GAN neural network multiple distortion suppression model based on coordinate attention mechanism
By using a GAN neural network model based on coordinate attention mechanism and a real neutron radiograph image training dataset, the problem of multiple distortions in neutron radiograph images was solved, and the sharpening and lightweight processing of neutron radiograph images was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTHEAST NORMAL UNIVERSITY
- Filing Date
- 2023-05-24
- Publication Date
- 2026-06-02
AI Technical Summary
Neutron radiographs suffer from severe image degradation due to multiple distortions introduced by factors such as neutron sources and collimators, which affects subsequent detection and recognition research.
A GAN neural network model based on coordinate attention mechanism is adopted to recover clear neutron radiographs through feature extraction, residual unit learning, and upsampling reconstruction. The model is trained using a training dataset of real neutron radiographs and the Huber loss function.
It effectively suppresses multiple distortions in neutron radiography images, improves image clarity, shortens training time and resource consumption, and is suitable for miniaturized neutron radiography systems.
Smart Images
Figure CN116629337B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of neutron photography, specifically relating to a neutron photography image degradation model, a GAN neural network multi-distortion suppression model based on coordinate attention mechanism, and a suppression method. Background Technology
[0002] Neutron radiography is an important non-destructive testing technique using radiation, with wide applications in aviation, aerospace, military, archaeology, and agriculture. Its principle is similar to X-ray radiography, utilizing a neutron beam to penetrate the object under test and capture its intensity distribution to obtain information about the object's internal structure. Due to the unique nature of neutron imaging, image degradation is introduced to varying degrees by the neutron source, collimator, and imaging detector. This is particularly true in small, portable neutron radiography systems, where the neutron yield of the neutron source is much lower than that of a reactor neutron source, and the collimation ratio of the collimation system is limited. The resulting neutron radiograph images are severely degraded, typically exhibiting a mixture of distortions such as geometrical unsharpness (approximated by Gaussian blur), Poisson noise, Gaussian noise, white spot noise, and gamma noise. This image degradation significantly hinders subsequent image detection and recognition research. Therefore, researching a multi-distortion suppression method suitable for neutron radiography image degradation models has significant theoretical and practical value. Summary of the Invention
[0003] The purpose of this invention is to solve the above-mentioned problems by providing a neutron radiograph image degradation model, a GAN neural network multi-distortion suppression model based on coordinate attention mechanism, and a suppression method.
[0004] A neutron radiograph image degradation model is represented as follows:
[0005] (1)
[0006] in These are real neutron radiographs containing multiple distortions. This is the clear neutron radiograph image recovered using this method. For Gaussian blur kernel, This represents the convolution operation. Represents Poisson noise interference. With a mean of 0 and a variance of Additive Gaussian noise, b is white spot noise, and multiplicative gamma noise is selectively added according to different images to be processed. In the above formula, G is used to represent it.
[0007] The GAN neural network multi-distortion suppression model based on coordinate attention mechanism uses a GAN neural network with coordinate attention mechanism to take a simulated real neutron radiograph image containing multiple distortions as input. Through a series of feature extraction, learning on residual units and upsampling reconstruction, a clear neutron radiograph image is obtained after restoration.
[0008] Specifically, the process is as follows: First, initial features are extracted from the noisy input using a 7*7 convolution. Next, two consecutive downsampling operations are performed. The resulting feature maps are then further learned through four cascaded RAB units. The first part of the RAB is a Context Block, using four dilated convolutions with dilation rates of 1, 2, 3, and 4 to increase the receptive field size. Following the Context Block are two residual connections. The final part of the RAB is a coordinate attention module, which uses two 1D global pooling operations to aggregate the input features into two independent orientation-aware feature maps along the vertical and horizontal directions, respectively. These two feature maps, embedding orientation-specific information, are encoded into two attention maps. Each attention map captures the long-range dependencies of the input feature map along a spatial direction; therefore, positional information can be stored in the generated attention map. The two attention maps are then applied to the input feature map through multiplication to emphasize the representation of interest. This operation can distinguish spatial directions and generate coordinate-aware attention maps. The output of the CoordinateAttention block can be explicitly represented as:
[0009] (2)
[0010] in and They represent the first Inputs and outputs of the attention module on each channel. For the corresponding coordinate positions; and Attention weights that encode positional information along the horizontal and vertical directions are used respectively. In order to recover the reconstructed neutron radiograph image, two transposed convolutions are used in the network for upsampling operations. Skip connections are used to link feature maps of the same size during upsampling and downsampling to adapt to inputs of arbitrary size. Finally, the recovered neutron radiograph image is obtained by passing it through a 7*7 convolution.
[0011] Training a GAN neural network multi-distortion suppression model based on coordinate attention mechanism:
[0012] Clear, real-world neutron radiograph images are selected as labels. Based on the aforementioned neutron radiograph image degradation model, Gaussian blur, Gaussian noise, and Poisson noise are added to the clear, real-world neutron radiograph images. Noise following a gamma distribution is selected to simulate real-world neutron radiograph images with multiple distortions, thereby constructing a neutron radiograph image training dataset. The multiple distortion suppression model is then trained.
[0013] The construction of the neutron photography image training dataset includes:
[0014] (1) For each reference image, 30 random selections were made from integers between 1 and 15, which were used as the size of the Gaussian blur kernel. The standard deviation of this two-dimensional Gaussian kernel function was set as the kernel size in the first 15 blurring operations. The value was set to this in the last 15 processing steps. This process serves to expand the training dataset and the blurry yet clear neutron photographic images;
[0015] (2) 20 intervals were generated. The uniformly distributed random numbers on the image are used as the variance of the Gaussian noise, and Gaussian noise with a mean of 0 and a variance of the selected value is added to the image obtained in the previous step.
[0016] (3) Add noise following a Poisson distribution to the neutron radiograph image generated in the previous step. The Poisson distribution follows the following probability density function.
[0017] (8)
[0018] In the formula:
[0019] —Pixels under Poisson noise interference;
[0020] —Original pixels without Poisson noise;
[0021] (4) For different neutron photographic images to be processed, noise following a gamma distribution is selectively added to the training dataset.
[0022] The GAN neural network is trained using real white spot noise to simulate white spot noise, which is then randomly added to clear, real neutron radiograph images, i.e., the training dataset.
[0023] The aforementioned real white spot noise refers to white spot data collected from real neutron radiographs containing multiple distortions using a magnification and cropping method.
[0024] A method for suppressing multiple distortions in neutron radiographs based on GAN neural networks is proposed. The method employs Huber loss and uses the aforementioned neutron radiograph training dataset, with 90% of the dataset used as the training set and 10% as the test set. The training set is used to train the GAN neural network multiple distortion suppression model based on the coordinate attention mechanism, and the test set is used for testing.
[0025] The real neutron radiograph image containing multiple distortions is used as input and fed into a pre-trained GAN neural network multiple distortion suppression model based on coordinate attention mechanism to predict the original image of the real neutron radiograph image containing multiple distortions, thus obtaining the target result.
[0026] A method for suppressing multiple distortions in neutron radiographs based on GAN neural networks.
[0027] It includes:
[0028] (1) A better degradation model for neutron radiographs:
[0029] Based on the background information above, analysis reveals that the degradation of neutron radiograph images follows a mathematical model:
[0030] (1)
[0031] in These are real neutron radiographs containing multiple distortions. This is the clear neutron radiograph image recovered using this method. For Gaussian blur kernel, This represents the convolution operation. Represents Poisson noise interference. With a mean of 0 and a variance of Additive Gaussian noise. b is white spot noise, and multiplicative gamma noise is selectively added according to different images to be processed, denoted by G in the above formula.
[0032] (2) Construction of a better dataset:
[0033] A dataset is constructed by selecting clear, real-world neutron radiograph images. Based on the neutron radiograph image degradation model described in the previous section, noise and blur of different distortion levels are added to simulate real-world neutron radiograph images with multiple distortions, thereby constructing the training dataset. The specific implementation method is as follows:
[0034] 1. Gaussian blur:
[0035] For each reference image, 30 random selections were made from integers between 1 and 15, and this value was used as the size of the Gaussian blur kernel. The standard deviation of this two-dimensional Gaussian kernel function was set as the kernel size during the first 15 blurring operations. The value was set to this in the last 15 processing steps. The fuzzy kernel was obtained and the fuzzy process was implemented using MATLAB functions. and Finish.
[0036] 2. Gaussian noise
[0037] 20 sub-intervals were generated. The variance of the Gaussian noise is represented by uniformly distributed random numbers, and the result is obtained using a MATLAB function. Gaussian noise with a mean of 0 and a variance of the selected values above was added to the obtained image.
[0038] 3. Poisson noise
[0039] Using the image after the two blurring operations as input, and through MATLAB functions... Obtain the image under Poisson noise interference.
[0040] 4. Gamma noise
[0041] Using the image after three blurring operations as input, and through MATLAB functions... Obtain a noisy image that follows a gamma distribution (gamma noise is not needed for processing most neutron radiographs, but it works well for a few neutron radiographs).
[0042] 5. Noise from vitiligo
[0043] To simulate more realistic neutron radiograph images, a method using a Generative Adversarial Network (GAN) to simulate real white spots was employed to compensate for the lack of realism in simulated neutron radiograph images caused by adding white spots of fixed shapes. GANs treat the generation problem as an adversarial game between a discriminator and a generator: the generator produces synthetic data from given noise (generally uniformly or normally distributed), while the discriminator distinguishes the generator's output from real data. The former attempts to generate data closer to reality, while the latter attempts to perfectly distinguish between real and generated data. Thus, the two networks improve through this adversarial process, and continue to compete after improvement, resulting in increasingly perfect data from the generative network, closely approximating real data, and ultimately enabling the generation of the desired data. Since GANs were designed to simulate real data, the white spot noise simulated by GAN networks is more realistic and has diverse shapes, effectively meeting the requirements for simulating real neutron radiograph images. When training the GAN network to generate white spots, to more closely resemble real white spots, the original dataset used was a white spot training dataset collected from real neutron radiograph images containing multiple distortions using a magnification and cropping method. In the GAN network, the generator continuously performs linear transformations, followed by ReLU activation after each transformation, with Tanh as the final activation function. The discriminator also continuously performs linear transformations, with Sigmoid as the final activation function. White patches simulated during 14000 epochs of GAN network training are randomly added to the image generated in process 4.
[0044] (3) A better GAN network architecture is proposed:
[0045] 1. Generator Architecture:
[0046] Using a simulated real-world neutron radiograph image with multiple distortions as input, the reconstructed neutron radiograph image is obtained through a series of feature extractions, learning on residual units, and upsampling reconstruction. Specifically, an initial feature map is first extracted from the noisy input using a 7x7 convolution, followed by two consecutive downsampling operations (convolution-pooling). The resulting feature map is then further learned through four cascaded Residual Attention Blocks (RABs). To reconstruct the neutron radiograph image, two transposed convolutions are used for upsampling. Skip connections are used to link feature maps of the same size during upsampling and downsampling to accommodate inputs of arbitrary sizes. Finally, another 7x7 convolution outputs the reconstructed neutron radiograph image. The first part of the Residual Attention Block (RAB) is a Context Block. Since excessive pooling operations reduce image resolution, which is detrimental to neutron radiograph image reconstruction, the Context Block uses four dilated convolutions (dilated rates 1, 2, 3, and 4) to increase the size of the receptive field without further reducing the neutron radiograph image resolution. Following the Context Block are two residual connections. The final part of the RAB is a coordinate attention module that uses two 1D global pooling operations to aggregate the input features into two independent orientation-aware feature maps along the vertical and horizontal directions, respectively. These two feature maps, embedding orientation-specific information, are encoded into two attention maps, each capturing a long-range dependency of the input feature map along a spatial direction. Therefore, positional information can be preserved in the generated attention map. The two attention maps are then applied to the input feature map through multiplication to emphasize the representation of interest. This operation can distinguish spatial directions (i.e., coordinates) and generate coordinate-aware attention maps. The output of the Coordinate Attention Block can be explicitly represented as:
[0047] (2)
[0048] in and They represent the first Inputs and outputs of the attention module on each channel. For the corresponding coordinate positions; and Attention weights are used to encode positional information along the horizontal and vertical directions, respectively.
[0049] 2. Discriminator structure:
[0050] The discriminator network consists of eight convolutional layers. As the network depth increases, the number of features increases while the feature size decreases. The LeakyReLU activation function is selected. Finally, the probability of predicting a noise-free neutron radiograph image is obtained through two fully connected layers and a final sigmoid activation function. The generator and discriminator structures are as follows: Figure 5 As shown.
[0051] Model training:
[0052] For each neutron photographic image in the training set, it is randomly cropped into a series of sizes. The blocks are used for training. Indicates noise input. Representing the corresponding ground truth, the training pair in each iteration is represented as... ,in This is the batch size, set to 16 during training. The choice of loss function is crucial for model training. Currently, the two most commonly used losses in neural networks are L1 and L2 loss. L2 loss has higher confidence in Gaussian noise, while L1 loss is more robust to outliers. L1 loss and L2 loss can be represented as follows:
[0053] (3)
[0054] (4)
[0055] Considering that the images in the training set contain various types of distortion, Huber loss is used as the loss function in the generator model training. The difference between Huber loss and L1 loss is that the derivative of L1 loss is not unique at 0, which may affect convergence. When the difference between the predicted value f(xi) and the true value yi is small (absolute difference less than 1), L2 loss is used; when the difference is large, a shift of L1 loss is used. Therefore, Huber loss is actually a combination of L1 loss and L2 loss, possessing some of the advantages of both: when the difference between the true and predicted values is small (absolute difference less than 1), the gradient is also relatively small (the loss function is smoother than ordinary L1 loss in this position); when the difference between the true and predicted values is large, the gradient value is small enough (ordinary L2 loss has a large gradient value in this position, which is prone to gradient explosion).
[0056] Huber loss can be represented as follows: (5)
[0057] The discriminator uses the BCE loss function, which can be represented as follows:
[0058] (6)
[0059] During training, the optimizer was set to ADAM, the initial learning rate was 0.001, and a total of 100 training loops were performed.
[0060] The beneficial effects of this invention are:
[0061] Firstly, this invention proposes a novel neutron radiograph image degradation model. Most existing degradation models only consider Gaussian and Poisson noise in neutron radiograph images, and few take into account blurring and white spots. This invention analyzes the factors that may cause neutron radiograph image degradation and comprehensively represents them in the degradation model.
[0062] Secondly, this invention constructs a novel dataset based on real neutron radiograph images. Most existing datasets are based on natural images, but the imaging principles of natural images and neutron radiograph images are completely different. In natural images, light reflected from the subject passes through a camera lens (objective lens) and a shutter that controls the exposure, forming a latent image of the subject on photosensitive material inside a dark chamber. In neutron radiographs, however, only a portion of neutrons can penetrate a layer of matter due to scattering or absorption by atomic nuclei. The number of neutrons that can penetrate depends on the density of the matter and the thickness of the layer, falling under the category of X-ray imaging. Due to the scarcity of available neutron sources and the high cost of neutron radiographs, neutron radiograph images are difficult to obtain. Therefore, many neutron radiograph image distortion suppression methods choose to use natural image datasets as a substitute for neutron radiograph image datasets. However, such substitution lacks scientific basis. Therefore, we use a training dataset based on real neutron radiographs to improve the applicability of our methods to neutron radiographs. Meanwhile, a method for simulating white spots in neutron radiographs using GAN networks was proposed. White spots were extracted from real neutron radiographs as a dataset, and after training, the simulation of white spots was successfully achieved.
[0063] Thirdly, this invention proposes a novel GAN network architecture based on a coordinate attention mechanism. After training, it can be verified that adding the coordinate attention mechanism to the image multiple distortion suppression network is entirely feasible. Unlike the very large models used in the field of natural image processing, our model ensures the efficiency of image multiple distortion suppression while also meeting the requirements of lightweight model, greatly shortening the training time and resource consumption, and providing an effective image sharpening and restoration method for the miniaturization of neutron photography in the future. Attached Figure Description
[0064] Figure 1 This is a schematic diagram of a neutron radiography device (not shown to scale).
[0065] Figure 2 This is a schematic diagram of intensity decay when a neutron beam probes a analyte;
[0066] Figure 3 This is the white spot training dataset collected from neutron radiographs containing multiple distortions, used in the generation of white spots in this invention;
[0067] Figure 4 This is a white spot image generated using a GAN network in this invention;
[0068] Figure 5 This is an image from the dataset constructed in this invention;
[0069] Figure 6 This is a diagram of the GAN network framework for suppressing distortion proposed in this invention;
[0070] Figures 7-13 The results are compared between real neutron radiographs containing multiple distortions processed using the method of this invention and clear neutron radiographs restored by this method. Detailed Implementation
[0071] A method for suppressing multiple distortions in neutron radiographs based on generative adversarial networks (GANs) includes the following steps:
[0072] The first step: Obtaining raw neutron radiographic images from the neutron source. As one of the earliest research institutions in China to conduct research on neutron radiography technology, Northeast Normal University developed a portable neutron tube neutron source and applied it to a small-scale neutron radiography device, making significant exploratory contributions to the miniaturization of neutron radiography. Generally, a neutron radiography system mainly consists of a neutron source system and a detection system. A typical neutron radiography device is as follows: Figure 1 As shown. During neutron radiography, the neutron beam perpendicularly incident on the analyte exhibits three different states upon entering the analyte: some neutrons are scattered and deviate from their original direction; the second portion is absorbed by atoms in the analyte, simultaneously releasing... X-rays; a portion of the neutron beam will also penetrate directly into the analyte and be captured. Thus, the neutron beam passing through the analyte will exhibit different distributions due to variations in the internal structure of the analyte. A schematic diagram illustrating the intensity attenuation of the neutron beam as it passes through the analyte is shown below. Figure 2 As shown, by recording the intensity distribution of these transmitted neutron beams, the internal structure of the sample under test can be obtained. According to Lambert-Beer's law, the intensity of the transmitted neutron beam and the intensity of the incident neutron beam satisfy the following formula:
[0073] (7)
[0074] In the formula:
[0075] —The neutron attenuation coefficient of the neutron beam scattered after it is incident on the analyte;
[0076] —The neutron attenuation coefficient absorbed after a neutron beam is incident on the analyte;
[0077] —The thickness of the neutron beam passing through the object being measured.
[0078] The second step is to select clear, real-world neutron radiographs as labels, and then degrade these clear, real-world neutron radiographs to simulate real-world neutron radiographs with multiple distortions.
[0079] Step 3: Because some neutrons deviating from the central axis pass through the same point on the object at different angles when the collimator collimates the neutron beam, they diffuse in a circular pattern upon reaching the detector, causing blurring in the final neutron photographic image. Gaussian blurring is applied to the neutron image to achieve the effect of adding Gaussian blur and expanding the dataset.
[0080] Step 4: When acquiring low-flux neutron imaging images, the CCD camera must undergo a long exposure time. Due to its own dark current, Gaussian noise is introduced. The interaction between neutrons and the object under test, as well as the absorption of neutrons and the emission of secondary ions by the neutron conversion screen, make the entire film exposure process somewhat random, which also generates Gaussian noise. Therefore, Gaussian noise is added to the resulting image based on step 3.
[0081] Step 5: The statistical fluctuations of neutrons produce noise distributed in a Poisson distribution. The Poisson distribution follows the probability density function:
[0082] (8)
[0083] In the formula:
[0084] —Pixels under Poisson noise interference;
[0085] —The original pixels without Poisson noise.
[0086] Therefore, we add Poisson noise to the image based on step four;
[0087] Step 6: Because neutron radiography is a very complex process, the types of noise generated during the imaging process are diverse. To enrich the types and randomness of noise, and to simulate neutron radiograph images containing various noises as closely as possible, noise following a gamma distribution is introduced. Experimental data shows that it has a very good effect on processing certain neutron radiograph images. Therefore, we add noise following a gamma distribution (to process certain neutron radiograph images) based on step 5.
[0088] Step 7: White spot noise in neutron radiographs is caused by the accumulation of energy from high-energy particles such as gamma rays on the CCD camera, forming large white spots with randomly distributed shapes, sizes, and intensities. White spot noise is the most prominent among the various distortion types in neutron radiographs and causes the greatest damage to the image's true content information. Therefore, a realistic shape of white spot noise was simulated through extensive training of a GAN network. The training dataset used was a white spot training dataset collected using magnification and cropping methods from real-world neutron radiographs containing multiple distortions, such as... Figure 3 As shown. Based on step six or five, add a white patch image generated by a GAN network. The generated white patch image is shown below. Figure 4 As shown, the randomness of the noise is increased by randomly adding the size and number of white spot noises.
[0089] Step 8: Use the final simulated real-world neutron radiograph image with multiple distortions generated in Step 7 as the training set, and use clear, real-world neutron radiograph images as training labels. Use 90% of the training set to train the GAN network proposed in this invention for suppressing distortion, and use the remaining 10% for testing.
[0090] The training results were used to perform multi-distortion suppression processing on real neutron radiographs containing multiple distortions. The results of processing the noisy neutron radiograph of a water bottle are as follows: Figure 7 As shown; neutron radiographs of severely affected vitiligo patients are as follows. Figure 8 , Figure 9 As shown; Figure 10 The result of processing blurred neutron radiograph images; Figures 11-13 The results of processing other neutron radiograph images.
Claims
1. A GAN neural network multi-distortion suppression model based on coordinate attention mechanism, characterized by: A GAN neural network employing a coordinate attention mechanism takes a simulated real-world neutron radiograph image containing multiple distortions as input, and obtains a restored, clear neutron radiograph image through a series of feature extractions, learning on residual units, and upsampling reconstruction. First, initial features are extracted from the noisy input using a 7x7 convolution, followed by two consecutive downsampling operations. The resulting feature maps are then further learned through four cascaded RAB units. The first part of the RAB is a ContextBlock, using four dilated convolutions with dilation rates of 1, 2, 3, and 4 to increase the receptive field size. Following the Context Block are two residual connections. The final part of the RAB is a coordinate attention module that uses two 1D global pooling operations to aggregate the input features into two independent orientation-aware feature maps along the vertical and horizontal directions, respectively. These two feature maps, embedding orientation-specific information, are encoded into two attention maps. Each attention map captures a long-range dependency of the input feature map along a spatial direction; therefore, positional information can be stored in the generated attention map. The two attention maps are then applied to the input feature map through multiplication to emphasize the representation of interest. This operation can distinguish spatial orientation and generate coordinate-aware attention maps. The output of the Coordinate Attention Block can be explicitly represented as: （2） wherein the respectively represent the input and output of the attention module on the is the corresponding coordinate position; and are attention weights encoding position information in horizontal and vertical directions, respectively, in order to obtain the reconstructed neutron radiograph image, two transposed convolutions are used for up-sampling operation in the network, skip connections are used to connect the feature maps with the same size in the up-sampling and down-sampling processes to adapt to the input of any size, and finally, a 7*7 convolution is used to output the recovered neutron radiograph image. 2. The GAN neural network multi-distortion suppression model based on coordinate attention mechanism according to claim 1, characterized in that: Clear, real-world neutron radiograph images were selected as labels. Based on a neutron radiograph image degradation model, Gaussian blur, Gaussian noise, and Poisson noise were added to the clear, real-world neutron radiograph images. Noise following a gamma distribution was selected to simulate real-world neutron radiograph images with multiple distortions, thereby constructing a neutron radiograph image training dataset. The multiple distortion suppression model was then trained. The aforementioned neutron radiograph image degradation model is expressed as follows: （1） in These are real neutron radiographs containing multiple distortions. For the restored, clear neutron radiograph image, For Gaussian blur kernel, This represents the convolution operation. Represents Poisson noise interference. With a mean of 0 and a variance of Additive Gaussian noise, b is white spot noise, and multiplicative gamma noise is selectively added according to different images to be processed. In the above formula, G is used to represent it.
3. The GAN neural network multiple distortion suppression model based on coordinate attention mechanism according to claim 2, characterized in that: The construction of the neutron photography image training dataset includes: (1) For each reference image, 30 random selections were made from integers between 1 and 15, which were used as the size of the Gaussian blur kernel. In the first 15 blurring operations, the standard deviation of the two-dimensional Gaussian kernel function was set as the kernel size. The value was set to this in the last 15 processing steps. This process serves to expand the training dataset and the blurry yet clear neutron photographic images; (2) 20 intervals were generated. The uniformly distributed random numbers on the image are used as the variance of the Gaussian noise, and Gaussian noise with a mean of 0 and a variance of the selected value is added to the image obtained in the previous step. (3) Add noise following a Poisson distribution to the neutron radiograph image generated in the previous step. The Poisson distribution follows the following probability density function. （8） In the formula: —Pixels under Poisson noise interference; —The original pixels without Poisson noise; (4) For different neutron photographic images to be processed, noise following a gamma distribution is selectively added to the training dataset.
4. The GAN neural network multi-distortion suppression model based on coordinate attention mechanism according to claim 2 or 3, characterized in that: The GAN neural network is trained using real white spot noise to simulate white spot noise, which is then randomly added to clear, real neutron radiograph images, i.e., the training dataset.
5. The GAN neural network multi-distortion suppression model based on coordinate attention mechanism according to claim 4, characterized in that: The real white spot noise refers to white spot data collected from real neutron radiographs containing multiple distortions using a magnification and cropping method.
6. The GAN neural network multi-distortion suppression model based on coordinate attention mechanism according to claim 5, characterized in that: Using Huber loss, the aforementioned neutron radiograph image training dataset is used, with 90% as the training set and 10% as the test set. The GAN neural network multi-distortion suppression model based on coordinate attention mechanism is trained using the training set and tested using the test set. The real neutron radiograph image containing multiple distortions is used as input and fed into a pre-trained GAN neural network multiple distortion suppression model based on coordinate attention mechanism to predict the original image of the real neutron radiograph image containing multiple distortions, thus obtaining the target result.