Array detector nuclear waste bin detection image reconstruction method based on SA-ResNeXt

Abstract

The invention relates to the field of nuclear waste bin transmission image reconstruction. The invention discloses an SA-ResNeXt-based method for reconstructing a nuclear waste bin detection image of an array detector, which can accurately reconstruct a nuclear waste bin chromatography TGS transmission image under sparse transmission scanning. The method comprises the following steps: S1, performing a simulation experiment on the nuclear waste bin based on Geant4 simulation software to obtain training data, and performing a real experiment to obtain verification data; establishing a database for the training data and the verification data; s2, constructing a ResNeXt neural network model according to the transmission image characteristics of the nuclear waste bin; s3, training data in the database are selected to train the ResNeXt neural network model; an optimized ResNeXt neural network model is obtained; and selecting verification data in the database, inputting the optimized ResNeXt neural network model, and realizing reconstruction of the nuclear waste bin transmission image through the optimized ResNeXt neural network model. By adopting the method, the reconstruction accuracy of the TGS image can be greatly improved, and the reconstruction speed can be improved.

Description

Technical Field

[0001] The invention relates to nuclear waste barrel transmission image reconstruction, in particular to a nuclear waste barrel detection image reconstruction method of an array detector based on SA-ResNeXt. Background Art

[0002] Because radioactive waste is different from general waste, it is necessary to accurately detect and classify the composition, content and distribution of radionuclides in waste barrels before temporary storage, transportation and disposal. Accurate and rapid reconstruction of the spatial radioactivity distribution of nuclear waste barrels is of great significance for the implementation of radiation protection.

[0003] At present, nuclear waste barrel image scanning technologies mainly include segmented gamma scanning and tomographic gamma scanning. There are two main types of transmission image reconstruction algorithms: analysis and iteration.

[0004] Iterative algorithms are the most widely used. The efficiency and image boundary quality of this algorithm are not as good as those based on analysis, but it can be applied to irregular geometric sampling and incomplete noise data sets. It has become the mainstream algorithm for the reconstruction of transmission images of tomographic gamma scanning, but when the measured projection has a low signal-to-noise ratio, it is difficult to obtain high-quality reconstructed images through iterative methods. Therefore, scholars have proposed simultaneous algebraic reconstruction technology (SART), adaptive algebraic reconstruction technology (AART) and discrete algebraic reconstruction technology (DART) to improve the quality of reconstructed images, eliminate artifacts, accelerate iterative convergence, and achieve significant performance improvements. Tomographic gamma scanning technology improves the detection accuracy of radioactive waste barrels through fine voxel segmentation, but it also increases the number of scans and time required for detection. Therefore, how to accurately and efficiently reconstruct the transmission image of nuclear waste barrels is an urgent need in the field of nuclear radiation protection, and related technical bottlenecks need to be broken through.

[0005] The essence of nuclear waste barrel image reconstruction is to iteratively derive the attenuation coefficient matrix of each voxel block by scanning projection data at different angles. If the number of scans is reduced and the nonlinear relationship between the blurred image and the high-definition image is found, the image of the nuclear waste barrel can be quickly and accurately established. As a method of artificial intelligence research, neural networks have strong nonlinear mapping capabilities by constructing multi-hidden layer learning models and performing feature learning on massive training data, thereby ultimately improving calculation accuracy. Therefore, they perform well in fitting fields where it is difficult to determine functional relationships. Some scholars have tried to use convolutional neural networks to reconstruct gamma scanning transmission images from sparse projection data. However, traditional convolutional neural networks are prone to local optimal solutions, poor prediction accuracy, and instability. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a nuclear waste barrel detection image reconstruction method based on SA-ResNeXt array detector, which can accurately reconstruct the nuclear waste barrel tomography TGS transmission image under sparse transmission scanning and increase the reconstruction accuracy and reconstruction speed of the TGS transmission image.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a method for reconstructing nuclear waste barrel detection images using an array detector based on SA-ResNeXt, comprising the following steps:

[0008] S1. Conduct simulation experiments on nuclear waste barrels based on Geant4 simulation software to obtain training data, and conduct real experiments to obtain verification data; establish a database with training data and verification data;

[0009] S2. Construct a ResNeXt neural network model based on the transmission image characteristics of the nuclear waste barrel;

[0010] S3. Select training data from the database to train the ResNeXt neural network model; obtain the optimized ResNeXt neural network model; select verification data from the database, input the optimized ResNeXt neural network model, and reconstruct the transmission image of the nuclear waste barrel through the optimized ResNeXt neural network model.

[0011] Furthermore, in step S2, constructing the ResNeXt neural network model includes the following steps:

[0012] S21. Use Geant4 software to simulate a real nuclear waste barrel model and obtain projection data at different angles;

[0013] S22, using the ART iterative algorithm to iterate the acquired projection data and the projection data of the empty bucket to obtain a low-definition image of 96 voxels;

[0014] S23, upsampling the low-definition image with 96 voxels to obtain a low-definition image with 1536 voxels;

[0015] S24. For the nuclear waste barrel model established in Geant4, the attenuation coefficient value corresponding to each voxel is determined according to the proportion of voxels occupied by the object to establish a high-definition image of 1536 voxels;

[0016] S25, determining the number of input layer nodes, the number of hidden layer layers, the number of hidden layer nodes and the number of output layer nodes of the ResNeXt neural network according to the characteristics of the nuclear waste barrel image;

[0017] S26. Use the low-definition image of 1536 voxels in step S23 as input data of the neural network, and use the high-definition image of 1536 voxels in step S24 as output data of the neural network, so as to establish a ResNeXt neural network model.

[0018] Furthermore, in step S2, the ResNeXt neural network model consists of four Convs, namely Conv1, Conv2, Conv3, and Conv4, wherein Conv1 is used to extract image features, and the image features are batch normalized using the BatchNormalization batch processing tool;

[0019] Conv2 expands the neural network channels to 32, and then uses the PReLu function to increase the nonlinearity of the network;

[0020] Conv3 is used to learn the mapping function from low-definition images to high-definition images. It contains 6 ResNeXt Blocks. Each ResNeXt Block contains 2 convolutional layers. After each convolutional layer, there is a PReLu function to increase the nonlinear ability of the network.

[0021] At the same time, each ResNeXt Block uses the ReLu function to sum the shortcut connections, and its expression is as follows:

[0022]

[0023] in, is the output of Conv3, is the input of Conv3; is the maximum value in the set, that is means The maximum value between 0 and 0;

[0024] The ResNeXt Block parameter is set to Cardinality = 32;

[0025] A spatial attention mechanism module is introduced after Conv3, which instructs the neural network model to assign different attention weights to different locations in the image, thereby more accurately restoring image details.

[0026] Conv4 is used to merge channels and complete image reconstruction through the Sigmoid function; its expression is as follows:

[0027] ;

[0028] in is the output of Conv4, is a natural constant, It is the input quantity of Conv4.

[0029] The beneficial effects of the present invention are as follows: the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method of the present invention replaces the traditional image reconstruction method with a neural network model method, so that the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method of the present invention can accurately reconstruct the TGS transmission image even under sparse transmission scanning;

[0030] Secondly, the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method described in the present invention converts the detector into an array detector, thereby increasing the scanning speed; therefore, it has a significant reconstruction speed advantage and can complete the reconstruction of the TGS transmission image in a very short time.

[0031] Again, the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method described in the present invention can greatly increase the reconstruction accuracy and reconstruction speed of TGS images, and also provide strong technical support for the reconstruction of spatial radioactivity distribution images of nuclear waste barrels. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram and block diagram of the principle of the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method in an embodiment of the present invention;

[0033] Figure 1 The arrows on the right side of the middle indicate the process steps of the method in a real experiment; Figure 1 The arrow direction on the left side indicates the process steps of the method under simulation state;

[0034] Figure 2 Schematic diagram of voxel segmentation of a low-definition image with 96 voxels in an embodiment of the present invention.

[0035] Figure 3 is a schematic diagram of the ResNeXt neural network structure in an embodiment of the present invention;

[0036] In the figure, Conve1 represents the first convolutional layer, Conve2 represents the second convolutional layer, Conve3 represents the third convolutional layer, Conve4 represents the fourth convolutional layer, BatchNorm represents batch normalization, PReLu and Sigmoid are the names of activation functions, Shortcut Connection represents jump connection, Spatial attention mechanism represents spatial attention mechanism, and Feature Maps represents feature maps.

[0037] Figure 4 Result diagrams of nuclear waste barrel reconstruction using different methods under a simulated environment in an embodiment of the present invention.

[0038] In the figure, a, b, c, d, and e represent the results of different methods; a is ART, b is ART-NLM, c is ART-TV, d is ResNeXt, that is, the array detector nuclear waste barrel detection image reconstruction method based on SA-ResNeXt described in the present invention, and e is DenseNet; the numbers 0.662, 1.173, and 1.332 after the letters respectively refer to: experiments using γ rays with energies of 0.662 MeV, 1.173 MeV, and 1.332 MeV; specifically, for example: a-0.662 refers to the experimental results of the ART method using γ rays with an energy of 0.662 MeV; a-1.173 refers to the experimental results of the ART method using γ rays with an energy of 1.173 MeV; a-1.332 refers to the experimental results of the ART method using γ rays with an energy of 1.332 MeV.

[0039] Figure 5 is a comparison chart of evaluation indicators of nuclear waste barrel reconstruction using different methods under a simulated environment in an embodiment of the present invention;

[0040] In the figure, Figure 5 (a) is a comparison of the mean square error of nuclear waste barrel reconstruction under different simulation environments; Figure 5 (b) is a comparison of the structural similarities of nuclear waste barrels reconstructed under different simulated environments.

[0041] Figure 6 is a result diagram of nuclear waste barrel reconstruction using different methods in a real environment in an embodiment of the present invention;

[0042] In the figure, a, b, c, d, and e represent the results of different methods; a is ART, b is ART-NLM, c is ART-TV, d is ResNeXt, that is, the array detector nuclear waste barrel detection image reconstruction method based on SA-ResNeXt described in the present invention, and e is DenseNet. The numbers 0.662, 1.173, and 1.332 after the letters respectively refer to: experiments using γ-rays with energies of 0.662 MeV, 1.173 MeV, and 1.332 MeV; specifically, for example: a-0.662 refers to the experimental results of the ART method using γ-rays with an energy of 0.662 MeV; a-1.173 refers to the experimental results of the ART method using γ-rays with an energy of 1.173 MeV; a-1.332 refers to the experimental results of the ART method using γ-rays with an energy of 1.332 MeV.

[0043] Figure 7 is a comparison chart of evaluation indicators of nuclear waste barrel reconstruction using different methods under real environment in an embodiment of the present invention;

[0044] In the figure, Figure 7 (a) is a comparison of the mean square error of nuclear waste barrel reconstruction using different methods in real environments; Figure 7 (b) is a comparison of the structural similarities of nuclear waste barrels reconstructed by different methods in real environments. DETAILED DESCRIPTION

[0045] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0046] like Figure 1 As shown, the SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method of the present invention refers to an array detector nuclear waste barrel detection image reconstruction method combined with the attention mechanism ResNeXt; specifically, it includes the following steps:

[0047] S1. Conduct simulation experiments on nuclear waste barrels based on Geant4 simulation software to obtain training data, and conduct real experiments to obtain verification data; establish a database with training data and verification data;

[0048] S2. Construct a ResNeXt neural network model based on the transmission image characteristics of the nuclear waste barrel;

[0049] The ResNeXt neural network model consists of 4 Convs, where Conv1 is used to extract image features and batch normalize the image features using the BatchNormalization batch processing tool;

[0050] Conv2 expands the neural network channels to 32, and then uses the PReLu function to increase the nonlinearity of the network;

[0051] Conv3 is used to learn the mapping function from low-definition images to high-definition images. It contains 6 ResNeXt Blocks. Each ResNeXt Block contains 2 convolutional layers. After each convolutional layer, there is a PReLu function to increase the nonlinear ability of the network.

[0052] At the same time, each ResNeXt Block uses the ReLu function to sum the shortcut connections, and its expression is as follows:

[0053]

[0054] in is the output of Conv3, is the input of Conv3; is the maximum value in the set, that is means The maximum value between 0 and 0;

[0055] The ResNeXt Block parameter is set to Cardinality = 32;

[0056] A spatial attention mechanism module is introduced after Conv3, which instructs the neural network model to assign different attention weights to different locations in the image, thereby more accurately restoring image details.

[0057] Conv4 is used to merge channels and complete image reconstruction through the Sigmoid function; its expression is as follows:

[0058] ;

[0059] in, is the output of Conv4, is a natural constant, It is the input quantity of Conv4.

[0060] Building a ResNeXt neural network model involves the following steps:

[0061] S21. Use Geant4 software to simulate a real nuclear waste barrel model and obtain projection data at different angles;

[0062] S22, using the ART iterative algorithm to iterate the acquired projection data and the projection data of the empty bucket to obtain a low-definition image of 96 voxels;

[0063] S23, upsampling the low-definition image with 96 voxels to obtain a low-definition image with 1536 voxels;

[0064] S24. For the nuclear waste barrel model established in Geant4, the attenuation coefficient value corresponding to each voxel is determined according to the proportion of voxels occupied by the object to establish a high-definition image of 1536 voxels;

[0065] S25, determining the number of input layer nodes, the number of hidden layer layers, the number of hidden layer nodes and the number of output layer nodes of the ResNeXt neural network according to the characteristics of the nuclear waste barrel image;

[0066] S26. Use the low-definition image of 1536 voxels in step S23 as input data of the neural network, and use the high-definition image of 1536 voxels in step S24 as output data of the neural network, so as to establish a ResNeXt neural network model.

[0067] S3. Select training data from the database to train the ResNeXt neural network model; obtain the optimized ResNeXt neural network model; select verification data from the database, input the optimized ResNeXt neural network model, and reconstruct the transmission image of the nuclear waste barrel through the optimized ResNeXt neural network model.

[0068] The SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method described in this application uses the ResNeXt neural network to solve the problems of traditional convolutional neural networks that are prone to local optimal solutions, poor prediction accuracy, instability, etc.; it has a better balance between local and global, wider shortcut connections, and better model width.

[0069] The SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method described in the present application uses a neural network to establish a mapping relationship from low-definition to high-definition images, and can reconstruct low-definition images reconstructed under sparse transmission scanning into high-definition images through the mapping relationship. Therefore, the tomographic gamma scanning transmission image can also be accurately reconstructed under sparse transmission scanning. Secondly, since the mapping relationship between images is established through a neural network to convert low-definition images into high-definition images, there is no need to use traditional image reconstruction algorithms for complex calculations, so the tomographic gamma scanning transmission image can be completed in a very short time.

[0070] Example

[0071] In this embodiment, the array detector tomography gamma scanning transmission image reconstruction method combined with the attention mechanism ResNeXt includes the following steps:

[0072] 1. Conduct simulation experiments on nuclear waste barrels based on Geant4 simulation software to obtain training data, and conduct real experiments to obtain verification data; establish a database with training data and verification data;

[0073] In the Geant4 simulation, experiments were conducted using γ-rays with energies of 0.662 MeV, 1.173 MeV, and 1.332 MeV, respectively, and the number of particles emitted was 1e 8 In the real experiment, the self-developed array detector TGS system was used to conduct the experiment. 137 Cs and 60 Co as a radioactive source, and its activity is 3.73 × 10 5 Bq and 3.94 × 10 5 Bq, the gamma ray energies are also 0.662 MeV, 1.173 MeV and 1.332 MeV respectively. In this example, there are 8064 simulation data in total in Geant4.

[0074] 2. Construct a ResNeXt neural network model based on the transmission image characteristics of nuclear waste barrels;

[0075] Since the image size of the array detector TGS reconstructed in this article is small, fewer features need to be learned. In order to reduce the hardware computing burden and speed up the neural network training, the network depth of ResNeXt is set to 15, and the output image size is always kept at 16×96 during the convolution process. The ResNeXt neural network consists of 4 Convs. Conv1 is used to extract image features, and the image features are batch normalized using the BatchNormalization batch processing tool; Conv2 expands the neural network channel to 32, and then uses the PReLu function to increase the nonlinear ability of the network; Conv3 is used to learn the mapping function from low-definition images to high-definition images, which contains 6 ResNeXt Blocks. Each ResNeXt Block contains 2 convolutional layers, and each convolutional layer is followed by a PReLu function to increase the nonlinear ability of the network. At the same time, each ResNeXt Block uses the ReLu function to sum the Shortcut Connection, and its expression is as follows:

[0076] ;

[0077] in, is the output of Conv3, is the input of Conv3; is the maximum value in the set, that is means The maximum value between 0 and 0;

[0078] The ResNeXt Block parameter sets Cardinality = 32. Conv4 is used to merge channels, and the image reconstruction is completed through the Sigmoid function. Its expression is as follows:

[0079] ;

[0080] is the output of Conv4, is a natural constant, It is the input quantity of Conv4.

[0081] Building a ResNeXt neural network model involves the following steps:

[0082] 2.1. Use Geant4 software to simulate the real nuclear waste barrel model and obtain projection data at different angles;

[0083] 2.2. Then use the ART iterative algorithm to iterate the acquired projection data and the projection data of the empty bucket to obtain a low-definition image of 96 voxels; Figure 2 shown.

[0084] 2.3. Upsample the low-resolution image with 96 voxels to obtain a low-resolution image with 1536 voxels;

[0085] 2.4. For the nuclear waste barrel model established in Geant4, the attenuation coefficient value corresponding to each voxel is determined according to the proportion of voxels occupied by the object to establish a high-definition image of 1536 voxels;

[0086] 2.5. Determine the number of input layer nodes, hidden layer nodes, hidden layer nodes and output layer nodes of the ResNeXt neural network according to the characteristics of the nuclear waste barrel image;

[0087] 2.6. Use the low-definition image of 1536 voxels in step 2.3 as the input data of the neural network, and the high-definition image of 1536 voxels in step 2.4 as the output data of the neural network to establish the ResNeXt neural network model.

[0088] The ResNeXt neural network structure is shown in Table 1 and Figure 3 shown.

[0089] Table 1 Neural network parameters

[0090]

[0091] 3. Select the training data in the database to train the ResNeXt neural network model; obtain the optimized ResNeXt neural network model; select the verification data in the database, input the optimized ResNeXt neural network model, and reconstruct the transmission image of the nuclear waste barrel through the optimized ResNeXt neural network model.

[0092] In this example, the nuclear waste barrel image is a 1536-voxel image. If a traditional algorithm is used, data from 1536 scanning angles are required to reconstruct a good quality image. However, the present invention only uses data from 12 scans.

[0093] 4. Evaluate the effect of the above reconstruction model.

[0094] Taking mean square error and structural similarity as evaluation indicators, the mean square error is calculated as follows:

[0095] (1)

[0096] In formula (1), image X and image Y have the same size. is the total pixel value of the image, , are the pixel values ​​of the corresponding pixels in image X and image Y. MSE stands for mean square error.

[0097] SSIM captures the structural features of images by introducing the human visual system (HVS) model, which can better judge the structural similarity of images. SSIM is structural similarity, an indicator to measure the similarity between two images; the formula of SSIM is as follows:

[0098] (2)

[0099] , and Respectively represent the comparison of the brightness, contrast and size of images X and Y. , and are all constants greater than 0. In general, .in , and is defined as follows:

[0100] (3)

[0101]

[0102]

[0103] in, and are the average pixel values ​​of images X and Y respectively, and are the variances of image X and Y pixel values, respectively. is the covariance of the image X and Y pixel values. , and is a constant. Usually, = 6.5025, = 58.5225, = 29.26125. , and The calculation formula is as follows:

[0104]

[0105]

[0106]

[0107] Where H is the height of the image, and W is the width of the image.

[0108] The maximum SSIM value is 1. The larger the SSIM value between two images, the higher the similarity between the two images.

[0109] Moreover, in the Geant4 simulation process, experiments were conducted using γ-rays with energies of 0.662 MeV, 1.173 MeV, and 1.332 MeV, respectively, and the number of particles emitted was 1e 8 , and conduct experiments on ART method, ART-NLM method, ART-TV method and DenseNet method. In the real experiment, the self-developed array detector TGS system was used to conduct experiments. 137 Cs and 60 Co as a radioactive source, and its activity is 3.73 × 10 5 Bq and 3.94 × 10 5 Bq, and the γ-ray energies are also 0.662 MeV, 1.173 MeV, and 1.332 MeV respectively; experiments were conducted using the ART method, ART-NLM method, ART-TV method, and DenseNet method.

[0110] The results of the two experiments are as follows Figure 4 ,5 and Figure 6 ,7. For the image reconstruction algorithm, not only the accuracy of the reconstructed image needs to be considered, but also the time required for the method to reconstruct the image in the actual environment. The time for reconstructing TGS images by each method in the two experiments is shown in Table 2 below.

[0111] Table 2 TGS image reconstruction time loss table

[0112]

[0113] From Table 2 and Figure 4 As can be seen from Figures 5, 6, and 7, the nuclear waste barrel transmission image reconstructed by the method proposed in the present invention is basically consistent with the actual nuclear waste barrel transmission image, and a smaller error and the lowest time loss are achieved in both the simulation experiment and the real experiment, indicating that the nuclear waste barrel transmission image method proposed in the present invention is fast, effective and accurate.

Claims

1. An array detector nuclear waste barrel detection image reconstruction method based on SA-ResNeXt, characterized in that: The following steps are involved: S1. Conduct simulation experiments on nuclear waste barrels based on Geant4 simulation software to obtain training data, and conduct real experiments to obtain verification data; establish a database with training data and verification data; S2. According to the transmission image characteristics of nuclear waste barrels, a ResNeXt neural network model combined with attention mechanism is constructed; S3. Select training data from the database to train the ResNeXt neural network model; obtain the optimized ResNeXt neural network model; select verification data from the database, input the optimized ResNeXt neural network model, and reconstruct the transmission image of the nuclear waste barrel through the optimized ResNeXt neural network model.

2. The SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method according to claim 1, characterized in that: In step S2, building a ResNeXt neural network model includes the following steps: S21. Use Geant4 software to simulate a real nuclear waste barrel model and obtain projection data at different angles; S22, using the ART iterative algorithm to iterate the acquired projection data and the projection data of the empty bucket to obtain a low-definition image of 96 voxels; S23, upsampling the low-definition image with 96 voxels to obtain a low-definition image with 1536 voxels; S24. For the nuclear waste barrel model established in Geant4, the attenuation coefficient value corresponding to each voxel is determined according to the proportion of voxels occupied by the object to establish a high-definition image of 1536 voxels; S25, determining the number of input layer nodes, the number of hidden layer layers, the number of hidden layer nodes and the number of output layer nodes of the ResNeXt neural network according to the characteristics of the nuclear waste barrel image; S26. Use the low-definition image of 1536 voxels in step S23 as input data of the neural network, and use the high-definition image of 1536 voxels in step S24 as output data of the neural network, so as to establish a ResNeXt neural network model.

3. The SA-ResNeXt-based array detector nuclear waste barrel detection image reconstruction method according to claim 1, characterized in that: In step S2, the ResNeXt neural network model consists of 4 Convs, namely Conv1, Conv2, Conv3, and Conv4; Conv1 is used to extract image features, and the BatchNormalization batch processing tool is used to perform batch normalization on image features; Conv2 expands the neural network channels to 32, and then uses the PReLu function to increase the nonlinearity of the network; Conv3 is used to learn the mapping function from low-definition images to high-definition images. It contains 6 ResNeXt Blocks. Each ResNeXt Block contains 2 convolutional layers. After each convolutional layer, there is a PReLu function to increase the nonlinear ability of the network. At the same time, each ResNeXt Block uses the ReLu function to sum the shortcut connections, and its expression is as follows: ； in, is the output of Conv3, is the input of Conv3; is the maximum value in the set, that is means The maximum value between 0 and 0; Among them, the ResNeXt Block parameter sets Cardinality = 32; A spatial attention mechanism module is introduced after Conv3, which instructs the neural network model to assign different attention weights to different locations in the image, thereby more accurately restoring image details; Conv4 is used to merge channels and complete image reconstruction through the Sigmoid function; its expression is as follows: ; in, is the output of Conv4, is a natural constant, The input quantity of Conv4.

Images