem may become an obstacle in making a standard for feature compression of a deep learning model.
- an intermediate deep feature compression method exists as a method for reducing the computational load in the cloud server while maintaining feature availability. This approach is considered a compromise between the extreme 'compression after analysis' and 'analysis after compression' as described above, and can provide a good balance between computational load, communication cost and generalization capability.
- a deep learning model can be considered as a combination of stacked feature extractors rather than a single feature extractor.
- the features of the upper layer are abstract and task-specific based on a large receptive field, whereas the features of the lower layer have a small receptive field and encode a lot of location information in the 2D feature map, so it can be suitable for a wide range of analysis. Therefore, in the case of compressing and transmitting the characteristics of the middle layer instead of the final layer, the cloud server can flexibly request the appropriate characteristics according to the task performed at the front end.
- Intermediate feature coding is very important in the pipeline of the VCM standard, and can be applied in the order of feature transformation, feature encoding, and feature decoding in the structure of the VCM coding apparatus illustrated in FIG. 1 .
- intermediate feature coding will be described. Meanwhile, in the illustration of FIG. 4 , components that perform video encoding and video decoding of an input image are not shown because they are not highly related to the present embodiment.
- the feature map represents the intermediate feature map that is the output of the middle layer of the deep learning model.
- FIG. 4 is a block diagram conceptually illustrating a VCM coding apparatus for performing intermediate feature compression, according to an embodiment of the present disclosure.
- the VCM encoding apparatus includes all or a part of a feature extractor 410 , a feature converter 420 , and a feature encoder 430 .
- the VCM decoding apparatus includes all or a part of a feature decoder 450 and a feature inverse converter 460 .
- the feature encoder 430 and the feature decoder 450 may be implemented using an existing video coding apparatus.
- the feature extractor 410 in the VCM encoding device extracts a feature map of a task for machine vision from an input image based on a deep learning model. Additionally, the deep learning model in the feature extractor 410 may perform a machine vision task, for example, to generate an analysis result of the machine vision task.
- the deep learning model in the feature extractor 410 any neural network model capable of generating a feature map suitable for analysis of machine vision tasks may be used.
- the deep learning model may be a CNN-based model implemented with multiple encoding layers.
- one encoding layer includes a convolutional layer and a pooling layer.
- the feature modifier 420 includes a pre-quantizer 422 and a repacker 424 for data type conversion for the feature map, channel reordering of the feature map, and the like.
- the feature map extracted by the deep learning model may have a number type that is not compatible with the input of the existing video coding device. Accordingly, in order to match the input compatibility with the feature encoder 430 , the pre-quantizer 422 performs pre-quantization first.
- feature maps of a general deep learning model are of a float32 type, whereas an existing video coding apparatus uses an integer input of 8 or more bits.
- another pre-quantization method may be used according to the distribution analysis of the intermediate feature map.
- the repacker 424 converts the pre-quantized N (here, N is a natural number) feature maps R N ⁇ H ⁇ W ⁇ C to R H′ ⁇ W′ , which is a format such as a video sequence, to fit the input format of the video coding device.
- Repack with ⁇ NC .
- H and W are the height and width of the 2D feature map
- C is the number of channels.
- H' ⁇ W' means that the size of the original feature map H ⁇ W is expanded by padding to fit the input frame size of the video coding apparatus.
- R represents a real number.
- N ⁇ H ⁇ W ⁇ C corresponds to the feature maps extracted from N frames.
- the feature map is represented by F ⁇ R H ⁇ W ⁇ C
- the 2D feature map constituting the feature map is represented by f ⁇ R H ⁇ W .
- the repacker 424 may repack the pre-quantized N (here, N is a natural number) feature maps R N ⁇ H ⁇ W ⁇ C into one frame.
- N is a natural number
- the feature maps for each channel may be repacked into one super frame. For these repacked super frames, better compression efficiency can be achieved using only intra prediction of the video coding apparatus.
- this repacking process will be referred to as frame packing.
- the feature encoder 430 generates a corresponding bitstream by encoding the transformed feature map.
- the feature decoder 450 in the VCM decoding apparatus decodes the transformed feature map from the bitstream.
- the feature inverse transformer 460 applies a feature map inverse transformation, which is an inverse process of the feature map transformation by the VCM encoding apparatus, to the bitstream, a derepacker 462 and a post-dequantizer 464 . ) is included.
- the de-repacker 462 transforms the repacked feature map in a format such as a video sequence into an original feature map size.
- the post inverse quantizer 464 inversely quantizes the feature map having an integer value into a floating-point type. This reconstructed feature map may be used for machine vision task analysis as described above.
- the compressed intermediate feature map may be larger than the video/image stream when transmitting the intermediate feature map instead of the visual signal.
- the features of the upper layer of the deep learning model generally have a smaller compression size than the features of the lower layer. That is, some kinds of features may use less bandwidth than the video/video bitstream. Therefore, in this case, transmitting the intermediate feature map is meaningful in terms of transmission bandwidth efficiency.
- the intermediate feature map cannot be compressed to a size smaller than the image/video bit stream, it is still meaningful to transmit the feature map. Since computational resources are consumed more than bandwidth in many application cases, it may be more important to distribute the computational load of the cloud server to the front-end equipment using intermediate feature map transmission.
- FIG. 5 is a block diagram conceptually illustrating a VCM encoding apparatus using intermediate feature map reduction according to an embodiment of the present disclosure.
- the feature reducer 512 additionally performs various feature map reduction methods to By reducing the size of the feature map, it is possible to improve the encoding efficiency of the feature map.
- the feature reducer 512 may use two methods: feature map sparsification-based feature map reduction, and tensor decomposition-based feature map reduction. . By using these two methods individually or simultaneously, the feature reducer 512 can maximize feature map reduction. In addition, when the two methods are used together, there is no restriction on the order of use.
- the method for reducing the feature map according to the present embodiment may be used in combination with a separate pre-processing method such as frame packing of the feature map as described above.
- FIG. 6 is a block diagram conceptually illustrating a VCM decoding apparatus using intermediate feature map reconstruction according to an embodiment of the present disclosure.
- the feature reconstructor 612 corresponding to the feature reducer 512 is a post inverse quantized feature map, that is, a reduced feature map. Restore the original feature map from the feature map.
- a specific feature map restoration method may depend on a feature map reduction method used in the VCM encoding apparatus.
- the feature map reduction methods performed by the feature reducer 512 and the feature restorer 612 will be described. 5 and 6 , it is assumed that an existing video coding apparatus is used for feature encoding/decoding, but the feature map reduction methods described below are not necessarily limited thereto. That is, the feature map reduction methods according to the present embodiment may be utilized to reduce the size of the extracted feature map regardless of a deep learning model for feature extraction or a coding method.
- feature map reduction based on feature map sparsity utilizes the sparsity existing in the space or channel side of the feature map.
- some regions and some channels may be greatly activated according to spatial characteristics of an input image/frame or characteristics of a trained filter.
- a small activated region induces small activations continuously even if it is input to the next layer, and finally has a small influence on the judgment in the last layer of the deep learning model. Therefore, even when the value of the region outputting such small activations is sparsed to 0, the size of the feature map itself can be reduced without significantly lowering the judgment performance of the deep learning model. That is, the encoding efficiency of the feature map can be improved based on the feature map sparsity that sparse the region within the feature map activated to be smaller than a predetermined reference value.
- FIG. 7 is an exemplary diagram illustrating a feature map reduction method based on feature map sparsity, according to an embodiment of the present disclosure.
- the feature reducer 512 sparses a region that outputs small activations according to the degree of activation within a single 2D feature map f ⁇ R H ⁇ W constituting the intermediate feature map F ⁇ R H ⁇ W ⁇ C .
- the feature reducer 512 when the area (eg, pixel) of the 2D feature map f of a specific channel of the feature map is expressed numerically according to the feature value, the feature reducer 512 completely sets the feature value of the area smaller than the specific threshold to 0. do.
- the threshold is set to 1.5, regions marked as 1 may be set to 0. Even if an appropriate threshold is set and the feature value of a region smaller than this threshold is set to 0, the performance of the deep learning model is not significantly affected.
- the feature reducer 512 is configured for C 2D maps ⁇ f 1 , f 2 , ...,f C ⁇ corresponding to all channels of the feature map, or some 2D maps ⁇ f 1 , f 2 , ..., For f C' ⁇ (C' ⁇ C), this spatial aspect feature map sparsation can be performed.
- FIG. 8 is an exemplary diagram illustrating a feature map reduction method based on feature map sparsity, according to another embodiment of the present disclosure.
- the feature reducer 512 reduces redundancy between two 2D feature maps when the 2D feature maps of two channels adjacent by a predetermined channel interval S are similar, that is, when the distance between the two feature maps is smaller than a specific threshold. For this purpose, by setting one 2D feature map to 0 or completely deleting one channel, the number of channels in the original feature map can be reduced, thereby reducing the size of the entire feature map.
- the feature reducer 512 calculates a distance between two 2D feature maps with a constant channel interval S ?f i +S - f i ? (1 ⁇ i ⁇ CS).
- a norm such as L1, L2, etc. may be used for the distance.
- the channel interval S is set to 1.
- the feature reducer 512 considers two 2D feature maps with a distance less than a certain threshold d as very similar feature maps.
- the feature reducer 512 selects all pairs of channels having similar 2D feature maps, and sets all values of the 2D feature maps of the rear channel (or the front channel) to 0 among the two 2D feature maps for each pair.
- the feature reducer 512 may completely delete one channel for each pair, as illustrated in FIG. 8 .
- the feature reducer 512 may generate an intermediate feature map F' in which the total number of channels is reduced by using only 2D feature maps of dissimilar channels.
- the VCM encoding apparatus When sparse channels exist, the VCM encoding apparatus encodes a channel interval S value and a 2D feature map of each sparse channel, and then transmits it to the VCM decoding apparatus. Since all values of the 2D feature map of each sparse channel are 0, the VCM encoding apparatus can easily encode the 2D feature map of each sparse channel, thereby greatly improving the encoding efficiency.
- the VCM encoding apparatus may encode the channel interval S value and the index of each deleted channel, and then transmit it to the VCM decoding apparatus. Instead of encoding the H ⁇ W size 2D feature map for the deleted channel, the VCM encoding apparatus encodes only an integer value corresponding to an index, thereby greatly improving encoding efficiency.
- the feature restorer 612 in the VCM decoding apparatus may refer to the restored reduced feature map, the restored channel stride S value, or the deleted channel index information.
- the VCM decoding apparatus restores the original feature map before reduction by copying the 2D feature map of the channel before (or after) the S interval at the sparse channel position.
- the VCM decoding apparatus restores the original feature map before reduction by copying the 2D feature map of the channel before (or after) the S interval to the deleted channel position with reference to the deleted channel index information.
- the reconstructed feature map is an array having the same size as the original feature map, and thereafter, the VCM decoding apparatus can use the reconstructed feature map to analyze the machine vision task.
- the VCM encoding apparatus may compress and transmit the feature map after multiplying the element value c i included in the 1 ⁇ C vector representing the weight of each channel by the corresponding feature map f i .
- the VCM decoding apparatus may generate a restored feature map by dividing the received feature map by c i when decoding the feature map.
- c i may be obtained from an SE block of a deep learning-based attention model or a deep learning-based SE (Squeeze and Excitation) model.
- the feature map in the form of a multidimensional array as described above may be expressed by decomposing several low-dimensional simple tensors using tensor decomposition. In this case, the total size of the decomposed tensors may be reduced than the size of the original feature map.
- the feature map reduction is not limited to a specific tensor decomposition method, and in addition to the examples described below, matrix decomposition (eg, singular value decomposition (SVD), CUR decomposition, QR decomposition, etc.) and HOSVD Various tensor decomposition methods such as (Higher Order SVD) may be applied. After the tensor decomposition method is applied, the VCM encoding apparatus may utilize decomposed low-dimensional tensors instead of the original feature map for feature encoding.
- matrix decomposition eg, singular value decomposition (SVD), CUR decomposition, QR decomposition, etc.
- HOSVD Various tensor decomposition methods such as (Higher Order SVD) may be applied.
- the VCM encoding apparatus may utilize decomposed low-dimensional tensors instead of the original feature map for feature encoding.
- FIG. 9 is an exemplary diagram illustrating a feature map reduction method using Tucker decomposition according to an embodiment of the present disclosure.
- the feature reducer 512 uses Tucker decomposition to create an intermediate feature map for a single image. is a single kernel tensor. , and three factor matrices It can be decomposed as (i ⁇ 1,2,3 ⁇ ). At this time, r i is expressed as i-rank, and the feature map is In the case of generating a matrix in the form of a matrix, it indicates the rank of this matrix. Using one kernel tensor and three factor matrices, the feature map F can be approximated by operation of tensors, as shown in Equation 1.
- Equation (2) the operators ⁇ 1 , ⁇ 2 and ⁇ 3 may be expressed as in Equation (2).
- the data size n 1 n 2 n 3 of the original feature map is reduced to 'r 1 r 2 r 3 + r 1 n 1 + r 2 n 2 + r 3 n 3 '.
- the VCM encoding apparatus encodes each of the decomposed tensors K, F (1) , F (2) and F (3) using a video encoder, and then multiplexes the generated bitstreams to the VCM decoding apparatus. can be transmitted
- the VCM decoding apparatus decodes the bitstream of decomposed tensors using a video decoder, and then applies the decoded tensors to the derepacker 462 and the post inverse quantizer 464 as illustrated in FIG. 4 . .
- the VCM decoding apparatus may reconstruct the original feature map according to Equation 1 using the dequantized dequantized tensors.
- a method of reducing the feature map by the feature reducer 512 based on Canonical Polyadic (CP) decomposition is described using the example of FIG. 10 .
- FIG. 10 is an exemplary diagram illustrating a feature map reduction method using CP decomposition according to an embodiment of the present disclosure.
- the feature reducer 512 uses CP decomposition, as illustrated in FIG. 10 , for the intermediate feature map. can be decomposed into a linear combination of rank 1 tensors. That is, as shown in Equation 3, the feature map F may be approximated based on P (here, P is a natural number) rank 1 tensors.
- the operator o represents the cross product.
- the VCM encoding apparatus generates tensors A, B, and C by concatenating rank 1 tensors having the same size, and then encodes them.
- the data size n 1 n 2 n 3 of the original feature map is reduced to '(n 1 + n 2 + n 3 )P'.
- the VCM encoding apparatus may encode the decomposed P rank 1 tensors.
- the VCM encoding apparatus transmits the bitstream obtained by encoding the decomposed tensors to the VCM decoding apparatus.
- the VCM decoding apparatus decodes the bitstream of decomposed tensors using a video decoder, and then applies the decoded tensors to the derepacker 462 and the post inverse quantizer 464 as illustrated in FIG. 4 . .
- the VCM decoding apparatus may reconstruct the original feature map according to Equation 3 by using the dequantized dequantized tensors.
- the VCM encoding apparatus may encode tensors obtained by combining only P' ( ⁇ P) rank 1 tensors, and may transmit the encoded tensors to the VCM decoding apparatus.
- the VCM decoding apparatus may reconstruct the original feature map by using the decoded P' rank 1 tensors.
- VCM encoding method using intermediate feature map reduction and a VCM decoding method using intermediate feature map reconstruction will be described using the illustrations of FIGS. 11 and 12 .
- FIG. 11 is a flowchart illustrating a VCM encoding method using intermediate feature map reduction according to an embodiment of the present disclosure.
- the VCM encoding apparatus extracts a feature map from the input image using a deep learning model (S1100).
- the feature map represents an intermediate feature map generated from the middle layer of the deep learning model.
- the feature map includes as many 2D feature maps as the number of channels having the same height and width.
- the VCM encoding apparatus generates a reduced feature map by reducing the size of the feature map (S1102).
- the VCM encoding apparatus reduces the size of the feature map based on the feature map sparsity in terms of space or channel of the feature map.
- the VCM encoding apparatus may reduce the feature map based on tensor decomposition.
- the VCM encoding apparatus compares the feature value of a region with a preset threshold in the 2D feature map constituting the feature map.
- a preset threshold the feature value of the corresponding region may be set to 0.
- the VCM encoding apparatus calculates a distance between two 2D feature maps having a preset channel interval. After selecting all pairs of 2D feature maps corresponding to the case where the calculated distance is smaller than a preset threshold, the VCM encoding apparatus sparses all values of one 2D feature map to 0 for each of the selected pairs, or Delete one 2D feature map.
- the VCM encoding apparatus may decompose the feature map into one kernel tensor and three factor matrices.
- the feature map can be approximated by one kernel tensor and three factor matrices.
- the VCM encoding apparatus may decompose the feature map into P rank 1 tensors.
- the feature map may be approximated based on P rank 1 tensors.
- the VCM encoding apparatus quantizes the reduced feature map to transform the data type of the reduced feature map (S1104).
- the VCM encoding apparatus generates a converted feature map by rearranging the quantized feature map in the form of a video sequence (S1106).
- the VCM encoding apparatus generates a bitstream by encoding the transformed feature map using a video encoder (S1108).
- the VCM encoding apparatus when using the channel sparsity of the feature map, the VCM encoding apparatus generates a bitstream as follows. With respect to the sparse feature map, the VCM encoding apparatus may encode the 2D feature map and a preset channel interval, and then transmit it to the VCM decoding apparatus. Alternatively, for the deleted feature map, the VCM encoding apparatus may encode the index of the deleted 2D feature map and a preset channel interval, and then transmit it to the VCM decoding apparatus.
- the VCM encoding apparatus may encode each of the kernel tensor and factor matrices to generate corresponding bitstreams, and then transmit the generated bitstreams to the VCM decoding apparatus.
- the VCM encoding apparatus may encode each of the P rank 1 tensors to generate corresponding bitstreams, and then transmit the generated bitstreams to the VCM decoding apparatus.
- the VCM encoding apparatus may generate tensors in which rank 1 tensors having the same size are combined.
- the VCM encoding apparatus may generate corresponding bitstreams by encoding them, and then transmit them to the VCM decoding apparatus.
- FIG. 12 is a block diagram illustrating a VCM decoding method using intermediate feature map reconstruction according to an embodiment of the present disclosure.
- the intermediate feature map represents a feature map generated from the middle layer of the deep learning model in the machine vision encoding apparatus.
- the feature map includes as many 2D feature maps as the number of channels having the same height and width.
- the VCM decoding apparatus decodes the transformed feature map from the bitstream using a video decoder (S1200).
- the VCM decoding apparatus decodes the bitstream as follows. For the sparse 2D feature map, the VCM decoding apparatus decodes the sparse 2D feature map with a preset channel interval. Alternatively, for the deleted 2D feature map, the VCM decoding apparatus decodes a preset channel interval and an index of the deleted 2D feature map.
- the VCM decoding apparatus rearranges the transformed feature map in the form of a video sequence (S1202).
- the VCM decoding apparatus restores the reduced feature map by inverse quantizing the rearranged transformed feature map in order to transform the data type of the rearranged transformed feature map (S1204).
- the VCM decoding apparatus generates a restored feature map by extending the size of the reduced feature map (S1206).
- the restored feature map corresponds to the above-described intermediate feature map.
- the VCM decoding apparatus When the reduced feature map is reduced based on channel-side feature map sparsity, the VCM decoding apparatus generates a restored feature map as follows. For the sparse 2D feature map, the VCM decoding apparatus may generate the restored feature map by copying the restored 2D feature map before or after a predetermined channel interval to the location of the sparse 2D feature map. Alternatively, for the deleted 2D feature map, the VCM decoding apparatus refers to the decoded index and copies the restored 2D feature map before or after a predetermined channel interval to the location of the deleted 2D feature map, thereby generating the restored feature map. can do.
- the VCM decoding apparatus approximates the restored feature map using the kernel tensor and factor matrices constituting the reduced feature map.
- the VCM decoding apparatus approximates the restored feature map using P rank 1 tensors constituting the reduced feature map. can do.
- non-transitory recording medium includes, for example, any type of recording device in which data is stored in a form readable by a computer system.
- the non-transitory recording medium includes a storage medium such as an erasable programmable read only memory (EPROM), a flash drive, an optical drive, a magnetic hard drive, and a solid state drive (SSD).
- EPROM erasable programmable read only memory
- SSD solid state drive
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Biodiversity & Conservation Biology (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
Description
Claims (19)
- 머신 비전 부호화 장치가 수행하는, 특징맵을 부호화하는 부호화 방법에 있어서,딥러닝(deep learning) 모델을 이용하여 입력 영상으로부터 상기 특징맵을 추출하는 단계, 여기서, 상기 특징맵은 상기 딥러닝 모델의 중간 계층으로부터 생성됨;상기 특징맵의 크기를 축소하여 축소(reduced) 특징맵을 생성하는 단계;상기 축소 특징맵의 데이터 타입을 변환하고, 상기 축소 특징맵을 재배열하여 변형(converted) 특징맵을 생성하는 단계; 및비디오 부호화기를 이용하여 상기 변형 특징맵을 부호화함으로써 비트스트림을 생성하는 단계를 포함하는 것을 특징으로 하는, 부호화 방법.
- 제1항에 있어서,상기 특징맵은,동일한 높이와 너비를 갖는 2D 특징맵을 채널의 개수만큼 포함하는 것을 특징으로 하는, 부호화 방법.
- 제1항에 있어서,상기 축소 특징맵을 생성하는 단계는,상기 특징맵의 공간 또는 채널 측면에서의 특징맵 희소화(feature map sparsification)를 기반으로 상기 특징맵을 축소하는 것을 특징으로 하는, 부호화 방법.
- 제3항에 있어서,상기 축소 특징맵을 생성하는 단계는,상기 특징맵 희소화 및 텐서 분해를 결합하여 상기 특징맵을 축소하는 것을 특징으로 하는, 부호화 방법.
- 제3항에 있어서,상기 축소 특징맵을 생성하는 단계는,상기 특징맵을 구성하는 2D 특징맵 내에서, 영역의 특징 값이 기설정된 임계치보다 작은 경우, 상기 영역의 특징 값을 0으로 설정하는 것을 특징으로 하는, 부호화 방법.
- 제3항에 있어서,상기 축소 특징맵을 생성하는 단계는,기설정된 채널 간격을 갖는 두 개의 2D 특징맵 간의 거리를 산정하는 단계;상기 거리가 기설정된 임계치보다 작은 경우에 해당하는 2D 특징맵의 짝들(pairs)을 모두 선택하는 단계; 및상기 선택된 짝들 각각에 대해, 하나의 2D 특징맵의 모든 값을 0으로 희소화하거나 상기 하나의 2D 특징맵을 삭제하는 단계를 포함하는 것을 특징으로 하는, 부호화 방법.
- 제6항에 있어서,상기 비트스트림을 생성하는 단계는,상기 하나의 2D 특징맵이 희소화된 경우, 희소화된 2D 특징맵, 및 상기 기설정된 채널 간격을 부호화하고, 상기 하나의 2D 특징맵이 삭제된 경우, 삭제된 2D 특징맵의 인덱스, 및 상기 기설정된 채널 간격을 부호화하는 것을 특징으로 하는, 부호화 방법.
- 제1항에 있어서,상기 축소 특징맵을 생성하는 단계는,텐서 분해(tensor decomposition)를 기반으로 상기 특징맵을 축소하는 것을 특징으로 하는, 부호화 방법.
- 제8항에 있어서,상기 축소 특징맵을 생성하는 단계는,터커 분해(Tucker decomposition)를 이용하여, 상기 특징맵을 하나의 커널 텐서(kernel tensor), 및 3 개의 인자 행렬들(factor matrices)로 분해하는 것을 특징으로 하는, 부호화 방법.
- 제8항에 있어서,상기 축소 특징맵을 생성하는 단계는,CP(Canonical Polyadic) 분해를 이용하여, 상기 특징맵을 P(여기서, P는 자연수) 개의 랭크(rank) 1 텐서들로 분해하는 것을 특징으로 하는, 부호화 방법.
- 딥러닝(deep learning) 모델을 이용하여 입력 영상으로부터 특징맵을 추출하는 특징 추출기, 여기서, 상기 특징맵은 상기 딥러닝 모델의 중간 계층으로부터 생성됨;상기 특징맵의 크기를 축소하여 축소(reduced) 특징맵을 생성하는 특징 축소기(feature reducer);상기 축소 특징맵의 데이터 타입을 변환하기 위해 상기 축소 특징맵을 양자화하는 사전 양자화기(pre-quantizer);상기 양자화된 특징맵을 비디오 시퀀스의 형태로 재배열하여 변형(converted) 특징맵을 생성하는 리패커(repacker); 및비디오 부호화기를 이용하여 상기 변형 특징맵을 부호화함으로써 비트스트림을 생성하는 특징 부호화기를 포함하는 것을 특징으로 하는, 머신 비전 부호화 장치.
- 제11항에 있어서,상기 특징 축소기는,상기 특징맵의 공간 또는 채널 측면에서의 특징맵 희소화(feature map sparsification)를 기반으로 상기 특징맵을 축소하는 것을 특징으로 하는, 머신 비전 부호화 장치.
- 제11항에 있어서,상기 특징 축소기는,텐서 분해(tensor decomposition)를 기반으로 상기 특징맵을 축소하는 것을 특징으로 하는, 머신 비전 부호화 장치.
- 머신 비전 복호화 장치가 수행하는 복호화 방법에 있어서,비디오 복호화기를 이용하여 비트스트림으로부터 변형 특징맵을 복호화하는 단계;상기 변형 특징맵을 재배열하고, 상기 재배열된 변형 특징맵의 데이터 타입을 변환하여 축소(reduced) 특징맵을 복원하는 단계; 및상기 축소 특징맵의 크기를 확장하여 복원 특징맵을 생성하는 단계를 포함하되,상기 복원 특징맵은 머신 비전 부호화 장치 내 딥러닝 모델의 중간 계층으로부터 생성된 특징맵에 대응하는 것을 특징으로 하는, 복호화 방법.
- 제14항에 있어서,상기 복원 특징맵은,동일한 높이와 너비를 갖는 2D 특징맵을 채널의 개수만큼 포함하는 것을 특징으로 하는, 복호화 방법.
- 제14항에 있어서,상기 축소 특징맵이 채널 측면에서의 특징맵 희소화(feature map sparsification)를 기반으로 축소된 경우, 기설정된 채널 간격 및 희소화된 2D 특징맵을 복호화하거나 상기 기설정된 채널 간격 및 삭제된 2D 특징맵의 인덱스를 복호화하는 단계를 더 포함하는 것을 특징으로 하는, 복호화 방법.
- 제16항에 있어서,상기 복원 특징맵을 생성하는 단계는,상기 희소화된 2D 특징맵에 대해 상기 기설정된 채널 간격의 앞 또는 뒤의 복원 2D 특징맵을 상기 희소화된 2D 특징맵의 위치에 복사하거나, 상기 삭제된 2D 특징맵에 대해 상기 인덱스를 참조하여 상기 기설정된 채널 간격의 앞 또는 뒤의 복원 2D 특징맵을 상기 삭제된 2D 특징맵의 위치에 복사함으로써, 상기 복원 특징맵을 생성하는 것을 특징으로 하는, 복호화 방법.
- 제14항에 있어서,상기 복원 특징맵을 생성하는 단계는,상기 축소 특징맵이 터커 분해(Tucker decomposition)를 이용하여 축소된 경우, 상기 축소 특징맵을 구성하는 커널 텐서(kernel tensor) 및 인자 행렬들(factor matrices)을 이용하여 상기 복원 특징맵을 생성하는 것을 특징으로 하는, 복호화 방법.
- 제14항에 있어서,상기 복원 특징맵을 생성하는 단계는,상기 축소 특징맵이 CP(Canonical Polyadic) 분해를 이용하여 축소된 경우, 상기 축소 특징맵을 구성하는 P(여기서, P는 자연수) 개의 랭크(rank) 1 텐서들을 이용하여 상기 복원 특징맵을 생성하는 것을 특징으로 하는, 복호화 방법.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202280024083.0A CN117063201A (zh) | 2021-03-31 | 2022-03-28 | 使用特征图减小来编解码机器视觉数据的方法和设备 |
| US18/371,730 US12561846B2 (en) | 2021-03-31 | 2023-09-22 | Method and apparatus for coding machine vision data using feature map reduction |
Applications Claiming Priority (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR20210041562 | 2021-03-31 | ||
| KR10-2021-0041562 | 2021-03-31 | ||
| KR1020220037097A KR20220136176A (ko) | 2021-03-31 | 2022-03-25 | 특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 |
| KR10-2022-0037097 | 2022-03-25 |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| US18/371,730 Continuation US12561846B2 (en) | 2021-03-31 | 2023-09-22 | Method and apparatus for coding machine vision data using feature map reduction |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2022211409A1 true WO2022211409A1 (ko) | 2022-10-06 |
Family
ID=83456516
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/KR2022/004312 Ceased WO2022211409A1 (ko) | 2021-03-31 | 2022-03-28 | 특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 |
Country Status (2)
| Country | Link |
|---|---|
| US (1) | US12561846B2 (ko) |
| WO (1) | WO2022211409A1 (ko) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2024080825A1 (ko) * | 2022-10-14 | 2024-04-18 | 엘지전자 주식회사 | 영상 데이터 처리 방법, 장치, 비트스트림을 저장한 기록 매체 및 비트스트림 전송 방법 |
| WO2024196160A1 (ko) * | 2023-03-22 | 2024-09-26 | 엘지전자 주식회사 | 피쳐 부호화/복호화 방법, 장치, 비트스트림을 저장한 기록 매체 및 비트스트림 전송 방법 |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119854503B (zh) * | 2024-12-27 | 2025-11-18 | 中国电信股份有限公司技术创新中心 | 机器视频特征编解码方法、装置、电子设备及存储介质 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110874631A (zh) * | 2020-01-20 | 2020-03-10 | 浙江大学 | 一种基于特征图稀疏化的卷积神经网络剪枝方法 |
| KR20200044171A (ko) * | 2018-10-05 | 2020-04-29 | 재단법인대구경북과학기술원 | 딥러닝 네트워크 모델에 의한 실시간 객체 검출 방법 및 장치 |
| CN107516129B (zh) * | 2017-08-01 | 2020-06-02 | 北京大学 | 基于维度自适应的Tucker分解的深度网络压缩方法 |
| CN111429463A (zh) * | 2020-03-04 | 2020-07-17 | 北京三快在线科技有限公司 | 实例分割方法、装置、电子设备和存储介质 |
| KR20200130105A (ko) * | 2019-05-10 | 2020-11-18 | 삼성전자주식회사 | 비디오 프레임 보간을 위한 cnn-기반의 시스템 및 방법 |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10417525B2 (en) * | 2014-09-22 | 2019-09-17 | Samsung Electronics Co., Ltd. | Object recognition with reduced neural network weight precision |
| US10366292B2 (en) * | 2016-11-03 | 2019-07-30 | Nec Corporation | Translating video to language using adaptive spatiotemporal convolution feature representation with dynamic abstraction |
| EP3596449A4 (en) * | 2017-03-14 | 2021-01-06 | University of Manitoba | DETECTION OF STRUCTURAL DEFECTS USING MACHINE LEARNING ALGORITHMS |
| US10095977B1 (en) * | 2017-10-04 | 2018-10-09 | StradVision, Inc. | Learning method and learning device for improving image segmentation and testing method and testing device using the same |
| US10402977B1 (en) * | 2019-01-25 | 2019-09-03 | StradVision, Inc. | Learning method and learning device for improving segmentation performance in road obstacle detection required to satisfy level 4 and level 5 of autonomous vehicles using laplacian pyramid network and testing method and testing device using the same |
| CN119027713A (zh) * | 2019-07-30 | 2024-11-26 | 华为技术有限公司 | 图像分类方法及装置 |
| US20220083807A1 (en) * | 2020-09-14 | 2022-03-17 | Nvidia Corporation | Generating labels for synthetic images using one or more neural networks |
-
2022
- 2022-03-28 WO PCT/KR2022/004312 patent/WO2022211409A1/ko not_active Ceased
-
2023
- 2023-09-22 US US18/371,730 patent/US12561846B2/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107516129B (zh) * | 2017-08-01 | 2020-06-02 | 北京大学 | 基于维度自适应的Tucker分解的深度网络压缩方法 |
| KR20200044171A (ko) * | 2018-10-05 | 2020-04-29 | 재단법인대구경북과학기술원 | 딥러닝 네트워크 모델에 의한 실시간 객체 검출 방법 및 장치 |
| KR20200130105A (ko) * | 2019-05-10 | 2020-11-18 | 삼성전자주식회사 | 비디오 프레임 보간을 위한 cnn-기반의 시스템 및 방법 |
| CN110874631A (zh) * | 2020-01-20 | 2020-03-10 | 浙江大学 | 一种基于特征图稀疏化的卷积神经网络剪枝方法 |
| CN111429463A (zh) * | 2020-03-04 | 2020-07-17 | 北京三快在线科技有限公司 | 实例分割方法、装置、电子设备和存储介质 |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2024080825A1 (ko) * | 2022-10-14 | 2024-04-18 | 엘지전자 주식회사 | 영상 데이터 처리 방법, 장치, 비트스트림을 저장한 기록 매체 및 비트스트림 전송 방법 |
| WO2024196160A1 (ko) * | 2023-03-22 | 2024-09-26 | 엘지전자 주식회사 | 피쳐 부호화/복호화 방법, 장치, 비트스트림을 저장한 기록 매체 및 비트스트림 전송 방법 |
Also Published As
| Publication number | Publication date |
|---|---|
| US12561846B2 (en) | 2026-02-24 |
| US20240013448A1 (en) | 2024-01-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| KR20220136176A (ko) | 특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 | |
| WO2022211409A1 (ko) | 특징맵 축소를 이용하는 머신 비전 데이터 코딩 장치 및 방법 | |
| KR100763178B1 (ko) | 색 공간 스케일러블 비디오 코딩 및 디코딩 방법, 이를위한 장치 | |
| CN116803079A (zh) | 视频和相关特征的可分级译码 | |
| US6847735B2 (en) | Image processing system, image processing apparatus, image input apparatus, image output apparatus and method, and storage medium | |
| WO2019054561A1 (ko) | 360도 영상 부호화 장치 및 방법, 이를 수행하기 위한 기록 매체 | |
| WO2022075754A1 (ko) | 머신비전을 위한 영상의 처리 방법 및 장치 | |
| BR102019000922A2 (pt) | Método para comprimir dados de campo de luz usando transformadas de quatro dimensões de tamanho de bloco variável e decomposição por planos de bits | |
| WO2023193629A1 (zh) | 区域增强层的编解码方法和装置 | |
| US7245775B2 (en) | Image processing apparatus for compositing images | |
| CN117376564B (zh) | 数据编解码方法及相关设备 | |
| WO2019151808A1 (ko) | 이미지 센서를 이용한 이미지 획득 과정에서 생성된 압축 속성을 이용하여 이미지를 압축하는 전자 장치 및 그 동작 방법 | |
| US20260067508A1 (en) | Encoding method and apparatus, and decoding method and apparatus | |
| CN105430406A (zh) | 一种分布式视频编码框架 | |
| WO2019088435A1 (ko) | 로우-퀄리티 부호화 모드에 따라 영상을 부호화하는 방법 및 장치, 및 영상을 복호화하는 방법 및 장치 | |
| CN117063201A (zh) | 使用特征图减小来编解码机器视觉数据的方法和设备 | |
| Naidu et al. | A novel framework for JPEG image compression using baseline coding with parallel process | |
| WO2025121685A1 (ko) | 암시적 신경망 비디오 표현에 기반하는 비디오 코딩을 위한 방법 및 장치 | |
| WO2014107070A1 (ko) | 가역 정보 은닉 시스템 및 방법 | |
| WO2021261950A1 (ko) | 뉴럴 네트워크의 압축 및 학습을 위한 방법 및 장치 | |
| CN119484818B (zh) | 高光谱图像处理方法、装置、存储介质及计算机设备 | |
| US20260046423A1 (en) | Picture encoding method and apparatus, and picture decoding method and apparatus | |
| WO2024107003A1 (ko) | 머신 비전을 위한 영상의 특징 맵의 처리 방법 및 장치 | |
| Vijay et al. | A review of image transmission using real time technique over wmsn | |
| US20240259597A1 (en) | A method, an apparatus and a computer program product for video encoding and video decoding |
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: 22781528 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202280024083.0 Country of ref document: CN |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202317064642 Country of ref document: IN |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 22781528 Country of ref document: EP Kind code of ref document: A1 |


