Video data compression and reconstruction method, system, equipment and medium

Abstract

The invention discloses a video data compression and reconstruction method, system and device and a medium, and relates to the technical field of video data compression, and the method comprises the steps: extracting a feature vector of a video frame in video data; performing sparse coding by taking the feature vector as a linear combination of a base vector in a dictionary to obtain a sparse coefficient of the video frame; coding the sparse coefficient of the video frame by using entropy coding to obtain a coded character string; reading the coded character string, searching each bit of the character string, moving the corresponding node according to the bit of the current character string, and adding the corresponding sparse coefficient feature vector into the decompressed sparse coefficient feature vector when the node is a leaf node; and reconstructing the sparse coefficient feature vector into an original feature vector by using inverse sparse coding, and outputting a reconstructed video frame. According to the method, the compression ratio far higher than that of a traditional compression standard is achieved, and the characteristic enables the higher compression ratio to be achieved on the premise that the video quality is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of video data compression, and in particular to a video data compression and reconstruction method, system, device and medium. Background Art

[0002] With the increasing popularity of high-definition and ultra-high-definition video content, how to effectively reduce the storage space and transmission bandwidth requirements of video data while ensuring video quality has become a pressing issue that needs to be addressed.

[0003] Traditional video compression standards, such as H.264 / AVC and H.265 / HEVC, mainly rely on block-based motion compensation and transform coding technologies to reduce spatial redundancy in video sequences and improve compression ratios.

[0004] Although these methods have achieved remarkable achievements in compression efficiency, they still face challenges such as high computational complexity and high requirements for hardware resources. The discontinuous block effect introduced between blocks seriously affects the visual quality, especially when processing high-resolution and high-frame rate videos. Summary of the invention

[0005] The purpose of the present invention is to provide a video data compression and reconstruction method, system, device and medium to address the deficiencies of the above-mentioned prior art, so as to solve the challenges faced by the prior art, such as high computational complexity and high requirements on hardware resources, and the non-continuous block effect introduced between blocks, which seriously affects the visual quality, especially when processing high-resolution and high frame rate videos.

[0006] The present invention specifically provides the following technical solution: a video data compression and reconstruction method, comprising the following steps: Obtain video data, and extract feature vectors of video frames in the video data through a convolutional neural network; At the compression end, when a dictionary is used to store video frames, the feature vector is used as a linear combination of a set of basis vectors in the dictionary, and an orthogonal matching algorithm is used to sparsely encode the feature vector to obtain a sparse coefficient of the video frame, and the sparse coefficient of the video frame is encoded using entropy coding to obtain an encoded string; At the decompression end, the encoded string is read, and each bit of the string is searched, the node corresponding to the string is moved according to the number of bits of the current string, and when the node after the move is a leaf node, the corresponding sparse coefficient is added to the decompressed sparse coefficient; the node is composed of sparse coefficients; Based on the dictionary storing the video frame, the sparse coefficient feature vector is reconstructed into the original feature vector by using inverse sparse coding, and the video frame reconstructed by the original feature vector is output.

[0007] Preferably, the extracting of the feature vector of the video frame in the video data comprises the following steps: Combine multiple convolution kernels of the convolutional neural network and convolve each convolution kernel with a local area of ​​the input video frame to generate a feature map; Performing average pooling on the feature map through a pooling layer to convert the feature map into a feature vector of a fixed length; Traverse the collection of video frames, input each frame into the convolutional neural network, and use the forward propagation of the convolutional neural network to obtain the feature vector of each frame image.

[0008] Preferably, before extracting the feature vector of the video frame in the video data, the following steps are also included: Performing video denoising on the video data, and performing color space conversion on the denoised video data; Perform color space conversion on the video data to obtain video frames.

[0009] Preferably, the feature vector is used as a linear combination of a set of basis vectors in a dictionary, and an orthogonal matching algorithm is used to perform sparse coding on the feature vector to obtain a sparse coefficient of the video frame, comprising the following steps: Initialize a dictionary D∈R through the K-SVD algorithm m*k , get the sparse representation a corresponding to the feature vector i , so that the coefficient eigenvector x i ≈Da i , each feature vector is represented as a linear combination of a set of basis vectors in the dictionary; where m is the dimension of the feature space, k is the size of the dictionary, and R is a set of real numbers; Use the orthogonal matching pursuit (OMP) algorithm to solve sparse coding: , where λ is the regularization parameter, is the square of the 2-norm of the matrix “•”, To take the 1-norm of the matrix “•”; Select a coefficient eigenvector x i and its corresponding sparse code a i , update a basis vector d in the dictionary j , so that x i and a i Minimize the error between The basis vector d j The update expression is: ; Repeatedly update a basis vector d in the dictionary jThe step of obtaining an optimized dictionary D is performed, and for each input video frame feature, the learned dictionary is used for sparse coding to obtain a set of sparse coefficients of the video frame.

[0010] Preferably, using entropy coding to encode the sparse coefficients of the video frame to obtain a coded string comprises the following steps: Get the probability of each coefficient in the sparse coefficient group, and construct a Huffman tree with the probability of each coefficient sorted from small to large; Traverse in a depth-first search manner, add a '0' character to the end of the path string each time the left child node is traversed, and add a '1' character to the end of the path string each time the right child node is traversed. When a leaf node is reached, the path string is associated with the sparse coefficient feature vector corresponding to the leaf node to form a Huffman coding table. Traverse each value of the sparse coefficient group, use the Huffman coding table to replace the corresponding Huffman code, and generate a coded string.

[0011] Preferably, the step of reconstructing the sparse coefficient feature vector into an original feature vector by using inverse sparse coding and outputting a video frame reconstructed by the original feature vector comprises the following steps: Based on the orthogonal matching pursuit (OMP) algorithm, the sparse representation a is calculated using the sparse coefficients. i , using the above coefficients to represent the optimized dictionary D of the module, by reconstructing the formula x i ≈Da i To get the original feature vector x i ; Through the generator network GA, the feature vector is used as input to the generator network. The deconvolution layer inside the generator network adjusts the input feature vector to an output image with image structure and details. Add a batch normalization layer after the convolution layer, and apply the ReLU activation function after the batch normalization layer. The activation function introduces nonlinearity so that the generator learns the mapping relationship; The generator network converts the mapping relationship of features into the final generated image through an output layer, provides each feature vector as input to the generator network GAN, and obtains the video frame through the output layer.

[0012] Preferably, after outputting the video frame reconstructed by the original feature vector, all reconstructed video frames are merged and color space conversion is performed.

[0013] The present invention provides a video data compression and reconstruction system, comprising: A data extraction module, used to obtain video data and extract feature vectors of video frames in the video data through a convolutional neural network; A compression module, for, at the compression end, when using a dictionary to store video frames, using the feature vector as a linear combination of a set of basis vectors in the dictionary, and using an orthogonal matching algorithm to sparsely encode the feature vector to obtain sparse coefficients of the video frame, and using entropy coding to encode the sparse coefficients of the video frame to obtain a coded string; A decompression module, for reading the encoded string at the decompression end, searching each bit of the string, moving the node corresponding to the string according to the number of bits of the current string, and adding the corresponding sparse coefficient to the decompressed sparse coefficient when the node after the move is a leaf node; the node is composed of sparse coefficients; The reconstruction module is used to reconstruct the sparse coefficient feature vector into the original feature vector by using inverse sparse coding based on the dictionary storing the video frame, and output the video frame reconstructed by the original feature vector.

[0014] The present invention provides a computer device, comprising a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned video data compression and reconstruction method.

[0015] Compared with the prior art, the present invention has the following significant advantages: The present invention uses an orthogonal matching algorithm to perform sparse coding on the feature vector extracted from the video frame in the video data, obtains the sparse coefficient of the video frame, and encodes the sparse coefficient to obtain the encoded string. When reading, the node corresponding to the string is moved by searching the number of bits of each bit of the string, and when the node is a leaf node, the corresponding sparse coefficient feature vector is added to the decompressed sparse coefficient feature vector, and the sparse coefficient feature vector is reconstructed into the original feature vector by inverse sparse coding, and the reconstructed video frame generated by the original feature vector is output. The present invention can reconstruct the original video frame by identifying and utilizing the sparsity of the video signal in a specific transformation domain, and only a small amount of non-zero coefficients need to be stored or transmitted, so as to achieve a compression ratio far higher than that of the traditional compression standard. This feature enables a higher compression ratio to be achieved while ensuring the quality of the video. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the principle of a method for efficient compression and reconstruction of video data based on sparse representation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0018] The present invention aims to address the fact that traditional video compression still faces considerable challenges in processing high-resolution and high-frame-rate videos, and in particular proposes a method for efficiently compressing and reconstructing video data based on sparse representation.

[0019] like Figure 1 As shown, an embodiment of the present invention provides a video data compression and reconstruction method, comprising the following steps: Step S1: Obtain video data and extract feature vectors of video frames in the video data through a convolutional neural network.

[0020] The process of extracting the feature vector of a video frame from the video data by using a convolutional neural network includes the following steps: A feature map is generated by combining multiple convolution kernels of the convolutional neural network and convolving each convolution kernel with a local area of ​​the input video frame.

[0021] The feature map is averaged through the pooling layer to convert the feature map into a feature vector of fixed length.

[0022] Traverse the collection of video frames, input each frame into the convolutional neural network, and use the forward propagation of the convolutional neural network to obtain the feature vector of each frame image.

[0023] Before extracting the feature vector of the video frame in the video data through the convolutional neural network, the following steps are also included: Perform video denoising on the video data, and perform color space conversion on the denoised video data.

[0024] Perform color space conversion on the video data to obtain video frames.

[0025] By adjusting the parameters of the Gaussian function, the degree of video smoothing can be controlled, and then a suitable Gaussian template (convolution kernel) is selected, and then the image is convolved with this template. The result of the convolution operation is a denoised video. This can remove noise and unnecessary details in the video, thereby reducing redundant information in the data. It can improve the effect of sparse representation, making the compressed data more compact and efficient.

[0026] Using OpenCV technology, we first create an FFmpegFrameGrabber example to open the video file. Then, we start the grabber and get some basic information of the video, such as frame rate, width, and height. Using a while loop, we continuously call grabber.grabFrame() to get the video frames, get the video frames, and convert these frames into OpenCV's Mat objects. Then, we use the Mat.cvtColor method to convert the video from RGB color space to YUV color space frame by frame, reduce the resolution or sampling rate of the chroma component to reduce the amount of data, thereby achieving higher compression efficiency, and finally convert these frames into OpenCV's Mat objects. Then, we use the Mat.imwrite() method to save each frame as a JPEG image file. The file names are numbered using the frame index to ensure that they are unique. Dividing the video into separate frames makes it easier to process, analyze, and store each frame independently.

[0027] Step S2: At the compression end, when a dictionary is used to store video frames, the feature vector is taken as a linear combination of a set of basis vectors in the dictionary, and the feature vector is sparsely encoded using an orthogonal matching algorithm to obtain the sparse coefficients of the video frame, and the sparse coefficients of the video frame are encoded using entropy coding to obtain an encoded string.

[0028] This step specifically includes the following steps: Initialize a dictionary D∈R through the K-SVD algorithm m*k , get the sparse representation a corresponding to the feature vector i , so that the coefficient eigenvector x i ≈Da i , each feature vector is represented as a linear combination of a set of basis vectors in the dictionary; where m is the dimension of the feature space, k is the size of the dictionary, and R is a set of real numbers.

[0029] Use the orthogonal matching pursuit (OMP) algorithm to solve sparse coding: , where λ is the regularization parameter, controlling sparsity; is the square of the 2-norm of the matrix “•”, To take the 1-norm of the matrix “•”.

[0030] Select a coefficient eigenvector x i and its corresponding sparse code a i , update a basis vector d in the dictionary j , so that x i and a i Minimize the error between them.

[0031] The basis vector d j The update expression is: ; Repeatedly update a basis vector d in the dictionary j The step of obtaining an optimized dictionary D is performed, and for each input video frame feature, the learned dictionary is used for sparse coding to find a set of sparse coefficients of the video frame.

[0032] For each input video frame feature, sparse coding is performed using the learned dictionary to find a set of sparse coefficients (i.e., only a few non-zero coefficients), which are usually some small non-zero values ​​and most of the coefficients are zero.

[0033] Traverse the coefficients of the sparse coefficient group and count the number of occurrences of each non-zero coefficient, calculate the number of non-zero coefficients, and obtain the frequency of occurrence of each coefficient through the formula. The specific expression is: ; in x i is the coefficient eigenvector, count ( x i ) is the count of the coefficient eigenvector, x j is the j coefficient eigenvector.

[0034] Encode the sparse coefficients of the video frame using entropy coding to obtain a coded string. This step specifically includes the following steps: Get the probability of each coefficient in the sparse coefficient group, and construct a Huffman tree with the probability of each coefficient sorted from small to large. Construct a Huffman tree according to the frequency of the coefficient feature vector, usually using a priority queue or a heap to construct the Huffman tree. First, take each coefficient feature vector and its frequency as a node, sort these nodes from small to large according to frequency, construct the initial Huffman tree, repeat the following steps until only one node is left, take the two nodes with the smallest frequency as the left and right child nodes, and generate a new parent node, whose frequency is the sum of the frequencies of the left and right child nodes. Insert the newly generated parent node into the priority queue, keep the frequency sorted from small to large, and save the Huffman tree corresponding to each video frame for subsequent decoding.

[0035] The traversal is performed in a depth-first search manner. Each time the left child node is traversed, the character '0' is added to the end of the path string. Each time the right child node is traversed, the character '1' is added to the end of the path string. When a leaf node is reached, the path string is associated with the sparse coefficient feature vector corresponding to the leaf node to form a Huffman coding table.

[0036] Traverse each value of the sparse coefficient group, use the Huffman coding table to replace the corresponding Huffman code, and generate a coded string.

[0037] Data compression using Huffman coding table: traverse each value in the sparse coefficient set and replace it with the corresponding Huffman code using the Huffman coding table. This will generate a coded string consisting of '0' and '1'. Store or transfer the coded string of all video frames to a local file for subsequent decoding. This coded string is the compressed data of the video frame.

[0038] Step S3: At the decompression end, read the encoded string and search each bit of the string, move the node corresponding to the string according to the number of bits of the current string, and add the corresponding sparse coefficient to the decompressed sparse coefficient when the moved node is a leaf node; the node is composed of sparse coefficients.

[0039] By reading the generated encoded string, starting from the beginning of the encoded string, using the Huffman tree to perform a depth-first search, traversing each bit in the encoded string ('0' or '1'), if the current bit is '0', move the pointer to the left child of the current node, if the current bit is '1', move the pointer to the right child of the current node. After each move, check whether the current node is a leaf node. If it is a leaf node, it means that a complete Huffman code is found, and the sparse coefficient feature vector corresponding to the leaf node is added to the decompressed sparse coefficient set. Then, reset the pointer to the root node of the Huffman tree to start searching for the next code. If the current node is not a leaf node, continue to traverse the next bit in the encoded string. Repeat this process until the entire encoded string is decoded, resulting in a set of sparse coefficients.

[0040] Step S4: Based on the dictionary storing the video frame, the sparse coefficient feature vector is reconstructed into the original feature vector by using inverse sparse coding, and the video frame reconstructed by the original feature vector is output.

[0041] This step specifically includes the following steps: Based on the orthogonal matching pursuit (OMP) algorithm, the sparse representation a is calculated using the sparse coefficients. i , using the above coefficients to represent the optimized dictionary D of the module, by reconstructing the formula x i ≈Da i To get the original feature vector x i .

[0042] Through the generator network GA, the feature vector is used as input to the generator network. The deconvolution layer inside the generator network adjusts the input feature vector to an output image with image structure and details. A batch normalization layer is added after the convolution layer, and a ReLU activation function is applied after the batch normalization layer. The activation function introduces nonlinearity so that the generator learns the mapping relationship.

[0043] The generator network converts the mapping relationship of features into the final generated image through an output layer, provides each feature vector as input to the generator network GAN, and obtains the video frame through the output layer.

[0044] After outputting the reconstructed video frames generated by the original feature vectors, the method further includes the steps of merging all the reconstructed video frames and performing color space conversion processing.

[0045] Using OpenCV technology, use the VideoWriter class to create a VideoWriter object. This object is used to write a frame sequence to a video file. Use a while loop to traverse the frame sequence, create a Mat object, use the Mat.cvtColor method to convert the video from the YUV color space to the RGB color space frame by frame, and use to write each frame to the video file one by one. In the loop, you can use the write method to write the image frame to the video file. When all frames are written, an original video will be generated and the original video will be output to the local computer.

[0046] Based on the above method, the present invention provides a video data compression and reconstruction system, including: a data extraction module, a compression module, a decompression module and a reconstruction module.

[0047] Among them, the data extraction module is used to obtain video data, and extract the feature vector of the video frame in the video data through a convolutional neural network; the decompression module is used to use the feature vector as a linear combination of a set of basis vectors in the dictionary when using a dictionary to store the video frame at the compression end, and use the orthogonal matching algorithm to sparsely encode the feature vector to obtain the sparse coefficient of the video frame, and use entropy coding to encode the sparse coefficient of the video frame to obtain the encoded string; the decompression module is used to read the encoded string at the decompression end, and search for each bit of the string, move the node corresponding to the string according to the number of bits of the current string, and add the corresponding sparse coefficient to the decompressed sparse coefficient when the moved node is a leaf node; the node is composed of sparse coefficients; the reconstruction module is used to reconstruct the sparse coefficient feature vector into the original feature vector based on the dictionary storing the video frame using inverse sparse coding, and output the video frame reconstructed by the original feature vector.

[0048] The present invention also provides a computer device, including a memory and a processor. The memory stores a program. When the program is executed by the processor, the processor executes the steps of a video data compression and reconstruction method.

[0049] According to the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth communications, etc.), or with any device (e.g., routers, modems, etc.) that enables a computing device to communicate with one or more other computing devices.

[0050] The present invention also provides a storage medium, which includes a stored computer program. The computer program is executed by at least one processor to enable the at least one processor to execute the steps of a video data compression and reconstruction method.

[0051] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, such as but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0052] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. For technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A video data compression and reconstruction method, characterized in that: The steps include: Acquire video data and extract feature vectors of video frames in the video data; At the compression end, when a dictionary is used to store video frames, the feature vector is used as a linear combination of a set of basis vectors in the dictionary, and an orthogonal matching algorithm is used to sparsely encode the feature vector to obtain a sparse coefficient of the video frame, and the sparse coefficient of the video frame is encoded using entropy coding to obtain an encoded string; At the decompression end, the encoded string is read, and each bit of the string is searched, the node corresponding to the string is moved according to the number of bits of the current string, and when the node after the move is a leaf node, the corresponding sparse coefficient is added to the decompressed sparse coefficient; The nodes are composed of sparse coefficients; Based on the dictionary storing the video frame, the sparse coefficient feature vector is reconstructed into the original feature vector by using inverse sparse coding, and the video frame reconstructed by the original feature vector is output.

2. A video data compression and reconstruction method as claimed in claim 1, characterized in that: The step of extracting a feature vector of a video frame from video data comprises the following steps: Combine multiple convolution kernels of the convolutional neural network and convolve each convolution kernel with a local area of ​​the input video frame to generate a feature map; Performing average pooling on the feature map through a pooling layer to convert the feature map into a feature vector of a fixed length; Traverse the collection of video frames, input each frame into the convolutional neural network, and use the forward propagation of the convolutional neural network to obtain the feature vector of each frame image.

3. A video data compression and reconstruction method as claimed in claim 2, characterized in that: Before extracting the feature vector of the video frame in the video data, the following steps are also included: Performing video denoising on the video data, and performing color space conversion on the denoised video data; Perform color space conversion on the video data to obtain video frames.

4. A video data compression and reconstruction method as claimed in claim 1, characterized in that: The feature vector is used as a linear combination of a set of basis vectors in a dictionary, and an orthogonal matching algorithm is used to perform sparse coding on the feature vector to obtain a sparse coefficient of a video frame, including the following steps: Initialize a dictionary D∈R through the K-SVD algorithm m*k , get the sparse representation a corresponding to the feature vector i , so that the coefficient eigenvector x i ≈Da i , each feature vector is represented as a linear combination of a set of basis vectors in the dictionary; Where m is the dimension of the feature space, k is the size of the dictionary, and R is a set of real numbers; Use the orthogonal matching pursuit (OMP) algorithm to solve sparse coding: , where λ is the regularization parameter, is the square of the 2-norm of the matrix "•", To take the 1-norm of the matrix "•"; Select a coefficient eigenvector x i and its corresponding sparse code a i , update a basis vector d in the dictionary j , so that x i and a i Minimize the error between The basis vector d j The update expression is: ； Repeatedly update a basis vector d in the dictionary j The step of obtaining an optimized dictionary D is performed, and for each input video frame feature, the learned dictionary is used for sparse coding to obtain a set of sparse coefficients of the video frame.

5. A video data compression and reconstruction method as claimed in claim 1, characterized in that: The method of encoding the sparse coefficients of the video frame by using entropy coding to obtain a coded string comprises the following steps: Get the probability of each coefficient in the sparse coefficient group, and construct a Huffman tree with the probability of each coefficient sorted from small to large; Traverse in a depth-first search manner, add a '0' character to the end of the path string each time the left child node is traversed, and add a '1' character to the end of the path string each time the right child node is traversed. When a leaf node is reached, the path string is associated with the sparse coefficient feature vector corresponding to the leaf node to form a Huffman coding table. Traverse each value of the sparse coefficient group, use the Huffman coding table to replace the corresponding Huffman code, and generate a coded string.

6. A video data compression and reconstruction method as claimed in claim 1, characterized in that: The method of reconstructing the sparse coefficient feature vector into the original feature vector by using inverse sparse coding and outputting the video frame reconstructed by the original feature vector comprises the following steps: Based on the orthogonal matching pursuit (OMP) algorithm, the sparse representation a is calculated using the sparse coefficients. i , using the above coefficients to represent the optimized dictionary D of the module, by reconstructing the formula x i ≈Da i To get the original feature vector x i ; Through the generator network GA, the feature vector is used as input to the generator network. The deconvolution layer inside the generator network adjusts the input feature vector to an output image with image structure and details. Add a batch normalization layer after the convolution layer, and apply the ReLU activation function after the batch normalization layer. The activation function introduces nonlinearity so that the generator learns the mapping relationship; The generator network converts the mapping relationship of features into the final generated image through an output layer, provides each feature vector as input to the generator network GAN, and obtains the video frame through the output layer.

7. A video data compression and reconstruction method as claimed in claim 1, characterized in that: After the video frames reconstructed by the original feature vectors are output, all reconstructed video frames are merged and color space conversion is performed.

8. A video data compression and reconstruction system, characterized in that: include: A data extraction module, used to obtain video data and extract feature vectors of video frames in the video data through a convolutional neural network; A compression module, for, at the compression end, when using a dictionary to store video frames, using the feature vector as a linear combination of a set of basis vectors in the dictionary, and using an orthogonal matching algorithm to sparsely encode the feature vector to obtain sparse coefficients of the video frame, and using entropy coding to encode the sparse coefficients of the video frame to obtain a coded string; A decompression module, for reading the encoded string at the decompression end, searching each bit of the string, moving the node corresponding to the string according to the number of bits of the current string, and adding the corresponding sparse coefficient to the decompressed sparse coefficient when the node after the move is a leaf node; The nodes are composed of sparse coefficients; The reconstruction module is used to reconstruct the sparse coefficient feature vector into the original feature vector by using inverse sparse coding based on the dictionary storing the video frame, and output the video frame reconstructed by the original feature vector.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a video data compression and reconstruction method as claimed in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium includes a stored computer program, and the computer program is executed by at least one processor to enable the at least one processor to perform the steps of a video data compression and reconstruction method as claimed in any one of claims 1 to 7.

Images