A neutron image denoising method and system based on a SUNet algorithm

Abstract

The application discloses a neutron image denoising method and system based on a SUNet algorithm, and the method comprises the following steps: obtaining an original neutron image from a neutron source; generating a noise image by using a Monte Carlo calculation method; constructing a denoising model based on the SUNet algorithm; training the denoising model based on the SUNet algorithm by using a perception loss function; constructing a training set based on the Waterloo Exploration Database and training the denoising model based on the SUNet algorithm; and reading in the original neutron image to generate a target image; the model after training can efficiently remove high-density gamma white spots in the neutron image, and the target image after denoising processing has higher resolution than the original image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of neutron imaging, in particular to a neutron image denoising method and system based on SUNet algorithm. BACKGROUND

[0002] Neutron radiography provides an effective tool for non-destructive testing research. Neutron radiography can deliver structural information of an object by measuring the intensity attenuation, scattering and other parameters of the neutron beam. Neutron radiography can be divided into many types according to different neutron sources, such as thermal neutron radiography, cold neutron radiography, fast neutron radiography and epithermal neutron radiography. The original neutron image obtained by the neutron detector based on CMOS sensor and CCD sensor contains noise and blur. In addition to camera dark current and sensor readout noise, the main noise source that reduces image quality in neutron imaging experiments is high-energy gamma particles generated by neutron nuclear reactions inside and around the experimental device. When these high-energy gamma particles hit the detector, they will produce bright white spots. Gamma white spots are a kind of random impulse noise. These gamma white spots of the original neutron image will further interfere with the analysis results of 2D and 3D tomographic reconstruction. Therefore, it is necessary to use a denoising method to eliminate the interference of gamma white spots while preserving the details of the neutron image.

[0003] In order to eliminate gamma white spot noise, the most commonly used techniques are median filter, convolution filter and low-pass filter (such as Gaussian filter) and improved filter methods, but using these methods cannot distinguish the edges of noisy images and flat areas, which will cause image edge smoothing and image detail blur, and when the gamma white spot is very large (sometimes as large as 150 pixels), it cannot be completely removed. In addition, some hybrid domain-based algorithms (such as Block-Matching and 3D filtering, BM3D) will over-smooth the original image and cannot remove various morphologies of gamma white spots, and a new SwinTransformer-based framework has recently been proposed for image denoising with good performance. In addition, this method has also been increasingly widely used in computer vision tasks such as image segmentation, object detection, inpainting and super-resolution. The existing convolutional neural network (CNNs) based methods cannot establish long-range dependencies and global context connections due to the limitations of the receptive field of convolution operation, therefore, a neutron image denoising method and system based on SUNet algorithm is now proposed. SUMMARY

[0004] In order to solve the problems mentioned in the background art, the purpose of the present application is to provide a neutron image denoising method and system based on SUNet algorithm.

[0005] The object of the application can be achieved by the following technical solutions: a neutron image denoising method based on a SUNet algorithm, the method comprising the following steps:

[0006] Obtaining an original neutron image from a neutron source;

[0007] Using a Monte Carlo calculation method to simulate the interaction between gamma rays and image sensors, normalizing and converting the simulated energy deposition data into 8-bit grayscale image pixel values to generate a noise image, and the noise image has the same characteristics as the noise in the original neutron image;

[0008] Constructing a denoising model based on the SUNet algorithm;

[0009] Training the denoising model based on the SUNet algorithm using a perception loss function;

[0010] Based on the Waterloo Exploration Database, a training set is constructed, and the denoising model based on the SUNet algorithm is trained using the training set, wherein 80% of the image pairs in the training set are used to train the denoising model, and 20% of the image pairs are used as a validation set;

[0011] Reading in the original neutron image as the input of the denoising model, and predicting the target image from the pre-trained denoising model according to the original image.

[0012] Preferably, the process of obtaining the original neutron image from the neutron source comprises the following steps:

[0013] Obtaining the original neutron image from a neutron source imaging system device, the neutron source imaging system device being a direct current / pulse dual-mode deuterium-tritium fusion neutron source composed of a high-voltage power supply, an ion source, a low-energy transmission section, an acceleration tube, a high-energy transmission section, a pulsing device, and a tritium target system;

[0014] After the neutron beam emitted by the neutron source imaging system device passes through the collimator and is collimated, it penetrates the object to be detected and attenuates, and the spatial distribution of the attenuated neutrons is received by the image detector to form a transmission projection image of the object to be detected. The neutron ray attenuation law follows the exponential attenuation law:

[0015]

[0016] In the formula, I is the transmission neutron intensity, I0 is the incident neutron intensity, T is the thickness of the sample object in the radiation direction, and μ x is the attenuation coefficient of the sample material to neutrons.

[0017] Preferably, the neutron source imaging system device can detect the light signal transmitted by the neutron source system and display the image through a series of processing, using CCD as the photosensitive element, including:

[0018] The mirror can change the original path direction of the light signal converted by the conversion screen, protect the camera, and avoid direct damage of the subsequent photosensitive system in the linear direction by scattered neutrons and radiation particles;

[0019] The focusing system is composed of high-flux large-aperture lenses and other optical components, and the scintillation screen sensor scans the light signal and reflects and focuses it on the CCD;

[0020] The refrigeration CCD camera is used to convert the light signal into a digital signal and collect information;

[0021] The radiation shielding device protects the imaging device during the entire imaging process because the various particle radiations emitted by the reaction have high energy, and there are still high-energy neutrons in the propagation process;

[0022] The optical fiber interface transmits image data information to the computer terminal.

[0023] Preferably, the process of simulating a noise image using the Monte Carlo calculation method includes the following steps:

[0024] Defining the energy of gamma rays, the material of the image sensor, and the geometry of the camera;

[0025] The noise image simulation data is processed using a logarithmic normalization method, and the normalization formula is as follows:

[0026]

[0027] Where x and x' correspond to the data before and after logarithmic normalization, respectively, and x' ∈ [0, 1].

[0028] Preferably, the process of constructing a denoising model based on the SUNet algorithm is as follows:

[0029] SUNet: SwinTransformerUNet adopts the skip connection structure in UNet and SwinTransformerblock to obtain more rich multi-scale image information and reduce the number of model parameters; the denoising model is constructed based on the SwinTransformer network, and the denoising model includes: a shallow feature extraction module, a UNet feature extraction module, and a reconstruction module.

[0030] Preferably, the shallow feature extraction module, the SUNet structure adopts a 3 × 3 convolutional layer M SFE (·) to obtain shallow information F shallowIn the UNet feature extraction module, SwinTransformer block is used instead of traditional convolution layer as the basic extraction module and high-level semantic information is obtained, and the resolution (H, W) and channel number (C) of the output feature are controlled to be the same as the convolution operation, F shallow Through the main feature extraction module M UFE (·) obtains F deep , M UFE (·) in a single STB contains 8 STLs; in the reconstruction module, a 3x3 convolution M R (·) is still used to generate a noise-free image from F deep , and the whole process is represented as:

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, LN(·) represents Layer Normalization, and MLP is a multilayer perceptron with two fully connected layers, wherein the fully connected layers have a Gaussian Error Linear Unit activation function.

[0036] Preferably, the process of training the denoising model based on the SUNet algorithm using the perceptual loss function is as follows:

[0037] The pre-trained VGG network is used for feature extraction, the perceptual loss function is used instead of the pixel-based loss, and the ReLU activation layer based on the pre-trained 19-layer VGG network, so the perceptual loss function has greater invariance to the change of pixel space, Φ i,j is represented as the feature map obtained after the jth convolution activation before the ith max pooling layer in the VGG19 network, and the VGG loss is defined as the reconstructed image The Euclidean distance between the feature representation of the reference image X and the formula is defined as follows:

[0038]

[0039] In the formula, W i,j and H i,j describe the width and height of each feature map in the VGG network.

[0040] Preferably, a neutron image denoising system based on the SUNet algorithm comprises:

[0041] image acquisition module: for acquiring original neutron image;

[0042] image denoising module: using SUNet algorithm to process original neutron image to generate target image.

[0043] Preferably, an apparatus comprising:

[0044] one or more processors;

[0045] a memory for storing one or more programs;

[0046] when one or more of the programs are executed by the one or more processors, the one or more processors implement a SUNet algorithm-based neutron image denoising method as described above.

[0047] Preferably, the computer executable instructions, when executed by a computer processor, are for performing a SUNet algorithm-based neutron image denoising method as described above.

[0048] Advantages of the present application:

[0049] In the first aspect, the present application provides a SUNet algorithm-based neutron image denoising method, which uses sample images with high-density gamma white spots collected in a neutron photographic device to perform neutron image denoising experiments. The Monte Carlo calculation method is used on the Super Monte Carlo nuclear simulation software system SuperMC to simulate the noise characteristics of real gamma white spots. The simulated noise image is used to make a data set, which is used for training the SUNet image restoration model. In addition, the perceptual loss function is used instead of the MSE-based loss function, and the trained model can efficiently remove high-density gamma white spots in the neutron image, and the target image after denoising processing has higher resolution than the original image.

[0050] In the second aspect, the present application provides a neutron image processing system, which comprises an original image acquisition module and an image denoising module. The original image of the image acquisition module is a sample image sampled in the neutron imaging system device of the HINEG neutron source. The image denoising module reads the original noise image through a data interface, and then loads the trained image denoising model to obtain a target image. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings;

[0052] Figure 1 is a simplified structural schematic diagram of a neutron imaging system in the present application;

[0053] Figure 2 is a structural schematic diagram of a digital imaging system using a CCD as a photosensitive element in the present application;

[0054] Figure 3 is a structural schematic diagram of a SUNet in the present application;

[0055] Figure 4 is a structural schematic diagram of a STB and a Swin Transformer Layer in the present application;

[0056] Figure 5 is a process schematic diagram of the entire denoising model from input to output in the present application;

[0057] Figure 6 is a generated image schematic diagram in the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0059] As shown in Figure 1 , a neutron image denoising method based on a SUNet algorithm, the method comprising the following steps:

[0060] obtaining an original neutron image from a neutron source;

[0061] obtaining an original neutron image from a neutron source;

[0062] using a Monte Carlo calculation method to simulate the interaction between gamma rays and image sensors, normalizing the simulated energy deposition data and converting them into 8-bit grayscale image pixel values, thereby generating a noise image, and the noise image has the same characteristics as the noise in the original neutron image;

[0063] constructing a denoising model based on a SUNet algorithm;

[0064] training the denoising model based on the SUNet algorithm using a perception loss function;

[0065] The training set is constructed based on the Waterloo Exploration Database, and the denoising model based on the SUNet algorithm is trained using the training set, wherein 80% of the image pairs in the training set are used to train the denoising model, and 20% of the image pairs are used as a verification set;

[0066] The original neutron image is read as the input of the denoising model, and the target image is output according to the pre-trained denoising model.

[0067] 1. Image acquisition module, from the neutron source imaging system device to obtain the original neutron image. The neutron source imaging system device is a direct current / pulse dual mode deuterium-tritium fusion neutron source, which mainly consists of a high-voltage power supply, an ion source, a low-energy transmission section, an acceleration tube, a high-energy transmission section, a pulsing device and a tritium target system. Figure 1 The simplified structure of the neutron imaging system is shown.

[0068] The neutron beam emitted by the neutron source imaging system device is collimated by the collimator, penetrates the detected object and attenuates. The spatial distribution of the attenuated neutrons is received by the image detector to form a transmission projection image of the detected object. The neutron ray attenuation law follows the exponential attenuation law:

[0069]

[0070] In the formula, I is the transmission neutron intensity, I0 is the incident neutron intensity, T is the thickness of the sample object in the radiation direction, μ x is the attenuation coefficient of the sample material to the neutron.

[0071] The detection imaging system after neutron photography is also called a digital imaging system, which can detect the light signal transmitted by the neutron source system and display the image through a series of processing. The structure of the digital imaging system using CCD as the photosensitive element is shown in Figure 2 The main components and functions of the digital imaging system are as follows:

[0072] The mirror can change the original path direction of the light signal converted by the conversion screen, so as to well protect the camera and avoid the direct irradiation damage of scattered neutrons and radiation particles to the subsequent photosensitive system in the straight line direction;

[0073] The focusing system is composed of high-flux large-aperture lenses and other optical parts, and the scintillation screen sensor scans the light signal and reflects and focuses it on the CCD;

[0074] The refrigeration CCD camera is used to convert the light signal into a digital signal and collect information.

[0075] The radiation shielding device protects the imaging device during the entire imaging process because the various particle radiation energies emitted by the reaction are large and high-energy neutrons still exist during propagation;

[0076] An optical fiber interface transmits image data information to a computer terminal.

[0077] 2. An image denoising module, which uses a SUNet algorithm to denoise the original neutron image. The original image obtained by the digital imaging system of neutron photography has a degradation phenomenon. The main reasons are as follows: on the one hand, the conversion screen efficiency of fast neutron photography is low, and a long time exposure is needed to obtain a clearer image, so more additional noise is brought. On the other hand, the CCD camera is affected by the self dark current during work due to long time exposure, thereby introducing Gaussian noise and electronic noise. During the process of CCD output digital signal, random noise interference is also mixed due to the circuit structure of itself, which affects the image quality. In the present application, the SwinTransformer model is combined with the U-net network model to obtain a denoising network model which can learn global information and is suitable for small sample data sets.

[0078] A. The Monte Carlo calculation method is used to simulate noise data. The real original neutron image has a high density of gamma white spots, and the calculation of the interaction between gamma rays and the sensor can simulate the generation of these gamma white spots. The generated noise image is added to the data set as the training set and test set for model training.

[0079] In the process of using the Monte Carlo calculation method, the energy of the gamma rays, the material of the image sensor and the geometry of the camera are defined. According to the decay mode of the cobalt 60 (Co60) source, the energy of the rays is determined to be 1.17 MeV and 1.33 MeV, respectively. Since the standard CMOS and CCD image sensors are both MOSFET materials, the material of the image sensor is set to silicon dioxide in the simulation experiment. The image sensor has a size of lw x lh x ld and is divided into a grid or pixel of W x H. lw, lh and ld represent the width, height and thickness of the sensor, respectively. W and H represent the width and height of the image resolution. The direction of the gamma rays emitted by the radiation source is uniformly distributed. Since the distance between the source and the sensor is much larger than the size of the sensor, most of the gamma rays will not affect the sensor. Therefore, we limit the range of the parallel gamma ray beam so that it only reaches the area of the sensor. Therefore, this method reduces the number of emitted gamma particles and the calculation time.

[0080] The Monte Carlo calculation method estimates the cumulative energy deposition of each grid. We finally want to obtain the noise image with GWS characteristics as an 8-bit grayscale image. It is crucial to select a suitable normalization algorithm to effectively convert the energy deposition corresponding to each grid into a pixel value. Considering that there are many orders of magnitude difference in energy deposition between different grids, we choose to log-normalize the simulation data. The normalization formula is as follows:

[0081]

[0082] Where x and x' correspond to the data before and after log-normalization respectively, and x' ∈ [0, 1]. The entire process of the Monte Carlo calculation method is summarized in the following algorithm, and the final result I noise ∈ [0, 255] is a WxH matrix.

[0083] B. Constructing an image denoising model. SwinTransformerUNet (SUNet) adopts the skip connection structure in UNet and Swin Transformer block to obtain more rich multi-scale image information and reduce the number of model parameters. The image denoising model of the present application is constructed based on the SwinTransformer network, Figure 3 The structure of SUNet is shown. Combined Figure 3 , the model uses SwinTransformer to construct a shallow feature extraction module, a UNet feature extraction module, and a reconstruction module. In the shallow feature extraction module, the SUNet structure adopts a 3x3 convolutional layer M SFE (·) to obtain shallow information F shallow . In the UNet feature extraction module, a Swin Transformer block (STB) is used instead of a traditional convolutional layer as the basic extraction module and obtains high-level semantic information. It can also control the resolution (H, W) and channel number (C) of the output features, which is the same as convolution operation. F shallow is obtained through the main feature extraction module M UFE (·). deep Figure 4 The STB and SwinTransformerLayer (STL) are shown. A single STB in M UFE (·) contains 8 STLs. In the reconstruction module, a 3x3 convolution M R (·) is still used to generate a noise-free image from F deep . Figure 5 is the process of the entire model from input to output. Taking Figure 4 as an example, the entire process can be represented as:

[0084] ​

[0085]

[0086]

[0087]

[0088] LN(·) represents Layer Normalization, and MLP is a multilayer perceptron with two fully connected layers, in which the fully connected layers have a Gaussian Error Linear Unit (GELU) activation function.

[0089] C. Training the model using a perceptual loss function. Using a pre-trained VGG network for feature extraction outperforms low-level pixel-wise error measures. Therefore, we adopted a custom perceptual loss function instead of a pixel-wise based loss. It is based on the ReLU activation layer of the pre-trained 19-layer VGG network, so the loss function has greater invariance to changes in the pixel space. Φ i,j is represented as the jth convolutional (after activation) feature map before the ith max-pooling layer in the VGG19 network. Define VGGloss as the reconstructed image Euclidean distance between the feature representation of the reference image X and the input image X, defined as follows:

[0090]

[0091] W i,j and H i,j Describe the width and height of each feature map in the VGG network.

[0092] D. Train the image denoising model using the modified training dataset. Some parameter settings of the network model during training are as follows: batchsize is 16, and the initial learning rate is 0.0001. To ensure that the entire network can converge rapidly and stably throughout the training process, the learning rate is step-decayed by 0.95 as the training iteration progresses. Save the model that performs best on the validation set, which has the lowest loss value. Evaluate the performance of the model on the internal test set at the end of the model training.

[0093] E. Input the original noisy image to be processed, and output the result as a neutron image without gamma white spots, as shown in Figure 6The present application does not need a large number of or a plurality of real neutron images as prior information, only needs to train the model in the simulated noise data set, and finally uses the pre-trained model to predict the noise image. This method can reduce the time cost of the experiment. In addition, the denoised image can provide more accurate samples for further data analysis and 2D and 3D tomographic reconstruction.

[0094] Based on the same inventive concept, the present application further provides a computer device, which comprises one or more processors and a memory for storing one or more computer programs; the program comprises program instructions, and the processor is configured to execute the program instructions stored in the memory. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are configured to implement one or more instructions, and are specifically configured to load and execute one or more instructions in the computer storage medium to implement the above method.

[0095] It needs to be further explained that, based on the same inventive concept, the present application further provides a computer storage medium, which stores a computer program, and the computer program is executed by the processor to perform the above method. The storage medium can adopt any combination of one or more computer readable media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples (non-exhaustive list) of the computer readable storage medium include: electrical connections with one or more conductive wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present application, the computer readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, device or component.

[0096] In the description of the specification, the description of the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in an appropriate manner.

[0097] The basic principles, main features and advantages of the present disclosure are shown and described above. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, various changes and improvements of the present disclosure can be made, which all fall within the scope of the claimed present disclosure.

Claims

1. A method for denoising a neutron image based on a SUNet algorithm, characterized in that, The method comprises the following steps: obtaining an original neutron image from a neutron source; using a Monte Carlo calculation method to simulate the interaction between gamma rays and an image sensor, normalizing the simulated energy deposition data and converting it into 8-bit grayscale pixel values to generate a noise image, and the noise image has the same characteristics as the noise in the original neutron image; constructing a denoising model based on the SUNet algorithm; The process of constructing the denoising model based on the SUNet algorithm is as follows: SUNet: SwinTransformerUNet adopts a skip connection structure in UNet and a SwinTransformerblock to obtain richer multi-scale image information and reduce the number of model parameters; The denoising model is constructed based on the SwinTransformer network, and the denoising model comprises a shallow feature extraction module, a UNet feature extraction module, and a reconstruction module; The shallow feature extraction module, the SUNet structure adopts 3 × 3 convolution layer M SFE (·) to obtain shallow information F shallow In the UNet feature extraction module, SwinTransformer block is adopted instead of the traditional convolution layer as the basic extraction module and high-level semantic information is obtained, and the resolution (H, W) and the channel number (C) of the output feature F shallow are controlled to be the same as the convolution operation. UFE (·) F deep is obtained. UFE (·) A single STB in M R (·) generates a noise-free image from F deep , and the whole process is represented as: In the formula, LN(·) represents Layer Normalization, and MLP is a multilayer perceptron with two fully connected layers, wherein the fully connected layers have a Gaussian Error Linear Unit activation function; The denoising model based on the SUNet algorithm is trained using a perception loss function; A training set is constructed based on the Waterloo Exploration Database, and the training set is used to train the denoising model based on the SUNet algorithm, wherein 80% of the image pairs in the training set are used to train the denoising model, and 20% of the image pairs are used as a validation set; The original neutron image is read in as input to the denoising model, and the pre-trained denoising model is used to predict the target image from the original image.

2. The method of claim 1, wherein, The process of obtaining the original neutron image from the neutron source comprises the following steps: The original neutron image is obtained from a neutron source imaging system device, which is a direct current / pulse dual-mode deuterium-tritium fusion neutron source composed of a high-voltage power supply, an ion source, a low-energy transmission section, an acceleration tube, a high-energy transmission section, a pulsing device, and a tritium target system; After the neutron beam emitted by the neutron source imaging system device is collimated by a collimator, it penetrates the object to be detected and attenuates, and the spatial distribution of the attenuated neutrons is received by an image detector to form a transmission projection image of the object to be detected, and the neutron ray attenuation law follows an exponential attenuation law: where I is the transmitted neutron intensity, I0is the incident neutron intensity, T is the thickness of the sample object in the direction of the radiation, and μ x is the attenuation coefficient of the sample material for neutrons.

3. The method of claim 2, wherein the SUNet algorithm is based on a neural network. The neutron source imaging system device can detect the optical signal transmitted by the neutron source system and display the image through a series of processing, and uses a CCD as a photosensitive element, comprising: A mirror can change the original path direction of the optical signal converted by the conversion screen, protect the camera, and avoid direct damage to the subsequent photosensitive system by scattered neutrons and radiation particles in the straight line direction; A focusing system composed of a high-flux large-aperture lens and other optical parts scans and reflects the optical signal of the scintillation screen sensor and focuses it on the CCD; A refrigeration CCD camera is used to convert the optical signal into a digital signal and collect information; The radiation shielding device protects the imaging device during the entire imaging process because the various particle radiation energies emitted by the reaction are large, and there are still high-energy neutrons during the propagation process; The optical fiber interface transmits image data information to a computer terminal.

4. The method of claim 1, wherein the SUNet algorithm is based on, The process of generating the noise image includes the following steps: Defining the energy of the gamma rays, the material of the image sensor, and the geometry of the camera; The noise image simulation data is processed using a log normalization method, and the normalization formula is as follows: Where x and x' correspond to the data before and after log normalization, respectively, and x' ∈ [0, 1].

5. The method of claim 1, wherein the SUNet algorithm is based on a neural network. The process of training the denoising model based on the SUNet algorithm using the perception loss function is as follows: Feature extraction is performed using a pre-trained VGG network, and a perceptual loss function is used instead of a pixel-based loss. The perceptual loss function is based on the ReLU activation layer of the pre-trained 19-layer VGG network, so it has greater invariance to changes in the pixel space, Φ i,j is represented as the feature map obtained after the jth convolutional activation before the ith max-pooling layer in the VGG19 network, and VGGloss is defined as the reconstructed image The Euclidean distance between the feature representation of the reference image X and the feature representation of the image to be reconstructed is defined as follows: wherein W i,j and H i,j Describes the width and height of each feature map in the VGG network.

6. A SUNet algorithm-based neutron image denoising system, adopting the SUNet algorithm-based neutron image denoising method of any one of claims 1-5, characterized in that, It includes: An image acquisition module for acquiring an original neutron image; An image denoising module for processing the original neutron image using the SUNet algorithm to generate a target image.

7. An apparatus, comprising: It includes: One or more processors; Memory for storing one or more programs; When one or more of the programs are executed by one or more of the processors, the one or more processors implement a neutron image denoising method based on the SUNet algorithm as claimed in any one of claims 1-5.

8. A storage medium containing computer-executable instructions, wherein: The computer executable instructions, when executed by a computer processor, are used to perform a neutron image denoising method based on the SUNet algorithm as claimed in any one of claims 1-5.

Images