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

Abstract

The invention discloses an image compression method and system based on variable code rate distribution and a storage medium, and the method comprises the steps: obtaining a to-be-compressed image, inputting the to-be-compressed image into a trained image continuous variable code rate image compression network, and outputting an image compression result; the method for encoding and decoding the image to be compressed comprises the following steps of: at an encoding end, extracting an image characteristic value y, and carrying out channel dimension and space dimension joint quantization on the image characteristic value y by utilizing a joint quantization module to obtain a quantized characteristic value; encoding the quantized feature value into bit stream data by using an arithmetic encoder; at a decoding end, after received bit stream data is decoded by an arithmetic decoder, joint inverse quantization of a channel dimension and a space dimension is carried out, and a decoded image characteristic value y'is obtained; according to the image compression method and device, the bit rate during image compression can be dynamically controlled, the bit error rate can be reduced, and the quality of the compressed image can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image compression, and in particular relates to an image compression method, system and storage medium based on variable bit rate allocation. Background Art

[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 compress the images appropriately to adapt to the changing channel bandwidth. Traditional image compression methods, such as JPEG and JPEG 2000, have been widely adopted, but have limitations in adaptability and flexibility. Recent research has focused on end-to-end image compression frameworks that use deep learning to achieve better compression ratios and image quality. These end-to-end frameworks typically use neural networks to encode and decode images, enabling them to learn optimal representations that balance perceptual quality and compression efficiency.

[0003] Despite these advances, a significant challenge remains in achieving effective bitrate control, which is critical to adapting to varying network conditions and user demands. Bitrate control involves dynamically adjusting the compression rate to ensure that the encoded image meets a specific quality target while efficiently utilizing available resources. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an image compression method, system and storage medium based on variable bit rate allocation.

[0005] In a first aspect, the present invention proposes an image compression method based on variable bit rate allocation, the method comprising: obtaining an image to be compressed, inputting the image 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 compressed image. The specific process includes:

[0007] On the encoding side:

[0008] S1: Use the feature extractor to extract features from the image and obtain the image feature value y;

[0009] S2: Using the joint quantization module, the image eigenvalue y is jointly quantized in the channel dimension and the spatial dimension to obtain the quantized eigenvalue;

[0010] S3: Encode the quantized feature value into bit stream data using an arithmetic encoder;

[0011] On the decoding side:

[0012] S4: After the received bit stream data is decoded by the arithmetic decoder, the channel dimension and the spatial dimension are jointly dequantized to obtain the decoded image feature value y';

[0013] S5: Reconstruct the decoded image feature value y' to obtain a reconstructed image.

[0014] In a second aspect, the present invention provides an image compression system based on variable bit rate allocation, the system comprising:

[0015] An acquisition module, used for acquiring images and obtaining input images;

[0016] A feature encoding module is used to perform feature encoding on the image and extract image feature values;

[0017] A joint quantization module is used to jointly quantize the channel dimension and the spatial dimension of the image eigenvalue y to obtain a quantized eigenvalue;

[0018] an arithmetic coding module, used for encoding the quantized feature values ​​into bit stream data;

[0019] An arithmetic decoding module, used for performing arithmetic decoding on the received bit stream data;

[0020] A joint dequantization module is used to dequantize the eigenvalues ​​after arithmetic decoding to obtain dequantized image eigenvalues ​​y';

[0021] A feature reconstruction module is used to perform feature reconstruction on the inverse quantized image feature value y' to obtain a reconstructed image;

[0022] The output module is used to output the final reconstructed image.

[0023] In a third aspect, the present invention provides a computer storable medium having a computer program stored thereon, wherein the computer program, when processed and executed, implements the method of the first aspect of the present invention.

[0024] Beneficial effects of the present invention:

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

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

[0027] 2. The present invention can significantly improve the RD performance of image compression.

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

[0029] 3. The image continuous variable bit rate image compression network proposed by the present invention has stronger adaptability.

[0030] The present invention can dynamically adjust quantization parameters according to image content and target bit rate, is applicable to different types of images and a variety of network conditions, and overcomes the problems of insufficient adaptability and flexibility of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0033] Figure 3 A flowchart of the steps of an image compression method based on variable bit rate allocation according to an embodiment of the present invention;

[0034] Figure 4 Schematic diagram of the process of image encoding and decoding in an embodiment of the present invention;

[0035] Figure 5 Schematic diagram of the processing flow of the space estimation network in an embodiment of the present invention;

[0036] Figure 6 Schematic diagram of the training process of the quantization parameter estimation network QSE in an embodiment of the present invention;

[0037] Figure 7 A schematic diagram of the structure of an image compression system based on variable bit rate allocation according to an embodiment of the present invention;

[0038] Figure 8 This is a comparison chart of the results of the experimental test between the present invention and the existing image compression method. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 are within the scope of protection of the present invention.

[0040] The image compression process mainly includes:

[0041] At the encoding end, the encoder extracts image features, quantizes the image features, and encodes them into bit stream data, which is then transmitted to the decoding end through a transmission channel or medium.

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

[0043] Figure 1 Schematic diagram of the block index-peak signal-to-noise ratio (PNSR) result after image quantization operation. Figure 1 In the example, the sample image is divided into 16 blocks of equal size, and the same quantization operation is taken for each block to obtain different peak signal-to-noise ratios (PNSRs).

[0044] Figure 2 Schematic diagram of the channel index-peak signal-to-noise ratio PNSR results after quantization operations on different channels of the image. Figure 2 In the example image, the sample image is divided into 16 blocks of equal size, and the same quantization operation is performed on different channels (all positions) to obtain different peak signal-to-noise ratios (PNSR).

[0045] It has been proved by experiments that Figure 1 , 2 As shown in the figure, in the process of image compression and reconstruction, the channel dimension and spatial dimension of the image feature map are both related to image reconstruction. The eigenvalues ​​of different channels and different spatial positions of the same channel play different roles in reconstructing the image. If the same quantization parameter is used for all positions in the same channel, the optimal rate-distortion (RD) performance cannot be obtained. Therefore, in order to adapt to different bandwidths, it is important to quantize the channel dimension and spatial dimension of the image features to different degrees (i.e., the joint quantization scheme of the channel dimension and the spatial dimension), which plays an important role in improving the quality of image compression.

[0046] The present invention aims to propose a new bit rate allocation scheme, which realizes accurate and fast bit rate control, improves compressed image quality, reduces bit rate distortion, and adapts to different network conditions and user needs by combining channel and spatial quantization schemes and quantization step estimation networks.

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

[0048] An embodiment of the present invention proposes an image compression method based on variable bit rate allocation, which includes: obtaining an image to be compressed, inputting the image into a trained image continuous variable bit rate image compression network, and outputting an image compression result.

[0049] Reference Figure 4 As shown, the image continuous variable bit rate image compression network 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 The figure is a schematic diagram of the process of image encoding and decoding in an embodiment of the present invention.

[0051] exist Figure 4 In the example, at the input end, the process of encoding the input image includes: inputting the input image into the CNN-based encoder to perform feature encoding (i.e., extracting features) to obtain the image feature value y; inputting the image feature value y into the channel dimension quantization module, multiplying the image feature value y and the gain vector to obtain the channel quantization feature value y. qc , the gain vector is obtained according to the quantization parameter, which is provided by the quantization parameter estimation network QSE; the channel quantization eigenvalue y qc Input the super prior network and the spatial estimation network to extract the spatial information qs of the input image; quantize the channel eigenvalue y qc And spatial information qs is input into the spatial dimension quantization module, and the channel quantization feature value y is used qc Divide by the spatial information qs to get the channel quantization eigenvalue y qs ; Quantify the channel eigenvalue y qs The bit stream data is rounded and then input into the arithmetic encoder to obtain the bit stream data (i.e. the compressed bit stream data); the bit stream data is transmitted to the decoding end through the transmission channel or medium. Decoding is the inverse process of encoding. The decoding process includes: inputting the received bit stream data into the arithmetic decoder, and after arithmetic decoding, inputting into the joint inverse coding module, respectively performing spatial dimension inverse quantization and channel dimension inverse quantization, and obtaining the decoded image feature value y'; inputting the image feature value y' into the CNN-based decoder, decoding, feature reconstruction, and obtaining the reconstructed image, which is the image compression result.

[0052] Reference Figure 3 , 4 As shown in FIG. 1 , the trained image continuous variable bit rate image compression network is used to encode and decode the compressed image. The specific process includes:

[0053] On the encoding side:

[0054] S1: Use the 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, the image feature value y is obtained, y∈R c×w×h , 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: Use the joint quantization module to jointly quantize the channel dimension and the spatial dimension of the image eigenvalue y to obtain the quantized eigenvalue.

[0057] In the embodiment of the present invention, the joint quantization module includes a channel dimension quantization module and a space dimension quantization module.

[0058] In a preferred embodiment, referring to Figure 4 As shown, the image feature value y is jointly quantized in the channel dimension and the spatial dimension, and the specific process includes:

[0059] S201: Multiply the image eigenvalue y by the gain vector to obtain the channel quantization eigenvalue y qc , the gain vector m l Obtained according to the quantization parameter l. The gain vector is a k-dimensional vector, where k is equal to the number of channels of the eigenvalue y. The quantization parameter l is optimized during the model training phase. In order to obtain the bit rate under multiple point bandwidths, multiple gain vectors need to be set. Each gain vector is multiplied by the eigenvalue to obtain the quantized eigenvalue under the corresponding bandwidth.

[0060] S202: Extract the spatial information qs of the image using the hyper-prior network and the spatial estimation network.

[0061] Specifically, the channel quantization eigenvalue y qc Input the super prior network, what kind of processing is performed, and the side information is obtained. Since the side information needs to be encoded into a byte stream to assist the decoder in decoding, it is visible at both the decoding end and the encoding end.

[0062] It should be noted that the super prior network can be regarded as a network for estimating the image bit rate. At the same time, its output also contains information such as the texture complexity of different positions of the image (collectively referred to as side information), which is used for subsequent spatial dimension quantization.

[0063] Figure 5 Schematic diagram of the processing flow of the space estimation network in an embodiment of the present invention.

[0064] Reference Figure 5 As shown in Figure 1, the side information is input into the spatial estimation network, and the spatial information qs of the image is obtained after three layers of convolutional layers and two layers of residual convolution. The spatial information qs of the image includes information such as the texture complexity of different spatial positions in the image.

[0065] S203: quantizing the channel characteristic value y qc Divided by the spatial information qs, we get the spatial quantization eigenvalue y qs That is, the quantized eigenvalue.

[0066] Through the above-mentioned quantization of the channel dimension and the spatial dimension, the information redundancy in the two dimensions (space and channel) can be removed at the same time, while reducing the bit rate and minimizing the impact on the quality of the reconstructed image, so that the bit rate under different bandwidths can be quantized.

[0067] In order to obtain the bit stream 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 spatial quantization eigenvalue y must be qs Round up.

[0068] In a preferred embodiment, referring to Figure 4 As shown, the quantized eigenvalues ​​are rounded and then input into an arithmetic encoder for encoding.

[0069] S3: Use the arithmetic encoder to convert the quantized eigenvalue (i.e., the spatial quantization eigenvalue y qs ) is encoded as bit stream data.

[0070] Arithmetic coding is an efficient lossless data compression technique that achieves data compression 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, but dynamically assigns the code length based on the content of the entire message.

[0071] The encoding end encodes the image to obtain bit stream data, and the bit stream data is transmitted to the decoding end through a transmission channel or medium. The decoding process is the inverse process of the encoding process.

[0072] On the decoding side:

[0073] S4: After the received bit stream data is decoded by the arithmetic decoder, the channel dimension and the spatial dimension are jointly dequantized to obtain the decoded image feature value y'.

[0074] Reference Figure 4 As shown, the decoding of the arithmetic decoder is the inverse process of the encoding of the arithmetic encoder, and the joint dequantization of the channel dimension and the spatial dimension is the inverse process of the joint quantization of the channel dimension and the spatial dimension. Specifically, after arithmetic decoding, the decoded spatial quantization feature value y is obtained. qs ', multiply it with the spatial information qs to obtain the decoded channel quantization feature value y qc '; Quantize the channel feature value y qc'Multiply the inverse gain vector to obtain the decoded image feature value y'. The inverse gain vector corresponds to the gain vector and can be obtained according to the quantization parameter.

[0075] S5: Reconstruct the decoded image feature value y' to obtain a reconstructed image.

[0076] Specifically, the decoded image feature value y' is input into a CNN-based decoder for decoding and feature reconstruction to obtain a reconstructed image, which is the image compression result.

[0077] The image continuous variable bit rate image compression network is based on end-to-end training. In the model, N gain vectors and inverse gain vectors are defined, for example, N=4; both the gain vector and the inverse gain vector are initialized to unit vectors, and the gradient is set to be differentiable for update optimization. During the model training process, a gain vector (and its corresponding inverse gain vector) is randomly selected, and then the loss function is back-propagated to optimize the model.

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

[0079] Loss=mse(Img,recoImg)+λR

[0080] Among them, mse(.) represents the mean square error function, Img represents the input image, reconImg represents the reconstructed image, R represents the bit rate, λ represents the trade-off factor between bit rate and image quality, and each λ corresponds to a gain vector (and its corresponding inverse gain vector). For example, the optional values ​​of λ are 256, 512, 1024, and 2048.

[0081] Through the above method, discrete variable bit rate can be achieved.

[0082] Considering that bandwidth is a continuously variable value, a continuously variable bit rate must be implemented. The embodiment of the present invention uses an exponential interpolation method to implement 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) m l 、m l+1 , where the middle value m v It can be calculated using the following formula:

[0084] m v =m l s +m l+1 1-s

[0085] Among them, ml represents the gain vector corresponding to the lth quantization level (and its corresponding inverse gain vector), m l+1 Represents the gain vector corresponding to the l+1th quantization level (and its corresponding inverse gain vector), l takes an integer such as 0, 1, or 2; s represents the interpolation coefficient, and its value range is 0 to 1.

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

[0087] After completing the training of the image continuous variable bit rate image compression network, the image compression under any bandwidth can be achieved by specifying the appropriate quantization coefficient (l, s), where l represents the quantization level and each quantization level corresponds to the gain vector m l , for example, the quantization level l takes values ​​of integers such as 0, 1, 2, 3, etc.; s represents the interpolation coefficient. For a given image, its bit rate is determined by the quantization parameters (QPs), which include the quantization level and the interpolation coefficient. This means that a mapping from quantization parameters to bit rates can be established. However, the quantization parameters cannot be directly obtained based on a given bit rate. To solve this problem, the present invention adopts a quantization parameter estimation QSE network. The QSE network aims to establish a mapping between a target bit rate and quantization parameters (QPs). Considering the information content in different images, the potential representation of the image is input into the QSE network together with the target bit rate. Since the continuous variable rate image compression network (Continuous Variable Rate Net, CVR) can allocate continuous bit rates based on the quantization parameters, and the QSE network can accurately estimate the quantization parameters, combining the CVR network and the QSE network can achieve fast and accurate bit rate control.

[0088] Specifically, for a given image, since there is a one-to-one correspondence between its bitrate and quantization parameter, accurate bitrate control can be achieved as long as the QSE network learns the mapping from target bitrate to quantization parameter.

[0089] Figure 6 Schematic diagram of the training process of the quantization parameter estimation network QSE in an embodiment of the present invention.

[0090] In a preferred embodiment, referring to Figure 6 As shown, the training process of the quantization parameter estimation network QSE is:

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

[0092] S102: Randomly select a target quantization parameter (l target ,s target) and the image sample, compress the image sample to get the actual bit rate R target ;

[0093] S103: The actual bit rate R target As the target bit rate, it is input into the quantization parameter estimation network QSE together with the image sample, and the estimated quantization parameter (l real ,s real );

[0094] S104: Calculate the target quantization parameter (l target ,s target ) and the estimated quantization parameter (l real ,s real ), through which the loss is back-propagated to optimize the quantization parameter estimation network QSE.

[0095] The absolute variance of the target quantization coefficient and the estimated quantization coefficient is used as the loss function L used by the quantization parameter estimation network QSE in the training stage, which is specifically:

[0096] L=|l target -l real |+|s target -s real |.

[0097] Based on the same inventive concept, the embodiment of the present invention also provides an image compression system based on variable bit rate allocation. The image compression system has the same or similar technical features as the above-mentioned image compression method; the same or similar technical features will not be described in detail below.

[0098] Reference Figure 7 As shown, the image compression system based on variable bit rate allocation includes:

[0099] An acquisition module, used for acquiring images and obtaining input images;

[0100] A feature encoding module is used to perform feature encoding on the image and extract image feature values;

[0101] A joint quantization module is used to jointly quantize the channel dimension and the spatial dimension of the image eigenvalue y to obtain a quantized eigenvalue;

[0102] an arithmetic encoder, used for encoding the quantized feature values ​​into bit stream data;

[0103] an arithmetic decoder, for performing arithmetic decoding on the received bit stream data;

[0104] A joint dequantization module is used to dequantize the eigenvalues ​​after arithmetic decoding to obtain dequantized image eigenvalues ​​y';

[0105] A feature reconstruction module is used to perform feature reconstruction on the inverse quantized image feature value y' to obtain a 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, wherein the channel dimension quantization module is used to realize channel dimension quantization of the image eigenvalue y, and the spatial dimension quantization module is used to realize spatial dimension quantization in combination with the spatial information of the image.

[0108] The embodiment of the present invention further provides a computer storable medium on which a computer program is stored. The computer program is processed and executed to implement the image compression method based on variable bit rate allocation in the above embodiment.

[0109] Experimental test:

[0110] Experimental platform or software: integrated development tools, such as Pycharm, Ubuntu operating system, and pytorch neural network framework.

[0111] Test dataset: Kodak dataset.

[0112] Test results:

[0113] Figure 8 The figure is a comparison chart of the results of the present invention and the existing image compression method during the experimental test. Among them, "only channel quantization scheme" means that only the channel dimension quantization is adopted in the continuous variable bit rate image compression model proposed by the present invention, "channel and space joint quantization scheme" means that the continuous variable bit rate image compression model proposed by the present invention considers the quantization of both channel and space dimensions at the same time, and "basic model" means that the "using discrete Gaussian mixture likelihood and attention module to learn image compression" proposed in the document [1] is adopted as the basic image compression model (referred to as "basic model"). The multiple points in each curve in the figure represent PSNRs under multiple bit rates. The present invention realizes multiple bit rates on the same model, while the "basic model" needs to train multiple models to realize multiple bit rates.

[0114] Reference Figure 7 As shown, the PSNR of the image compression performed by the present invention is not lower than that of the basic model, and the present invention realizes multiple bit rates on one image compression model, without the need to train multiple models, thus saving training costs.

[0115] The present invention is tested on a Kodak data set and obtains a bit rate error of less than 3%, while the image quality does not decrease.

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

[0117] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

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

Claims

1. An image compression method, system and storage medium based on variable bit rate allocation, characterized in that: include: Obtain the image to be compressed, input it into the trained image continuous variable bit rate image compression network, and output the image compression result; The trained image continuous variable bit rate image compression network is used to encode and decode the compressed image. The specific process includes: At the encoding end: use the feature extractor to extract features from the image and obtain the image feature value y; Using the joint quantization module, the image eigenvalue y is jointly quantized in the channel dimension and the spatial dimension to obtain the quantized eigenvalue; Using an arithmetic encoder, the quantized feature values ​​are encoded into bit stream data; On the decoding side: After the received bit stream data is decoded by the arithmetic decoder, the channel dimension and the spatial dimension are jointly dequantized to obtain the decoded image feature value y'; The decoded image feature value y' is subjected to feature reconstruction to obtain a reconstructed image.

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

3. The image compression method based on variable bit rate allocation according to claim 1, characterized in that: The trained quantization parameter estimation network QSE provides quantization parameters for the joint quantization of channel dimension and spatial dimension.

4. The image compression method based on variable bit rate allocation according to claim 1, characterized in that: The image eigenvalue y is jointly quantized in terms of channel dimension and spatial dimension. The specific process includes: Multiply the image eigenvalue y by the gain vector to obtain the channel quantization eigenvalue y qc , the vector gain is obtained according to the quantization parameter; The spatial information qs of the image is extracted using the hyper-prior network and the spatial estimation network; The channel quantized feature value y qc Divided by the spatial information qs, we get the spatial quantization eigenvalue y qs That is, the quantized eigenvalue.

5. The image compression method based on variable bit rate allocation according to claim 1, characterized in that: The loss function of the image continuous variable bit rate image compression network during training is: Loss=mse(Img,recoImg)+λR Among them, mse(.) represents the mean square error function, Img represents the input image, reconImg represents the reconstructed image, R represents the bit rate, λ represents the trade-off factor between bit rate and image quality, and each λ corresponds to a gain vector.

6. The image compression method based on variable bit rate allocation according to claim 1, characterized in that: The training process of the quantization parameter estimation network QSE is: Initialize the parameters of the quantization parameter estimation network QSE; Randomly select a target quantization parameter (l target ,s target ) and the image sample, compress the image sample to get the actual bit rate R target ; The actual bit rate R target As the target bit rate, it is input into the quantization parameter estimation network QSE together with the image sample, and the estimated quantization parameter (l real ,s real ); Calculate the target quantization parameter (l target ,s target ) and the estimated quantization parameter (l real ,s real ), through which the loss is back-propagated to optimize the quantization parameter estimation network QSE.

7. An image compression system based on variable bit rate allocation, characterized in that: The system includes: An acquisition module, used for acquiring images and obtaining input images; A feature encoding module is used to perform feature encoding on the image and extract image feature values; A joint quantization module is used to jointly quantize the channel dimension and the spatial dimension of the image eigenvalue y to obtain a quantized eigenvalue; an arithmetic coding module, used for encoding the quantized feature values ​​into bit stream data; An arithmetic decoding module, used for performing arithmetic decoding on the received bit stream data; A joint dequantization module is used to dequantize the eigenvalues ​​after arithmetic decoding to obtain dequantized image eigenvalues ​​y'; A feature reconstruction module is used to perform feature reconstruction on the inverse quantized image feature value y' to obtain a reconstructed image; The output module is used to output and display the final reconstructed image.

8. The image compression system based on variable bit rate allocation according to claim 7, characterized in that: The joint quantization module includes a channel dimension quantization module and a space dimension quantization module, wherein the channel dimension quantization module is used to realize channel dimension quantization of the image eigenvalue y, and the space dimension quantization module is used to realize space dimension quantization in combination with the spatial information of the image.

9. A computer storable medium having a computer program stored thereon, characterized in that: When the computer program is processed and executed, the method according to any one of claims 1 to 6 is implemented.

Images