SAR image intelligent refocusing method

Abstract

The invention discloses an SAR intelligent refocusing method based on deep learning. The method mainly solves the problems that in traditional SAR image refocusing, phase error modeling is complex, the image resolution is low, and details are lost. According to the implementation scheme, the method comprises the following steps: acquiring a data set of an SAR pure image and a corresponding defocused image, and dividing the data set into a training set, a verification set and a test set; constructing a deep learning network comprising an encoder and a decoder; using the training set to carry out forward and back propagation iteration to train the network, and evaluating and updating model parameters on the verification set to obtain a trained network model; and inputting a defocused image in a test set into the trained network model, extracting image features through an encoder, and reconstructing a high-resolution image through a decoder to obtain a restored image result. According to the method, the deep features of the SAR image can be effectively extracted, gradient disappearance is relieved, features of different levels are effectively fused, the reconstruction quality and resolution of the SAR image are improved, the detail performance of the restored image is enhanced, and the method can be used for radar imaging and evaluation.

Description

Technical Field

[0001] The invention belongs to the field of signal and information processing, and in particular relates to a radar refocusing method, which can be used for radar imaging and evaluation. Background Art

[0002] Synthetic aperture radar SAR is an active microwave remote sensing system with all-day and all-weather observation capabilities. With the increasing demand for military and civilian use, higher requirements are also placed on the resolution of SAR images. The higher the resolution of SAR images, the richer the image information they contain. During the image acquisition process, factors such as system errors caused by the hardware facilities of the imaging equipment and the relative motion between the detector target and the imaging device will cause image quality degradation, which is manifested as image blur, obvious noise and reduced resolution. Therefore, it is necessary to refocus the defocused image to improve the clarity of the image. There are usually two reasons for the defocusing of SAR images. The first is atmospheric interference, such as changes in the atmospheric refractive index affecting the propagation path and phase of the wave. The second is motion error. The complex motion state of the imaging target during SAR imaging, the change in relative motion speed, the non-ideal motion of the platform, small accelerations, speed changes or inaccurate attitude control will all produce phase errors, resulting in image defocusing. Phase error will reduce image resolution, blur the edge of the target, and affect the target positioning accuracy and shape recovery. This error will have a significant impact in high-precision detection applications such as target recognition or navigation. Therefore, it is very necessary to use advanced signal processing technology to compensate for defocused images and improve the resolution of SAR images.

[0003] Existing SAR image refocusing methods include visual misalignment MDA, phase gradient method PGA, minimum entropy algorithm MEA and deep learning DL. Among them, visual misalignment has weak detection capabilities for static targets, will produce false targets in high-speed targets or complex scenes, and is sensitive to noise; the phase gradient algorithm is sensitive to large-scale scene changes and background interference, has high computational complexity, and poor real-time performance. If there are two adjacent strong scattering points, phase aliasing will occur, resulting in the inability to accurately distinguish the true position of the target; the minimum entropy algorithm may converge to a local optimal solution rather than a global optimal solution, which is particularly obvious in complex environments, resulting in unsatisfactory imaging effects, and the algorithm performance often depends on parameter settings and needs to be optimized through experiments. Deep learning can perform automatic feature learning and high-precision imaging. The model can improve the overall performance of the algorithm through large-scale data training and optimization, and overcome the limitations of traditional algorithms.

[0004] The patent document with application number: CN202311764464.5 discloses a "synthetic aperture radar image self-focusing method, device and equipment", which determines the target focus position corresponding to the offset pixel point from the preset pixel point set according to the initial position of the offset pixel point of the image to be processed, and performs phase shift on the offset pixel point according to the target focus position of the offset pixel point and the initial position of the offset pixel point to obtain the target image. Since this method needs to process multiple offset pixel points and preset pixel points, the amount of calculation is large, which will lead to a longer processing time; and the selection of preset step size and other parameters will affect the focusing result.

[0005] The patent document with application number: CN202311651855.6 discloses "a method for adaptively judging whether to perform PGA processing for high squint SAR imaging", which uses radar echoes to adaptively judge whether to perform phase gradient autofocus PGA processing. Before improving image quality through PGA phase compensation, this method needs to adaptively judge whether to perform PGA processing based on actual radar echo data after processing and analyzing the echo data. This invention collects radar echo data based on real-time terrain, adaptively judges whether to perform PGA processing, and obtains the best image quality. The threshold selected by this method may have a significant impact on the judgment result. If the detection of strong scattering points is inaccurate, it will affect the final image quality.

[0006] In recent years, with the development of deep learning, many radar refocusing algorithms based on deep learning have been proposed, and the effectiveness of deep learning methods in the field of SAR refocusing has been verified one after another. In current research, radar images with known phase errors and corresponding ideal images are usually used as training data, and mainstream generative models such as CNN, GAN, and Transformer are used to extract features, and then the network is optimized through a specific loss function. For example, the patent document with application number: CN202110772455.5 discloses a "SAR self-focusing method based on deep learning", and proposes a network based on deep expansion and convolutional neural network for learning, which is used to solve the problem of poor focusing effect and long time consumption of SAR images with space-varying errors in traditional methods. The implementation scheme is: first obtain multiple coarsely focused SAR images, and randomly divide them into training sets and test sets according to a certain ratio to facilitate model training and verification; then design and build a SAR self-focusing neural network model composed of eight convolutional layers and four deconvolutional layers in cascade, and train it; input the coarsely focused SAR images in the test set into the trained network model to obtain the focused SAR image, and evaluate the focusing effect and focusing speed to ensure that the phase error caused by motion error can be accurately estimated. Although the present invention can improve the imaging quality and focusing efficiency of SAR images. However, since the phase error must first be explicitly modeled during the above processing, and then the error model parameters are estimated by self-supervision, there is a problem that the parameters are relatively single and can only focus on specific types of defocused images, and more complex phase errors cannot be processed. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a SAR intelligent refocusing method based on deep learning to avoid explicit modeling of radar phase error, directly learn the mapping from defocused data to focused imaging, realize the processing of more complex phase errors, make the image refocusing effect more accurate, and thus improve the image resolution.

[0008] To achieve the above object, the technical solution of the present invention includes the following:

[0009] 1. A SAR non-parametric intelligent generation refocusing method, characterized in that it includes the following steps:

[0010] (1) Obtain a SAR image dataset, preprocess it, and divide it into a training set, a validation set, and a test set;

[0011] (2) Building a deep learning network

[0012] (2a) establishing an encoder including a stack of multiple residual blocks for feature extraction of defocused images;

[0013] (2b) establishing a decoder including multiple upsampling layers and residual blocks alternately stacked to perform feature fusion and obtain a reconstructed image;

[0014] (2c) selecting a single-layer residual block, connecting the encoder and the decoder through the single-layer residual block, and connecting the output of each residual block in the encoder with the upsampling layer output of its corresponding decoder to form a deep learning network, and setting its loss function;

[0015] (3) Input the training set data into the deep learning network, perform forward propagation and back propagation, iteratively update the network parameters, and obtain a trained deep learning network;

[0016] (4) The test set data is input into the trained deep learning network to extract the defocused image features and generate a restored SAR image to complete the refocusing of the defocused image.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] First, traditional SAR imaging often relies on complex mathematical modeling, while the present invention adopts a non-parametric approach to learn defocused image features through a deep learning network, which is independent of a specific imaging model and simplifies the modeling process.

[0019] Secondly, the encoder of the present invention adopts a multi-layer residual block stacking structure, which can effectively extract the deep features of the SAR image, alleviate the gradient disappearance problem, and improve the accuracy and speed of SAR image refocusing.

[0020] Thirdly, the decoder of the present invention can effectively fuse features at different levels and improve the reconstruction quality and resolution of SAR images by alternately stacking upsampling layers and residual blocks.

[0021] Fourthly, the present invention utilizes skip connections and multi-level feature splicing, which not only retains important shallow detail features, but also effectively avoids the loss of high-level information, making the restored SAR image clearer and richer in details.

[0022] Fifthly, the structures of the encoder and decoder of the present invention fully utilize the advantages of deep feature extraction and multi-level information fusion, thereby greatly improving the accuracy of SAR image refocusing and restoration effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart for realizing the present invention;

[0024] Figure 2 This is a schematic diagram of the deep learning network structure in the present invention;

[0025] Figure 3 This is a schematic diagram of the structure of the head residual block of the Chinese encoder of the present invention;

[0026] Figure 4 Schematic diagram of the residual block structure of the decoder and the residual block structure of the encoder excluding the header;

[0027] Figure 5 This is a diagram showing the result of refocusing a defocused image using the present invention. DETAILED DESCRIPTION

[0028] The focus of the present invention is to use a deep learning network to directly learn the mapping of defocused data to focused imaging without explicitly modeling the phase error, and to propose an effective radar refocusing method.

[0029] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0030] Reference Figure 1 The implementation steps of this example include the following:

[0031] Step 1: Obtain a SAR defocused-focused image dataset.

[0032] 1.1) Obtain echo signals s(t) of river, ship and coast scenes from public websites:

[0033]

[0034] Where A is the amplitude of the signal, 2πf 0 t is the fundamental frequency part of the signal, φ(t) is the time-dependent phase;

[0035] 1.2) Use the range Doppler algorithm to convert echo data into a pure image:

[0036] First, perform a range-direction fast Fourier transform on each echo signal s(t) to obtain the distance information of each target and generate a range-slow time two-dimensional matrix;

[0037] Then, the range-slow time matrix is ​​subjected to fast Fourier transform in the Doppler direction to obtain a range-Doppler image;

[0038] Finally, the range-Doppler matrix is ​​interpolated to obtain a clear image, and the echo data is converted into a two-dimensional image of range-azimuth, i.e., a pure SAR image;

[0039] 1.3) Introduce phase interference Δφ(t)=C into the echo signal 0 t 2 +C 1 t 3 +C 2 t 4The interference echo signal is obtained, where the azimuth frequency t∈[-0.5,0.5] and the coefficient C 0 , C 1 , C 2 ∈[-5,5]×30, respectively representing the contribution of the quadratic error term, cubic error term, and quartic error term to the phase error;

[0040] 1.4) Load the phase interference Δφ(t) into the phase of the received echo signal s(t) to obtain the interference echo signal s err (t):

[0041]

[0042] 1.5) Using the range Doppler algorithm to image the interference echo data, a defocused SAR image is obtained;

[0043] 1.6) The data of the clean image and the corresponding defocused image are combined into a dataset, and divided into training set, validation set and test set according to 6:2:2. The clean image in the dataset is the label during network training, which is used to guide the network to learn the ideal output. The defocused image in the dataset is the input of network training, which consists of real part and imaginary part and is normalized to help the network better learn and fit the mapping relationship between defocused image and clean image.

[0044] Step 2: Build a deep learning network.

[0045] Reference Figure 2 , the specific implementation of this step is as follows:

[0046] 2.1) Establish an encoder consisting of multiple residual blocks stacked together:

[0047] 2.1.1) Create a head residual block consisting of a convolutional block and a residual connection, such as Figure 3 As shown in the figure. The convolution block contains two layers of 3×3 convolution, and batch normalization and ReLU activation operations are performed after the first layer of convolution; the residual connection performs 3×3 convolution on the input and adjusts the output dimension through batch normalization to ensure that the output dimension of the convolution block and the residual connection is consistent; the output of the convolution block and the residual connection is added to form the output of the head residual block, which is the input of the next layer of residual blocks;

[0048] This structure can initially extract features and effectively transmit information, laying the foundation for the subsequent feature extraction process; after the head residual block processing, the number of feature channels of the input image increases to 32;

[0049] 2.1.2) Establish two layers of residual blocks with the same structure, such as Figure 4As shown in the figure. Each residual block contains a convolution block and a residual connection. Each convolution block contains two 3×3 convolution layers. Batch normalization and ReLU activation operations are performed before each convolution layer. The convolution block processes the image through a 3×3 convolution kernel and downsamples it with a stride of 2 to reduce the spatial size and extract higher-level features.

[0050] The first layer of residual blocks has 32 input channels and 64 output channels. Through convolution block processing, the spatial size of the image is reduced and the number of feature channels is increased, thereby extracting deeper features.

[0051] The function of the second layer residual block is similar to that of the first layer, but the spatial size is further reduced and the number of channels continues to increase to 128, which can extract deeper features;

[0052] The parameters and output size of each convolutional layer in the above residual block are shown in Table 1:

[0053] Table 1 Parameters and output size of each convolutional layer of the encoder

[0054]

[0055] The residual connection of each residual block performs a 3×3 convolution on the input and adjusts the dimension through batch normalization. Finally, the outputs of the convolutional blocks and residual connections are added together to obtain the final output of each layer.

[0056] 2.1.3) The above-mentioned head residual block and the two residual blocks are connected layer by layer, and the output of each residual block is used as the input of the next residual block to form a deep hierarchical encoder structure. The encoder can effectively restore the spatial resolution of the image through upsampling operations and extract shallow features at the same time, thereby providing rich feature representation for subsequent tasks.

[0057] 2.2) Establish a decoder containing multiple upsampling layers and residual blocks stacked alternately:

[0058] 2.2.1) The upsampling layer is constructed by a transposed convolution operation with a stride of 2, doubling the spatial size;

[0059] 2.2.2) Create a residual block consisting of a convolutional block and a residual connection, such as Figure 4 As shown in the figure. The convolution block contains two 3×3 convolution layers, and batch normalization and ReLU activation operations are performed before each convolution layer; the residual connection performs 3×3 convolution on the input and adjusts the dimension through batch normalization; finally, the outputs of the convolution block and the residual connection are added to obtain the output of each residual block, which is the input of the next upsampled layer;

[0060] 2.2.3) The processed output features of the encoder are used as the input of the first layer of upsampling. The shallow features in the encoder are concatenated with the corresponding upsampled features through skip connections to form the input of the first layer of residual blocks. The output features of the residual blocks are passed to the next layer of upsampling, and are processed by the second layer of upsampling and residual blocks, and the third layer of upsampling and residual blocks in turn. Through further processing and information fusion of these residual convolution blocks, the spatial details and high-resolution features of the image are gradually restored, and the number of feature map channels is gradually reduced to 32 through three upsampling blocks.

[0061] Each residual block is expressed as:

[0062] O i =h(I i )+F(I i ; W i )

[0063] Where Ii, Oi are the input and output of the ith residual block, respectively, F(·) is the mapping function of the residual block, W is the parameter of the ith residual block, and h(·) is the unit mapping function.

[0064] The parameters and output size of each convolutional layer in the above residual block are shown in Table 2:

[0065] Table 2 Parameters and output size of each convolutional layer of the decoder

[0066]

[0067] 2.2.4) The above three upsampling layers and three residual blocks are alternately connected, and the output of the third residual block passes through a 1×1 convolutional layer and a Sigmoid activation function to reduce the number of feature channels to 1, forming the overall structure of the decoder.

[0068] 2.3) Establish a bridge part consisting of a convolutional block and a residual connection. The convolutional block contains two 3×3 convolutional layers, and batch normalization and ReLU activation operations are performed before each convolutional layer, where:

[0069] In the first convolution process, downsampling is performed with a stride of 2, reducing the size of the space by half;

[0070] In the second convolution process, the number of channels of the output feature map is kept unchanged by operating with a stride of 1, and deeper features are further extracted;

[0071] 2.4) The encoder, bridge part and decoder are cascaded in sequence to form a deep learning network. The input defocused image first passes through the encoder to extract image features, then passes through the bridge part to further extract and fuse the shallow features of the encoder, and finally inputs it into the decoder to gradually restore the details of the image and generate the final output restored image.

[0072] In this example, the number of feature map channels of the defocused image after the encoder extracts image features is 128, the number of feature map channels after the bridging part is processed is 256, and finally the number of feature map channels after the decoder is processed is 1.

[0073] Step 3: Set up the deep learning network L.

[0074] This example sets the existing mean absolute error MAE as the loss function L, which is expressed as follows:

[0075]

[0076] Among them, H and M are the height and width of the image respectively, and y(i,j) represents the pixel value of the pure image at position (i,j). Represents the pixel value of the restored image at position (i, j); is the absolute error between the clean image and the restored image at position (i, j).

[0077] Step 4: Train the deep learning network.

[0078] 4.1) Input the defocused images in the training set into the deep learning network;

[0079] 4.2) Set the initial learning rate η of this network to 0.001 and the maximum number of iterations to 300;

[0080] 4.3) In the forward propagation stage, the defocused image input into the network is passed through the convolutional layer in the encoder to extract the image features x, and the skip connection is used to enhance the retention of details;

[0081] 4.4) The input feature x is calculated through the convolution layer to obtain the new feature x*=W×x+b, which is used as the input of the next convolution layer, where the weight matrix W and bias b are independent parameters of each convolution layer;

[0082] 4.5) Each convolution layer of the encoder and decoder repeats step 4.4), so that the network gradually extracts and reconstructs the features of the image. After multi-layer convolution calculations in the decoder, the network finally generates the image restoration result.

[0083] 4.6) Using the restored image results The absolute error between the pure image y and the calculated loss function L;

[0084] 4.7) In the back propagation phase, the gradient of the loss function L for each weight matrix W and bias b is calculated separately and The Adam optimizer is used to dynamically adjust the learning rate so that the learning rate changes dynamically according to the mean and variance of the gradient.

[0085] 4.8) Update the model parameters according to the gradient direction and the adjusted learning rate η:

[0086]

[0087] Where W′ is the updated weight matrix and t′ is the updated bias;

[0088] 4.9) Evaluate model performance on the validation set.

[0089] 4.9.1) Calculate the peak signal-to-noise ratio (PSNR) of each restored image and clean image:

[0090]

[0091] Where MAX represents the maximum value of the image pixels. For two n×n monochrome images, MSE is defined as:

[0092]

[0093] Among them, y(i,j) is the grayscale value of the pixel in the i-th row and j-th column of the original image. is the gray value of the pixel in the i-th row and j-th column of the restored image;

[0094] 4.9.2) Calculate the structural similarity SSIM of each restored image and the pure image:

[0095]

[0096] Among them, l, c, and s compare the brightness, contrast, and structure of the two images respectively; α, β, and γ represent the proportion of different features in the SSIM measurement, all greater than 0; μ I and μ O Represent the mean of the two images respectively; σ I and σ O Respectively represent the standard deviation of the two images; σ IO represents the covariance between the two images; C 1 , C 2 , C 3 are constants that keep l, c, and s stable; y(i,j) is the grayscale value of the pixel in the i-th row and j-th column of the original image. is the gray value of the pixel in the i-th row and j-th column of the restored image;

[0097] 4.9.3) Calculate the peak signal-to-noise ratio and structural similarity between the restored image and the corresponding clean image using the above formula. If the model has a higher peak signal-to-noise ratio and a structural similarity closer to 1 on the validation set of the current round, save the model parameters of this round.

[0098] 4.10) Clear the gradient of the previous batch of data loss and repeat steps 4.3)-4.9) until 300 rounds of training are completed. Select the best model parameters from all rounds of evaluation to obtain the trained refocusing network model.

[0099] Step 5: Use the trained deep learning network to refocus the defocused images in the test set.

[0100] The defocused image in the test set data is input into the trained deep learning network, the defocused image features are extracted and the restored SAR image is generated, the defocused image is refocused, and the output restored image is obtained.

[0101] The effect of the present invention can be further illustrated by the following measured data experiment.

[0102] 1. Experimental conditions

[0103] The SAR ground scene measured echo data is used for verification. The basic environment of the experiment is shown in Table 3, and the specific model parameters are shown in Table 4.

[0104] Table 3 Basic experimental environment

[0105]

[0106] Table 4 Model parameter settings

[0107]

[0108] 2. Experimental content

[0109] Under the above conditions, a deep learning network containing multiple residual blocks is constructed and trained. The defocused image of the test set is input into the trained deep learning network to refocus the defocused image and obtain the focusing result of the defocused image, such as Figure 5 As shown. Among them:

[0110] Figure 5 (a) is pure imaging, Figure 5 (b) is the defocused image. Figure 5 (c) is the refocusing result.

[0111] from Figure 5 (a) with Figure 5From the comparison of (c), it can be concluded that the focusing result of the defocused image by the present invention restores more structures and details, which is very close to the original pure image.

[0112] The structural similarity SSIM between the restored image and the pure image is shown in Table 5, and the peak signal-to-noise ratio PSNR is shown in Table 6.

[0113] Table 5 Structural similarity SSIM improvement

[0114] name Real part Imaginary part Amplitude Initial comparison 0.48062456 0.48026143 0.54541461 The present invention compares 0.78912158 0.78921288 0.5889053

[0115] Table 6 Peak signal-to-noise ratio PSNR improvement

[0116] name Real part Imaginary part Amplitude Initial comparison 26.4182728 26.37997636 25.58261405 The present invention compares 32.38408315 32.39016352 26.28383608

[0117] The initial comparison in the above table represents the structural similarity and peak signal-to-noise ratio of the real part, imaginary part and amplitude between the pure image and the defocused image, and the comparison of the present invention represents the structural similarity and peak signal-to-noise ratio of the real part, imaginary part and amplitude between the pure image and the focused image obtained by the present invention.

[0118] It can be seen from Tables 5 and 6 that the structural similarity SSIM and peak signal-to-noise ratio PSNR of the present invention on this data set are significantly improved, indicating that the present invention can effectively refocus the defocused image and restore the details of the image.

[0119] Experimental results show that the present invention can effectively process radar defocused images and ensure the restoration of high-resolution details. The deep learning network constructed by the present invention improves the feature extraction capability and training stability, and enhances the accuracy and efficiency of radar refocusing.

[0120] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. It is obvious that professionals in this field, after understanding the content and principles of the present invention, may make respective modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

[0121] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the implementation scheme of the present invention to facilitate understanding, and the sequence of the step numbers is not limited.

Claims

1. A SAR image intelligent refocusing method, characterized in that: The steps include: (1) Obtain a SAR image dataset, preprocess it, and divide it into a training set, a validation set, and a test set; (2) Building a deep learning network (2a) establishing an encoder including a stack of multiple residual blocks for feature extraction of defocused images; (2b) establishing a decoder including multiple upsampling layers and residual blocks alternately stacked to perform feature fusion and obtain a reconstructed image; (2c) selecting a single-layer residual block, connecting the encoder and the decoder through the single-layer residual block, and connecting the output of each residual block in the encoder with the upsampling layer output of its corresponding decoder to form a deep learning network, and setting its loss function; (3) Input the training set data into the deep learning network, perform forward propagation and back propagation, iteratively update the network parameters, and obtain a trained deep learning network; (4) The test set data is input into the trained deep learning network to extract the defocused image features and generate a restored SAR image to complete the refocusing of the defocused image.

2. The method according to claim 1, characterized in that The implementation of step (1) includes the following steps: Obtain echo data of river, ship and coast scenes from public websites and generate pure SAR images based on them; Then, phase interference is introduced into the echo data to obtain interference echo data, and imaging is performed to obtain a defocused SAR image; The data of clean images and the corresponding defocused images are combined to form a dataset, which is divided into training set, validation set and test set in a ratio of 6:2:

2.

3. The method according to claim 2, characterized in that The phase interference is introduced into the echo data to obtain the interference echo data, and its implementation includes the following steps: Set the error term Δφ(t) = C0t 2 +C1t 3 +C2t 4 , where the azimuth frequency t∈[-0.5,0.5], coefficients C0, C1, C2∈[-5,5]×30, represent the contribution of the quadratic error term, cubic error term, and quartic error term to the phase error, respectively; The error term Δφ(t) is added to the received echo signal: The interference echo data s is obtained in the phase err (t): Where A is the amplitude of the signal, 2πf0t is the fundamental frequency portion of the signal, and φ(t) is the time-dependent phase.

4. The method according to claim 1, characterized in that: Step (2a) establishes an encoder comprising a stack of multiple residual blocks, and its implementation includes the following: (2a1) Establish a head residual block including a convolution block and a residual connection. The convolution block contains two layers of 3×3 convolutions, and batch normalization and ReLU activation operations are performed after the first layer of convolution. The residual connection performs a 3×3 convolution on the input and adjusts the output dimension through batch normalization to ensure that the output dimensions of the convolution block and the residual connection are consistent. Finally, the outputs of the convolution block and the residual connection are added to form the output of the head residual block, which is the input of the next layer of residual blocks. (2a2) Establish two layers of residual blocks with the same structure. Each layer of residual blocks contains a convolution block and a residual connection. Each convolution block contains two 3×3 convolution layers. Batch normalization and ReLU activation operations are performed before each convolution layer. The residual connection performs 3×3 convolution on the input and adjusts the dimension through batch normalization. Finally, the outputs of the convolution block and the residual connection are added to obtain the final output of each layer. (2a3) The above three residual blocks are connected layer by layer, and the output of each residual block is used as the input of the next residual block to form a deep hierarchical encoder structure; Each residual block is expressed as: O i =h(I i )+F(I i ;W i ) Where Ii, Oi are the input and output of the ith residual block, respectively, F(·) is the mapping function of the residual block, W is the parameter of the ith residual block, and h(·) is the unit mapping function.

5. The method according to claim 1, characterized in that Step (2b) establishes a decoder including multiple upsampling layers and residual blocks alternately stacked, and its implementation includes the following (2b1) An upsampling layer is formed by a transposed convolution operation with a stride of 2; (2b2) ​​Create a residual block consisting of a convolutional block and a residual connection. The convolutional block contains two 3×3 convolutional layers, with batch normalization and ReLU activation operations before each convolutional layer. The residual connection performs a 3×3 convolution on the input and adjusts the dimension through batch normalization. Finally, the outputs of the convolutional block and the residual connection are added together to obtain the output of each residual block, which is the input for the next upsampled layer. (2b3) The processed output features of the encoder are used as the input of the first upsampling layer. The shallow features in the encoder are concatenated with the corresponding upsampled features in the decoder through skip connections to form the input of the first layer of residual blocks. The output features of the residual blocks are passed to the next upsampling layer, and are processed by the second and third layers of upsampling and residual blocks in turn. The output of the third layer of residual blocks is passed through a 1×1 convolutional layer and a Sigmoid activation function to obtain a decoder with upsampling and residual blocks stacked alternately. Each residual block is expressed as: O i =h(I i )+F(I i ;W i ) Where Ii, Oi are the input and output of the ith residual block, respectively, F(·) is the mapping function of the residual block, W is the parameter of the ith residual block, and h(·) is the unit mapping function.

6. The method according to claim 1, characterized in that In step (2c), the loss function of the deep learning network is set as follows: Among them, H and M are the height and width of the image respectively, and y(i,j) represents the pixel value of the pure image at position (i,j). Represents the pixel value of the restored image at position (i, j); is the absolute error between the clean image and the restored image at position (i, j).

7. The method according to claim 1, characterized in that Step (3) performs forward propagation and back propagation on the deep learning network and iteratively updates the network parameters. The implementation steps include the following: (3a) Set the initial learning rate η to 0.001 and the maximum number of iterations to 300; (3b) In the forward propagation stage, the defocused image input to the network is passed through the convolutional layer in the encoder to extract image features x, and skip connections are used to enhance the preservation of details; (3c) The input feature x is calculated through the convolution layer to obtain the new feature x*=W×x+b, which is used as the input of the next convolution layer, where the weight matrix W and bias b are independent parameters of each convolution layer; (3d) Each convolutional layer of the encoder and decoder repeats step (3c), allowing the network to gradually extract and reconstruct the features of the image. After multiple layers of convolution calculations in the decoder, the network finally generates the image restoration result. (3e) Using the restored image results The absolute error between the pure image y and the calculated loss function L; (3f) In the back-propagation phase, the gradient of the loss function L for each weight matrix W and bias b is calculated separately and The Adam optimizer is used to dynamically adjust the learning rate so that the learning rate changes dynamically according to the mean and variance of the gradient. (3g) Update the model parameters according to the gradient direction and the adjusted learning rate η: Among them, W′ is the updated weight matrix, b′ is the updated bias; (3h) Evaluate the model performance on the validation set, that is, calculate the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of each restored image. If the model has a higher peak signal-to-noise ratio and a structural similarity closer to 1 on the validation set of the current round, save the model parameters of this round. (3i) Clear the gradient of the previous batch of data loss and repeat steps (3b) to (3h) until 300 rounds of training are completed. The best model parameters are selected from all rounds of evaluation to obtain the trained refocusing network model.

Images