Mural restoration method and system based on pulse coupling neural network and diffusion model

Abstract

The invention discloses a mural restoration method and system based on a pulse coupling neural network and a diffusion model, and relates to the technical field of mural restoration. The method comprises the following steps of: firstly, enhancing an image by adopting a three-dimensional pulse coupled neural network enhancement module (SMSPCNN), removing noise and reserving original features; then global and local features are fused through the diffusion reconstruction module, the extraction module is responsible for extracting corresponding features, and finally the attention feature fusion module is used for fusing the features and enhancing expression, so that local and global information fusion is realized, and the model performance is improved. According to the method, the damage of the Dunhuang mural can be effectively repaired, details and structures are reserved, the sense of reality is enhanced, multiple indexes in a quantitative contrast experiment are superior to those of other methods, and a new way is provided for digital repair of the Dunhuang mural.

Description

Technical Field

[0001] The present invention relates to the technical field of mural restoration, and in particular to a mural restoration method and system based on a pulse coupled neural network and a diffusion model. Background Art

[0002] Dunhuang murals are facing varying degrees of damage and fading, necessitating urgent conservation and restoration efforts. Digital restoration of ancient murals using deep learning technology can significantly reduce the time and cost of manual restoration, effectively prevent secondary damage, and improve the overall accuracy and technical level of mural restoration. The pulse-coupled neural network (PCNN) originated from Eckhorn et al.'s research on the cat visual cortex. Its neuronal model is highly similar to biological neurons, accurately simulating the neural activity of the mammalian visual cortex and demonstrating unique advantages in image processing tasks based on neuronal activity. In recent years, the PCNN model has been widely used in image enhancement. Combining PCNN with deep learning methods not only fully utilizes the PCNN's neuronal activity characteristics in image processing, but also leverages the powerful feature extraction capabilities of deep learning to achieve more accurate and efficient digital restoration of ancient murals.

[0003] Deep learning-based image restoration methods can be categorized into unconditional image generation and conditional image generation. Conditional image generation methods require various conditions, such as text and images, to be satisfied. Unconditional image generation methods are relatively easy to train, but suffer from poor controllability, relying solely on randomly generated images based on a learned data distribution. Conditional image generation methods fully consider given conditions during both training and testing to generate images with specific requirements, enhancing the controllability of image generation and offering greater practical value for image restoration in specific scenarios. Dhariwal et al. designed a category guide to improve sample quality, adding category label information to strengthen model constraints and thus enhance image generation quality. However, explicit classifiers require additional training, increasing model complexity. Inspired by explicit classifiers, HO et al. proposed implicit classifier guidance, which eliminates the need for training additional classifiers and significantly reduces training difficulty. Furthermore, multimodal conditional guidance based on CLIP has been widely used, providing more flexible and richer constraints for the generation process. Zhu et al. utilized the global structure of text images as prior knowledge to restore clear text using an efficient diffusion model. Liu et al. proposed a new residual denoising diffusion model RDDM, which decouples the traditional single denoising diffusion process into two processes: residual diffusion and noise diffusion, and constructs a dual diffusion framework.

[0004] Chen et al. proposed the Simplified Pulse Coupled Neural Network (SPCNN) approach, which simplifies the PCNN model to reduce parameters and computational complexity. To further simplify the computational complexity of neurons, Lian et al. proposed a pulse-coupled neural network (FC-MSPCNN) with controllable firing times. This model overcomes the uncertainty inherent in traditional models due to the random nature of neuron firing by introducing a precise control mechanism for the firing times of neurons. The proposed stereoscopic pulse-coupled neural network (SMSPCNN) significantly improves the stability and reliability of PCNNs in image processing tasks, reducing computational complexity while maintaining high computational efficiency and processing accuracy.

[0005] Dunhuang murals possess a unique artistic style, characterized by complex line structures, delicate texture details, and rich historical and cultural connotations. Existing text-guided or multimodal methods still have limitations in capturing and reproducing these characteristics, resulting in inpainted images prone to local blurring or overall structural inconsistencies. Especially in complex image scenes, these methods still suffer from insufficient feature utilization, loss of semantic information in the inpainted image, poor global consistency, and overly smoothed inpainted results, making them difficult to meet the requirements of high-precision image restoration. To address this, this paper proposes a Dunhuang mural restoration method (ST-DiffNet) that integrates the SMSPCNN image enhancement mechanism with a diffusion model. This method uses dual conditions to guide the restoration process of Dunhuang murals. This method leverages the advantages of the SMSPCNN and combines it with the generative power of the diffusion model to optimize the restoration results in terms of structural consistency and texture integrity. Notably, this method simulates the spiking mechanism of biological neurons, enhancing the impulse response of local regions. This method preserves mural details while reducing oversmoothing, ensuring that the generated results are more consistent with the original content and artistic style of the murals, achieving high-precision and high-fidelity digital restoration. Summary of the Invention

[0006] The purpose of the present invention is to provide a mural restoration method and system based on a pulse coupled neural network and a diffusion model to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a mural restoration method based on a pulse coupled neural network and a diffusion model, comprising the following steps:

[0008] S1: The Dunhuang mural images are enhanced using the stereo pulse-coupled neural network enhancement module SMSPCNN. Based on the working mechanism of pulse-coupled neural network (PCNN), image pixels are processed to effectively suppress noise while preserving the original image features, making the enhanced image clearer and more stable.

[0009] S2: Using the diffusion reconstruction module to fuse dual-condition global and local features, it captures global structural information while preserving local texture details, improving the integrity and fineness of the restoration results;

[0010] S3: extracting global structural features and local texture features through a feature extraction module, wherein the feature extraction module includes a global feature extraction submodule for multi-scale attention and a local feature extraction submodule for cropping salient regions and reconstructing features;

[0011] S4: The extracted global features and local features are fused with the help of the attention feature fusion module, and channel attention, spatial attention and cross attention are combined to enhance feature expression and realize the fusion of local and global information.

[0012] Furthermore, the stereo pulse-coupled neural network enhancement module SMSPCNN adopts a new method to set the synaptic weight matrix, enhances the interaction force between the central neuron and its four adjacent neurons above, below, left and right, and adjusts the change in the number of each neuron during the burst process by improving parameters and adopting a method based on dynamic threshold amplitude.

[0013] Furthermore, the diffusion reconstruction module is based on a denoising diffusion probability model, which gradually adds Gaussian noise to the original image through a forward diffusion process, and gradually calculates the posterior distribution from the noisy image through a reverse denoising process to restore the original image, and fuses global structural features and local texture features respectively in the downsampling stage.

[0014] Furthermore, in the feature extraction module, the global feature extraction submodule uses convolution operations with different expansion rates to extract features at various scales, and splices the outputs according to the channel dimension to form a complete attention map; the local feature extraction submodule first extracts saliency features based on the low-resolution image, generates a saliency map to determine the region of interest, and then inputs the high-resolution cropped area into the feature reconstruction module to extract local features.

[0015] Furthermore, the attention feature fusion module combines channel attention and cross attention when processing local features, and combines spatial attention and cross attention when processing global features. It maps intermediate features to the latent space through cross attention to obtain attention scores, and dynamically weights and fuses different features accordingly.

[0016] Furthermore, a mural restoration system based on a pulse coupled neural network and a diffusion model includes:

[0017] The stereo pulse coupled neural network enhancement module SMSPCNN is used to enhance the Dunhuang mural images;

[0018] Diffusion reconstruction module, used to fuse bi-conditional global and local features for image restoration;

[0019] Feature extraction module, used to extract global structural features and local texture features;

[0020] Attention feature fusion module is used to fuse global features and local features.

[0021] Furthermore, the structures and functions of the stereo pulse coupled neural network enhancement module (SMSPCNN), diffusion reconstruction module, feature extraction module and attention feature fusion module correspond to the method described in any one of claims 1 to 5.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) Aiming at the specific task of restoring Dunhuang murals, a mural restoration method and system based on pulse coupled neural network and diffusion model (ST-DiffNet) was proposed. By guiding the restoration process of Dunhuang murals under dual conditions, the texture rationality and structural consistency of the restored images were achieved.

[0024] (2) A smooth structure graph is introduced into the dual-condition guided network to effectively improve the ability to restore structural details during image restoration. By adding a smooth structure graph to the network, the overall structural consistency of the image can be maintained while avoiding excessive noise and unnatural edges during detail restoration.

[0025] (3) A feature extraction module and an attention feature fusion module are added to the restoration network to effectively extract and fuse global structural features and local texture features, avoiding the problem of insufficient feature utilization.

[0026] (4) A stereo pulse-coupled neural network mural enhancement method (SMSPCNN) is proposed. This method processes image pixels based on the working mechanism of pulse-coupled neural networks (PCNN). It effectively suppresses noise while preserving the original image features, making the enhanced image clearer and more stable. The SMSPCNN model does not require training and can be applied to most mural enhancement scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of trajectory tracking control of an emergency material transport robot based on fuzzy model predictive control and extended state observer of the present invention;

[0028] Figure 2is a flow chart of the overall method of the present invention;

[0029] Figure 3 It is the overall framework of DDPM of the present invention;

[0030] Figure 4 It is a dual-branch feature extraction module of the present invention;

[0031] Figure 5 It is the cross-attention feature fusion module framework of the present invention;

[0032] Figure 6 This is a qualitative comparison of the damaged area of the Dunhuang murals of the present invention, which is 0-20%;

[0033] Figure 7 This is a qualitative comparison of the damaged area of the Dunhuang murals of the present invention, which is 20%-40%;

[0034] Figure 8 This is a qualitative comparison of the damaged area of the Dunhuang murals of the present invention, which is 40%-60%. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0036] The technical solution process of the technical solution of the present invention is as follows Figure 1 The method of the present invention mainly consists of a stereo pulse coupled neural network enhancement module (SMSPCNN), a diffusion reconstruction module, a feature extraction module, and an attention feature fusion module.

[0037] ① This paper proposes a stereo pulse-coupled neural network mural enhancement method (SMSPCNN). This method processes image pixels based on the working mechanism of pulse-coupled neural network (PCNN) neurons, effectively suppressing noise while retaining the original image features, making the enhanced image clearer and more stable. The SMSPCNN model does not require training and can be applied to most mural enhancement scenarios. Existing PCNN models have problems such as many parameters, difficult coefficients to determine, large amount of calculations, and complex calculations. We have improved it and designed an SMSPCNN image enhancement module, the corresponding expression is as follows:

[0038]

[0039]

[0040]

[0041] Compared with the PCNN model, the SMSPCNN model uses a new method to set the synaptic weight matrix , and proposed a new synaptic weight matrix For the sake of simplicity, we combine the synaptic weight and the decay index Combined and define new amplitude parameters To improve the rationality of the dynamic threshold change of the SMSPCNN model. The network structure of the SMSPCNN model is as follows Figure 2 shown.

[0042] Among them, the synaptic weight matrix 、 A new setting method was adopted, which enhanced the interaction between the central neuron and its four adjacent neurons above, below, left and right. 、 The setting method is as follows:

[0043]

[0044]

[0045] On this basis, we improved the parameters and adopted a method based on dynamic threshold amplitude to adjust the changes in the number of neurons in the burst process. The parameters we modified are as follows:

[0046]

[0047]

[0048]

[0049] ② Based on the generative capability of the diffusion model, we designed a diffusion reconstruction module that integrates dual-condition global and local features. It captures global structural information while retaining local texture details, thereby improving the integrity and fineness of the restoration results and making the restored image more realistic. The denoising diffusion probability model is divided into two processes: forward diffusion and reverse denoising. The overall framework is as follows: Figure 3 As shown. The forward process is also called the diffusion process, which is to Gaussian noise is gradually added in, and the standard deviation of the noise is a fixed value The mean is a fixed value. and the data at time t-1 OK. With Continuously increasing, eventually distributed Obeys an isotropic Gaussian distribution. In the forward process, we start from a real image At the beginning, we randomly generate some Gaussian noise at each step, and then gradually add the generated noise to the input image. When T is large enough, the noisy image we get is close to a Gaussian noise image. The diffusion process The distribution of is as follows:

[0050]

[0051]

[0052] In order to restore the original image from random noise, we need to do inverse denoising, that is, gradually calculate the posterior distribution from the standard Gaussian distribution In the backward process, we start with a noisy image and learn through a neural network arrive The noise added by training the parameterized model To predict arrive Intermediate accumulated noise To obtain , its training objectives are:

[0053]

[0054] Then the final image to be generated is obtained by gradually denoising, and the definition , .

[0055]

[0056] Specifically, we fuse the structural features composed of global features and the texture features composed of local features into the downsampling stage. We fuse the structural features in the first layer, where the spatial resolution of the feature map is high, effectively capturing global structural information and ensuring the overall contour and morphological consistency of the repaired image. We fuse the texture features in the third layer, where the spatial resolution of the feature map is low, but still retains a certain amount of local detail. By enhancing the expression of texture and microscopic features, the repaired image becomes more delicate and realistic. Ultimately, both the global consistency and local details of the image are improved, making the repaired image more natural and rich in both morphology and texture.

[0057] ③ Feature extraction module: This module contains two submodules, one is the global feature extraction module using multi-scale attention, and the other is the local feature extraction module using cropping salient regions and reconstructing features. The overall architecture is as follows Figure 4 As shown. For the global flow, the original mural image data (256×256) is first input into the global encoder composed of multi-scale attention , generate global feature representation, such as Figure 4 (a). Global encoder Convolution operations with different dilation rates are used to extract features at various scales. Specifically, the process consists of three convolutional layers, each with a convolution kernel size of 3×3 and a stride of 1, but with different dilation rates (1, 2, 4) for convolution. The first convolutional layer is responsible for extracting basic features, the second convolutional layer captures a wider range of contextual information, and the third convolutional layer further expands the field of view to obtain a wider range of contextual features. Finally, the outputs of these three convolutional layers are concatenated according to the channel dimension to form a complete attention map, thereby obtaining the global features of the input image. 、 For the local flow, we further crop the regions of interest in the saliency map and pass these regions as input to the local feature extraction module, such as Figure 4 (b). In the process of extracting local features, the input image is first downsampled to a low resolution (8×8), and salient features are extracted based on the low-resolution image to generate a saliency map to determine the important regions of interest in the image. Subsequently, the cropping position is automatically derived based on the information in the saliency map, and the position of the salient region is mapped back to the high-resolution image (256×256) through a nonlinear coordinate transformation. Next, the adjusted high-resolution cropped region is input into the feature reconstruction module, and the multi-scale attention mechanism is used to further extract local features, thereby obtaining refined local features of the input image. 、 .

[0058] ④ Attention feature fusion module: For different features, we designed a feature fusion module based on cross attention. Its overall framework is as follows Figure 5 shown.

[0059] The attention feature fusion module combines channel attention and cross attention when processing local features, and spatial attention and cross attention when processing global features. Channel attention selectively highlights important channels related to texture and detail in the feature map, while cross attention further integrates local features with their contextual information, thereby improving the expressiveness of local regions, especially capturing subtle differences. Spatial attention focuses on spatial relationships within the global scope. Combined with cross attention, it can effectively model the interaction between global structure and semantic features, improving understanding of the overall context. This module specifically strengthens feature expression, enabling the fusion of local and global information in the execution of various tasks, improving model performance and robustness.

[0060] Generally speaking, spatial attention focuses more on structural information, while channel attention focuses more on texture information. Based on this, before using cross attention for feature fusion, we first apply the spatial and channel attention mechanisms to the global features extracted by the feature extraction module. and local features After this allocation, channel attention can effectively enhance the detail representation of texture features, while spatial attention can better preserve and emphasize the spatial relationship of structural features, enabling better interactive fusion of texture and structural information at the same level.

[0061]

[0062]

[0063] In the above formula, and Represent structural features and texture features respectively. refers to spatial attention, Refers to channel attention, Refers to crossed attention.

[0064] Unlike the self-attention mechanism that only focuses on the internal relationship of a single dimension, the attention feature fusion module focuses more on the correlation between different scales and selects more important features in the feature fusion process. The cross attention plays a key role in mapping the intermediate features to the latent space through a 1x1 convolution filter, thereby obtaining the attention scores of two different feature spaces. 、 For example ( and The processing is the same as above). and are the global features of different images. Their mathematical formulas are:

[0065]

[0066]

[0067] in and is a 1×1 convolution filter. The output of the attention feature fusion module can be obtained as follows

[0068]

[0069] in,

[0070]

[0071] Also in the above formula, is a 1×1 convolution filter. The attention feature fusion module dynamically weights and fuses different features through a cross-level attention mechanism to ensure a balance between local details and global semantics.

[0072] A Dunhuang mural restoration system based on pulse coupled neural network and diffusion model, including:

[0073] The stereo pulse coupled neural network enhancement module SMSPCNN is used to enhance the Dunhuang mural images;

[0074] Diffusion reconstruction module, used to fuse bi-conditional global and local features for image restoration;

[0075] Feature extraction module, used to extract global structural features and local texture features;

[0076] Attention feature fusion module is used to fuse global features and local features.

[0077] Comparison of experimental data and evaluation results

[0078] The dataset used in this experiment includes 4,564 Dunhuang mural images, including those of the Flying Buddha statues, the Donor murals, and the Dunhuang landscape murals. A corresponding smoothed structure map dataset is also available. The experimental portion uses an irregular mask dataset, containing 5,000 irregular mask images. The irregular mask dataset is divided into three regions based on mask percentage: the first region covers areas where the mask occupies 0-20% of the total image area, the second region covers areas where the mask occupies 20%-40%, and the third region covers areas where the mask occupies 40%-60%. The effectiveness of our method is verified by comparing it with six representative methods: RePaint, ADF, MagConv, DEST, StrDiffusion, and TSBGNet. Image quality evaluation uses several common metrics used in image restoration tasks: peak signal-to-noise ratio (PSNR), structural similarity (SSIM), learning-perceptual image patch similarity (LPIPS), L1 error, and Frechette distance (FID). The quantitative results of the comparative experiments on the Dunhuang mural dataset are shown in Table 1.

[0079] Table 1 Quantitative comparison results of different methods in different mask areas on the Dunhuang murals dataset

[0080]

[0081] As can be seen from Table 1, when the mask ratio is 0-20%, the proposed method outperforms other comparison methods in terms of PSNR, L1, and FID. When the mask ratio is 20-40%, the proposed method outperforms other comparison methods in terms of LPIPS, L1, and FID. When the mask ratio is 40-60%, the proposed method outperforms other comparison methods in all metrics except SSIM. This demonstrates that the proposed method generates images of high quality and good diversity, and the image distribution is more similar to real images.

[0082] In order to simulate different degrees of damage to the Dunhuang murals, different mask areas were set to simulate the damage on the murals. Figure 6 、 Figure 7 、 Figure 8 The following figure shows the qualitative comparison results of the simulated restoration of Dunhuang murals using the method of the present invention and six advanced methods at different mask ratios. The experimental image library is a Dunhuang mural dataset. The figure shows, from top to bottom, the restoration results of two Dunhuang flying Buddha murals, two Dunhuang donor murals, and two Dunhuang landscape paintings. Among them, the first column of images is the original image; the second column of images is the defective image; the third column of images is the restored image obtained by the RePaint method; the fourth column of images is the restored image obtained by the ADF method; the fifth column of images is the restored image obtained by the MagConv method; the sixth column of images is the restored image obtained by the DEST method; the seventh column of images is the restored image obtained by the StrDiffusion method; the eighth column of images is the restored image obtained by the TSBGNet method; and the ninth column of images is the restored image obtained by the method of the present invention.

[0083] from Figure 6 , Figure 7 , Figure 8From the results, it can be seen that the seven methods including the method of the present invention can fill the masked area with texture. However, different methods differ in the details of filling and the consistency of texture. Specifically, when MagConv is repairing damaged areas, the image repair effect is poor, and some areas appear blurred or incomplete, especially in areas with more serious damage, the repair results have color distortion and detail loss. The RePaint method and the DEST method have some flaws and distortions, especially at the edges, where the distortion is more obvious, and the overall visual effect is far from ideal. The image effect after repair by the ADF method is not satisfactory, and the details and structure of the image are not well restored. It is still insufficient in restoring the original color and texture, and the subtle features in the image are not clearly presented. The repair result presents a blurred feeling. Although the StrDiffusion method and the TSBGNet method have been repaired, the image effect is still not fine enough, many details are missing, and the repair results are relatively rough. This shows that the six compared methods still have deficiencies in restoring the details of the Dunhuang murals. The method of the present invention excels in restoring local details and original textures, particularly when the masking ratio is large, resulting in more realistic and clear images of the restored Dunhuang murals. The method of the present invention achieves relatively better results in Dunhuang mural restoration under different masking area ratios. Regardless of the size of the masked area, the method of the present invention better preserves the details and structure of the murals, demonstrating more stable restoration results compared to other methods.

[0084] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0085] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A mural restoration method based on pulse coupled neural network and diffusion model, characterized in that: The following steps are involved: S1: The Dunhuang mural images are enhanced using the stereo pulse-coupled neural network enhancement module SMSPCNN. Based on the working mechanism of pulse-coupled neural network (PCNN), image pixels are processed to effectively suppress noise while preserving the original image features, making the enhanced image clearer and more stable. S2: Using the diffusion reconstruction module to fuse dual-condition global and local features, it captures global structural information while preserving local texture details, improving the integrity and fineness of the restoration results; S3: extracting global structural features and local texture features through a feature extraction module, wherein the feature extraction module includes a global feature extraction submodule for multi-scale attention and a local feature extraction submodule for cropping salient regions and reconstructing features; S4: The extracted global features and local features are fused with the help of the attention feature fusion module, and channel attention, spatial attention and cross attention are combined to enhance feature expression and realize the fusion of local and global information.

2. The mural restoration method based on pulse coupled neural network and diffusion model according to claim 1 is characterized in that: The stereo pulse-coupled neural network enhancement module SMSPCNN adopts a new method to set the synaptic weight matrix, enhances the interaction force between the central neuron and its four adjacent neurons above, below, left and right, and adjusts the change of the number of each neuron in the burst process by improving parameters and adopting a method based on dynamic threshold amplitude.

3. The mural restoration method based on pulse coupled neural network and diffusion model according to claim 1 is characterized in that: The diffusion reconstruction module is based on a denoising diffusion probability model. It gradually adds Gaussian noise to the original image through a forward diffusion process, and gradually calculates the posterior distribution from the noisy image through a reverse denoising process to restore the original image. It also fuses global structural features and local texture features in the downsampling stage.

4. The mural restoration method based on pulse coupled neural network and diffusion model according to claim 1 is characterized in that: In the feature extraction module, the global feature extraction submodule uses convolution operations with different dilation rates to extract features at various scales, and splices the outputs according to the channel dimension to form a complete attention map; the local feature extraction submodule first extracts salient features based on the low-resolution image, generates a saliency map to determine the region of interest, and then inputs the high-resolution cropped area into the feature reconstruction module to extract local features.

5. The mural restoration method based on pulse coupled neural network and diffusion model according to claim 1 is characterized in that: The attention feature fusion module combines channel attention and cross attention when processing local features, and combines spatial attention and cross attention when processing global features. It maps intermediate features to the latent space through cross attention to obtain attention scores, and dynamically weights and fuses different features accordingly.

6. A mural restoration system based on pulse coupled neural network and diffusion model, characterized in that: include: The stereo pulse coupled neural network enhancement module SMSPCNN is used to enhance the Dunhuang mural images; Diffusion reconstruction module, used to fuse bi-conditional global and local features for image restoration; Feature extraction module, used to extract global structural features and local texture features; Attention feature fusion module is used to fuse global features and local features.

7. The mural restoration system based on pulse coupled neural network and diffusion model according to claim 6 is characterized in that: The structure and function of the stereo pulse coupled neural network enhancement module SMSPCNN, diffusion reconstruction module, feature extraction module and attention feature fusion module correspond to the method described in any one of claims 1 to 5.

Images