Backlight X-ray image multilayer noise removal method based on deep learning

Abstract

The invention provides a backlight X-ray image multi-layer noise removal method based on deep learning, and the method comprises the steps: collecting a spherical shell contour in a backlight X-ray image, and carrying out the modeling through a computer, and obtaining a noise-free image data set; simulating image noise, and adding the image noise into the noiseless image data set to obtain a noisy image data set; constructing an image noise reduction model taking the FR-Unet architecture as a core, performing image enhancement by taking the noisy image data set as input, and outputting a noisy image after noise reduction; comparing the image with a noiseless image, calculating mean square error loss (MSE), and training for multiple times to obtain an optimal image noise reduction model; and applying the optimal image noise reduction model to a real experimental image, verifying the noise reduction effect of the model, and ensuring image quality and detail retention. According to the method, multi-layer noise can be effectively removed, high-efficiency noise reduction is realized while high image fidelity is kept, loss of important information is avoided, and the quality and accuracy of inertial confinement fusion experimental data analysis are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of X-ray imaging noise reduction in inertial confinement fusion (ICF) experiments, and in particular relates to a method for removing multi-layer noise from backlit X-ray images based on deep learning. Background Art

[0002] Back-illuminated X-ray imaging plays a vital role in inertial confinement fusion (ICF) experiments. Traditionally, the method of capturing back-illuminated X-ray signals using film packages has significant time inefficiency and limited resolution. To address these issues, digital detectors such as charge-coupled devices (CCDs) and complementary metal-oxide semiconductor (CMOS) cameras have been introduced, which are capable of providing fast and high-resolution imaging data. However, these advanced digital detectors also bring new challenges, namely, they are susceptible to shot noise, periodic speckle noise, and electronic noise, which originate from radiation dose, the scintillator used, and the imaging system itself. The presence of these noises seriously hinders the accurate observation and analysis of the internal evolution of matter in real experiments.

[0003] At present, X-ray image denoising technology is mainly divided into two categories: traditional methods and deep learning-based methods: traditional methods include spatial domain filtering (such as median filtering, mean filtering), frequency domain filtering (such as Wiener filtering, Fourier transform) and various statistical methods (such as non-local mean denoising). These methods can remove noise to a certain extent, but may cause edge blur or detail loss.

[0004] Deep learning-based methods, especially convolutional neural networks (CNNs), have made significant progress in the field of image processing in recent years. By learning from a large amount of annotated data, deep learning models can automatically learn and extract features in images, achieving more accurate noise reduction while preserving image details. The advantage of this method is that it can handle more complex noise patterns, but it requires a lot of computing resources and training data. Summary of the invention

[0005] In view of the above problems, this paper proposes a deep learning-based multi-layer noise removal method for backlit X-ray images to solve the challenges brought by complex noise in X-ray backlit images in the ICF experimental environment. The present invention adopts the frequency residual U-Net (FR-UNet) model, simulates the formation process of different types of physical noise, and combines the image synthesis algorithm to generate training data, aiming to effectively remove multi-layer noise including shot noise, periodic speckle noise and electronic noise.

[0006] Based on the above technical objectives, the present invention provides the following technical solutions:

[0007] A method for removing multi-layer noise from backlit X-ray images based on deep learning, which specifically includes:

[0008] S1. Design an image acquisition module to collect the spherical shell contour in the backlit X-ray image during the inertial confinement fusion experiment (ICF) through a digital detector CCD; then calibrate the optical and physical parameters through the spherical shell contour, and use a computer program to obtain the projection image of the spherical shell illuminated by X-rays as a noise-free image data set;

[0009] S2, design a noise simulation module, by simulating the image noise distribution similar to the distribution in the real experimental scene, adding the simulated image noise to the noise-free image dataset to obtain the noisy image dataset;

[0010] S3, construct an image denoising model with FR-Unet architecture as the core, take the noisy image dataset obtained in step S2 as input, perform image enhancement, and output the denoised noisy image;

[0011] S4, image denoising model loss function calculation; the image mean square error loss MSE is calculated on the denoised noisy image and the noise-free image data set obtained in step S1 to evaluate the denoising effect of the model and train the model; through multiple iterations and optimization, a trained optimal image denoising model is obtained;

[0012] S5. Apply and verify; apply the optimal image denoising model to real inertial confinement fusion experimental images, verify the denoising effect of the model, and ensure image quality and detail retention.

[0013] Furthermore, step S2 specifically includes:

[0014] S21, intercepting the non-spherical shell projection part of the backlit X-ray image as the experimental image noise, and converting it into a double-precision floating point type;

[0015] S22, design a circular sampling noise image block, traverse the experimental image noise from the center of the image with a row and column step size of fixed pixel intervals to extract the image block; first perform detrending processing on each image block, calibrate the experimental image noise, and then calculate the radial average power spectrum of the experimental image noise;

[0016] S23, radially rotating the radial average power spectrum obtained in step S22 to obtain an average noise power spectrum, then performing an inverse Fourier transform to obtain a spatial convolution kernel of the average noise power spectrum, and performing spatial convolution with Gaussian white noise to obtain simulated shot noise;

[0017] S24, converting the experimental image noise from the spatial domain to the frequency domain through Fourier transform, wherein the periodic structure in the collected image appears as discrete and sharp peaks in the frequency domain, and the corresponding frequency components are separated to obtain simulated periodic speckle noise;

[0018] S25, adding the simulated shot noise obtained in step S23 and the simulated periodic speckle noise obtained in step S24 to the noise-free image data set, so as to obtain a noisy image data set.

[0019] More specifically, in step S22: the detrending process includes: removing abnormally bright image areas through wavelet transform, and subtracting the mean of the image block to obtain a zero-mean image block; the calculation of the radial mean power spectrum of the experimental image noise specifically includes: first performing a two-dimensional discrete Fourier transform on the calibrated experimental image noise to obtain a noise power spectrum, then performing a radial analysis on the noise power spectrum to obtain a radial NPS, and accumulating the radial NPS of each image block and then calculating the average to obtain the radial mean power spectrum of the experimental image noise.

[0020] Furthermore, step S3 specifically includes:

[0021] S31, preprocessing the input noisy image data set, and adjusting the image pixel values ​​to between [-1, 1] by normalization;

[0022] S32, constructing a FR-Unet network model, wherein the FR-Unet network model includes three parts: a backbone feature extraction part, an enhanced feature extraction part, and an image reconstruction part;

[0023] S33, inputting the normalized noisy image data set into the FR-Unet network for multi-layer denoising to obtain a denoised noisy image.

[0024] Furthermore, the specific structures of the three parts of the FR-Unet network model are:

[0025] Backbone feature extraction part: It includes n backbone feature extraction layers, each of which includes a self-learning frequency domain module, a dense residual module and a maximum pooling layer; the backbone feature extraction module is used to extract and stack image features to obtain a series of preliminary effective image features;

[0026] Enhanced feature extraction part: including m enhanced feature extraction layers, each layer includes a self-learning frequency domain module, a dense residual module and an upsampling layer; used to fuse preliminary effective image features and enhance the expression ability of image features;

[0027] Image reconstruction part: Through a convolution layer, the output features of the enhanced feature extraction part are resized to reconstruct a noise-free X-ray spherical shell projection image.

[0028] More specifically, the self-learning frequency domain module is used to achieve noise reduction and image enhancement, and the specific implementation process includes:

[0029] In the self-learning frequency domain module, Fourier transform is first performed to transform the image spatial features input into the self-learning frequency domain module into image frequency domain features;

[0030] The image frequency domain features are then multiplied by a complex convolution kernel to obtain a complex weight matrix that matches the dimension of the image frequency domain features. The real and imaginary parts of the complex weights in the matrix are recorded as trainable parameters R1 and R2, respectively. These two parameters are iteratively updated through back propagation during the training process.

[0031] The frequency spectrum of the input image feature is multiplied element by element by the complex weight matrix to complete the linear transformation, thereby enhancing the low-frequency signal and weakening the high-frequency signal;

[0032] Finally, an inverse Fourier transform is performed to convert the image features from the frequency domain back to the spatial domain to obtain the image spatial features after image enhancement.

[0033] Furthermore, step S4 specifically includes:

[0034] S41, define the mean square error MSE between the noise-reduced noisy image I and the noise-free image K of size m*n as:

[0035]

[0036] Among them, I(i,j) and K(i,j) represent the grayscale value of the image at (i,j) respectively;

[0037] S42. Driven by the Adam optimization algorithm, the learning rate is adaptively adjusted according to the mean square error (MSE), so that the model dynamically explores and converges to a set of optimal network parameters θ; through multiple cycles of iteration and optimization, each iteration aims to refine the network weights and gradually improve the model's ability to restore clear details from the noisy image after denoising, and finally output the trained optimal image denoising model.

[0038] Based on the above technical solution, the present invention has the following beneficial effects:

[0039] The method proposed in the present invention designs a denoising model that combines frequency domain processing with densely connected residual learning. While enhancing key information in the image, it also improves the nonlinear expression capability and alleviates the gradient vanishing problem. It greatly improves the quality and accuracy of inertial confinement fusion experimental data analysis and promotes the development of nuclear fusion research. Compared with traditional methods, the solution provided by the present invention can achieve more efficient noise reduction while maintaining high image fidelity, avoiding the loss of important information. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the schematic diagram of the backlit X-ray imaging system and the result diagram;

[0041] Figure 2 Schematic diagram of modeling backlight projection on a spherical shell;

[0042] Figure 3 Obtain flow charts for noise-free image datasets and noisy image datasets;

[0043] Figure 4 This is a simulation process diagram of experimental image noise;

[0044] Figure 5 This is a schematic diagram of the FR-Unet network model structure;

[0045] Figure 6 This is a comparison chart of the noise reduction effect and pixel value curve;

[0046] Figure 7 The flowchart of the method proposed by the present invention is shown in FIG. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0049] like Figure 7 As shown, the present invention proposes a method for removing multi-layer noise from backlit X-ray images based on deep learning, which specifically includes the following steps:

[0050] S1. Design an image acquisition module to collect the spherical shell contour in the backlit X-ray image during the inertial confinement fusion experiment ICF through a digital detector CCD;

[0051] In this embodiment, Figure 1 As shown: Figure 1(a) describes a laser beam with a duration of 2.5 nanoseconds and an energy of 10 kilojoules shining on a metal strip made of molybdenum material. The metal strip measures 1.5 mm by 1.5 mm and is 5 microns thick. The X-rays emitted from the back of the metal strip will be as follows Figure 1 The geometry of the hollow spherical shell shown in (b) is projected onto the detector. The hollow sphere is about 1.5 mm away from the backlight source, while the detector is placed about 15 mm away from the backlight source, thus producing a magnification effect of about 10 times. The hollow spherical shell is filled with deuterium gas with a density of 2.0±0.4 mg / cc; the spherical shell is made of CH plastic with a density of 1.05 g / cc. On the detector side, a scintillator is used to convert X-rays into visible light, which is then coupled and recorded by a frame camera equipped with a charge-coupled device (CCD);

[0052] Then the optical and physical parameters are calibrated by the spherical shell profile, such as Figure 2 As shown, the projection image of the spherical shell irradiated by X-rays is obtained by modeling using a computer program as a noise-free image data set;

[0053] S2, design a noise simulation module, by simulating the image noise distribution similar to the distribution in the real experimental scene, adding the simulated image noise to the noise-free image dataset to obtain the noisy image dataset;

[0054] In order to effectively train the deep learning denoising model, a sufficiently large image dataset is required; however, in an experimental environment, it is limited to obtain a large number of noisy image resources. Therefore, in order to solve this problem, the present invention generates more training data by simulating the noise in the experiment; and as a preferred implementation, Figure 3 , 4 As shown, step S2 specifically includes:

[0055] S21, intercepting the non-spherical shell projection part of the backlit X-ray image as the experimental image noise, and converting it into a double-precision floating point type;

[0056] S22, design a circular sampling noise image block, traverse the experimental image noise from the center of the image with a row and column step size of fixed pixel intervals to extract the image block; first perform detrending processing on each image block, calibrate the experimental image noise, and then calculate the radial average power spectrum of the experimental image noise;

[0057] The detrending process includes: removing abnormally bright image areas by wavelet transform, subtracting the mean value of the image block to obtain a zero-mean image block;

[0058] The radial average power spectrum of the experimental image noise is calculated specifically as follows: firstly, a two-dimensional discrete Fourier transform is performed on the calibrated experimental image noise to obtain a noise power spectrum, then a radial analysis is performed on the noise power spectrum to obtain a radial NPS, and the radial NPS of each image block is accumulated and then averaged to obtain a radial average power spectrum of the experimental image noise;

[0059] S23, radially rotating the radial average power spectrum obtained in step S22 to obtain an average noise power spectrum, and then performing inverse Fourier transform to obtain the spatial convolution kernel of the average noise power spectrum, and performing spatial convolution with Gaussian white noise to obtain simulated shot noise; as described in step 1, there is also periodic dark spot noise in the imaging system; therefore, periodic speckle noise can be added to the noisy image. Since there is granular noise in the experimental image, it is a periodic speckle noise. We add periodic speckle noise based on the experimental image. The purpose of doing so is to make the simulated image noise closer to the experimental noise, so that the network model can better remove the noise in the experiment;

[0060] S24. The experimental image noise is converted from the spatial domain to the frequency domain through Fourier transform. The periodic structure in the collected image appears as discrete and sharp peaks in the frequency domain. The corresponding frequency components are separated to obtain simulated periodic speckle noise.

[0061] S25, adding the simulated shot noise obtained in step S23 and the simulated periodic speckle noise obtained in step S24 to the noise-free image dataset, thus obtaining a noisy image dataset. Figure 4 China X-ray backlit spherical shell projection image dataset.

[0062] In this embodiment, in order to perform supervised training on the FR-Unet network model, paired training data sets must be used to constrain and guide model learning, that is, each pair contains a noisy image and its corresponding noise-free image. Therefore, a paired data set required for network model training is constructed: one side is the original, clear X-ray backlit spherical shell projection image (noise-free), and the other side is the processed image with simulated noise added (noise). Through such a data set pair, we can effectively train the network model to learn to distinguish and remove noise and restore the clarity of the image.

[0063] S3, construct an image denoising model with FR-Unet architecture as the core, take the noisy image dataset obtained in step S2 as input, perform image enhancement, and output the denoised noisy image;

[0064] As a preferred embodiment, Figure 5 As shown, step S3 specifically includes:

[0065] S31, preprocessing the input noisy image data set, and adjusting the image pixel values ​​to between [-1, 1] by normalization;

[0066] S32, constructing a FR-Unet network model, wherein the FR-Unet network model includes three parts: a backbone feature extraction part, an enhanced feature extraction part, and an image reconstruction part;

[0067] The backbone feature extraction part includes n backbone feature extraction layers, each of which includes a self-learning frequency domain module, a dense residual module and a maximum pooling layer; the backbone feature extraction module is used to extract and stack image features to obtain a series of preliminary effective image features; Figure 5 As shown, modules 1-3 are the main feature extraction parts. The input image feature is F0, whose size is (H0, W0, C0). After module 1, the image feature F1 with a size of (H1, W1, C1) is obtained. H, W, C represent the image feature height, width and channel dimension respectively; the formula is expressed as:

[0068] F1=DownSample(LFT(DCR(F0)));

[0069] Among them, DCR() represents the dense residual module, LFT() represents the self-learning frequency domain module; DownSample() represents downsampling; similarly, image features F2 and F3 are obtained in turn; module 4 is used as the connection between the backbone feature extraction part and the enhanced feature extraction part, and its output image feature F4 has the same height and width as F3, and only the channel dimension is different;

[0070] In this embodiment, the self-learning frequency domain module is used to achieve noise reduction and image enhancement. The specific implementation process includes:

[0071] In the self-learning frequency domain module, Fourier transform is first performed to transform the image spatial features input into the self-learning frequency domain module into image frequency domain features;

[0072] The image frequency domain features are then multiplied by a complex convolution kernel to obtain a complex weight matrix that matches the dimension of the image frequency domain features. The real and imaginary parts of the complex weights in the matrix are recorded as trainable parameters R1 and R2, respectively. These two parameters are iteratively updated through back propagation during the training process.

[0073] The frequency spectrum of the input image feature is multiplied element by element by the complex weight matrix to complete the linear transformation, thereby enhancing the low-frequency signal and weakening the high-frequency signal;

[0074] Finally, an inverse Fourier transform is performed to convert the image features from the frequency domain back to the spatial domain to obtain the image spatial features after image enhancement;

[0075] Enhanced feature extraction part: includes m enhanced feature extraction layers, each layer includes a self-learning frequency domain module, a dense residual module and an upsampling layer; used to fuse preliminary effective image features and enhance the expression ability of image features; such as Figure 5 As shown, modules 5-7 are the enhanced feature extraction parts; taking module 5 as an example, the image feature F4 is first upsampled and concatenated with the image feature F3 on the channel dimension C4, and then passes through the dense residual module and the self-learning frequency domain module in sequence to obtain the image feature F5 of size (H5, W5, C5), which is expressed as follows:

[0076] F5=DCR(LFT(Contact(F3,UpSample(F4))));

[0077] Among them, Contact() represents concatenation, and UpSample() represents upsampling; similarly, image features F6 and F7 are obtained in sequence;

[0078] Image reconstruction part: Reconstruct the noise-free X-ray spherical shell projection image using the output features of the enhanced feature extraction part; that is, by performing a convolution operation on the output feature F7 of module 7 to convert the image feature dimension into (H0, W0, C0) to be consistent with the input image feature dimension, the X-ray backlit spherical shell projection image after noise reduction is reconstructed;

[0079] S33. According to the above steps, the normalized noisy image data set is input into the FR-Unet network for multi-layer denoising, so as to obtain a denoised noisy image.

[0080] S4, image denoising model loss function calculation; the image mean square error loss MSE is calculated on the denoised noisy image and the noise-free image data set obtained in step S1 to evaluate the denoising effect of the model and train the model; through multiple iterations and optimization, a trained optimal image denoising model is obtained;

[0081] As a preferred implementation, step S4 specifically includes:

[0082] S41, define the mean square error MSE between the noise-reduced noisy image I and the noise-free image K of size m*n as:

[0083]

[0084] Among them, I(i,j) and K(i,j) represent the grayscale value of the image at (i,j) respectively;

[0085] S42. Driven by the Adam optimization algorithm, the learning rate is adaptively adjusted according to the mean square error (MSE), so that the model dynamically explores and converges to a set of optimal network parameters θ; through multiple cycles of iteration and optimization, each iteration aims to refine the network weights and gradually improve the model's ability to restore clear details from the noisy image after denoising, and finally output the trained optimal image denoising model.

[0086] S5. Application and verification: Apply the optimal image denoising model to real inertial confinement fusion experimental images to verify the denoising effect of the model and ensure image quality and detail retention. The final denoising effect, image quality and detail retention are as follows: Figure 6 As shown:

[0087] The first and third columns in the figure respectively represent the experimental data image and the image after denoising by the FR-Unet network, and the second column represents the effect comparison in the local frame. It can be clearly seen that the method proposed in the present invention has a good denoising effect.

[0088] The fourth column shows the pixel value distribution curves of the actual X-ray backlit image and the denoised X-ray backlit image in the central area of ​​the image, comparing the pixel changes in the region of interest (ROI) between the original image and the denoised image. The red curve represents the pixel value distribution of the actual X-ray image, which shows a large fluctuation, meaning that there are significant differences between adjacent pixel values. In contrast, the blue curve represents the pixel value distribution after denoising using the FR-UNet model proposed in the present invention, which appears smoother, indicating that the difference between adjacent pixel values ​​is small.

[0089] In summary, the present invention proposes a frequency residual U-Net (FR-UNet) model, which can effectively remove multi-layer noise including shot noise, periodic speckle noise and electronic noise by simulating the formation process of different types of physical noise and combining image synthesis algorithm to generate training data. Compared with traditional methods, the solution provided by the present invention can achieve more efficient noise reduction while maintaining high image fidelity, avoiding the loss of important information.

[0090] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0091] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for removing multi-layer noise from backlit X-ray images based on deep learning, characterized in that: The specific steps include: S1. Design an image acquisition module to collect the spherical shell contour in the backlit X-ray image during the inertial confinement fusion experiment (ICF) through a digital detector CCD; then calibrate the optical and physical parameters through the spherical shell contour, and use a computer program to model the calibrated parameters to obtain the X-ray irradiated spherical shell projection image as a noise-free image data set; S2. Design a noise simulation module, by simulating image noise distribution similar to that in a real experimental scene, and adding simulated image noise to a noise-free image dataset to obtain a noisy image dataset; the simulated image noise includes simulated shot noise and simulated periodic speckle noise; S3, construct an image denoising model with FR-Unet architecture as the core, take the noisy image dataset obtained in step S2 as input, perform image enhancement, and output the denoised noisy image; S4, image denoising model loss function calculation; the image mean square error loss MSE is calculated on the denoised noisy image and the noise-free image data set obtained in step S1 to evaluate the denoising effect of the model and train the model; through multiple iterations and optimization, a trained optimal image denoising model is obtained; S5. Apply and verify; apply the optimal image denoising model to real inertial confinement fusion experimental images, verify the denoising effect of the model, and ensure image quality and detail retention.

2. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 1, characterized in that: Step S2 specifically includes: S21, intercepting the non-spherical shell projection part of the backlit X-ray image as the experimental image noise, and converting it into a double-precision floating point type; S22, design a circular sampling noise image block, traverse the experimental image noise from the center of the image with a row and column step size of fixed pixel intervals to extract the image block; first perform detrending processing on each image block, calibrate the experimental image noise, and then calculate the radial average power spectrum of the experimental image noise; S23, radially rotating the radial average power spectrum obtained in step S22 to obtain an average noise power spectrum, then performing an inverse Fourier transform to obtain a spatial convolution kernel of the average noise power spectrum, and performing spatial convolution with Gaussian white noise to obtain simulated shot noise; S24, converting the experimental image noise from the spatial domain to the frequency domain through Fourier transform, wherein the periodic structure in the collected image appears as discrete and sharp peaks in the frequency domain, and the corresponding frequency components are separated to obtain simulated periodic speckle noise; S25, adding the simulated shot noise obtained in step S23 and the simulated periodic speckle noise obtained in step S24 to the noise-free image data set, so as to obtain a noisy image data set.

3. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 2, characterized in that: In step S22: The detrending process includes: removing abnormally bright image areas by wavelet transform, subtracting the mean value of the image block to obtain a zero-mean image block; The radial average power spectrum of the experimental image noise is calculated specifically as follows: first, a two-dimensional discrete Fourier transform is performed on the calibrated experimental image noise to obtain a noise power spectrum, then the noise power spectrum is radially analyzed to obtain a radial NPS, the radial NPS of each image block is accumulated and then the average is calculated to obtain a radial average power spectrum of the experimental image noise.

4. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 1, characterized in that: Step S3 specifically includes: S31, preprocessing the input noisy image data set, and adjusting the image pixel values ​​to between [-1, 1] by normalization; S32, constructing a FR-Unet network model, wherein the FR-Unet network model includes three parts: a backbone feature extraction part, an enhanced feature extraction part, and an image reconstruction part; S33, inputting the normalized noisy image data set into the FR-Unet network for multi-layer denoising to obtain a denoised noisy image.

5. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 4, characterized in that: The specific structure of the three parts of the FR-Unet network model is: Backbone feature extraction part: It includes n backbone feature extraction layers, each of which includes a self-learning frequency domain module, a dense residual module and a maximum pooling layer; the backbone feature extraction module is used to extract and stack image features to obtain a series of preliminary effective image features; Enhanced feature extraction part: including m enhanced feature extraction layers, each layer includes a self-learning frequency domain module, a dense residual module and an upsampling layer; used to fuse preliminary effective image features and enhance the expression ability of image features; Image reconstruction part: Through a convolution layer, the output features of the enhanced feature extraction part are resized to reconstruct a noise-free X-ray spherical shell projection image.

6. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 5, characterized in that: The self-learning frequency domain module is used to achieve noise reduction and image enhancement, and the specific implementation process includes: In the self-learning frequency domain module, Fourier transform is first performed to transform the image spatial features input into the self-learning frequency domain module into image frequency domain features; The image frequency domain features are then multiplied by a complex convolution kernel to obtain a complex weight matrix that matches the dimension of the image frequency domain features. The real and imaginary parts of the complex weights in the matrix are recorded as trainable parameters R1 and R2, respectively. These two parameters are iteratively updated through back propagation during the training process. The frequency spectrum of the input image feature is multiplied element by element by the complex weight matrix to complete the linear transformation, thereby enhancing the low-frequency signal and weakening the high-frequency signal; Finally, an inverse Fourier transform is performed to convert the image features from the frequency domain back to the spatial domain to obtain the image spatial features after image enhancement.

7. The method for removing multi-layer noise from backlit X-ray images based on deep learning according to claim 1, characterized in that: Step S4 specifically includes: S41, define the mean square error MSE between the noise-reduced noisy image I and the noise-free image K of size m*n as: Among them, I(i,j) and K(i,j) represent the grayscale value of the image at (i,j) respectively; S42. Driven by the Adam optimization algorithm, the learning rate is adaptively adjusted according to the mean square error (MSE), so that the model dynamically explores and converges to a set of optimal network parameters θ; through multiple cycles of iteration and optimization, each iteration aims to refine the network weights and gradually improve the model's ability to restore clear details from the noisy image after denoising, and finally output the trained optimal image denoising model.

Images