Neutron image denoising method based on deep learning
Through a multi-stage denoising processing method based on deep learning, combined with harmonic mean filtering and U-Net framework, the problem of unsatisfactory denoising effect caused by the complexity of neutron image noise distribution is solved, and the effect of efficient denoising and image detail retention is achieved.
Patent Information
- Application Number
- CN202510089968.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-05-16
AI Technical Summary
The prior art is difficult to effectively remove noise in neutron images, especially the spatial non-uniformity and randomness of the noise distribution, making the noise denoising effect unsatisfactory.
A neutron image denoising method based on deep learning is adopted, including harmonic mean filtering in the preprocessing stage and multi-stage denoising processing of deep learning networks. The specific steps include: extracting noise characteristic parameters, using harmonic mean filtering for preliminary denoising, inputting the image to the image denoising network of the U-Net framework, feature extraction and fusion through the gated cycle unit, and finally outputting the denoised neutron image.
This method can adaptively process the non-uniform noise of neutron images, effectively remove noise and preserve image details, improve image quality and clarity, and is suitable for real-time neutron image processing.
Smart Images
Figure CN120013799A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image denoising, and in particular to a neutron image denoising method based on deep learning. Background Art
[0002] Neutron imaging is an important nondestructive testing technology that uses the property of neutrons to penetrate materials to obtain information about the internal structure of samples. Unlike X-ray imaging, neutrons are highly sensitive to light elements (such as hydrogen) and have strong penetration ability for heavy metals, which makes them widely used in the fields of nuclear industry, materials science, energy, aerospace, etc. Neutron imaging usually includes two forms: two-dimensional projection images and three-dimensional tomographic reconstruction images, which can provide information such as density distribution, internal defects, and dynamic behavior of samples. However, due to the low intensity of the neutron source, limited detector sensitivity, and noise interference in the experimental environment, neutron imaging data often contains a lot of noise. These noises not only reduce the quality of the image, but also have a serious impact on the subsequent image analysis and the accuracy of the results. Therefore, it is very important to effectively denoise the neutron image.
[0003] Image denoising is one of the basic problems in the field of image processing. Its goal is to remove noise interference as much as possible while retaining the structure and detail information of the image. Common image denoising methods include transform domain-based methods, spatial domain filtering methods, and deep learning methods. Transform domain methods such as wavelet transform denoising convert the image to the frequency domain and suppress noise using threshold operations; spatial domain methods such as median filtering and bilateral filtering directly operate on pixel values to remove isolated noise points or smooth regional noise; deep learning methods have developed rapidly in recent years, and have achieved good results in many applications by training neural networks to achieve efficient denoising. However, these methods have certain limitations when applied to neutron images, mainly because the noise distribution of neutron images has the characteristics of spatial non-uniformity and randomness, and existing denoising methods are difficult to adapt to their unique needs.
[0004] Existing image denoising methods have the following defects when processing neutron images. First, the transform domain method requires accurate modeling of the noise distribution, but the noise components of neutron images are complex and irregularly distributed, resulting in large modeling errors. Second, traditional spatial domain filtering methods such as median filtering and mean filtering will significantly lose image detail information, especially for edges and textures in neutron images. Although deep learning methods perform well on some standard data sets, their model training requires a large amount of labeled data, and it is very difficult to obtain high-quality denoised samples in neutron image experiments. In addition, these methods are generally not optimized for the non-uniform noise characteristics of neutron images, resulting in unsatisfactory denoising effects. Therefore, designing an efficient denoising method that can adapt to the noise distribution characteristics of neutron images, which can both retain image details and have high computational efficiency, is one of the key research issues in the current field of neutron imaging data processing.
[0005] In summary, this application proposes a neutron image denoising method based on deep learning. Summary of the invention
[0006] The purpose of the present invention is to propose a neutron image denoising method based on deep learning to address the problem that the optimization of the non-uniform noise characteristics of neutron images in the background technology leads to unsatisfactory denoising effect.
[0007] The technical solution of the present invention is a neutron image denoising method based on deep learning, comprising the following steps:
[0008] Preprocessing the noisy neutron image to extract noise characteristic parameters, wherein the noise characteristic parameters include noise distribution characteristics and noise intensity;
[0009] Performing preliminary denoising on the noisy neutron image using a harmonic mean filtering module to smooth the noise and effectively preserve the edge and texture details of the image;
[0010] The image after preliminary denoising is input into a first-stage image denoising network, which is based on the U-Net deep learning framework, gradually extracts information of different layers through multi-layer downsampling, and then restores it to the original size through multiple upsampling, and uses skip connections to avoid information loss;
[0011] Based on the gated recurrent unit, the original image and the output result of the first-stage image denoising network are extracted and fused to fuse the information of the two stages;
[0012] The fused feature map is used as the input of a two-stage denoising network, which further extracts image information, identifies noise in the image and removes the noise, and finally outputs a denoised neutron image.
[0013] Optionally, the step of preprocessing the noisy neutron image and extracting noise characteristic parameters also includes using the noisy neutron image and its noise characteristic parameters to train a deep learning noise parameter estimation network GFDNet, inputting the noise slice into the GFDNet, outputting the noise parameters through the deep learning model, adding simulated virtual noise to the neutron image to generate a noisier image, inputting the neutron image and the noisier image into a denoising deep learning model in pairs, and training the denoising deep learning model.
[0014] Optionally, in the step of performing preliminary denoising on the noisy neutron image using a harmonic mean filter module, the output of the harmonic mean filter is calculated using the following formula:
[0015]
[0016] Among them, (x, y) represents the coordinates of a pixel in the image to be filtered, I filitered (x, y) is the output of the harmonic mean filter at (x, y), n is the number of neighborhood pixels, I(x i ,y i ) is the value of the neighborhood pixel.
[0017] Optionally, in the one-stage image denoising network, the image is first preliminarily processed through a 3×3 convolution layer to convert the original image into a 64-channel feature map to extract basic features of the image.
[0018] Optionally, in the step of extracting and fusing features of the original image and the output result of the one-stage image denoising network based on the gated recurrent unit, the original image information is used as an early input, and the result of the one-stage denoising network is used as a new input. The gating mechanism dynamically adjusts the fusion method of the original image information and the one-stage denoising result according to different requirements of image features.
[0019] Optionally, the noisy image is a neutron image, and the neutron image is obtained by a neutron imaging device.
[0020] Optionally, the deep learning denoising model is trained by an L2 loss function to minimize the difference between the output image and the true denoised image.
[0021] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:
[0022] The present invention can adaptively process the non-uniform distribution and randomness of neutron image noise, thus overcoming the problem that the existing methods have poor adaptability to neutron image noise.
[0023] Harmonic mean filtering can remove noise pixel by pixel and retain edge texture details. The U-Net framework and gated feature fusion module can optimize the quality of denoised images and enhance images for subsequent analysis and application.
[0024] The model is trained using a variety of data to improve efficiency and effectiveness, solve the problem of difficult sample acquisition, and has low computational complexity, making it suitable for real-time neutron image processing;
[0025] The present invention effectively removes noise from neutron images and improves the quality and clarity of images by combining filtering methods, feature fusion, deep learning and other technologies. The combination of these steps can not only reduce the impact of noise on the image, but also maintain the detailed information of the image, especially playing an important role in the denoising process of complex images. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 This is a flowchart of a neutron image denoising method based on deep learning in an embodiment. DETAILED DESCRIPTION
[0027] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.
[0028] Examples, such as Figure 1 As shown, the present invention proposes a neutron image denoising method based on deep learning, which includes: using a noisy neutron image and its noise characteristic parameters to train a deep learning noise parameter estimation network GFDNet; inputting the noise slice into the GFDNet, and outputting noise parameters through a deep learning model; adding the simulated virtual noise to the neutron image to generate a noisier image; inputting the neutron image and the noisier image into a denoising deep learning model in pairs, and training the denoising deep learning model; and outputting a denoised image obtained by removing noise from the neutron image through the trained denoising deep learning model. The method is described in detail below.
[0029] Figure 1 FIG. 2 shows an example processing flow chart of a neutron image denoising method according to an embodiment of the present invention. Figure 1 As shown, the processing flow starts from step S110, firstly, the noisy neutron image to be processed is pre-denoised to obtain a preliminary clear image, which specifically includes using a harmonic mean filtering algorithm.
[0030] Then, step S120 is executed to input the preprocessed image into a first-stage image denoising network. Specifically, the image is converted into a 64-channel feature map through a 3*3 convolution, and then input into a U-shaped denoising network. After multiple layers of downsampling, information of different layers is gradually extracted, and then it is restored to its original size through multiple upsampling, and skip connections are used to avoid information loss.
[0031] Next, step S130 is performed. In this step, different image features are fused based on the gated recurrent unit. Specifically, the context information of the image is extracted based on the gated thermal fusion, the original image information is used as the early input, and the result of the first-stage denoising network is used as the new input, effectively fusing the information of the two stages.
[0032] In step S140, the fused features are processed, and the size of the fused feature map is the same as the input feature map size of the first-stage denoising network, which is a 64-channel feature map and can be directly used as the input of the second-stage denoising network. In this stage, image information is further extracted to better identify the noise in the image and remove the noise.
[0033] In the first step of image denoising, we need to preprocess the noisy sub-image to be processed and estimate the noise characteristics first. Noise estimation is a key step in image denoising, which helps us identify the type and distribution of noise. A suitable filter window size is hung. Next, a filtering method based on harmonic mean filtering is used. Harmonic mean filtering is a nonlinear filtering method that can smooth image noise while retaining the edge information of the image. Suppose I(x,y) is the value of a pixel in the image to be filtered, and its neighborhood pixel value is I(x i ,y i ), the output of the harmonic mean filter can be calculated by the following formula:
[0034]
[0035] Where n represents the number of neighboring pixels, I(x i ,y i ) is the value of the neighborhood pixel. Harmonic mean filtering can effectively suppress the noise in the image by taking a weighted average of the reciprocals of the neighborhood pixel values, while retaining the image details to the maximum extent. After filtering, the image noise is suppressed and the preliminary denoising result is obtained.
[0036] Then, the preprocessed image is input into the first stage of the image denoising network. Specifically, the image is first processed by a 3×3 convolution layer, the original image is converted into a 64-channel feature map, and the basic features of the image are extracted. Subsequently, the feature map is input into the denoising network based on the U-Net architecture. The network uses multi-layer downsampling operations to gradually extract the multi-scale features of the image and capture information at different levels, thereby enhancing the network's ability to distinguish details and noise. The downsampling operation of each layer not only reduces the spatial dimension, but also enhances the perception of global information. The downsampled feature map will be passed to the upsampling part for multiple operations to gradually restore to the original image size. During the upsampling process, the network directly passes the low-level features extracted in the downsampling process to the corresponding upsampling layer through skip connections, which can effectively avoid information loss and ensure that the image details are retained during the denoising process, thereby improving the denoising effect and the quality of image restoration. Finally, after this stage of processing, a denoised image close to the original size is output, providing a clear input for the subsequent stages.
[0037] In the process of image denoising, a method based on gated feature fusion is used to extract contextual information of the image and enhance the denoising effect. Specifically, the original image information is input in the early stages of the network to provide complete visual context and ensure that the network can utilize global features in the image. At the same time, the output results of the first-stage denoising network are passed as new input to subsequent network layers to provide feature information after preliminary denoising. This input method is effectively fused through a gating mechanism, which dynamically adjusts the fusion method of the original image information and the first-stage denoising results according to the different requirements of image features, thereby enhancing the network's ability to process information at different levels.
[0038] After feature fusion, the fused feature map will be further processed. The size of the fused feature map is consistent with the input feature map size of the first-stage denoising network, both of which are 64-channel feature maps, which can ensure the consistency and continuity of information transmission. Therefore, the fused feature map can be directly used as the input of the second-stage denoising network, avoiding additional resizing operations and ensuring efficient information flow. In the second stage, the denoising network further extracts multi-level information of the image. The network structure of this stage is more focused on refining the noise features of the image, and through further feature learning, it can more accurately identify potential noise areas in the image. Compared with the first stage, the network in the second stage can better capture the subtle differences in noise and effectively remove it from the image. Through the refined processing of features, the denoising network can optimize the image quality at a deeper level, not only removing noise, but also retaining the key details and structure of the image, achieving a higher denoising effect.
[0039] The present invention effectively removes noise from neutron images and improves image quality and clarity by combining filtering methods (especially harmonic mean filtering), feature fusion, deep learning and other technologies. The combination of these steps can not only reduce the impact of noise on the image, but also maintain the detailed information of the image, especially playing an important role in the denoising process of complex images (such as neutron images).
[0040] It is worth noting that the noise of neutron images has the characteristics of spatial non-uniformity and randomness, and the present invention is designed specifically for this characteristic. By extracting noise characteristic parameters such as non-uniform noise distribution and noise intensity in neutron images, and using a deep learning model to assume that the noise is non-Gaussian and modeling according to the noise characteristics of neutron images, the noise of neutron images can be processed adaptively, which improves the adaptability and flexibility of denoising methods to neutron image noise and overcomes the problem that existing denoising methods are difficult to adapt to the unique noise requirements of neutron images.
[0041] In this embodiment, the harmonic mean filter function is used to perform pixel-by-pixel denoising on the neutron image. This filtering method can smooth the noise while effectively retaining the edge and texture details of the image and reducing information loss. Based on the U-Net deep learning framework and combined with the gated feature fusion module, the multi-level features of the image are further extracted and fused, and the information of different layers is gradually extracted through multi-layer downsampling, and then restored to the original size through multiple upsampling. The jump connection is used to avoid information loss, and the gate mechanism dynamically adjusts the fusion mode according to the image features, etc., which can better retain the key details and structure of the image while removing the noise, further optimize the image quality after denoising, avoid the defects of traditional spatial domain filtering methods such as median filtering and mean filtering that will significantly lose image detail information, and overcome the problem that the denoising effect of the existing method is not ideal. The neutron image after denoising is enhanced, and the optimized denoised image is output, which further improves the image quality, makes the image clearer, and is more conducive to subsequent image analysis and application.
[0042] It should be noted that the present invention trains a deep learning denoising model by taking multiple learning data including virtual neutron image data with different noise characteristic parameters as input, and uses noisy neutron images and their noise characteristic parameters to train a deep learning noise parameter estimation network GFDNet, and generates simulated virtual noise based on the noise parameters, adds it to the neutron image to generate a noisier image, and then inputs the neutron image and the noisier image into the denoising deep learning model in pairs for training, thereby improving the efficiency and effect of model training, solving the problem that it is difficult for deep learning methods to obtain high-quality denoising samples in neutron image experiments, and also enabling the model to better learn and process the noise of neutron images. The denoising method of the present invention has a low amount of calculation, can efficiently process neutron imaging data, is suitable for real-time neutron image processing, meets the requirements for processing efficiency in practical applications, and has obvious advantages over some methods that are complex and time-consuming.
[0043] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.
Claims
1. A neutron image denoising method based on deep learning, characterized in that: The following steps are involved: Preprocessing the noisy neutron image to extract noise characteristic parameters, wherein the noise characteristic parameters include noise distribution characteristics and noise intensity; Performing preliminary denoising on the noisy neutron image using a harmonic mean filtering module to smooth the noise and effectively preserve the edge and texture details of the image; The image after preliminary denoising is input into a first-stage image denoising network, which is based on the U-Net deep learning framework, gradually extracts information of different layers through multi-layer downsampling, and then restores it to the original size through multiple upsampling, and uses skip connections to avoid information loss; Based on the gated recurrent unit, the original image and the output result of the first-stage image denoising network are extracted and fused to fuse the information of the two stages; The fused feature map is used as the input of a two-stage denoising network, which further extracts image information, identifies noise in the image and removes the noise, and finally outputs a denoised neutron image.
2. The neutron image denoising method based on deep learning according to claim 1, characterized in that: The step of preprocessing the noisy neutron image and extracting noise characteristic parameters also includes using the noisy neutron image and its noise characteristic parameters to train a deep learning noise parameter estimation network GFDNet, inputting a noise slice into the GFDNet, outputting noise parameters through a deep learning model, adding simulated virtual noise to the neutron image to generate a noisier image, inputting the neutron image and the noisier image into a denoising deep learning model in pairs, and training the denoising deep learning model.
3. The neutron image denoising method based on deep learning according to claim 1, characterized in that: In the step of performing preliminary denoising on the noisy neutron image using a harmonic mean filter module, the output of the harmonic mean filter is calculated using the following formula: Among them, (x, y) represents the coordinates of a pixel in the image to be filtered, I filitered (x, y) is the output of the harmonic mean filter at (x, y), n is the number of neighborhood pixels, I(x i ,y i ) is the value of the neighborhood pixel.
4. The neutron image denoising method based on deep learning according to claim 1, characterized in that: In the one-stage image denoising network, the image is first preliminarily processed through a 3×3 convolution layer to convert the original image into a 64-channel feature map to extract the basic features of the image.
5. The neutron image denoising method based on deep learning according to claim 1, characterized in that: In the step of extracting and fusing features of the original image and the output result of the one-stage image denoising network based on the gated recurrent unit, the original image information is used as an early input, and the result of the one-stage denoising network is used as a new input. The gating mechanism dynamically adjusts the fusion method of the original image information and the one-stage denoising result according to different requirements of image features.
6. The neutron image denoising method based on deep learning according to claim 1, characterized in that: The noisy image is a neutron image, and the neutron image is obtained by a neutron imaging device.
7. The neutron image denoising method based on deep learning according to claim 1, characterized in that: The deep learning denoising model is trained with an L2 loss function to minimize the difference between the output image and the true denoised image.
8. The neutron image denoising method based on deep learning according to claim 2, characterized in that: The GFDNet is trained by taking virtual noise data consisting of random noise parameters with Gaussian distribution noise as input and taking corresponding noise parameters as output, wherein the noise parameters include noise mean and standard deviation.