Image compression method, system, and storage medium based on continuous variable bit rate allocation

Abstract

The application discloses an image compression method, system and storage medium based on continuous variable code rate allocation, and the method comprises the following steps: acquiring an image to be compressed, inputting the image into a trained image continuous variable code rate image compression network, and outputting an image compression result; the image to be compressed is encoded and decoded, and the encoding end comprises the following steps: extracting an image feature value y, using a joint quantization module to perform joint quantization on the channel dimension and the spatial dimension of the image feature value y, and obtaining a quantized feature value; using an arithmetic encoder to encode the quantized feature value into bit stream data; the decoding end comprises the following steps: after the received bit stream data is decoded by an arithmetic decoder, joint dequantization is performed on the channel dimension and the spatial dimension, and a decoded image feature value y is obtained ’ ; the decoded image feature value y ’ is subjected to feature reconstruction, and a reconstructed image is obtained. The application can dynamically control the bit rate during image compression, reduce the bit error rate, and improve the quality of the compressed image.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image compression, and particularly relates to an image compression method, system and storage medium based on continuous variable bit rate allocation. BACKGROUND

[0002] Image compression has always been a valuable research direction in the field of computer vision. In many scenarios, the transmission and storage of images are often limited by bandwidth, such as network services, so it is necessary to properly compress the images to adapt to the changing channel bandwidth. Traditional image compression methods, such as JPEG and JPEG 2000, have been widely adopted, but there are limitations in adaptability and flexibility. Recent research focuses on end-to-end image compression frameworks that use deep learning to achieve better compression ratio and image quality. These end-to-end frameworks usually use neural networks to encode and decode images, enabling them to learn the optimal representation that balances perceptual quality and compression efficiency.

[0003] Despite these advances, there is still a major challenge in achieving effective bit rate control, which is crucial for adapting to different network conditions and user needs. Bit rate control involves dynamically adjusting the compression rate to ensure that the encoded image meets specific quality targets while effectively utilizing available resources. SUMMARY

[0004] To solve the above technical problems, the present application proposes an image compression method, system and storage medium based on continuous variable bit rate allocation.

[0005] In a first aspect, the present application proposes an image compression method based on continuous variable bit rate allocation, which comprises: obtaining a to-be-compressed image, inputting it into a trained image continuous variable bit rate image compression network, and outputting an image compression result.

[0006] The trained image continuous variable bit rate image compression network is used to encode and decode the to-be-compressed image, and the process specifically comprises:

[0007] At the encoding end:

[0008] S1: using a feature extractor to extract features of the image to obtain image feature values y;

[0009] S2: using a joint quantization module to jointly quantize the channel dimension and spatial dimension of the image feature values y to obtain quantized feature values;

[0010] S3: using an arithmetic encoder to encode the quantized feature values into bit stream data;

[0011] At the decoding end:

[0012] S4: the received bit stream data is decoded by an arithmetic decoder, and then joint dequantization of channel dimension and spatial dimension is performed to obtain decoded image feature values y ’ ;

[0013] S5: feature reconstruction is performed on the decoded image feature values y ’ , to obtain a reconstructed image.

[0014] In a second aspect, the present application provides an image compression system based on continuous variable bit rate allocation, which is used to implement the image compression method based on continuous variable bit rate allocation as described in the first aspect of the present application, and the system comprises:

[0015] a collection module, configured to collect an image to obtain an input image;

[0016] a feature encoding module, configured to perform feature encoding on the image to extract image feature values;

[0017] a joint quantization module, configured to perform joint quantization of channel dimension and spatial dimension on the image feature values y to obtain quantized feature values;

[0018] an arithmetic encoding module, configured to encode the quantized feature values into bit stream data;

[0019] an arithmetic decoding module, configured to perform arithmetic decoding on the received bit stream data;

[0020] a joint dequantization module, configured to perform dequantization processing on the feature values after arithmetic decoding to obtain dequantized image feature values y ’ ;

[0021] a feature reconstruction module, configured to perform feature reconstruction on the dequantized image feature values y ’ to obtain a reconstructed image;

[0022] an output module, configured to output the final reconstructed image.

[0023] In a third aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method as described in the first aspect of the present application.

[0024] The present application has the following beneficial effects:

[0025] 1. The present application can realize accurate rate control during image compression.

[0026] The proposed quantization step estimation network can accurately estimate the quantization parameter according to the target code rate, so that the quantization result is close to the target code rate. Experiments show that the average code rate error of the present application is about 3.6%, and the code rate can be effectively controlled under different target code rates, such as 1.0 bpp, 0.9 bpp and other target code rates, and the actual code rate has a small deviation from the target code rate.

[0027] 2. The present application can significantly improve the R-D performance of image compression.

[0028] The joint channel and spatial quantization scheme has better rate-distortion performance than only channel quantization at high code rate, which can improve the image reconstruction quality while ensuring the compression efficiency. In the experiment on Kodak dataset, compared with only channel quantization and other image compression methods, the joint quantization method has more advantages at high code rate.

[0029] 3. The image continuous variable code rate image compression network proposed in the present application has stronger adaptability.

[0030] The present application can dynamically adjust the quantization parameter according to the image content and the target code rate, and is suitable for different types of images and various network conditions, which overcomes the problems of insufficient adaptability and flexibility of traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a schematic diagram of the block index-peak signal-to-noise ratio (PNSR) result after image quantization operation;

[0032] Figure 2 It is a schematic diagram of the channel index-peak signal-to-noise ratio (PNSR) result after image quantization operation on different channels;

[0033] Figure 3 It is a step flowchart of the image compression method based on variable code rate allocation in the embodiment of the present application;

[0034] Figure 4 It is a flowchart of image encoding and decoding in the embodiment of the present application;

[0035] Figure 5 It is a processing flowchart of the spatial estimation network in the embodiment of the present application;

[0036] Figure 6 It is a schematic diagram of the training process of the quantization parameter estimation network (QSE) in the embodiment of the present application;

[0037] Figure 7 It is a structural schematic diagram of the image compression system based on variable code rate allocation in the embodiment of the present application;

[0038] Figure 8 It is a comparison diagram of the results of the present application and existing image compression methods in experimental test. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The image compression process mainly includes:

[0041] At the encoding end, image features are extracted by the encoder, quantized, and then encoded into bitstream data. The bitstream data is then transmitted to the decoding end through a transmission channel or medium.

[0042] At the decoding end, the received bitstream data is decoded and reconstructed by the decoder to obtain the reconstructed image (i.e., the image after compression and reconstruction).

[0043] Figure 1 This is a schematic diagram of the block index-peak signal-to-noise ratio (PNSR) result after image quantization. Figure 1 In this study, the example image was divided into 16 equally sized blocks, and the same quantization operation was performed on each block to obtain different peak signal-to-noise ratios (PNSRs).

[0044] Figure 2 This is a schematic diagram of the Channel Index-Peak Signal-to-Noise Ratio (PNSR) result after quantization operations on different channels of an image. Figure 2 In this study, the example image was divided into 16 equally sized blocks, and the same quantization operation was performed on different channels (all positions) to obtain different peak signal-to-noise ratios (PNSRs).

[0045] Experiments have proven that, referring to Figure 1 , 2 As shown, in the image compression and reconstruction process, both the channel dimension and spatial dimension of the image feature map are related to image reconstruction. Feature values ​​from different channels and different spatial locations within the same channel play different roles in the reconstructed image. Using the same quantization parameter for all locations within the same channel will not yield optimal rate-distortion (RD) performance. Therefore, to adapt to different bandwidths, applying varying degrees of channel and spatial quantization to image features (i.e., a joint quantization scheme of channel and spatial dimensions) plays a crucial role in improving image compression quality.

[0046] This invention aims to propose a novel bitrate allocation scheme that achieves precise and rapid bitrate control by combining channel and spatial quantization schemes with a quantization step size estimation network, thereby improving compressed image quality and reducing bitrate distortion to adapt to different network conditions and user needs.

[0047] Figure 3 This is a flowchart illustrating the steps of an image compression method based on continuous variable bit rate allocation according to an embodiment of the present invention.

[0048] This invention proposes an image compression method based on continuous variable bitrate allocation. The method includes: acquiring the image to be compressed, inputting it into a trained image compression network with continuous variable bitrate, and outputting the image compression result.

[0049] Reference Figure 4 As shown, the image compression network with continuously variable bit rate includes a feature extractor (including an encoder based on a convolutional neural network CNN), a joint quantization module, a quantization parameter estimation network QSE, an arithmetic encoder, an arithmetic decoder, a joint inverse quantization module, and a decoder (including a decoder based on a convolutional neural network CNN).

[0050] Figure 4 This is a schematic diagram illustrating the image encoding and decoding process in an embodiment of the present invention.

[0051] exist Figure 4 In the process, at the input end, the encoding of the input image includes: inputting the input image into a CNN-based encoder for feature encoding (i.e., feature extraction) to obtain image feature values ​​y; inputting the image feature values ​​y into a channel dimension quantization module, multiplying the image feature values ​​y with the gain vector to obtain channel quantization feature values ​​y. qc The gain vector is obtained based on the quantization parameters, which are provided by the quantization parameter estimation network QSE; the channel quantization feature values ​​y are then used. qc Input a hyperprior network and a spatial estimation network to extract the spatial information qs from the input image; then quantize the channel feature values ​​y. qc The spatial information qs is input into the spatial dimension quantization module, and the feature value y is quantized using channels. qc Dividing by the spatial information qs, we obtain the channel quantization feature value y. qs ; for the channel quantization eigenvalue y qs The data is rounded down and then input into the arithmetic encoder to obtain the bitstream data (i.e., the compressed bitstream data). This bitstream data is then transmitted to the decoder via a transmission channel or medium. Decoding is the reverse process of encoding. The decoding process includes: inputting the received bitstream data into the arithmetic decoder; after arithmetic decoding, the data is input into the joint inverse encoding module, where spatial dimension inverse quantization and channel dimension inverse quantization are performed respectively to obtain the decoded image feature values ​​y. ’ ; to extract image feature values ​​y ’The input is a CNN-based decoder, which performs decoding and feature reconstruction to obtain a reconstructed image, which is the image compression result.

[0052] Reference Figure 3 , 4 As shown, the trained image compression network with continuously variable bitrate encodes and decodes the image to be compressed. The specific process includes:

[0053] At the encoding end:

[0054] S1: Use a feature extractor to extract features from the image and obtain the image feature value y.

[0055] For example, the feature extractor is an encoder based on a convolutional neural network (CNN). After encoding, image feature values ​​y, y∈ Where R represents a real number, c represents the number of channels of the image feature, w represents the width of the image feature, and h represents the depth of the image feature.

[0056] S2: Using the joint quantization module, the image feature value y is jointly quantized in terms of channel dimension and spatial dimension to obtain the quantized feature value.

[0057] In this embodiment of the invention, the joint quantization module includes a channel-dimensional quantization module and a spatial-dimensional quantization module.

[0058] In a preferred embodiment, refer to Figure 4 As shown, the joint quantization of image feature values ​​y using both channel and spatial dimensions involves the following steps:

[0059] S201: Multiply the image feature value y with the gain vector to obtain the channel quantization feature value y. qc The gain vector m l The gain vector is obtained based on the quantization parameter l. It is a k-dimensional vector, where k equals the number of channels in the feature value y. The quantization parameter l is optimized during model training. To obtain the bitrate across multiple bandwidths, multiple gain vectors are set, and each gain vector is multiplied by the feature value to obtain the corresponding quantized feature value for that bandwidth.

[0060] S202: Extract spatial information qs from images using a super-prior network and a spatial estimation network.

[0061] Specifically, the channel quantization feature value y qc Given a hyperprior network, what processing is performed to obtain edge information? Since the edge information needs to be encoded into a byte stream to assist the decoder, it is visible at both the decoding and encoding ends.

[0062] It should be noted that the super-prior network can be regarded as a network for estimating the image bitrate. At the same time, its output also includes information such as the texture complexity at different locations of the image (collectively referred to as edge information). This edge information is used for subsequent spatial dimension quantization.

[0063] Figure 5 This is a schematic diagram of the processing flow of the spatial estimation network in an embodiment of the present invention.

[0064] Reference Figure 5 As shown, the edge information is input into the spatial estimation network, and after passing through three convolutional layers and two residual convolutional layers, the spatial information qs of the image is obtained. The spatial information qs of the image includes information such as the texture complexity at different spatial locations in the image.

[0065] S203: Quantize the channel feature value y qc Dividing by the spatial information qs yields the spatial quantization feature value y. qs That is, the quantized feature values.

[0066] The quantization of the channel and spatial dimensions described above can simultaneously remove information redundancy in both dimensions (space and channel), minimizing the impact on the quality of the reconstructed image while reducing the bitrate. This allows for quantization of bitrates under different bandwidths, making it possible to achieve the desired results.

[0067] To obtain the bitstream data, arithmetic coding is required. Arithmetic coding is a lossless compression method that can further reduce bandwidth requirements. However, this scheme can only compress discrete data, so the spatially quantized feature values ​​y must be processed. qs Round down.

[0068] In a preferred embodiment, refer to Figure 4 As shown, the quantized feature values ​​are rounded down and then input into the arithmetic encoder for encoding.

[0069] S3: Using an arithmetic encoder, the quantized feature values ​​(i.e., spatially quantized feature values ​​y) are converted into quantized feature values. qs It is encoded as bitstream data.

[0070] Arithmetic coding is a highly efficient lossless data compression technique that compresses data by encoding the entire message sequence into a single number between 0 and 1. Unlike other entropy coding methods, arithmetic coding does not assign a fixed-length code to each symbol; instead, it dynamically allocates the code length based on the entire message content.

[0071] The encoding end encodes the image to obtain bitstream data, which is then transmitted to the decoding end via a transmission channel or medium. The decoding process is the reverse of the encoding process.

[0072] On the decoding end:

[0073] S4: After the received bitstream data is decoded by the arithmetic decoder, joint inverse quantization of the channel dimension and spatial dimension is performed to obtain the decoded image feature value y. ’ .

[0074] Reference Figure 4 As shown, the arithmetic decoder decoding is the inverse process of the arithmetic encoder encoding, and the joint inverse quantization of the channel dimension and spatial dimension is the inverse process of the joint quantization of the channel dimension and spatial dimension. Specifically, after arithmetic decoding, the decoded spatial quantized feature value y is obtained. qs ’ Multiplying it with the spatial information qs yields the decoded channel quantization feature value y. qc ’ ; Quantize the channel feature value y qc ’ Multiplying the result with the inverse gain vector yields the decoded image feature value y. ’ The inverse gain vector corresponds to the gain vector and can be obtained from the quantization parameters.

[0075] S5: For the decoded image feature values ​​y ’ Feature reconstruction is performed to obtain the reconstructed image.

[0076] Specifically, the decoded image feature value y ’ The input is based on a CNN decoder, which performs decoding and feature reconstruction to obtain a reconstructed image, which is the image compression result.

[0077] The image compression network with continuously variable bitrate described is trained end-to-end. In this model, N gain vectors and inverse gain vectors are defined; for example, N=4. Both the gain vectors and inverse gain vectors are initialized as unit vectors, and their gradients are set to be differentiable for update optimization. During model training, a gain vector (and its corresponding inverse gain vector) is randomly selected, and then backpropagation using the loss function is employed to optimize the model.

[0078] In some preferred embodiments, the loss function of the continuously variable bitrate image compression network during training is:

[0079]

[0080] Where mse(.) represents the mean squared error function, Img represents the input image, reconImg represents the reconstructed image, and R represents the bit rate. Each represents a trade-off between bitrate and image quality. This corresponds to a gain vector (and its inverse gain vector). For example, The possible values ​​are 256, 512, 1024, and 2048.

[0081] Discrete variable code rate can be achieved through the above methods.

[0082] Since bandwidth is a continuously variable value, a continuously variable bit rate must be implemented. This invention employs exponential interpolation to achieve a continuously variable bit rate based on a discrete variable bit rate.

[0083] For any two adjacent gain vectors (and their corresponding inverse gain vectors) , The middle value It can be calculated using the following formula:

[0084]

[0085] In the formula, Indicates the first The gain vector corresponding to each quantization level Indicates the first +1 quantization level corresponding to the gain vector Take an integer, such as 0, 2, 3, etc. Indicates the interpolation coefficients. The value range is from 0 to 1.

[0086] By using the above interpolation method, a continuous gain vector (and its corresponding inverse gain vector) can be obtained, thereby achieving a continuous variable code rate.

[0087] After training the image compression network with continuously variable bitrate, it can be trained by specifying appropriate quantization coefficients. To achieve image compression at arbitrary bandwidth, l represents the quantization level, and each quantization level corresponds to a gain vector. For example, the quantization level l can take the value of an integer such as 0, 1, 2, 3, etc. The interpolation coefficients are represented. For a given image, its bit rate is determined by quantization parameters (QPs), which include the quantization level and interpolation coefficients. This means that a mapping from quantization parameters to bit rate can be established. However, the quantization parameters cannot be directly obtained based on a given bit rate. To address this issue, this invention employs a quantization parameter estimation (QSE) network. The QSE network aims to establish a mapping between the target bit rate and the quantization parameters (QPs). Considering the information content in different images, the latent representation of the image, along with the target bit rate, is input into the QSE network. Since Continuous Variable Rate (CVR) image compression networks can allocate continuous bit rates based on quantization parameters, and the QSE network can accurately estimate the quantization parameters, combining CVR and QSE networks enables fast and accurate bit rate control.

[0088] Specifically, for a given image, since there is a one-to-one correspondence between its bit rate and quantization parameters, as long as the QSE network learns the mapping from the target bit rate to the quantization parameters, it can achieve precise bit rate control.

[0089] Figure 6 This is a schematic diagram illustrating the training process of the Quantization Parameter Estimation Network (QSE) in an embodiment of the present invention.

[0090] In a preferred embodiment, refer to Figure 6 As shown, the training process of the Quantization Parameter Estimation Network (QSE) is as follows:

[0091] S101: Initialize the parameters of the quantization parameter estimation network QSE;

[0092] S102: Randomly select a target quantization parameter (l target s target The image samples are compressed to obtain the actual bitrate. ;

[0093] S103: The actual bitrate As the target bitrate, it is input along with image samples into the quantization parameter estimation network QSE, which outputs the estimated quantization parameters (l). real s real );

[0094] S104: Calculate the target quantization parameter (l) target s target ) and estimated quantization parameters (l real s real The loss between the two is used to backpropagate the quantization parameter estimation network (QSE) and optimize it.

[0095] The target quantization coefficients and the absolute variance of the estimated quantization coefficients are used as quantization parameters to estimate the loss function employed by the QSE network during the training phase. Specifically, it is as follows:

[0096] .

[0097] Based on the same inventive concept, embodiments of the present invention also provide an image compression system based on variable bitrate allocation. This image compression system has the same or similar technical features as the above-described image compression method; the same or similar technical features will not be repeated below.

[0098] Reference Figure 7 As shown, an image compression system based on continuously variable bit rate allocation is used in the aforementioned image compression method based on continuously variable bit rate allocation. The system includes:

[0099] The acquisition module is used to acquire images and obtain the input image;

[0100] The feature encoding module is used to encode the features of an image and extract its feature values.

[0101] The joint quantization module is used to perform joint quantization of the image feature value y in both channel and spatial dimensions to obtain the quantized feature value.

[0102] An arithmetic encoder is used to encode quantized feature values ​​into a bitstream data.

[0103] An arithmetic decoder is used to perform arithmetic decoding on the received bitstream data.

[0104] The joint dequantization module is used to dequantize the feature values ​​after arithmetic decoding to obtain the dequantized image feature values ​​y. ’ ;

[0105] The feature reconstruction module is used to reconstruct the feature values ​​y of the dequantized image. ’ Feature reconstruction is performed to obtain the reconstructed image;

[0106] The output module is used to output the final reconstructed image.

[0107] In a preferred embodiment, the joint quantization module includes a channel dimension quantization module and a spatial dimension quantization module. The channel dimension quantization module is used to perform channel dimension quantization on the image feature value y, and the spatial dimension quantization module is used to perform spatial dimension quantization by combining the spatial information of the image.

[0108] This invention also proposes a computer-storable medium storing a computer program thereon, which is processed and executed to implement the image compression method based on variable bitrate allocation described in the above embodiments.

[0109] Experimental test:

[0110] Experimental platform or software: Integrated development tools, such as PyCharm, Ubuntu operating system, PyTorch neural network framework.

[0111] Test dataset: Kodak dataset.

[0112] Test results:

[0113] Figure 8 This is a comparison chart of the results of the present invention and existing image compression methods during experimental testing. The term "channel-only quantization scheme" indicates that the continuously variable bitrate image compression model proposed in this invention only uses channel-level quantization. The term "channel and spatial joint quantization scheme" indicates that the continuously variable bitrate image compression model proposed in this invention considers both channel and spatial quantization. The term "base model" refers to the image compression model proposed in reference 1, which uses discrete Gaussian mixture likelihood and attention modules for learning image compression (referred to as the "base model"). Multiple points on each curve in the figure represent PSNR at multiple bitrates. The present invention achieves multiple bitrates on the same model, while the "base model" requires training multiple models to achieve multiple bitrates.

[0114] Reference Figure 7 As shown, the PSNR of image compression using this invention is no lower than that of the base model. Furthermore, this invention achieves multiple bitrates on a single image compression model, eliminating the need to train multiple models and saving training costs.

[0115] The present invention achieved a bitrate error of less than 3% on the Kodak dataset without any degradation in image quality.

[0116] Document 1: Cheng, Zhengxue et al. "Learned Image Compression WithDiscretized Gaussian Mixture Likelihoods and Attention Modules." 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020):7936-7945.

[0117] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An image compression method based on continuous variable bit rate allocation, characterized in that, include: The image to be compressed is obtained, input into the trained image compression network with continuously variable bitrate, and the image compression result is output. The trained image compression network with continuously variable bitrate is used to encode and decode the image to be compressed. The specific process includes: At the encoding end: a feature extractor is used to extract features from the image to obtain the image feature value y; Multiplying the image feature value y by the gain vector yields the channel quantization feature value y. qc The gain vector is a k-dimensional vector, where k equals the number of channels of the image feature value y. The gain vector m1 is provided by the trained quantization parameter estimation network QSE, which provides quantization coefficients (l, s), where l represents the quantization level, each quantization level corresponds to the gain vector m1, the quantization level l takes any integer value of 0, 1, 2 or 3, and s represents the interpolation coefficients. The training process of the Quantization Parameter Estimation Network (QSE) is as follows: Initialize the quantization parameters to estimate the parameters of the QSE network; Randomly select a target quantization parameter ( l target ，s target The image samples are compressed to obtain the actual bitrate. R target ; The actual bit rate R target As the target bitrate, it is input along with image samples into the quantization parameter estimation network QSE, which outputs estimated quantization parameters. l real ，s real ); Calculate the target quantization parameter ( l target ，s target ) and estimated quantization parameters ( l real ，s real The loss between the two parameters is used to backpropagate and optimize the quantization parameter estimation network QSE. quantize the channel feature value y qc The edge information is obtained by inputting the hyperprior network; the edge information is then input into the spatial estimation network, and after passing through three convolutional layers and two residual convolutional layers, the spatial information of the image is obtained. qs The spatial information of the image qs This includes texture complexity information at different spatial locations within the image; quantize the channel feature value y qc Spatial Information qs Divide to obtain the spatially quantized eigenvalue y qs That is, the quantized eigenvalues; The quantized feature values ​​are encoded into bitstream data using an arithmetic encoder. On the decoding end: After the received bitstream data is decoded by an arithmetic decoder, joint inverse quantization of the channel and spatial dimensions is performed to obtain the decoded image feature values ​​y. ’ ; For the decoded image feature value y ’ Feature reconstruction is performed to obtain the reconstructed image.

2. The image compression method based on continuous variable bit rate allocation according to claim 1, characterized in that, Also includes: At the decoding end, the quantized feature values ​​are rounded down and then input into the arithmetic encoder for decoding.

3. The image compression method based on continuous variable bit rate allocation according to claim 1, characterized in that, The image compression network with continuously variable bit rate includes a feature extractor, a joint quantization module, a quantization parameter estimation network (QSE), an arithmetic encoder, an arithmetic decoder, a joint inverse quantization module, and a decoder.

4. The image compression method based on continuous variable bit rate allocation according to claim 1, characterized in that, The loss function (Loss) of the continuously variable bitrate image compression network during training is: ， Where mse(.) represents the mean squared error function, Img represents the input image, and reconImg represents the reconstructed image. R Represents bitrate. Each represents a trade-off between bitrate and image quality. This corresponds to a gain vector.

5. An image compression system based on continuously variable bitrate allocation, the system being used to implement the image compression method based on continuously variable bitrate allocation as described in claim 1, characterized in that, The system includes: The acquisition module is used to acquire images and obtain the input image; The feature encoding module is used to encode the features of an image and extract its feature values. The joint quantization module is used to perform joint quantization of the image feature value y in both channel and spatial dimensions to obtain the quantized feature value. The arithmetic coding module is used to encode the quantized feature values ​​into bitstream data; The arithmetic decoding module is used to perform arithmetic decoding on the received bitstream data; The joint dequantization module is used to dequantize the feature values ​​after arithmetic decoding to obtain the dequantized image feature values ​​y. ’ ; The feature reconstruction module is used to reconstruct the feature values ​​y of the dequantized image. ’ Feature reconstruction is performed to obtain the reconstructed image; The output module is used to output and display the final reconstructed image.

6. The image compression system based on continuously variable bit rate allocation according to claim 5, characterized in that, The joint quantization module includes a channel-dimensional quantization module and a spatial-dimensional quantization module. The channel-dimensional quantization module is used to perform channel-dimensional quantization on the image feature value y, while the spatial-dimensional quantization module is used to perform spatial-dimensional quantization by combining the spatial information of the image.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is processed and executed, it implements the image compression method based on continuous variable bit rate allocation as described in any one of claims 1-4.

Images