Priori information enhanced image deblurring method

Abstract

The invention discloses a priori information enhanced image deblurring method, and belongs to the technical field of image processing. In the image edge information extraction and enhancement module, a Canny operator is adopted to extract clear edge information of the image, fuzzy image features are combined with the edge information, information enhancement is carried out through operations such as Fourier transform and global pooling, an edge enhancement feature graph is obtained, and edge high-frequency information of the image can be better recovered. In a foreground and background attention feature weighting module, the blurred image features are divided into foreground and background features, the foreground and background features are obtained by multiplying the foreground and background features with input fuzzy features, then a weighted graph of the foreground and background features is obtained by using pooling and convolution activation operation, and foreground and background information is subjected to weighted fusion. A structure information weighted feature map of different weights of the foreground and the background is obtained, the structure information weighted feature map comprises position prior information of the foreground and the background, and deblurring can be better achieved. And the two modules are combined, so that the deblurring network has better generalization.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to an image deblurring method with enhanced prior information. Background Art

[0002] Image deblurring is an important task in image processing and computer vision, aiming to recover a clear image from a blurred image. In real life, blurred images not only affect people's visual perception but also have an adverse impact on multiple fields such as medical imaging, surveillance, and autonomous driving. Therefore, the research on image deblurring is particularly important. Blur can be mainly divided into several types: motion blur, focus blur, Gaussian blur, etc. Motion blur usually occurs when photographing a fast-moving object, resulting in streaks in the image; focus blur is caused by incorrect focusing during shooting, resulting in loss of details; while Gaussian blur is a common image processing effect, usually used for image smoothing. The technology of image deblurring has gone through multiple development stages. Early deblurring methods were mostly based on classical mathematical models and filtering techniques, such as Wiener filtering and R-L filtering. These methods perform well in dealing with simple blur and single blur kernel blur, but for complex blur situations, the effects are often not good. With the rapid development of computer vision and machine learning technologies, more and more deep learning methods have been applied to image deblurring. Deep learning-based deblurring algorithms can handle complex blur patterns. Through training on a large amount of data, they learn the mapping relationship between blurred images and clear images, and achieve image deblurring end-to-end. These methods show strong capabilities in dealing with high-dynamic-range images, complex scenes, and multiple blur types. The application of deblurring technology has achieved remarkable results in multiple fields. In medical imaging, clear images are crucial for the diagnosis of lesions. Image deblurring technology can help doctors observe fine structures and lesions more accurately. In the field of security surveillance, blurred surveillance videos may lead to the loss of important evidence, and deblurring technology can effectively improve the usability of surveillance images and help law enforcement officers identify suspects.

[0003] Traditional image deblurring technologies are mainly based on prior knowledge and mathematical models. For example, Wiener filtering restores an image by assuming the statistical characteristics of the image, while blind image deblurring based on deconvolution attempts to estimate both the blur kernel and the clear image simultaneously. In real-world scenarios, the same image often contains different blur kernels. Estimating a single blur kernel for the image is not conducive to image restoration. For the inverse problem of solving the clear image, the solution space is not uniquely determined, which makes it difficult to use the blur kernel to solve the clear image. These methods have a certain theoretical basis, but in practical applications, especially when dealing with complex blur types, the effects are often not good.

[0004] In recent years, the rapid development of deep learning has brought new opportunities to the field of image deblurring. Deblurring methods based on convolutional neural networks can be trained with a large number of blurred and clear image pairs, adaptively learn the features of images, and thus more effectively restore the details of images. These deep learning models can capture complex patterns and texture information that are difficult to handle by traditional methods, resulting in a significant improvement in the deblurring effect. However, although deep learning methods perform well in many benchmark tests, there are still some challenges. Blur in real scenes is often non-uniform, with different regions containing different blur kernels, making it ineffective in restoring the edge textures of images. In addition, the training of the model requires a large amount of labeled data, and its generalization ability in different blur types and scenarios still needs to be improved. Summary of the Invention

[0005] In view of the above technical problems, the present invention proposes an image deblurring scheme with enhanced prior information.

[0006] The first aspect of the present invention discloses an image deblurring method with enhanced prior information, the method comprising:

[0007] Step S1: Invoke the image edge information extraction and enhancement module, extract the edge information of the input blurred image using the Canny operator, combine the blurred image features with the edge information, perform edge information enhancement through Fourier transform and global pooling operations, and obtain an edge-enhanced feature map for restoring the edge high-frequency information of the blurred image;

[0008] Step S2: Invoke the foreground and background attention feature weighting module, divide the blurred image features into foreground and background, multiply them with the blurred image features respectively to obtain foreground features and background features, obtain the weighted maps of the foreground features and background features through pooling and convolutional activation operations, and then obtain a structure information weighted feature map of the foreground and background with different weights through weighted fusion;

[0009] Step S3: Based on the input of the edge-enhanced feature map and the structure information weighted feature map of the foreground and background with different weights, use the deblurring network to perform deblurring processing on the input blurred image. The parameters of the deblurring network are pre-trained with blurred-clear image pairs, and the deblurring network is composed of residual convolutions.

[0010] According to the method of the first aspect of the present invention, the deblurring network is trained with blurred-clear image pairs to obtain the optimal parameters of the deblurring network; wherein, the blurred image in the blurred-clear image pair is used as the input of the deblurring network, and the loss function is calculated using the clear image and the deblurred image in the blurred-clear image pair.

[0011] According to the method of the first aspect of the present invention, the input blurred image is preprocessed to obtain a blurred image feature map with a size of 3×H×W, where H×W is the image size and 3 is the number of image channels.

[0012] According to the method of the first aspect of the present invention, the Canny operator is used to extract edge information from the input blurred image feature map to obtain an edge information map, and the edge information is combined with the blurred image feature, and an edge-enhanced feature map is obtained through Fourier transform, convolutional activation, global average pooling, and inverse Fourier transform.

[0013] According to the method of the first aspect of the present invention, the edge information map is F_edge, and the edge information map F_edge and the blurred image feature map F_b are fused and input into the global pooling and Fourier transform branches to obtain an edge-enhanced feature map F_s, and:

[0014] F_s = Conv(Avg(concate(F_b,F_edge)))+iFFT(Conv(FFT(concate(F_b,F_edge))))

[0015] Where Conv represents convolutional activation, Avg represents global average pooling, concate represents concatenation fusion, iFFT represents inverse fast Fourier transform, FFT represents fast Fourier transform, and Conv(·) = Conv 1×1 (BatchNorm(ReLU(Conv 1×1 (·)))), ReLU represents the activation function, and BatchNorm represents batch normalization.

[0016] According to the method of the first aspect of the present invention, the blurred image feature is divided into a foreground feature M1 and a background feature M2, and the overfitting effect is eliminated through global average pooling and max pooling processing. Among them, global average pooling extracts overall information and averages the information within the channels; after convolutional activation operation, the Sigmoid function is used to obtain a weighted map of the foreground feature and the background feature, and then weighted fusion is performed to obtain a structure information weighted feature map F_w of the foreground and the background with different weights:

[0017] F1 = F_b*Sigmoid(Conv(concate(Avg(F_b*M1),Max(F_b*M1))))

[0018] F2 = F_b*Sigmoid(Conv(concate(Avg(F_b*M2),Max(F_b*M2))))

[0019] F_w = F1+F2

[0020] Among them, the Sigmoid function maps the eigenvalue to (0, 1), Max represents global maximum pooling, F1 represents the foreground feature weighted graph, and F2 represents the background feature weighted graph.

[0021] According to the method of the first aspect of the present invention, the edge-enhanced feature map F_s, the structure information weighted feature map F_w of the foreground and background with different weights, and the blurred image are input into the deblurring network to obtain a deblurred image.

[0022] The second aspect of the present invention discloses an image deblurring system with enhanced prior information, characterized in that the system includes an image edge information extraction and enhancement module, a foreground and background attention feature weighting module, and a deblurring network, wherein:

[0023] The image edge information extraction and enhancement module is configured to: extract the edge information of the input blurred image by using the Canny operator, combine the blurred image features with the edge information, and perform edge information enhancement through Fourier transform and global pooling operations to obtain an edge-enhanced feature map for restoring the edge high-frequency information of the blurred image;

[0024] The foreground and background attention feature weighting module is configured to: divide the blurred image features into foreground and background, multiply them with the blurred image features respectively to obtain foreground features and background features, obtain the weighted graphs of the foreground features and background features through pooling and convolutional activation operations, and then obtain the structure information weighted feature map of the foreground and background with different weights through weighted fusion;

[0025] Using the deblurring network, based on the input of the edge-enhanced feature map and the structure information weighted feature map of the foreground and background with different weights, the input blurred image is deblurred. The parameters of the deblurring network are pre-trained by a blurred-clear image pair, and the deblurring network is composed of residual convolutions.

[0026] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. When the processor executes the computer program, it implements the method for enhancing image deblurring with prior information described in the first aspect of the present disclosure.

[0027] The fourth aspect of the present invention discloses a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, it implements the method for enhancing image deblurring with prior information described in the first aspect of the present disclosure.

[0028] In summary, the technical solution proposed by the present invention performs excellently in capturing edge frequency information and enhancing the understanding of image structure, enhancing the model's attention to different features, thus solving the problems of insufficient edge information processing and image structure information perception in the prior art, being able to better process non-uniform blur in images, achieving a better deblurring effect, and enhancing the generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0030] Figure 1 is a schematic flow chart of an image deblurring method with prior information enhancement;

[0031] Figure 2 is a schematic structural diagram of an image edge information extraction and information enhancement module;

[0032] Figure 3 is a schematic structural diagram of a foreground and background attention feature weighting module;

[0033] Figure 4 is a schematic structural diagram of a deblurring network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Image blur is the most common image information loss phenomenon in daily life. As an important information carrier, image blur not only affects people's visual perception but also affects the accuracy of downstream computer vision tasks. In real scenarios, the blur of images is non-uniform, and the solution space is very large, which makes image deblurring an ill-posed problem with an uncertain solution, and the image restoration effect still needs to be improved. At the same time, due to the large range of the solution space, the generalization of image deblurring algorithms is not good.

[0036] To address the limitations of existing methods, an image deblurring scheme based on structural priors is proposed. It performs excellently in capturing edge frequency information and enhancing the understanding of image structures, enhancing the model's attention to different features, thus solving the problems of insufficient edge information processing and image structure information perception in existing methods. It can better handle non-uniform blur in images, achieve better deblurring effects, and enhance the generalization of the model.

[0037] The first aspect of the present invention discloses an image deblurring method with enhanced prior information, and the method includes:

[0038] Step S1: Invoke the image edge information extraction and enhancement module, use the Canny operator to extract the edge information of the input blurred image, combine the blurred image features with the edge information, and perform edge information enhancement through Fourier transform and global pooling operations to obtain an edge-enhanced feature map for restoring the edge high-frequency information of the blurred image.

[0039] Step S2: Invoke the foreground and background attention feature weighting module, divide the blurred image features into foreground and background, multiply them with the blurred image features respectively to obtain foreground features and background features, obtain the weighted maps of the foreground features and background features through pooling and convolutional activation operations, and then obtain a structure information weighted feature map of the foreground and background with different weights through weighted fusion.

[0040] Step S3: Based on the input of the edge-enhanced feature map and the structure information weighted feature map of the foreground and background with different weights, use the deblurring network to perform deblurring processing on the input blurred image. The parameters of the deblurring network are pre-trained by blurred-sharp image pairs, and the deblurring network is composed of residual convolutions.

[0041] According to the method of the first aspect of the present invention, use the blurred-sharp image pairs to train the deblurring network to obtain the optimal parameters of the deblurring network; wherein, the blurred image in the blurred-sharp image pair is used as the input of the deblurring network, and the loss function is calculated using the sharp image and the deblurred image in the blurred-sharp image pair.

[0042] According to the method of the first aspect of the present invention, preprocess the input blurred image to obtain a blurred image feature map of size 3×H×W, where H×W is the image size and 3 is the number of image channels.

[0043] According to the method of the first aspect of the present invention, use the Canny operator to extract the edge information of the input blurred image feature map to obtain an edge information map, combine the edge information with the blurred image features, and obtain the edge-enhanced feature map through Fourier transform, convolutional activation, global average pooling, and inverse Fourier transform.

[0044] According to the method of the first aspect of the present invention, the edge information map is F_edge. The edge information map F_edge and the blurred image feature map F_b are fused and input into the global pooling and Fourier transform branches to obtain the edge-enhanced feature map F_s, and:

[0045] F_s = Conv(Avg(concate(F_b,F_edge))) + iFFT(Conv(FFT(concate(F_b,F_edge))))

[0046] Wherein, Conv represents convolutional activation, Avg represents global average pooling, concate represents concatenation fusion, iFFT represents inverse fast Fourier transform, FFT represents fast Fourier transform, and Conv(·) = Conv 1×1 (BatchNorm(ReLU(Conv 1×1 (·)))), ReLU represents the activation function, and BatchNorm represents batch normalization.

[0047] According to the method of the first aspect of the present invention, the blurred image features are divided into foreground features M1 and background features M2. The overfitting effect is eliminated through global pooling and max pooling processing. Among them, global pooling extracts overall information and averages the information within the channels; after convolutional activation operation, the Sigmoid function is used to obtain the weighted maps of the foreground and background features, and then the weighted fusion is used to obtain the structure information weighted feature map F_w of the foreground and background with different weights:

[0048] F1 = F_b * Sigmoid(Conv(concate(Avg(F_b * M1),Max(F_b * M1))))

[0049] F2 = F_b * Sigmoid(Conv(concate(Avg(F_b * M2),Max(F_b * M2))))

[0050] F_w = F1 + F2

[0051] Wherein, the Sigmoid function maps the feature values to (0,1), Max represents global max pooling, F1 represents the foreground feature weighted map, and F2 represents the background feature weighted map.

[0052] According to the method of the first aspect of the present invention, the edge-enhanced feature map F_s, the structure information weighted feature map F_w of the foreground and background with different weights, and the blurred image are input into the deblurring network to obtain the deblurred image.

[0053] The second aspect of the present invention discloses an image deblurring system with enhanced prior information, characterized in that the system includes an image edge information extraction and enhancement module, a foreground and background attention feature weighting module, and a deblurring network, where:

[0054] The image edge information extraction and enhancement module is configured to: extract the edge information of the input blurred image using the Canny operator, combine the blurred image features with the edge information, and perform edge information enhancement through Fourier transform and global pooling operations to obtain an edge-enhanced feature map for restoring the high-frequency edge information of the blurred image;

[0055] The foreground and background attention feature weighting module is configured to: divide the blurred image features into foreground and background, multiply them with the blurred image features respectively to obtain foreground features and background features, obtain the weighted maps of the foreground features and background features through pooling and convolution activation operations, and then obtain the structure information weighted feature map of the foreground and background with different weights through weighted fusion;

[0056] Using the deblurring network, based on the input of the edge-enhanced feature map and the structure information weighted feature map of the foreground and background with different weights, the input blurred image is deblurred. The parameters of the deblurring network are pre-trained by blurred-sharp image pairs, and the deblurring network is composed of residual convolutions.

[0057] First Embodiment

[0058] This embodiment relates to an image deblurring method with enhanced prior information for improving the image deblurring effect. Compared with general image deblurring algorithms, the method of this embodiment mainly acts on the image edge information extraction and enhancement module and the foreground and background attention feature weighting module. As Figure 1 shown, in the image edge information extraction and enhancement module, the Canny operator is used to extract the clear edge information of the image, the blurred image features are combined with the edge information, and information enhancement is performed through operations such as Fourier transform and global pooling to obtain an edge-enhanced feature map, which can better restore the high-frequency edge information of the image. In the foreground and background attention feature weighting module, the blurred image features are divided into foreground and background, and multiplied with the input blurred features to obtain foreground and background features. Then, pooling and convolution activation operations are used to obtain the weighted maps of the foreground and background features, and the foreground and background information is weighted and fused to obtain the structure information weighted feature map of the foreground and background with different weights, which contains the position prior information of the foreground and background and can better deblur. In addition, combining the two modules makes the deblurring network have better generalization.

[0059] Second Embodiment

[0060] This embodiment relates to the module composition of the image deblurring method with enhanced prior information.

[0061] 1. Input module

[0062] Use the clear-blurry image data pairs in the existing blurry image dataset as training data, where the blurry images are used as the input to the network, and the clear images are used for calculating the loss function of the de-blurred images.

[0063] 2. Image edge information extraction and information enhancement module

[0064] Extract the edge information of the blurry image using the Canny operator to obtain an edge information map. Input the edge information map and the blurry feature map into the image edge information extraction and information enhancement module to obtain an edge-enhanced feature map. The module structure is designed as Figure 2 shown.

[0065] 3. Foreground and background attention feature weighting module

[0066] Divide the input blurry image into foreground features and background features, and input them into the foreground and background attention feature weighting module to obtain a foreground and background weighted feature map. The structural composition of the module is as Figure 3 shown.

[0067] 4. De-blurring network

[0068] Fuse the information features of the above two modules and input them into the de-blurring network. According to the training parameters of the blurry image pair, output the de-blurred image. The de-blurring network is mainly composed of residual convolutions, as Figure 4 shown.

[0069] The third embodiment

[0070] This embodiment relates to the process design of an image de-blurring method with prior information enhancement, aiming to enhance the model's attention to different features, thereby solving the problems of insufficient processing of edge information and perception of image structure information by the model, and improving the generalization of the model while enhancing the de-blurring effect of the network. Therefore, the input of this device is a blurry image, and the output is a de-blurred image. The algorithm process is as follows:

[0071] Step S1: Obtain blurry-clear image data pairs, and preprocess the blurry images to obtain a blurry image feature map with an input size of 3×H×W, where H×W is the size of the input image, and 3 is the number of channels of the input blurry image.

[0072] Step S2: Perform edge detection on the blurry image using the Canny operator to obtain a relatively clear edge map F_edge. Fuse the edge map F_edge and the blurry image feature map F_b, and input them into the global pooling and Fourier transform branch to fuse the edge features into the features of the blurry image to obtain an edge-enhanced feature map F_s, as Figure 2As shown, the overall calculation process is as follows:

[0073] F_s = Conv(Avg(concate(F_b, F_edge))) + iFFT(Conv(FFT(concate(F_b, F_edge))))

[0074] Where iFFT represents the inverse fast Fourier transform, FFT represents the fast Fourier transform, and Conv represents the convolutional activation layer, which can be expressed as Conv(·) = Conv 1×1 (BatchNorm(ReLU(Conv 1×1 (·)))). Concate represents concatenating features in the channel dimension.

[0075] Step S3: Divide the blurred image features into foreground feature M1 and background feature M2, multiply the foreground feature and the background feature with the blurred image feature map respectively to obtain the foreground part and the background part of the features. Process both parts of the information through global pooling and max pooling. Global max pooling can highlight the most significant information in the feature map, thus alleviating the overfitting phenomenon to a certain extent. Global average pooling can extract the overall information of the feature map, average the information within the channels, and help reduce the sensitivity of the model to specific positions, thereby improving the generalization ability of the model. After passing through the convolutional activation layer, use the Sigmoid function to obtain the weighted features of the foreground and the background. Then fuse the foreground and background features weighted to obtain the foreground and background weighted feature map F_w, as Figure 3 As shown, the overall calculation process is as follows:

[0076] F1 = F_b * Sigmoid(Conv(concate(Avg(F_b * M1), Max(F_b * M1))))

[0077] F2 = F_b * Sigmoid(Conv(concate(Avg(F_b * M2), Max(F_b * M2))))

[0078] F_w = F1 + F2

[0079] Where the Sigmoid function maps the feature values to between (0, 1), Avg represents global average pooling, and Max represents global max pooling.

[0080] Step S4: Input the edge enhancement feature F_s and the foreground and background weighted feature F_w together with the blurred image into the Figure 4 deblurring network as shown to obtain the deblurred clear image.

[0081] In summary, the present invention proposes an image deblurring method with enhanced prior information. Based on the deblurring network, the proposed network in this method adds the feature enhancement and fusion of edge prior information and foreground-background position prior information. It can enhance the image deblurring effect while improving the generalization of the model. The present invention proposes an image edge information extraction and information enhancement module. An edge detection algorithm is used to extract the edges of the blurred image. Fourier transform and convolution pooling operations are used to fuse the frequency information of the edges, thereby improving the restoration effect of the edge contour information of the blurred image. The present invention proposes a foreground-background attention feature weighting module. The features of the blurred image are divided into foreground and background features. The foreground and background are processed separately, and the attention mechanism is used to assign different weights, thereby effectively fusing the structural prior information of the foreground and background to improve the image deblurring effect.

[0082] In the third aspect of the present invention, an electronic device is disclosed. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the image deblurring method with enhanced prior information described in the first aspect of the present disclosure.

[0083] In the fourth aspect of the present invention, a computer-readable storage medium is disclosed. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the image deblurring method with enhanced prior information described in the first aspect of the present disclosure.

[0084] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification. The above embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An image deblurring method with enhanced prior information, characterized in that: The method comprises: Step S1, calling the image edge information extraction and enhancement module, using the Canny operator to extract the edge information of the input blurred image, combining the blurred image features with the edge information, performing edge information enhancement through Fourier transform and global pooling operations, and obtaining an edge enhancement feature map for restoring the edge high-frequency information of the blurred image; Step S2, calling the foreground and background attention feature weighting module, dividing the blurred image features into foreground and background, multiplying them with the blurred image features respectively to obtain foreground features and background features, obtaining weighted maps of foreground features and background features through pooling and convolution activation operations, and then obtaining weighted feature maps of structural information of foreground and background at different weights through weighted fusion; Step S3, based on the edge enhancement feature map and the structural information weighted feature map input with different weights of the foreground and background, the input blurred image is deblurred using a deblurring network, the parameters of the deblurring network are pre-trained by blurred-clear image pairs, and the deblurring network is composed of residual convolution.

2. The image deblurring method with enhanced prior information according to claim 1, characterized in that: The deblurring network is trained using blurry-clear image pairs to obtain the optimal parameters of the deblurring network; wherein the blurry image in the blurry-clear image pair is used as the input of the deblurring network, and the loss function is calculated using the clear image and the deblurred image in the blurry-clear image pair.

3. The image deblurring method with enhanced prior information according to claim 2, characterized in that: The input blurred image is preprocessed to obtain a blurred image feature map of size 3×H×W, where H×W is the image size and 3 is the number of image channels.

4. The image deblurring method with enhanced prior information according to claim 3, characterized in that: The Canny operator is used to extract edge information from the input blurred image feature map to obtain an edge information map. The edge information is combined with the blurred image features, and the edge enhancement feature map is obtained through Fourier transform, convolution activation, global average pooling, and inverse Fourier transform.

5. The image deblurring method with enhanced prior information according to claim 4, characterized in that: The edge information map is F_edge. The edge information map F_edge and the blurred image feature map F_b are fused and input into the global pooling and Fourier transform branch to obtain the edge enhancement feature map F_s, and: F_s=Conv(Avg(concate(F_b,F_edge)))+iFFT(Conv(FFT(concate(F_b,F_edge)))) Where Conv stands for convolution activation, Avg stands for global average pooling, concate stands for concatenation, iFFT stands for inverse fast Fourier transform, FFT stands for fast Fourier transform, and Conv(·) = Conv 1×1 (BatchNorm(ReLU(Conv 1×1 (·)))), ReLU represents the activation function, and BatchNorm represents batch normalization.

6. The image deblurring method with enhanced prior information according to claim 5, characterized in that: The blurred image features are divided into foreground features M1 and background features M2. The effects of overfitting are eliminated through global pooling and maximum pooling. Global pooling extracts overall information and averages the information within the channel. After convolution activation, the weighted graph of foreground and background features is obtained using the Sigmoid function. Then, the weighted feature graph F_w of the structural information of the foreground and background at different weights is obtained through weighted fusion: F1=F_b*Sigmoid(Conv(concate(Avg(F_b*M1),Max(F_b*M1)))) F2=F_b*Sigmoid(Conv(concate(Avg(F_b*M2),Max(F_b*M2)))) F_w=F1+F2 Among them, the Sigmoid function maps the eigenvalue to (0,1), Max represents the global maximum pooling, F1 represents the foreground feature weighted map, and F2 represents the background feature weighted map.

7. The image deblurring method with enhanced prior information according to claim 6, characterized in that: The edge enhancement feature map F_s, the weighted feature map F_w of the structural information of the foreground and background with different weights, and the blurred image are input into the deblurring network to obtain the deblurred image.

8. An image deblurring system enhanced with prior information, characterized in that: The system includes an image edge information extraction and enhancement module, a foreground and background attention feature weighting module, and a deblurring network, wherein: The image edge information extraction and enhancement module is configured as follows: using the Canny operator to extract the edge information of the input blurred image, combining the blurred image features with the edge information, enhancing the edge information through Fourier transform and global pooling operations, and obtaining an edge enhancement feature map for restoring the edge high-frequency information of the blurred image; The foreground and background attention feature weighting module is configured as follows: the blurred image features are divided into foreground and background, and the blurred image features are multiplied respectively to obtain the foreground features and background features, and the weighted maps of the foreground features and background features are obtained through pooling and convolution activation operations, and then the weighted feature maps of the structural information of the foreground and background at different weights are obtained through weighted fusion; The deblurring network is used to deblur the input blurred image based on the edge enhancement feature map and the structural information weighted feature map input of the foreground and background with different weights. The parameters of the deblurring network are pre-trained by the blurred-clear image pairs, and the deblurring network is composed of residual convolution.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the image deblurring method with enhanced prior information as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the image deblurring method with enhanced prior information as described in any one of claims 1 to 7 is implemented.

Images