CN114663536B - Image compression method and device - Google Patents

Abstract

The present invention provides an image compression method and device. The method includes: acquiring an image to be compressed; dividing the image to be compressed into multiple image blocks based on preprocessing rules, and inputting all the image blocks to be compressed into a pre-stored In the target encoder of , to obtain the first hidden variable; input the first hidden variable into the pre-stored entropy model to obtain the second hidden variable; input the second hidden variable into the pre-stored target decoder , to obtain compressed image blocks, and obtain compressed images according to the compressed image blocks; the method of the present invention introduces a Transformer module in the image compression task and uses a symmetrical processing architecture to encode and decode images, improving image compression efficiency.

Description

An image compression method and device

technical field

The invention belongs to the field of computer vision, in particular to an image compression method and device.

Background technique

Image compression is the application of data compression technology on digital images. The purpose of image compression is to reduce redundant information in image data, so as to store and transmit data efficiently, that is, to obtain the best image quality at a given bit rate or compression ratio. .

Existing technologies usually design decoders and encoders based on convolutional neural networks to perform image compression tasks, but the image compression process based on convolutional neural networks cannot capture the semantic information of images, and the use of spatial redundancy information of images makes the image compression based on Global attention mechanisms perform poorly in image compression tasks, resulting in inefficient image compression.

Contents of the invention

An image compression method and device provided by the present invention are used to solve the rate distortion of the image compression caused by the inability to capture the semantic information of the image when the decoder and encoder are designed based on the convolutional neural network to perform the image compression task in the prior art The defect of poor performance improves the compression efficiency of images.

The invention provides an image compression method, the method comprising:

Obtaining an image to be compressed; dividing the image to be compressed into a plurality of image blocks based on preprocessing rules, and inputting all the image blocks to be compressed into a pre-stored target encoder to obtain a first latent variable, wherein the The target encoder includes a linear embedding layer module, a Transformer module and a block merging module; the first hidden variable is input into a pre-stored entropy model to obtain a second hidden variable; the second hidden variable is input into a pre-stored In the target decoder, the compressed image block is obtained, and the compressed image is obtained according to the compressed image block, wherein the target decoder includes a de-embedding layer module, the Transformer module and a block splitting module.

According to an image compression method provided by the present invention, the method further includes:

Input the first hidden variable into the entropy model, obtain the mean and variance of each element in the first hidden variable, and simulate the first hidden variable according to the mean and variance of each element normal distribution, to obtain a probability distribution function; perform arithmetic coding on the first hidden variable based on the probability distribution function, to obtain a target bit stream; perform arithmetic decoding on the target bit stream based on the probability distribution function, to obtain Obtaining a third hidden variable; obtaining a quantized residual loss of the third hidden variable through the entropy model, and acquiring the second hidden variable based on the third hidden variable and the quantized residual loss.

The global loss L is calculated using the following formula:

L=R+λD

Among them, λ is a hyperparameter, R is the size of the compressed bit stream, and D is a distortion item; the target image compression model is obtained according to the global loss.

Train the image compression model based on the BP algorithm, and adjust the bit stream size R and the distortion item D to reduce the global loss L to obtain the target hyperparameters; train the image compression model according to the target hyperparameters to obtain Image compression model.

The image to be compressed is subjected to normalization processing, and the processed image is equally divided into a plurality of image blocks according to a fixed division area.

The Transformer module includes a window-based attention layer, a multilayer perceptron, and a normalization layer.

The present invention also provides an image compression device, which includes:

An image acquisition module, configured to acquire an image to be compressed; a decoding module, configured to divide the image to be compressed into multiple image blocks based on preprocessing rules, and input all the image blocks to be compressed into a pre-stored target encoder , to obtain the first hidden variable; wherein, the target encoder includes a linear embedding layer module, a Transformer module, and a block merging module; a conversion module is used to input the first hidden variable into a pre-stored entropy model to obtain The second latent variable; a decoding module, configured to input the second latent variable into a pre-stored target decoder to obtain compressed image blocks, and obtain compressed images according to the compressed image blocks; wherein , the target decoder includes a de-embedding layer module, the Transformer module and a block splitting module.

The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the program, it realizes any one of the image compression methods described above. A step of.

The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the above-mentioned image compression methods are realized.

An image compression method and device provided by the present invention first obtains the image to be compressed; then divides the image to be compressed into multiple image blocks based on preprocessing rules, and inputs all the image blocks to be compressed to the pre-stored target In the encoder, to obtain the first hidden variable; then input the first hidden variable into the pre-stored entropy model to obtain the second hidden variable; finally, input the second hidden variable into the pre-stored target decoder , to obtain compressed image blocks, and obtain compressed images according to the compressed image blocks; the method of the present invention introduces a Transformer module in the image compression task and uses a symmetrical processing architecture to encode and decode images, improving image compression efficiency.

Description of drawings

In order to more clearly illustrate the present invention or the technical solutions in the prior art, the accompanying drawings that need to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings in the following description are the present invention. For some embodiments of the invention, those skilled in the art can also obtain other drawings based on these drawings without creative effort.

Fig. 1 is a schematic flow chart of an image compression method provided by an embodiment of the present invention;

Fig. 2 is a schematic structural diagram of a target encoder provided by an embodiment of the present invention;

Fig. 3 is a schematic structural diagram of a target decoder provided by an embodiment of the present invention;

FIG. 4 is a schematic flow diagram of obtaining a second hidden variable provided by another embodiment of the present invention;

Fig. 5 is a schematic structural diagram of a Transformer module provided by another embodiment of the present invention;

FIG. 6 is a schematic structural diagram of an image compression device provided by an embodiment of the present invention;

Fig. 7 is a schematic structural diagram of an electronic device provided by the present invention.

detailed description

In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention , but not all examples. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

The image compression method provided by the embodiment of the present invention is described below in conjunction with FIG. 1, including:

Step 101, acquire an image to be compressed.

It can be understood that image compression is the application of data compression technology on digital images. Its purpose is to reduce redundant information in image data and store and transmit data in a more efficient format. When the amount of image data is too large, image information Therefore, it is necessary to perform image compression on the image to be processed, so that the compressed image can be used effectively; this embodiment randomly extracts 1 million RGB color images from the open source database as the image to be processed To compress the image, the required image data to be compressed can also be obtained from other data acquisition platforms.

Step 102: Divide the image to be compressed into multiple image blocks based on preprocessing rules, and input all the image blocks to be compressed into a pre-stored target encoder to obtain a first latent variable; wherein, the target The encoder consists of a linear embedding layer module, a Transformer module, and a block merging module.

It can be understood that since redundant information in the image to be compressed needs to be removed to obtain the compressed image, this embodiment uses image processing software to normalize the acquired RGB color image to be compressed, and unifies the size; then each image is divided into multiple fixed-size image blocks according to a certain order, and the image block sequence is obtained; finally, the image block sequence is input into the target encoder for encoding processing. When the encoder is training, it can The image block is mapped to the parameters of the probability distribution subject to the hidden variable, and then the probability distribution parameter is sampled to obtain the first hidden variable, which is quantized by the feature output by the image block after the target encoder training obtained after rounding. As shown in Figure 2: the main structure of the target encoder includes a linear embedding layer module, a Transformer module, and a block merging module; among them, the linear embedding layer is composed of a multi-layer perceptron (MLP), and the linear embedding layer can be used for each image in the image. The pixel channel data is linearly transformed. The Transformer module is composed of a window-based attention layer, a multi-layer perceptron and a normalization layer. The number of Transformer modules in each layer is 2, 2, 6 and 2 respectively. The block merging module consists of multiple A layer perceptron and a normalization layer are used to downsample image features.

Step 103: Input the first hidden variable into a pre-stored entropy model to obtain a second hidden variable.

It is understandable that since the first hidden variable is obtained by quantizing and rounding the output features of the image block after the target encoder is trained, the rounding method is rounded, which will bring quantization loss. In order to make up for For the quantization loss of the image block passing through the target encoder, this embodiment inputs the obtained first hidden variable into the pre-stored entropy model to obtain the second hidden variable, and the second hidden variable makes up for the quantization loss of the first hidden variable Specifically, this embodiment builds a channel-by-channel autoregressive entropy model based on a convolutional neural network, and then inputs the first hidden variable into the entropy model to obtain a probability distribution function of each element in the first hidden variable, and then based on the probability distribution function Arithmetic coding and decoding are used to obtain the second hidden variable respectively.

Step 104: Input the second hidden variable into a pre-stored target decoder to obtain a compressed image block, and obtain a compressed image according to the compressed image block; wherein, the target decoder includes De-embedding layer module, the Transformer module and block splitting module.

It can be understood that decoding is the inverse process of encoding, and the obtained second latent variable is input into the target decoder for analysis to obtain the compressed image blocks corresponding to the RGB image, and then according to the ordering of the image blocks in the above-mentioned embodiment Information, re-splicing the compressed image blocks into a complete image; as shown in Figure 3: the main structure of the target decoder includes a de-embedding layer module, a Transformer module and a block splitting module; wherein, the de-embedding layer is composed of a multi-layer perceptron (MLP), the Transformer module is composed of a window-based attention layer, a multi-layer perceptron and a normalization layer, and the block merging module is composed of a multi-layer perceptron and a normalization layer for upsampling of image features .

The method of the invention introduces a Transformer module into the image compression task and adopts a symmetrical processing framework to encode and decode images, thereby improving image compression efficiency.

Optionally, input the first hidden variable into the entropy model, obtain the mean value and variance of each element in the first hidden variable, and simulate the first hidden variable according to the mean value and variance of each element a normal distribution of hidden variables to obtain a probability distribution function; arithmetically encode the first hidden variable based on the probability distribution function to obtain a target bit stream; perform the target bit stream on the basis of the probability distribution function Arithmetic decoding to obtain a third hidden variable; obtaining a quantized residual loss of the third hidden variable through the entropy model, and obtaining the second hidden variable based on the third hidden variable and the quantized residual loss .

接着将第一隐变量

输入到基于卷积神经网络构建逐通道自回归熵模型中，以获取

中各元素对应的均值μ和方差σ，并根据并由均值μ和方差σ模拟每个元素的高斯分布，得到其概率分布函数，再由算术编码根据概率分布函数将量化后的第一隐变量

无损压缩为比特流，比特流为二进制字符串；最后，由算术解码根据概率分布函数将比特流解析为量化后的第三隐变量y，同时获取逐通道自回归熵模型预测隐变量y的量化残差损失r，由损失公式

可以获取第二隐变量

Specifically, as shown in FIG. 4 , in this embodiment, the image to be compressed is first normalized, and each image is divided into a plurality of image blocks; then these image blocks are arranged in a certain order to obtain an image block sequence , and input the sequence of image blocks into the target encoder based on the Transformer architecture for training to obtain the first latent variable

Then the first latent variable

Input to the channel-by-channel autoregressive entropy model based on the convolutional neural network to obtain

The mean value μ and variance σ corresponding to each element in , and the Gaussian distribution of each element is simulated by the mean value μ and variance σ to obtain its probability distribution function, and then the quantized first hidden variable is calculated by arithmetic coding according to the probability distribution function

Lossless compression into a bit stream, the bit stream is a binary string; finally, the bit stream is parsed into the quantized third hidden variable y by the arithmetic decoding according to the probability distribution function, and the quantization of the hidden variable y predicted by the channel-by-channel autoregressive entropy model is obtained at the same time Residual loss r, by the loss formula

The second hidden variable can be obtained

It should be noted that the third hidden variable is obtained from the first hidden variable through lossless arithmetic coding and decoding, and the values of the two are the same.

This embodiment provides a method of inputting the hidden variables output by the encoder into the entropy model, and obtaining new hidden variables through arithmetic coding and decoding techniques respectively, which makes up for the quantized residuals lost by the quantized hidden variables, and can reduce Distortion of the image.

Optionally, use the following formula to calculate the global loss L:

L=R+λD

Among them, λ is a hyperparameter, R is the size of the compressed bit stream, and D is a distortion item; the target image compression model is obtained according to the global loss.

其中

是熵模型中的超隐变量，是一种先验信息，用于求取

的均值和方差；

表示在先验信息

条件下，

的正态分布概率值，

表示

的条件熵，

表示先验信息

的正态分布概率值，

表示先验信息

的信息熵，E_x～px[·]表示图像x在其正态分布px下的表达式内的期望值，x表示待压缩图像，

为所述压缩后的图像；D为失真项，表示所述压缩后的图像和所述待压缩图像之间的差异大小，

表示原图像 x和重建图像

之间的失真，常用评估标准为均方误差MSE；本实施例通过计算图像压缩模型的全局损失L来确定合适的超参数λ，并利用目标超参数λ来获取目标图像压缩模型。Specifically, λ is a hyperparameter, which is used to control the compressed bit rate and compression quality to generate a rate-distortion curve. The calculation formula of R is:

in

It is an ultra-hidden variable in the entropy model, and it is a kind of prior information, which is used to obtain

The mean and variance of ;

represented in the prior information

condition,

The normal distribution probability value of ,

express

The conditional entropy of

represent prior information

The normal distribution probability value of ,

represent prior information

The information entropy of , E _x～px [ ] represents the expected value of the image x in the expression under its normal distribution px, x represents the image to be compressed,

is the compressed image; D is a distortion item, representing the difference between the compressed image and the image to be compressed,

Represents the original image x and the reconstructed image

The common evaluation standard is the mean square error MSE; this embodiment determines the appropriate hyperparameter λ by calculating the global loss L of the image compression model, and uses the target hyperparameter λ to obtain the target image compression model.

The method described in this embodiment provides a calculation method for the global loss L of the image compression model, and the required hyperparameter λ is determined by reducing L, so as to obtain image compression models with different bit rates or reconstruction quality requirements.

Optionally, train an image compression model based on the BP algorithm, and adjust the bit stream size R and the distortion item D to reduce the global loss l, so as to obtain target hyperparameters; train the image compression model according to the target hyperparameters model to get the image compression model.

Specifically, this embodiment uses the backpropagation algorithm and the stochastic gradient descent method to reduce the overall prediction error l to train the image compression model, and obtain the final image compression model through multiple iterations of training; for example, the hyperparameter λ set in this embodiment The values are {0.0018, 0.0035, 0.0067, 0.0130, 0.025, 0.0483} to obtain multiple image compression models suitable for different scenarios; for different bit rates or reconstruction quality requirements, different image compression models are selected, and the reconstruction quality For scenes with high requirements, choose a larger λ such as 0.0483, and for scenes that require a lower bit rate, select a lower λ such as 0.0018.

This embodiment provides a method for determining a hyperparameter λ by using a BP training algorithm to meet image compression models with different bit rates or reconstruction quality requirements.

Optionally, the image to be compressed is subjected to normalization processing, and the processed image is equally divided into a plurality of image blocks according to a fixed division area.

Specifically, in this embodiment, the acquired 1 million RGB color images are used as images to be compressed, and after the images to be compressed are normalized, RGB color images of the same size are obtained, and the dimension of each image is 768×512× 3. Divide into image blocks with a unit size of 2 in the length (768) and width (512) dimensions, and then send these 98304 image blocks to the target encoder for training.

This embodiment provides a method for data preprocessing and image block division, which provides convenience for the subsequent process of inputting image blocks into a target encoder to obtain latent variables.

Optionally, the Transformer module includes a window-based attention layer, a multilayer perceptron, and a normalization layer.

As shown in Figure 5, the Transformer module is composed of a window-based (Windows) attention layer (W-MSA, SW-MSA), a multi-layer perceptron (MLP) and a normalization layer (LN); where, W- MSA and SW-MSA are used in pairs. Assuming that layer L uses W-MSA, then layer L+1 uses SW-MSA. According to the comparison of the left and right pictures, it can be found that the window has been shifted, and the shifted window enables the communication between the previous adjacent windows, which solves the problem that information exchange cannot be carried out between different windows. In this embodiment, the Transformer The attention mechanism is introduced into the module to form a window-based attention mechanism layer.

The method described in this embodiment provides a composition structure of the Transformer module, and introduces an attention mechanism in each window to make it pay more attention to the local structural information of the input image, that is, the correlation between spatially adjacent elements, to overcome The difficulty of missing semantic information in image compression improves the distortion rate performance of image compression.

An image compression device provided by an embodiment of the present invention is described in conjunction with FIG. 6 . An image compression device described below and an image compression method described above may be referred to in correspondence.

An image compression device provided by the present invention, said device comprising:

An image acquisition module 601, configured to acquire an image to be compressed; a decoding module 602, configured to divide the image to be compressed into multiple image blocks based on preprocessing rules, and input all the image blocks to be compressed into a prestored target code In the encoder, to obtain the first hidden variable; wherein, the target encoder includes a linear embedding layer module, a Transformer module and a block merging module; the conversion module 603 is used to input the first hidden variable into a pre-stored entropy model , to obtain the second latent variable; the decoding module 604 is used to input the second latent variable into the pre-stored target decoder to obtain the compressed image block, and obtain the compressed image block according to the compressed image block. image; wherein, the target decoder includes a de-embedding layer module, the Transformer module and a block splitting module.

An image compression device provided by the present invention first obtains the image to be compressed through the image acquisition module 601; then divides the image to be compressed into multiple image blocks through the decoding module 602 based on preprocessing rules, and divides all the images to be compressed The image block is input into the pre-stored target encoder to obtain the first hidden variable; then the first hidden variable is input into the pre-stored entropy model through the conversion module 603 to obtain the second hidden variable; finally through the decoding module 604 Input the second latent variable into the pre-stored target decoder to obtain the compressed image block, and obtain the compressed image according to the compressed image block; the device of the present invention introduces in the image compression task The Transformer module uses a symmetrical processing architecture to encode and decode images, which improves image compression efficiency.

FIG. 7 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 7, the electronic device may include: a processor (processor) 710, a communication interface (Communications Interface) 720, a memory (memory) 730, and a communication bus 740, Wherein, the processor 710 , the communication interface 720 , and the memory 730 communicate with each other through the communication bus 740 . The processor 710 can invoke logic instructions in the memory 730 to execute an image compression method, the method comprising: acquiring an image to be compressed; dividing the image to be compressed into multiple image blocks based on preprocessing rules, and dividing all The image block to be compressed is input into a pre-stored target encoder to obtain a first hidden variable, wherein the target encoder includes a linear embedding layer module, a Transformer module and a block merging module; the first hidden variable is input into In the pre-stored entropy model, to obtain the second hidden variable; input the second hidden variable into the pre-stored target decoder to obtain the compressed image block, and obtain the compressed image block according to the compressed image block image, wherein the target decoder includes a de-embedding layer module, the Transformer module and a block splitting module.

In addition, the above-mentioned logic instructions in the memory 730 may be implemented in the form of software functional units and when sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several The instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes. .

The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the image compression method provided by the above-mentioned methods is implemented. The method includes: obtaining the Compressing an image; dividing the image to be compressed into a plurality of image blocks based on preprocessing rules, and inputting all the image blocks to be compressed into a pre-stored target encoder to obtain a first latent variable, wherein the target The encoder includes a linear embedding layer module, a Transformer module and a block merging module; the first hidden variable is input into a pre-stored entropy model to obtain a second hidden variable; the second hidden variable is input into a pre-stored target decoding A decoder to obtain compressed image blocks, and obtain compressed images according to the compressed image blocks, wherein the target decoder includes a de-embedding layer module, the Transformer module and a block splitting module.

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without any creative efforts.

Through the above description of the implementations, those skilled in the art can clearly understand that each implementation can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, disk , CD, etc., including several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (7)

1. An image compression method, comprising:

acquiring an image to be compressed;

dividing the image to be compressed into a plurality of image blocks based on a preprocessing rule, and inputting all the image blocks to be compressed into a pre-stored target encoder to obtain a first hidden variable, wherein the target encoder comprises a linear embedding layer module, a Transformer module and a block merging module; normalizing the image to be compressed, and equally dividing the processed image into a plurality of image blocks according to a fixed division area, wherein the image blocks have the same size;

inputting the first hidden variable into a pre-stored entropy model to obtain a second hidden variable;

inputting the second hidden variable into a pre-stored target decoder to obtain a compressed image block, and obtaining a compressed image according to the compressed image block, wherein the target decoder comprises an embedding layer removing module, a transform module and a block splitting module;

after the compressed image block acquires a compressed image, the method further comprises:

the global loss L is calculated using the following formula:

L＝R+λD；

the method comprises the following steps that A is a hyper-parameter, and the A is used for obtaining a rate-distortion curve by controlling the bit rate and compression quality of compression;

r is the bit stream size obtained by compression, and the calculation formula of R is as follows:

wherein x is the image to be compressed,

in order to be a first hidden variable of said first type,

for a hyper-hidden variable in the entropy model,

for obtaining

Mean and variance of;

for prior information

Under the condition of

The probability value of the normal distribution of (c),

is composed of

The conditional entropy of (a) is,

as a priori information

The probability value of the normal distribution of (c),

as a priori information

Information entropy of (E) _x～px [·]Is the expected value of x within its expression under its normal distribution px;

d is a distortion term and is used for representing the difference between the compressed image and the image to be compressed, and the calculation formula of D is as follows:

wherein,

for the purpose of the compressed image, the image is,

denotes x and

distortion between;

and acquiring a target image compression model according to the global loss.

2. The image compression method according to claim 1, wherein inputting the first hidden variable into a pre-stored entropy model to obtain a second hidden variable specifically comprises:

inputting the first hidden variable into the entropy model, acquiring the mean value and the variance of each element in the first hidden variable, and simulating the normal distribution of the first hidden variable according to the mean value and the variance of each element to acquire a probability distribution function;

performing arithmetic coding on the first hidden variable based on the probability distribution function to obtain a target bit stream;

arithmetically decoding the target bit stream based on the probability distribution function to obtain a third hidden variable;

and obtaining the quantized residual loss of the third hidden variable through the entropy model, and obtaining the second hidden variable based on the third hidden variable and the quantized residual loss.

3. The image compression method of claim 1, wherein obtaining a target image compression model based on the global loss comprises:

training an image compression model based on a BP algorithm, and adjusting the bit stream size R and the distortion item D to reduce the global loss L so as to obtain a target hyper-parameter;

and training the image compression model according to the target hyper-parameter to obtain the image compression model.

4. The method of image compression of claim 1, wherein the transform module comprises a window-based attention layer, a multi-layer perceptron, and a normalization layer.

5. An image compression apparatus, characterized in that the apparatus comprises:

the image acquisition module is used for acquiring an image to be compressed;

the decoding module is used for dividing the image to be compressed into a plurality of image blocks based on a preprocessing rule and inputting all the image blocks to be compressed into a pre-stored target encoder to obtain a first hidden variable, wherein the target encoder comprises a linear embedding layer module, a Transformer module and a block merging module;

the decoding module is specifically configured to perform normalization processing on the image to be compressed, and equally divide the processed image into a plurality of image blocks according to a fixed division area, where the image blocks are the same in size;

the conversion module is used for inputting the first hidden variable into a pre-stored entropy model so as to obtain a second hidden variable;

the decoding module is used for inputting the second hidden variable into a pre-stored target decoder to obtain a compressed image block and obtaining a compressed image according to the compressed image block, wherein the target decoder comprises a de-embedding layer module, a Transformer module and a block splitting module;

the decoding module is further configured to, after the compressed image block obtains a compressed image, calculate a global loss L using the following formula:

L＝R+λD；

the method comprises the following steps that A is a hyper-parameter, and the A is used for obtaining a rate-distortion curve by controlling the bit rate and compression quality of compression;

r is the bit stream size obtained by compression, and the calculation formula of R is as follows:

wherein,

for the hyper-hidden variables in the entropy model,

for obtaining

Mean and variance of;

for prior information

Under the condition of

The probability value of the normal distribution of (c),

is composed of

The conditional entropy of (a) is,

is a priori information

The probability value of the normal distribution of (b),

as a priori information

Information entropy of (E) _x～px [·]Is the expected value of the image x within its expression under the normal distribution px, x being the image to be compressed,

the compressed image is obtained;

d is a distortion term and is used for representing the difference between the compressed image and the image to be compressed, and the calculation formula of D is as follows:

wherein,

denotes x and

distortion between;

and acquiring a target image compression model according to the global loss.

6. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the steps of the image compression method according to any of claims 1 to 4 are implemented when the processor executes the program.

7. A non-transitory computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the image compression method according to any one of claims 1 to 4.

Images