Image deblurring and reconstruction method and system based on image block knowledge accumulation and fusion

Abstract

The present invention belongs to the field of image deblurring and reconstruction. In order to solve the problem of dislocation often occurring at the joints when processing large images using the existing technical methods, an image deblurring and reconstruction method and system based on image block knowledge accumulation and fusion are provided. Among them, the image deblurring and reconstruction method based on image block knowledge accumulation and fusion includes dividing the blurred image into a number of blurred image blocks of set sizes; extracting shallow features of each blurred image block; performing multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then extracting and outputting the features of each blurred image block obtained in the last step as deep features; connecting the deep features and shallow features of each blurred image block through residual operations to obtain the deblurred reconstruction results of each blurred image block; and splicing the deblurred reconstruction results of all blurred image blocks to obtain a deblurred reconstructed image. It can achieve a true and clear visual experience while unifying the edges of image blocks and reducing the occurrence of dislocation.

Description

Technical Field

[0001] The present invention belongs to the field of image deblurring and reconstruction, and in particular relates to an image deblurring and reconstruction method and system based on image block knowledge accumulation and fusion. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] During the acquisition process of images (e.g., remote sensing images), they are usually blurred due to factors such as atmospheric disturbances, sensor motion, or focal misalignment, which is not conducive to accurate image analysis and application. Traditional image deblurring methods mainly include deconvolution methods, frequency domain methods, and deblurring based on prior knowledge. Among them, the deconvolution method is sensitive to noise in the image and requires an accurate estimation of the blur kernel, otherwise the opposite effect will be obtained; the frequency domain method is not ideal for complex or nonlinear blurring processes; and deblurring based on prior knowledge requires accurate prior assumptions.

[0004] In order to solve the defects of traditional image deblurring methods, the existing technology provides a deep learning deblurring model, which has improved effects on smaller images, but for larger images, they are usually cut into blocks before splicing. Since it is difficult to define a unified outline of the edges of objects in adjacent blurred image blocks, misalignment often occurs at the splicing point, affecting the application of images in the real world. Existing methods all alleviate the above problems by increasing the size of image blocks. However, while increasing the number of model parameters and improving the complexity of the model, it also poses a great challenge to computing resources. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an image deblurring and reconstruction method and system based on image block knowledge accumulation and fusion, which divides the blurred image into several blurred image blocks, processes each blurred image block and integrates the knowledge of the processed image blocks at the same time, so as to achieve the effect of unifying the edges of the image blocks and reducing the occurrence of dislocation while obtaining a real and clear visual experience.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides an image deblurring and reconstruction method based on image block knowledge accumulation and fusion.

[0008] In one or more embodiments, a method for image deblurring and reconstruction based on image block knowledge accumulation and fusion is provided, comprising:

[0009] Divide the blurred image into a number of blurred image blocks of set sizes;

[0010] Extract shallow features of each blurred image block;

[0011] Perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then use the features of each blurred image block obtained in the last step as the deep feature extraction output;

[0012] The deep features and shallow features of each blurred image block are connected through residual operation to obtain the deblurred reconstruction result of each blurred image block;

[0013] splicing the deblurred reconstruction results of all blurred image blocks to obtain a deblurred reconstructed image;

[0014] The process of knowledge accumulation and fusion at each step is as follows: the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step, and the features of the previous blurred image block in the current step are fused to obtain the features of the current blurred image block in the current step.

[0015] As an implementation method, a cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block.

[0016] As an implementation mode, the cyclic image block knowledge accumulation and fusion module is composed of knowledge accumulation and fusion units cascaded between different time steps and image blocks.

[0017] As an implementation mode, each knowledge accumulation and fusion unit includes a learnable feature extraction module, a cross attention module for knowledge conversion and fusion, a channel attention module, and a gated feedforward neural network module;

[0018] The learnable feature extraction module is used to map the input features into three difference feature expressions;

[0019] The cross attention module for knowledge conversion and fusion is used to model the dependency between the three difference feature expressions of the mapping and to fuse the knowledge of surrounding blurred image blocks into the current blurred image block using the cross attention mechanism;

[0020] The channel attention module is used to learn channel weight coefficients by using average pooling and convolution, and to assign different channel weight coefficients to the cross-attention features in the cross-attention mechanism;

[0021] The gated feedforward neural network module is used to filter redundant information and generate features of the current blurred image block at the current step.

[0022] As an implementation method, a shallow feature extraction and dependency building module is used to extract shallow features of each blurred image block; the shallow feature extraction and dependency building module is composed of a single convolution and an STB module.

[0023] As an implementation method, the blurred image is divided into a number of blurred image blocks of set sizes in a sliding window manner.

[0024] A second aspect of the present invention provides an image deblurring and reconstruction system based on image block knowledge accumulation and fusion.

[0025] In one or more embodiments, an image deblurring and reconstruction system based on image block knowledge accumulation and fusion includes:

[0026] A blurred image segmentation unit, which is used to segment the blurred image into a plurality of blurred image blocks of a set size;

[0027] A shallow feature extraction unit, which is used to extract shallow features of each blurred image block;

[0028] A knowledge accumulation and fusion unit is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then extract and output the features of each blurred image block obtained in the last step as deep features;

[0029] A deep-shallow feature fusion unit, which is used to connect the deep features and shallow features of each blurred image block through a residual operation to obtain a deblurred reconstruction result of each blurred image block;

[0030] A deblurring and reconstruction unit, which is used to splice the deblurring and reconstruction results of all blurred image blocks to obtain a deblurred and reconstructed image;

[0031] Among them, in the knowledge accumulation and fusion unit, the process of each step of knowledge accumulation and fusion is: fusing the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step and the features of the previous blurred image block in the current step to obtain the features of the current blurred image block in the current step.

[0032] As an implementation mode, in the knowledge accumulation and fusion unit, a cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block.

[0033] As an implementation mode, the cyclic image block knowledge accumulation and fusion module is composed of knowledge accumulation and fusion units cascaded between different time steps and image blocks.

[0034] As an implementation mode, each knowledge accumulation and fusion unit includes a learnable feature extraction module, a cross attention module for knowledge conversion and fusion, a channel attention module, and a gated feedforward neural network module;

[0035] The learnable feature extraction module is used to map the input features into three difference feature expressions;

[0036] The cross attention module for knowledge conversion and fusion is used to model the dependency between the three difference feature expressions of the mapping and to fuse the knowledge of surrounding blurred image blocks into the current blurred image block using the cross attention mechanism;

[0037] The channel attention module is used to learn channel weight coefficients by using average pooling and convolution, and to assign different channel weight coefficients to the cross-attention features in the cross-attention mechanism;

[0038] The gated feedforward neural network module is used to filter redundant information and generate features of the current blurred image block at the current step.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention is suitable for deblurring and reconstructing image blocks of large images. It accumulates and fuses the shallow features of each blurred image block in multiple steps. While processing each blurred image block, the prior knowledge of surrounding image blocks and processed image blocks is fused into the current image block. This is conducive to enhancing the directional texture of the image and restoring clear and natural local details. It ensures that the edges of the image blocks are aligned during splicing, reduces misalignment, and improves the effect of reconstructing the image.

[0041] (2) The present invention utilizes shallow feature extraction and dependency building modules to extract shallow features of each blurred image block, and first models the image block features through internal window attention and cross-window attention, so that the cyclic image block knowledge accumulation and fusion module can focus on processing the transfer and fusion of texture features between image blocks, thereby improving the real and clear visual experience of the deblurred reconstructed image. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] Figure 1 It is a flowchart of an image deblurring and reconstruction method based on image block knowledge accumulation and fusion according to an embodiment of the present invention;

[0044] Figure 2 It is the overall framework of the image deblurring and reconstruction process based on image block knowledge accumulation and fusion according to the embodiment of the present invention;

[0045] Figure 3 4 is a structural diagram of a shallow feature extraction (MSTB, ModifiedSwinTransformerBlock) module according to an embodiment of the present invention;

[0046] Figure 4 4 is a structural diagram of a recurrent image patch knowledge accumulation and fusion module (RIPKAF, Image Patch Knowledge Accumulation and Fusion) according to an embodiment of the present invention;

[0047] Figure 5 is a structural diagram of a knowledge accumulation and fusion unit (KAF) according to an embodiment of the present invention;

[0048] Figure 6 is a structural diagram of a learnable feature extraction module (FE) according to an embodiment of the present invention;

[0049] Figure 7 is a structural diagram of an attention module (CSA, Cross-Scale Attention) for knowledge conversion and fusion according to an embodiment of the present invention;

[0050] Figure 8 is a structural diagram of a channel attention module (CA) according to an embodiment of the present invention;

[0051] Fig. 9 is a structural diagram of a gate feedforward neural network module (SG, Simple Gate) of an embodiment of the present invention;

[0052] Fig.10 It is a schematic diagram of the structure of an image deblurring and reconstruction system based on image block knowledge accumulation and fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0054] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0055] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0056] Figure 1FIG. 1 is a flow chart of an image deblurring and reconstruction method based on image block knowledge accumulation and fusion in an embodiment of the present invention. Figure 1 The image deblurring and reconstruction method based on image block knowledge accumulation and fusion in the embodiment shown may include:

[0057] S101, dividing the blurred image into a plurality of blurred image blocks of set sizes;

[0058] S102, extracting shallow features of each blurred image block;

[0059] S103, performing multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then extracting and outputting the features of each blurred image block obtained in the last step as deep features;

[0060] S104, connecting the deep features and shallow features of each blurred image block through a residual operation to obtain a deblurred reconstruction result of each blurred image block;

[0061] S105, splicing the deblurred reconstruction results of all blurred image blocks to obtain a deblurred reconstructed image;

[0062] Among them, in step S103, the process of knowledge accumulation and fusion in each step is: the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step and the features of the previous blurred image block in the current step are fused to obtain the features of the current blurred image block in the current step.

[0063] This embodiment is suitable for deblurring and reconstruction of image blocks of large images. It accumulates and fuses the shallow features of each blurred image block in multiple steps, and fuses the knowledge priors of surrounding image blocks and processed image blocks into the current image block while processing each blurred image block. This is conducive to enhancing the directional texture of the image, restoring clear and natural local details, ensuring the consistent alignment of the image block edges during splicing, reducing misalignment, and improving the effect of reconstructing the image.

[0064] It should be noted here that the processes in step S102 to step S105 can be executed in the image deblurring reconstruction model Image Patch Knowledge Accumulation and Fusion Network (IPKAF-Net).

[0065] Specifically, the image deblurring and reconstruction model consists of a shallow feature extraction and dependency building module, a cyclic image block knowledge accumulation and fusion module, and an image reconstruction module. The overall process is as follows: Figure 2 shown.

[0066] Among them, the shallow feature extraction and dependency establishment module executes the process in S102.

[0067] The cyclic image block knowledge accumulation and fusion module executes the process in S103.

[0068] The image reconstruction module executes the process in S104~S105.

[0069] In the process of training the image deblurring and reconstruction model, a sample set of high-definition and blurred image data pairs is used for training.

[0070] The process of obtaining the HD-blurred image data pair is as follows:

[0071] The acquired high-definition image data , a second-order degradation pipeline simulating the real-world degradation process is used to obtain the corresponding low-quality blurred image data , each order of degradation includes low-pass filtering ,noise , downsampling operation as well as JPEG Compression, etc. Among them, the first-order degenerate pipeline process is shown in the following formula:

[0072]

[0073] In the process of training the image deblurring and reconstruction model, we use L 1 The loss function evaluates the pixel-level difference between the reconstructed result and the real result to ensure the consistency between the reconstructed result and the real result. The Canny edge detection algorithm is also used in the experiment to detect the edges of the reconstructed result and the real result respectively, aiming to ensure that adjacent reconstructed image blocks have a good degree of alignment at the edges.

[0074] use L 1 loss function evaluates the reconstruction result and the real result to reduce the pixel difference; at the same time, Canny The edge detection algorithm detects the edges of the reconstructed result and the real result and evaluates the detection results to reduce the possibility of misalignment. Real images and reconstruct the image , the loss function can be defined as:

[0075]

[0076]

[0077] in, express Canny Edge detection algorithm function; is the L1 loss function.

[0078] Specifically, in step S101, the blurred image is divided into a number of blurred image blocks of a set size in a sliding window manner. For example, the blurred image is divided into a number of 256 256-sized image blocks.

[0079] In this embodiment, in step S102, shallow features of each blurred image block are extracted using a shallow feature extraction and dependency building module; the shallow feature extraction and dependency building module (ModifiedSwinTransformerBlock, MSTB) is composed of a single convolution and a Swin Transformer Block (STB) based on window attention, such as Figure 3 Among them, Swin Transformer Block is a commonly used module, which is proposed in the paper Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.

[0080] For example, every 3 3 blurred image patches is a set of inputs for shallow feature extraction and dependency building modules. Each blurred image block is processed once convolution and STB module Get the corresponding shallow features :

[0081]

[0082] In step S103, the cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block. , and use them as the input of the Recurrent Image Patch Knowledge Accumulation and Fusion (RIPKAF) module to establish the dependency of adjacent image patches, such as Figure 4 shown.

[0083] The cyclic image block knowledge accumulation and fusion module RIPKAF is mainly composed of knowledge accumulation and fusion units (KAF) cascaded between different time steps and image blocks. The generation of each new image block feature at each time step is associated with the current image block feature at the previous time step. , the feature of the previous image block at the previous time step And the image block features before the current time step RIPKAF improves the current feature expression and ensures edge consistency through texture migration and fusion.

[0084]

[0085] Specifically, Figure 5 As shown, each knowledge accumulation and fusion unit KAF includes a learnable feature extraction module FE, a cross attention module CSA for knowledge conversion and fusion, a channel attention module CA, and a gated feedforward neural network module SG;

[0086] The learnable feature extraction module FE is used to map the input features into three difference feature expressions;

[0087] The cross attention module CSA for knowledge conversion and fusion is used to model the dependency between the three difference feature expressions of the mapping and use the cross attention mechanism to fuse the knowledge of surrounding blurred image blocks into the current blurred image block;

[0088] The channel attention module CA is used to learn channel weight coefficients by using average pooling and convolution, and to assign different channel weight coefficients to the cross-attention features in the cross-attention mechanism;

[0089] The gated feedforward neural network module SG is used to filter redundant information and generate features of the current blurred image block at the current step.

[0090] Specifically, the learnable feature extraction module FE adopts a multi-branch structure, each branch contains convolutions of different sizes, so that it can capture image details and cover a wider receptive field, which is crucial for remote sensing data. Figure 6 As shown, the FE module Mapping matrix and learn the optimal parameter configuration in end-to-end training. For example, first pass convolution , convolution , convolution Capture image features of different scales, and then use channel dimensions Operation to perform feature splicing, and finally pass once convolution and Activate to obtain the mapping result. This process can be expressed as:

[0091]

[0092]

[0093]

[0094]

[0095] The attention module CSA used for knowledge conversion and fusion takes the features of the current image block at the previous time step as the query. , the feature of the previous image block in the previous time step is the key value , the feature of the previous image block at the current time step is value ,like Figure 7 As shown in the figure, the knowledge prior of the processed image block is integrated into the features of the current image block by cross attention to obtain the output , which is conducive to reconstructing clear texture details and ensuring the coherence of image block edges.

[0096]

[0097] in, is the normalized exponential function; Represents the size of feature dimension; Relative position encoding.

[0098] like Figure 8 As shown, the channel attention module CA is implemented by average pooling And convolution operation To learn the channel weight coefficient , weight the importance of different channels, enhance key features, and suppress redundant information. Finally, the residual connection operation is used to prevent feature forgetting and obtain the output feature The process is shown in the formula.

[0099]

[0100]

[0101] The gated feedforward neural network SG dynamically adjusts the transmission of information flow by introducing gate control units, thereby improving the model's expressiveness and ability to process complex data. Fig. 9 As shown, SG mainly includes layer normalization operations , convolution operation And the gate control unit. The gate control unit uses the split operation The feature map is transformed along the channel dimension Divided into and Two parts, then fuse the two parts of features in the channel dimension, and finally combine them into one convolution And residual connection to promote the effective combination of deep features and shallow features, improve the accuracy of image block feature extraction and reconstruction. This process can be expressed as:

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] in, is the layer normalization operation Result; It is a convolution operation Result; It is fusion in the channel dimension and Results in two parts.

[0108] The multi-image patch features obtained in the last time step are As the output of the RIPKAF module, it passes through the MSTB module again And a convolution Strengthen the internal dependencies of image blocks to obtain deep features .

[0109] In step S104, the deep features With shallow features Perform the residual connection operation to obtain the deblurred reconstructed image block .

[0110]

[0111]

[0112] In step S105, all reconstructed image blocks are spliced ​​to obtain the deblurred remote sensing image. .

[0113] Fig.10 is a schematic diagram of the structure of an image deblurring and reconstruction system based on image block knowledge accumulation and fusion in an embodiment of the present invention. Figure 1 The image deblurring and reconstruction method based on image block knowledge accumulation and fusion corresponds to Fig.10 As shown, the image deblurring and reconstruction system based on image block knowledge accumulation and fusion in this embodiment may include:

[0114] A blurred image segmentation unit 1001 is used to segment the blurred image into a plurality of blurred image blocks of a set size;

[0115] A shallow feature extraction unit 1002, which is used to extract shallow features of each blurred image block;

[0116] The knowledge accumulation and fusion unit 1003 is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then extract and output the features of each blurred image block obtained in the last step as deep features;

[0117] A deep-shallow feature fusion unit 1004 is used to connect the deep features and the shallow features of each blurred image block through a residual operation to obtain a deblurred reconstruction result of each blurred image block;

[0118] A deblurring and reconstruction unit 1005 is used to stitch together the deblurring and reconstruction results of all blurred image blocks to obtain a deblurred and reconstructed image;

[0119] Among them, in the knowledge accumulation and fusion unit 1003, the process of each step of knowledge accumulation and fusion is: fusing the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step and the features of the previous blurred image block in the current step to obtain the features of the current blurred image block in the current step.

[0120] In the knowledge accumulation and fusion unit 1003, a cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block.

[0121] The cyclic image block knowledge accumulation and fusion module is composed of knowledge accumulation and fusion units cascaded between different time steps and image blocks.

[0122] Specifically, each knowledge accumulation and fusion unit includes a learnable feature extraction module, a cross attention module for knowledge conversion and fusion, a channel attention module, and a gated feedforward neural network module;

[0123] The learnable feature extraction module is used to map the input features into three difference feature expressions;

[0124] The cross attention module for knowledge conversion and fusion is used to model the dependency between the three difference feature expressions of the mapping and to fuse the knowledge of surrounding blurred image blocks into the current blurred image block using the cross attention mechanism;

[0125] The channel attention module is used to learn channel weight coefficients by using average pooling and convolution, and to assign different channel weight coefficients to the cross-attention features in the cross-attention mechanism;

[0126] The gated feedforward neural network module is used to filter redundant information and generate features of the current blurred image block at the current step.

[0127] In the shallow feature extraction unit 1002, shallow features of each blurred image block are extracted using a shallow feature extraction and dependency building module; the shallow feature extraction and dependency building module is composed of a single convolution and a STB module.

[0128] In the blurred image segmentation unit 1001, the blurred image is segmented into a plurality of blurred image blocks of a set size in a sliding window manner.

[0129] It should be noted here that the various units in the image deblurring and reconstruction system based on image block knowledge accumulation and fusion correspond one-to-one to the various steps in the above-mentioned image deblurring and reconstruction method based on image block knowledge accumulation and fusion, and the specific implementation process is the same, which will not be described in detail here.

[0130] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer readable medium, the computer program including a computer program for executing Figure 1 In such an embodiment, the computer program can be downloaded and installed from a network through a communication part, and / or installed from a removable medium. When the computer program is executed by a central processing unit, various functions defined in the device of the present application are executed.

[0131] in, Figure 1 The computer program instructions corresponding to the method shown may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0132] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An image deblurring and reconstruction method based on image block knowledge accumulation and fusion, characterized in that: include: Divide the blurred image into a number of blurred image blocks of set sizes; Extract shallow features of each blurred image block; Perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then use the features of each blurred image block obtained in the last step as the deep feature extraction output; The deep features and shallow features of each blurred image block are connected through residual operation to obtain the deblurred reconstruction result of each blurred image block; splicing the deblurred reconstruction results of all blurred image blocks to obtain a deblurred reconstructed image; The process of knowledge accumulation and fusion at each step is as follows: the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step, and the features of the previous blurred image block in the current step are fused to obtain the features of the current blurred image block in the current step; In the knowledge accumulation and fusion unit, in each step of knowledge accumulation and fusion, firstly, the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step, and the features of the previous blurred image block in the current step are mapped into three difference feature expressions; then, the dependency relationship between the three mapped difference feature expressions is modeled and the knowledge of the surrounding blurred image blocks is fused into the current blurred image block by using the cross-attention mechanism; then, average pooling and convolution are used to learn the channel weight coefficients and different channel weight coefficients are assigned to the cross-attention features in the cross-attention mechanism; finally, redundant information is filtered to generate the features of the current blurred image block in the current step.

2. The image deblurring and reconstruction method based on image block knowledge accumulation and fusion as claimed in claim 1, characterized in that: The cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block.

3. The image deblurring and reconstruction method based on image block knowledge accumulation and fusion as claimed in claim 2, characterized in that: The cyclic image block knowledge accumulation and fusion module is composed of knowledge accumulation and fusion units cascaded between different time steps and image blocks.

4. The image deblurring and reconstruction method based on image block knowledge accumulation and fusion as claimed in claim 1, characterized in that: A shallow feature extraction and dependency building module is used to extract shallow features of each blurred image block; the shallow feature extraction and dependency building module is composed of a single convolution and an STB module.

5. The image deblurring and reconstruction method based on image block knowledge accumulation and fusion as claimed in claim 1, characterized in that: The blurred image is divided into several blurred image blocks of set sizes in a sliding window manner.

6. An image deblurring and reconstruction system based on image block knowledge accumulation and fusion, characterized in that: include: A blurred image segmentation unit, which is used to segment the blurred image into a plurality of blurred image blocks of a set size; A shallow feature extraction unit, which is used to extract shallow features of each blurred image block; A knowledge accumulation and fusion unit is used to perform multi-step knowledge accumulation and fusion on the shallow features of each blurred image block, and then extract and output the features of each blurred image block obtained in the last step as deep features; A deep-shallow feature fusion unit, which is used to connect the deep features and shallow features of each blurred image block through a residual operation to obtain a deblurred reconstruction result of each blurred image block; A deblurring and reconstruction unit, which is used to splice the deblurring and reconstruction results of all blurred image blocks to obtain a deblurred and reconstructed image; Wherein, in the knowledge accumulation and fusion unit, the process of each step of knowledge accumulation and fusion is: fusing the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step, and the features of the previous blurred image block in the current step to obtain the features of the current blurred image block in the current step; In the knowledge accumulation and fusion unit, in each step of knowledge accumulation and fusion, firstly, the features of the current blurred image block in the previous step, the features of the previous blurred image block in the previous step, and the features of the previous blurred image block in the current step are mapped into three difference feature expressions; then, the dependency relationship between the three mapped difference feature expressions is modeled and the knowledge of the surrounding blurred image blocks is fused into the current blurred image block by using the cross-attention mechanism; then, average pooling and convolution are used to learn the channel weight coefficients and different channel weight coefficients are assigned to the cross-attention features in the cross-attention mechanism; finally, redundant information is filtered to generate the features of the current blurred image block in the current step.

7. The image deblurring and reconstruction system based on image block knowledge accumulation and fusion as claimed in claim 6, characterized in that: In the knowledge accumulation and fusion unit, a cyclic image block knowledge accumulation and fusion module is used to perform multi-step knowledge accumulation and fusion on the shallow features of each fuzzy image block.

8. The image deblurring and reconstruction system based on image block knowledge accumulation and fusion as claimed in claim 7, characterized in that: The cyclic image block knowledge accumulation and fusion module is composed of knowledge accumulation and fusion units cascaded between different time steps and image blocks.

Images