A decomposition model inspired image compression artifact removal method

Abstract

The application discloses a decomposed model inspired image compression artifact removal method, belongs to the field of image quality enhancement, and is committed to solving the low-quality compressed image recovery problem. First, an image compression artifact removal optimization model based on structure texture decomposition is constructed. Secondly, the model is expanded into an interpretable image compression artifact removal network STNet. The network mainly comprises a structure feature extraction module and a texture feature extraction module. The two modules are alternately iterated and updated and optimize the structure feature and the texture feature of the image in a progressive enhancement manner, so that the structure feature and the texture feature can be complementary. Finally, the structure image and the texture image output by the structure feature extraction module and the texture feature extraction module are added pixel by pixel to obtain the output enhanced image. The reconstruction performance of the method is better than that of multiple existing image compression artifact removal methods in terms of the image quality objective evaluation indexes PSNR, SSIM and PSNR-B.

Description

[0001] This invention belongs to the field of image quality enhancement, specifically relating to an image compression artifact removal method inspired by a decomposition model. Background Technology

[0002] With the popularity of multimedia social networking apps like WeChat, Facebook, TikTok, and Instagram for posting photos and videos, the number of photos posted online has increased dramatically, making it difficult for existing network communication equipment to handle the massive data transmission. To address this issue, lossy image compression standards and video compression standards such as JPEG, AVC, and HEVC are typically used to encode images, reducing the number of bits transmitted to alleviate the ever-increasing demand for network bandwidth caused by the rapid increase in data transmission. It's worth noting that JPEG is still widely used on smartphones and desktop computers. The main steps of JPEG lossy compression include: first, dividing the input image into several 8×8 blocks and performing a DCT transform on each 8×8 block to obtain DCT coefficients; then, quantizing the DCT coefficients of each block using a non-uniform quantization table to reduce storage space; finally, lossless encoding of the quantized coefficients. However, considering the irreversible nature of quantization during image compression, and the fact that coarse quantization can lead to the loss of high-frequency details and pixel discontinuities caused by block processing, compressed images often exhibit noticeable visual artifacts, such as block artifacts (discontinuous blocks), ringing artifacts, and color distortion artifacts. Since these artifacts not only affect the user's visual experience but also hinder subsequent image processing and various computer vision tasks, more and more researchers are attempting to address this problem using image compression artifact removal techniques. Many advanced image enhancement methods have been proposed to improve the quality of JPEG compressed images. Generally, image compression artifact removal methods can be classified into three types: image filtering-based methods, prior regularization-based optimization methods, and deep learning-based methods.

[0003] In early image filtering-based methods, researchers typically designed a filter operator manually in the pixel or frequency domain to remove compression artifacts. For example, Reeve et al. applied Gaussian filtering to pixels around the boundaries of coded blocks in the pixel domain to smooth block artifacts. Buades et al. proposed predicting the filter output by weighted averaging of neighboring pixels. Chen et al. applied a low-pass filter to the DCT coefficients of adjacent coded blocks in the frequency domain to remove high-frequency noise. Kauff et al. proposed a shape-adaptive discrete cosine transform algorithm that adaptively determines the optimal shape for each block. However, since most compression artifacts are introduced by highly nonlinear, irreversible quantization operations, and these filtering algorithms are complex and lack adaptability, this hand-designed approach cannot effectively address the problem of compression artifact removal in complex texture regions. To overcome these problems, many researchers have treated image compression artifact removal as an optimization problem and attempted to incorporate various prior constraints into the optimization model. For example, some typical compression artifact removal methods utilize sparse representation, low-rank priors, and nonlocal similarity to construct regularization constraints. Although this prior regularization optimization method can achieve better artifact removal results, most prior regularization-based algorithms require complex iterative optimization, which often makes these methods time-consuming.

[0004] In recent years, with the continuous development of deep learning, deep learning-based methods can directly learn end-to-end nonlinear mappings from compressed images to corresponding clean images. For example, Dong et al. proposed a four-layer shallow network, ARCNN, to learn the mapping function. Inspired by deep residual learning, Zhang et al. proposed a layered convolutional neural network, DnCNN, based on residual learning. Many researchers have also explored deep learning models based on dual domains. However, deep learning-based methods still have some drawbacks. For example, as the network becomes deeper, the roles of internal modules become difficult to understand and analyze, resulting in a lack of interpretability. Unlike black-box deep learning networks, Deep Unfolded Networks (DUNs) have received increasing attention in the field of image processing due to their excellent interpretability of network topology. These methods typically utilize degradation models to construct optimization models and then unfold them into interpretable networks. Deep Unfolded Networks (DUNs) can effectively combine the good interpretability of model-based methods with the powerful learning capabilities of CNNs. For example, Fu et al. used sparse convolutional dictionary learning to build a JPEG artifact removal model, solved the model through an iterative optimization algorithm, and unfolded it into an interpretable network. Although existing deep unrolling network (DUN) methods consider the prior properties of images and utilize these properties to optimize the unrolling network, these networks do not fully utilize the structural and textural complementarity of the images themselves, making it difficult for these methods to improve the quality of compressed images. Summary of the Invention

[0005] This invention presents an image compression artifact removal method inspired by a decomposition model, aiming to solve the problem of compressed image restoration. The network comprises two modules: a structural feature extraction module (SFEM) and a texture feature extraction module (TFEM). By reconstructing the initialized structural and texture images through these two modules, the structure and details of the reconstructed image can be made clearer.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a decomposition model-inspired image compression artifact removal method, comprising the following steps:

[0007] Step 1) Construct training data for the image compression artifact removal model;

[0008] All images in the DIV2K dataset are compressed using the JPEG standard encoder to obtain a compressed image training dataset that pairs with the original input images.

[0009] Step 2) Construct an image compression artifact removal optimization model based on structure texture decomposition;

[0010] First, the JPEG standard encoder compressed image is filtered using the L0 gradient minimization smoothing method to obtain a structure map. Subtracting the structure map from the compressed image yields the texture map. Next, the structure map and texture map are passed through the structure initialization subnetwork ISN-T and the texture structure initialization subnetwork ISN-T, respectively, to obtain initial feature maps. Then, the maximum a posteriori probability model is used to mathematically model the relationship between the input image X, the structure map S, and the texture map T. Finally, this probability model is transformed into an image compression artifact removal optimization model based on structure-texture decomposition.

[0011] Step 3) Expand the image compression artifact removal optimization model based on structure texture decomposition in Step 2) into an interpretable network;

[0012] The proximal gradient descent algorithm is used to solve the image compression artifact removal optimization model obtained in step 2), and it is expanded into an interpretable image compression artifact removal network STNet, which includes an initialization subnetwork and a depth reconstruction subnetwork.

[0013] Step 4) Construct the initialization subnetwork

[0014] The structure map and texture map obtained in step 2) are subjected to image smoothing filtering using 3×3 convolution kernels with all weights of 1 to obtain preliminary denoised structure map and texture map. Then, the preliminary denoised structure map and texture map are stacked along the channel dimension with the structure map and texture map obtained in step 2) and fed into the initialization sub-network. Next, the first three feature maps are extracted from the output features of the structure map initialization sub-network and the texture map initialization sub-network, respectively, as the initialization S0 and T0. The remaining feature maps... and This information will be sent to the next stage as supplementary information.

[0015] Step 5) Constructing a deep reconstruction subnetwork

[0016] The structural feature extraction module and texture feature extraction module of the deep reconstruction subnetwork obtained in step 3) are used to continuously update the structural features and texture features in an alternating iterative manner;

[0017] Step 6) Loss Function

[0018] The network's learning is supervised by using an L1-norm-constrained reconstruction loss, with images enhanced by the interpretable image compression artifact removal network STNet and the original lossless image as inputs to the loss function.

[0019] Furthermore, the specific method for constructing the training data for the compressed artifact removal model in step 1) is as follows:

[0020] First, the DIV2K dataset containing 1000 images is compressed using a standard JPEG encoder. 800 images are used for STNet training, 100 for STNet validation, and 100 for STNet testing. Each image in the 800-image dataset is then randomly cropped into 128×128 image patches, which are used as the training dataset.

[0021] Furthermore, the calculation process of the image compression artifact removal optimization model based on structure texture decomposition in step 2) is as follows:

[0022] The structure map is obtained by filtering the input compressed image X using the L0 gradient minimization smoothing method. The prior distribution of the low-quality compressed image Y is P(Y)∝e -αg(Y) The implicit prior distribution of X guided by the low-quality compressed graph Y is P(X|Y)∝e -μΩ(X|Y) Where g(Y) and Ω(X|Y) are energy functions of X and Y, and α and μ are two hyperparameters. The following equation can be obtained using Bayes' theorem:

[0023]

[0024] Here, P(Y) is the probability distribution of Y, independent of X, and P(Y|X) is a marginal distribution of Y. The desired X is decomposed into S and T, X = S + T, where S is the structure graph and T is the texture graph, and the prior distributions of S and T are... Assuming S and T are independent, substituting X = S + T into equation (1) yields the following equation:

[0025]

[0026] Using the maximum a posteriori probability estimation (MAP) formula to solve for the posterior probability P(S, T|Y), the above formula can be transformed into a maximization optimization problem:

[0027]

[0028] By performing a negative logarithmic transformation on equation (3), the above equation can be transformed into an energy minimization model, which can be expressed as:

[0029]

[0030] This model is an image compression artifact removal optimization model based on structure texture decomposition. The first term of formula (4) is the data term, and the second and third terms of formula (4) are two regularization constraint terms.

[0031] Furthermore, the specific steps of step 3) are as follows: First, the optimization problem of formula (4) is transformed into two sub-optimization problems using the proximal gradient descent algorithm; then, the gradient descent algorithm and the proximal mapping algorithm are used to optimize these two sub-problems respectively; finally, they are iteratively optimized in sequence, and the optimization solution formulas for these two sub-problems can be written as:

[0032] s (k) =S (k-1) -η S (S (k-1 )-(YT (k-1) (5)

[0033]

[0034] t (k) =T (k-1) -η T (T (k-1) -(YS k (7)

[0035]

[0036] Where k represents the index value of the iteration, ηS and η T The learnable step size parameters for updating S and T are respectively. and These are the proximal operators for the implicit priors f1(S) and f2(T), respectively. By iteratively solving equations (6) and (8), the iteratively updated structure graph S can be obtained. k and texture map T k Finally, the last updated S K and T K The enhanced image X is obtained by adding elements one by one.

[0037] Furthermore, in step 5), the deep reconstruction subnetwork consists of K stages, each stage consisting of a structural feature extraction module (SFEM) and a texture feature extraction module (TFEM).

[0038] The structural feature extraction module (SFEM) processes the initial structural map and texture map input to this module using the gradient descent algorithm shown in formula (5) to obtain a 3-channel image. For the approximate mapping of the near-end mapping algorithm shown in formula (6), the 3-channel image obtained by the gradient descent algorithm of formula (5) is first stacked with the remaining feature map of the previous stage in the channel dimension, and then processed through five residual blocks, keeping the number of channels unchanged. Finally, the first 3 feature maps of the output feature map with 67 channels are taken as the update, and the remaining 64 feature maps are used as the update. This information will be entered into the next stage as supplementary information.

[0039] Texture feature extraction module TFEM: The initial texture map input to this module and the structure map updated by the structure feature enhancement module SFEM are used to obtain a 3-channel image through the gradient descent algorithm shown in formula (7). The near-end mapping algorithm in formula (8) is composed of 5 concatenated residual blocks. The first 3 feature maps are used as the updated T, and the remaining 64 feature maps are used as the updated T. This information will be entered into the next stage as supplementary information.

[0040] Furthermore, in step 6), the image enhanced by STNet and the original distortion-free image X are compared. GT As input to the loss function, the reconstruction loss constrained by the L1 norm is used to supervise the learning of the interpretable image compression artifact removal network STNet. The total loss function can be written as:

[0041]

[0042] Where N is the image X GT The total number of pixels, K is the total number of stages in STNet, X0 is the initial augmented image, X i This is the enhanced image for the i-th stage.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This invention first proposes an image compression artifact removal method inspired by a decomposition model, then selects the proximal gradient descent algorithm to solve the model, and finally expands the algorithm model into an image compression artifact removal optimization model based on structure texture decomposition. Compared with other compression artifact removal methods based on black-box networks, this invention has the interpretability of network topology.

[0045] 2. The deep reconstruction subnetwork of the present invention has two sub-modules: the structural feature extraction module SFEM and the texture feature extraction module TFEM, which are used to optimize the structural features and texture features of the image, respectively. These two modules are updated alternately and progressively to optimize the structural features and texture features of the image, so that the structural features and texture features can complement each other, ultimately resulting in a higher quality reconstructed image. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of STNet, an interpretable image compression artifact removal network of the present invention.

[0047] Figure 2 This is a schematic diagram illustrating the alternating iteration of the structural feature extraction module and the texture feature extraction module of the present invention.

[0048] Figure 3 This is a schematic diagram of the network initialization for the structure and texture maps of this invention.

[0049] Figure 4 This is a schematic diagram of the structural feature enhancement module SFEM and the texture feature enhancement module TFEM in the kth stage of the present invention.

[0050] Figure 5 This is a schematic diagram of the proximal operator of the present invention.

[0051] Figure 6 This image shows a comparison of the visual quality of the method of this invention and three existing compression artifact removal methods on the "Barbara" image in the Classic5 dataset. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. To better understand the image enhancement method of the present invention, the network structure of the present invention will be described in detail first.

[0053] This invention provides an image compression artifact removal method inspired by a decomposition model, the network structure of which is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0054] Step 1) Construct training data for the image compression artifact removal model; compress all images in the DIV2K dataset using the JPEG standard encoder to obtain a compressed image training dataset that pairs with the original input image.

[0055] Step 2) Construct an image compression artifact removal optimization model based on structure texture decomposition;

[0056] First, the JPEG standard encoder compressed image is filtered using the L0 gradient minimization smoothing method to obtain the structure map. Subtracting the structure map from the compressed image yields the texture map. Next, a maximum a posteriori probability model is used to mathematically model the relationship between the input image X, the structure map S, and the texture map T. The specific process of mathematical modeling is as follows:

[0057] The input compressed image X is processed by L o The gradient minimization smoothing method is used to filter and obtain the structure graph; the prior distribution of the low-quality compressed graph Y is P(Y)∝e -αg(Y) The implicit prior distribution of X guided by the low-quality compressed graph Y is P(X|Y)∝e -μΩ(X|Y) Where g(Y) and Ω(X|Y) are energy functions related to X and Y, and α and μ are two hyperparameters. The following equation can be obtained using Bayes' theorem:

[0058]

[0059] Here, P(Y) is the probability distribution of Y, independent of X, and P(Y|X) is a marginal distribution of Y. We decompose the desired X into S and T, i.e., X = S + T, where S is the structure graph and T is the texture graph, and the prior distributions of S and T are... Assuming S and T are independent, substituting X = S + T into equation (1) yields the following equation:

[0060]

[0061] Here, the maximum a posteriori probability estimation (MAP) formula is used to solve for the posterior probability P(S, T|Y). Formula (2) can be transformed into the following maximization optimization problem:

[0062]

[0063] By performing a negative logarithmic transformation on equation (3), the above equation can be transformed into an energy minimization model, which can be expressed as:

[0064]

[0065] This model is an image compression artifact removal optimization model based on structured texture decomposition.

[0066] Step 3) Expand the image compression artifact removal optimization model based on structure texture decomposition into an interpretable network.

[0067] The optimized model obtained in step 2) is solved using the proximal gradient descent algorithm and expanded into an interpretable image compression artifact removal network STNet, which consists of an initialization subnetwork and a depth reconstruction subnetwork.

[0068] The optimization problem of formula (4) is transformed into two sub-optimization problems using the proximal gradient descent algorithm; then, the gradient descent algorithm and the proximal mapping algorithm are used to optimize these two sub-problems respectively; finally, the optimization is iteratively solved in sequence, and the optimization formulas for these two sub-problems can be written as:

[0069] s (k) =S (k-1) -η S (S (k-1) -(YT (k-1) (5)

[0070]

[0071] t (k) =T (k-1) -η T (T (k-1) -(YS k (7)

[0072]

[0073] Where k represents the index value of the iteration, η S and η T These are the learnable step size parameters for updating S and T, respectively. and These are the proximal operators corresponding to the implicit priors f1(S) and f2(T), respectively. By iteratively solving equations (6) and (8), the iteratively updated structure graph S can be obtained. k and texture map T k Finally, the last updated S K and T K The enhanced image X is obtained by adding elements one by one.

[0074] Step 4) Construct the initialization subnetwork

[0075] The structure map and texture map obtained in step 2) are respectively subjected to image smoothing filtering using 3×3 convolution kernels with all weights of 1 to obtain preliminary denoised structure map and texture map. Then, the preliminary denoised structure map and texture map are stacked with the structure map and texture map obtained in step 2) along the channel dimension and sent to the initialization module. The first 3 feature maps of the output of the structure map initialization module and the texture map initialization module are taken as the initialization S0 and T0, respectively, and the remaining feature maps are sent as supplementary information to the next stage.

[0076] Step 5) Constructing a deep reconstruction subnetwork

[0077] The deep reconstruction subnetwork is a network that expands Equations (6) and (8). In this algorithm, due to the two proximal operators... and Since regularization exists, f1(S) and f2(T) cannot be explicitly derived. Therefore, a deep convolutional neural network is used to learn and update s. (k) and t (k) The deep reconstruction subnetwork consists of K stages, each of which is composed of a structural feature enhancement module (SFEM) and a texture feature enhancement module (TFEM).

[0078] Structural Feature Extraction Module (SFEM): The network expanded by formulas (5) and (6) is called SFEM. This module is mainly used to update the structural features of the image. The initial structural map and texture map input to this module are processed by the gradient descent algorithm shown in formula (5) to obtain a 3-channel image. For the approximate mapping of the near-end mapping algorithm shown in formula (6), the 3-channel image obtained by the gradient descent algorithm of formula (5) is first stacked with the remaining feature map of the previous stage in the channel dimension, and then passed through five residual blocks, with the number of channels remaining unchanged. Finally, the first 3 feature maps of the output feature map with 67 channels are taken as the updated S, and the remaining 64 feature maps are used as the updated S. This information will be entered into the next stage as supplementary information.

[0079] Texture Feature Extraction Module (TFEM): The network expanded by formulas (7) and (8) is called TFEM. This module is mainly used to update the texture features of the image. The initial texture map input to this module and the structure map updated by the structure feature enhancement module (TFEM) are used to obtain a 3-channel image through the gradient descent algorithm shown in formula (7). The near-end mapping algorithm of formula (8) is composed of 5 residual blocks cascaded together. The first 3 feature maps are used as the updated T, and the remaining 64 feature maps are used as the updated T. This information will be entered into the next stage as supplementary information.

[0080] Step 6) Loss Function

[0081] The image obtained by fusing the structure map and texture map, along with the original lossless image, is used as input to the loss function. The reconstruction loss constrained by the L1 norm is used to supervise the learning of the interpretable deep learning image compression artifact removal network driven by the structure-texture decomposition task.

[0082] Network training parameter configuration

[0083] First, the DIV2K dataset containing 1000 images is compressed using a standard JPEG encoder. Here, 800 images are used for STNet training, 100 images for STNet validation, and 100 images for STNet testing. Each image in the 800-image dataset is cropped randomly into 128×128 image patches, and these patches are used as the training dataset. For data augmentation, random horizontal flips and 90-degree rotations were performed. The model was trained for a total of 100 epochs using the Adam optimizer with an initial learning rate of 1e-3, which was divided by 5 every 25 epochs. The STNet network had 64 channels for its convolutional filters, a kernel size of 3×3, 10 iterations, and a batch size of 64. This invention was implemented using the PyTorch deep learning framework, and all experiments were conducted using an NVIDIA RTX A6000 graphics card for both training and testing. To ensure fair testing, five widely used test datasets were used: LIVE1, Classic5, BSD68, Set14, and Set5.

[0084] Network training

[0085] X compares the STNet-enhanced image with the original lossless image. GT As input to the loss function, the reconstruction loss constrained by the L1 norm is used to supervise the learning of the interpretable image compression artifact removal network STNet. The total loss function can be written as:

[0086]

[0087] Where N is the image X GT The total number of pixels, K is the total number of stages in STNet. X0 is the initial augmented image, X i This is the enhanced image for the i-th stage.

[0088] The method of the present invention has been described in detail above. The enhancement method proposed in this invention will be evaluated below using three objective indicators, namely peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and PSNR-B, as well as subjective indicators.

[0089] 1. Comparison of objective indicators of network test results

[0090] To verify the feasibility of this patented method, it was compared with the JPEG compression algorithm and three representative methods for removing artifacts from compressed images: DNCNN-15, FBCNN, and RCDNet. Table 1 compares the average PSNR / SSIM / PSNR-B of these methods on the LIVE1, Classic5, BSD68, Set14, and other datasets, with the best results indicated in bold and the second-best results underlined. DNCNN-15 was trained using the BCDS500 training set, and FBCNN was trained using the DIV2K and Flick2K training sets. For a fairer comparison, the RCDNet method was retrained on an NVIDIA RTX A6000 device using the same training dataset as this patent. As can be observed from Table 1, the method of this patent outperforms the other five methods in terms of average PSNR / SSIM / PSNR-B across these five datasets. The main contribution of this patent method is the proposal of an image compression artifact removal method inspired by a decomposition model. By using the structure map enhancement module and texture map enhancement module of the deep reconstruction subnetwork to continuously enhance the structural and texture features in an alternating iterative manner, the quality of the reconstructed image is improved. Therefore, the method of this patent is superior to these five representative compressed image artifact removal methods.

[0091]

[0092]

[0093] 2. Comparison of Subjective Indicators in Network Test Results

[0094] To verify the effectiveness of the method of the present invention, it is compared with DNCNN-15, FBCNN, and RCDNe. Figure 6 Visual comparisons of different augmentation methods on the "Barbara" image in the Classic5 dataset are presented. These three figures demonstrate that the method described in this patent restores the image with higher visual quality, fully validating the superiority of the method presented in this invention.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A decomposition model-inspired method for removing image compression artifacts, characterized in that: Includes the following steps: Step 1) Construct training data for the image compression artifact removal model; All images in the DIV2K dataset are compressed using the JPEG standard encoder to obtain a compressed image training dataset that pairs with the original input images. Step 2) Construct an image compression artifact removal optimization model based on structure texture decomposition; First, using The gradient minimization smoothing method filters the JPEG standard encoder-compressed image to obtain a structure map. Subtracting the structure map from the compressed image yields the texture map. Then, the structure map and texture map are respectively passed through a structure initialization subnetwork. and texture structure initialization subnetwork The initial feature map is obtained, and then the maximum a posteriori probability model is used to process the input image. Structural diagram and texture map The relationship between them is mathematically modeled; finally, the probabilistic model is transformed into an image compression artifact removal optimization model based on structural texture decomposition. Step 3) Expand the image compression artifact removal optimization model based on structure texture decomposition in Step 2) into an interpretable network; The proximal gradient descent algorithm is used to solve the image compression artifact removal optimization model obtained in step 2), and it is unfolded into an interpretable image compression artifact removal network. The network consists of two parts: an initialization subnetwork and a deep reconstruction subnetwork. Step 4) Construct the initial subnetwork The structure map and texture map obtained in step 2) are respectively used with all weights of 1. The convolutional kernel performs image smoothing filtering to obtain preliminary denoised structure maps and texture maps. Then, the preliminary denoised structure maps and texture maps are stacked along the channel dimension with the structure maps and texture maps obtained in step 2) and fed into the initialization sub-network. Next, the first three feature maps are extracted from the output features of the structure map initialization sub-network and the texture map initialization sub-network, respectively, as the initialized structure maps. and texture map The remaining feature maps and This information will be sent to the next stage as supplementary information. Step 5) Constructing a deep reconstruction subnetwork The structural feature extraction module and texture feature extraction module of the deep reconstruction subnetwork obtained in step 3) are used to continuously update the structural features and texture features in an alternating iterative manner; Step 6) Loss Function The image enhanced by the interpretable image compression artifact removal network STNet and the original lossless image are used as inputs to the loss function. The reconstruction loss with norm constraints is used to supervise the learning of the network.

2. The image compression artifact removal method inspired by a decomposition model according to claim 1, characterized in that: The specific method for constructing the training data for the compressed artifact removal model in step 1) is as follows: First, the DIV2K dataset containing 1000 images is compressed using a standard JPEG encoder. Here, 800 images are used for STNet network training, 100 images for STNet network validation, and 100 images for STNet network testing. Each image in the 800-image dataset is cropped randomly into 128×128 image patches, and these patches are used as the training dataset. .

3. The image compression artifact removal method inspired by a decomposition model according to claim 1, characterized in that: The calculation process of the image compression artifact removal optimization model based on structure texture decomposition in step 2) is as follows: pass Gradient minimization smoothing method for input compressed images Filtering yields a structure diagram, resulting in a low-quality compressed image. The prior distribution is Low-quality compressed image guide The implicit prior distribution is ,in, and and The energy function, and These are two hyperparameters, which can be obtained using Bayes' theorem as follows: （1） in, yes The probability distribution, and Irrelevant yes A marginal distribution, which is what we are looking for here. Decomposed into and , ,in, It is a structural diagram. It is a texture map, and and The prior distribution is , ,assumed and Independent of each other, Substituting into equation (1), we get the following equation: （2） Using maximum a posteriori probability estimation Formula to solve for posterior probability The above equation can be transformed into a maximization optimization problem: （3） By performing a negative logarithmic transformation on equation (3), the above equation (3) is transformed into an energy minimization model, which can be expressed as: （4） This model is an image compression artifact removal optimization model based on structure texture decomposition. The first term of formula (4) is the data term, and the second and third terms of formula (4) are two regularization constraint terms.

4. The image compression artifact removal method inspired by a decomposition model according to claim 3, characterized in that: The specific steps of step 3) are as follows: First, the optimization problem of formula (4) is transformed into two sub-optimization problems using the proximal gradient descent algorithm; then, the gradient descent algorithm and the proximal mapping algorithm are used to optimize these two sub-problems respectively; finally, they are iteratively optimized in sequence, and the optimization solution formulas for these two sub-problems can be written as: （5） （6） （7） （8） in, Indicates the index value of the iteration. and Update respectively and Learnable step size parameter, and These are about implicit priors. and By iteratively solving formulas (6) and (8) using the proximal operator, the iteratively updated structure diagram can be obtained. and texture map Finally, the last update and The enhanced image is obtained by adding elements one by one. .

5. The image compression artifact removal method inspired by a decomposition model according to claim 4, characterized in that: In step 5), the deep reconstruction subnetwork is composed of It consists of several stages, each of which is composed of a structural feature extraction module (SFEM) and a texture feature extraction module (TFEM). The structural feature extraction module (SFEM) processes the initial structural map and texture map input to this module using the gradient descent algorithm shown in formula (5) to obtain a 3-channel image. For the approximate mapping of the near-end mapping algorithm shown in formula (6), the 3-channel image obtained by the gradient descent algorithm of formula (5) is first stacked with the remaining feature map of the previous stage in the channel dimension, and then processed through five residual blocks, keeping the number of channels unchanged. Finally, the first 3 feature maps of the output feature map with 67 channels are taken as the updated feature map. The remaining 64 feature maps are used as supplementary information and input into the next stage; Texture Feature Extraction Module (TFEM): The initial texture map input to this module and the structure map updated by the structure feature enhancement module (SFEM) are used to obtain a 3-channel image through the gradient descent algorithm shown in formula (7). The near-end mapping algorithm in formula (8) is composed of 5 concatenated residual blocks, and the first 3 feature maps are used as the updated texture map. The remaining 64 feature maps are then used as supplementary information and input into the next stage.

6. The image compression artifact removal method inspired by a decomposition model according to claim 1, characterized in that: In step 6), the image enhanced by the STNet network and the original lossless image are compared. As input to the loss function, the reconstruction loss constrained by the L1 norm is used to supervise the learning of the interpretable image compression artifact removal network STNet. The total loss function can be written as: (9) in, It is an image Total number of pixels, This is the total number of stages in STNet. For the initial enhanced image, For the first Enhanced images at each stage.

Images