X-ray image denoising domain gap bridging method based on physical information and data synthesis

Abstract

The invention discloses an X-ray image denoising domain gap bridging method based on physical information and data synthesis, and the method comprises the steps: obtaining a synthetic image of an X-ray image through physical data, and the synthetic image comprises a clean signal and synthetic noise; acquiring a real image of the X-ray image through an inertial confinement fusion (ICF) experiment; constructing a synthetic image data set and a real image data set; constructing an image noise reduction model, wherein the image noise reduction model comprises an enhancement module, a denoising module and three discriminators; the synthetic image data set and the real image data set are used as input of an image noise reduction model, and respectively pass through an enhancement module and a noise reduction module in sequence to obtain respective noise reduction results; a mixed loss function is designed, an image noise reduction model is guided to bridge a domain gap between a synthetic image and a real image, and the denoising quality is improved while the structural fidelity is kept. Compared with an existing baseline, the method has the advantages that the performance is better, and a denoising frame is more effective in a scientific imaging task with scarce data.

Description

Technical Field

[0001] This invention belongs to the field of X-ray imaging noise reduction technology in inertial confinement fusion (ICF) experiments, and particularly relates to a method for bridging the gap in the X-ray image noise reduction domain based on physical information and data synthesis. Background Technology

[0002] X-ray imaging is widely used in medical imaging applications, non-destructive industrial inspection, and diagnostics in high-energy-density physics (HEDP) research. The quality of the acquired X-ray images is a critical aspect of these applications, but it is often limited by permissible radiation dose and imaging system limitations. Noise is one of the most significant factors affecting image quality. Based on its origin, noise can be mainly classified into several types: shot noise, which is directly related to radiation dose and is significantly affected by scintillator characteristics; impulse noise, generated by more complex sources, such as the radiation environment; and electronic noise from digital recording equipment.

[0003] Currently, learning-based methods often achieve better results in enhancing X-ray image quality. Existing methods for recovering image signals in X-ray imaging systems have verified the superiority of the Unet method; however, they neglect the inter-domain differences between synthetic and real images, exhibiting low generalization and accuracy. Therefore, there is an urgent need to propose a hybrid denoising framework that combines physical simulation with generative modeling. This alleviates the problem caused by existing techniques ignoring the inter-domain differences between synthetic and real images, ultimately significantly improving SSIM and PSNR values ​​and achieving more effective denoising on real X-ray images. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a method for bridging the gaps in X-ray image denoising based on physical information and data synthesis, aiming to achieve more effective X-ray image denoising.

[0005] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution: A method for bridging the noise domain gap in X-ray images based on physical information and data synthesis, specifically including the following steps: S1. A composite image of X-ray images obtained from physical data, which includes clean signals and composite noise; S2. Real images of X-ray images acquired through inertial confinement fusion (ICF) experiments; S3. Construct synthetic image datasets and real image datasets; S4. Construct an image denoising model. The image enhancement model includes an enhancement module, a denoising module, and three discriminators. The synthetic image dataset and the real image dataset are used as inputs to the image denoising model. They are passed through the enhancement module and the denoising module in sequence to obtain their respective denoising results. During the enhancement and denoising process, joint adversarial learning is performed through the three discriminators in sequence to align the synthetic image and the real image in terms of image, features, and results. S5. Based on the denoising and alignment results in step S4, design a hybrid loss function to guide the image denoising model to bridge the domain gap between the synthesized image and the real image, thereby improving the denoising quality while maintaining structural fidelity.

[0006] Furthermore, step S1 specifically includes: S11. First, the synthesized image is specifically modeled as follows: ; in, Indicates a composite image; It is the clean signal of the synthesized image. K s and K n These represent the system fuzzy kernel and the scintillator fuzzy kernel, respectively. P (.) represents shot noise distributed in Poisson. This represents the impulse noise generated by neutron interactions. This represents the convolution operation; S12. The clean signal of the synthesized image is generated using the Legendre method. A 10th-order Legendre polynomial expansion is used to model the radial boundary of the clean signal hotspots, capturing the asymmetry and offset between the clean signal hotspots and the geometric center of the clean signal. The formula for the radial boundary is expressed as: ; in, It is a Legendre polynomial; c i It tracks the shape factor. c 0 indicates the overall size. c 1 indicates the degree of center offset, and the other higher-order terms indicate the degree of asymmetry; Is with angle The relevant radial boundary; Next, random sampling is performed by tracking the shape coefficient to synthesize a set of clean signal hotspot profiles with different degrees of asymmetry. The signal intensity distribution of pixels within the hotspot range is modeled as an exponential decay starting from the clean signal hotspot, modulated by the radial boundary. The formula for the signal intensity of pixels within the hotspot range is expressed as: ; in,s 0 is the peak intensity. ρ It is the radial distance from the hotspot. yes ρ Signal strength at the location; S13. After generating a clean signal dataset through step S12, the clean information is enhanced by introducing additional texture diversity using natural images. S14. The shot noise is modeled as a Poisson process, where the variance of the Poisson process is proportional to the expected signal strength. The shot noise texture is modeled as the convolution of white noise and the point spread function (PSF). The modeling formula is expressed as: ; in, N ( x, y The image shows white Gaussian noise. The point spread function is used; the shot noise level is controlled by a scaling factor α, which is estimated based on image patches with uniform luminosity; and a clean signal is used during the synthesis of shot noise. Given the pixel expectation, the shot noise is obtained. The formula is expressed as: ; S15. By analyzing real images obtained from ICF experiments, spike noise specks are synthesized, and a random walk-based algorithm is used to replicate the irregular shape and intensity decay of the specks. Then, linear stripes of spike noise are synthesized by placing noise clusters at random angles along a straight path to simulate the path of charged particles. Different shot parameters are randomly sampled from empirical distributions and embedded into clean signals. Shots are randomly added to clean signals to generate diverse spike noise. S16. Combining steps S14 and S15, a synthesized noise with both shot noise and impulse noise is obtained. Adding the synthesized noise to the clean signal yields a synthesized X-ray image.

[0007] Furthermore, step S13 specifically includes: First, the texture of the natural image is converted into a grayscale image, then reformatted to match the data structure of the clean signal. At the same time, the grayscale values ​​are normalized, and the background intensity and peak intensity ranges are set to match the dynamic range observed in the ICF experiment.

[0008] Furthermore, step S3 specifically includes: S31. Construct a synthetic image dataset consisting of three sets of paired images. Each set of paired images is an image pair of clean signal and synthetic noise. The first set uses only a small number of Legendre moments to represent hotspot asymmetry. The second set uses more Legendre moments with higher peak intensity and the same noise configuration than the first set to enhance diversity. The third set uses natural texture images and uses the same noise addition method as the second set. S32. Perform data augmentation on the real images to increase the number of samples and obtain a real image dataset.

[0009] Furthermore, the enhancement module described in step S4 consists of a Conv convolution layer, multiple dense blocks, and a fusion module. The Conv convolution and dense blocks are connected by residual jumps, and each dense block consists of multiple convolutions, an instantiation layer, and an activation function. The enhancement module takes synthetic noise as input and extracts high-level features through multiple dense blocks to enhance the synthetic noise.

[0010] Furthermore, the denoising module described in step S4 consists of two U-Net-like networks with shared weights. Each network has an encoder, a latent layer, and a decoder. The denoising process is as follows: The enhanced synthetic noise is added to the synthesized clean signal to obtain the enhanced synthetic image, while the real noise is added to the reconstructed real clean signal to obtain the real noisy image. The enhanced synthetic image and the real noisy image are respectively used as inputs to a U-Net-like network. They are first passed through convolutional layers and downsampling to extract high-level features in the encoder, and then decoded and denoised through upsampling layers and fusion layers. The two U-Net-like networks output the denoised tensor results of the enhanced synthetic image and the real noisy image, respectively. After inverse normalization of the denoised tensor results, they are saved as images of the same format to obtain the final visualized denoising result.

[0011] Furthermore, in step S4, the joint adversarial learning process, which sequentially uses three discriminators to align the synthetic image and the real image in terms of image, features, and result, is specifically described as follows: After enhancement, the enhanced synthetic noise and real noise are trained on image adversarial training through the first discriminator to align the enhanced synthetic noise and real noise from the image perspective. During the denoising process, the enhanced synthetic image and the real noisy image are subjected to feature adversarial training by the second discriminator in the latent layer of the denoising module; the enhanced synthetic image and the real noisy image are aligned in terms of features; after the two U-Net meshes output the denoising tensor results, the results are subjected to another adversarial training by the third discriminator to align the denoising tensor results; at the same time, the loss function is calculated for each discriminator, and the calculated loss function is included in the hybrid loss function in step S5.

[0012] Furthermore, the hybrid loss function described in step S5 is specifically expressed as follows: ; in, Represents the mixed loss function; This represents the image adversarial loss function set for the first discriminator; This indicates a loss of content retention for enhanced task settings; This represents the task loss set for the denoising task; This represents the feature adversarial loss set for the second discriminator; This represents the output adversarial loss set for the third discriminator; , , , , These are the weighting coefficients that control the contribution of each loss function.

[0013] Furthermore, the specific loss functions in step S5 are as follows: Image adversarial loss The synthesized image, enhanced by the image denoising model, is closer to the real image. The formula is expressed as follows: ; in, Represents real noise. This represents the enhanced synthesized noise. express , Interpolation between It is the gradient penalty coefficient; Indicates the first discriminator; Represents the L2 norm; Is the discriminator in the sample gradient at; This represents the mean; Loss of content retention The enhancement module is encouraged to retain high-level semantic results while reducing low-level adaptations, as expressed by the following formula: ; in, Indicates synthesized noise; It is the L1 norm. This indicates that features are extracted from the k-th layer of a pre-trained VGG19 network. For VGG19 network layers, The corresponding weights; Mission loss The formula for guiding the denoising module to restore visual fidelity and semantic structure is expressed as follows: ; in, y s The clean signal representing the synthesized image. y st Indicates the enhanced synthetic image x st The denoising results x st Depend on , y s Add them together to get; This is the set of network layers; weights β and γ control the contributions of image loss and perceptual loss, respectively. Feature-based adversarial loss The guiding image denoising model adjusts the stylistic attribute differences between synthetic and real images while preserving semantic content; its formula is expressed as: ; in, This indicates the shared encoder of the noise reduction module. It is a linear interpolation between the latent features of the enhanced synthetic image and the real noisy image. λ 2 represents the gradient penalty coefficient; This is a real, noisy image; This represents the second discriminator; Output to counter the loss To ensure that the synthesized image after denoising has visual realism close to that of the original denoised image, the formula is expressed as: ; in, This indicates the third discriminator. λ 3 represents the gradient penalty coefficient; This is the denoising result of the real image.

[0014] Based on the above technical solution, the present invention has at least the following beneficial effects: This invention proposes a high-performance hybrid X-ray image denoising framework that combines physics-based simulation with generative modeling to enhance the denoising of self-emission X-ray images in ICF experiments. By modeling signal and noise components according to physical principles and refining them through a generative enhancement module, the proposed method reduces the domain gap and generates realistic, statistically consistent synthetic data. A bi-branch, weight-shared denoising network guided by multi-level adversarial and perceptual losses enables robust domain-invariant learning. Experiments show that the framework outperforms traditional and learning-based baselines, highlighting its effectiveness in data-scarce scientific imaging tasks. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a backlit X-ray imaging system; Figure 2 Examples of different types of images in the dataset; Figure 3 This is a schematic diagram of the image denoising model proposed in this invention; Figure 4 A comparison chart showing the noise reduction results of different existing models with the model proposed in this invention; Figure 5 A comparison chart of noise reduction effects and pixel value curves for real images. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figure 1-5 The present invention will be further described in detail below with reference to specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0017] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0018] This invention proposes a method for bridging the gaps in the denoising domain of X-ray images based on physical information and data synthesis, which specifically includes: S1. A composite image of X-ray images obtained from physical data, which includes clean signals and composite noise; In a preferred embodiment, step S1 specifically includes: S11. First, the synthesized image is specifically modeled as follows: ; in, Indicates a composite image; It is the clean signal of the synthesized image. K s and K n These represent the system fuzzy kernel and the scintillator fuzzy kernel, respectively. P (.) represents shot noise distributed in Poisson. This represents the impulse noise generated by neutron interactions. This represents the convolution operation; S12. The signal originating from the central hotspot exhibits perfect rotational symmetry, with its peak intensity aligned with the geometric center. However, in practice, asymmetry and offset frequently occur. To capture these variations, the clean signal of the synthesized image in this application is generated using the Legendre method. A 10th-order Legendre polynomial expansion is used to model the radial boundary of the clean signal hotspot, capturing the asymmetry and offset between the clean signal hotspot and the clean signal's geometric center. The formula for the radial boundary is expressed as: ; in, It is a Legendre polynomial; c i It tracks the shape factor. c 0 indicates the overall size. c 1 indicates the degree of center offset, and the other higher-order terms indicate the degree of asymmetry; Is with angle The relevant radial boundary; Next, random sampling is performed by tracking the shape coefficient to synthesize a set of clean signal hotspot profiles with different degrees of asymmetry. The signal intensity distribution of pixels within the hotspot range is modeled as an exponential decay starting from the clean signal hotspot, modulated by the radial boundary. The formula for the signal intensity of pixels within the hotspot range is expressed as: ; in, s 0 is the peak intensity. ρ It is the radial distance from the hotspot. yes ρ Signal strength at the location; S13. During the training process, the model may overfit to a specific asymmetric pattern. Therefore, after generating a clean signal dataset in step S12, this application enhances the generalization ability by introducing additional texture diversity using natural images to enhance the clean information. Specifically, the textures of the natural images (such as rice grains and clothing textures) are first converted into grayscale images, then reformatted to adjust the natural textures to match the data structure of the clean signal. At the same time, the grayscale values ​​are normalized, that is, the background intensity strain is set to the range of 60-2000, and the peak signal value is set between 4000-24000, which is consistent with the dynamic range observed in the experiment. S14. The shot noise is modeled as a Poisson process, where the variance of the Poisson process is proportional to the expected signal strength. The shot noise texture is modeled as the convolution of white noise and the point spread function (PSF). The modeling formula is expressed as: ; in, N ( x, yThe image shows white Gaussian noise. The point spread function is used; the shot noise level is controlled by a scaling factor α, which is estimated based on image patches with uniform luminosity; and a clean signal is used during the synthesis of shot noise. Given the pixel expectation, the shot noise is obtained. The formula is expressed as: ; S15. By analyzing the real images obtained from the ICF experiment, speckles of impulse noise are synthesized, and a random walk-based algorithm is used to replicate the irregular shape and intensity decay of the speckles. Then, linear stripes of impulse noise are synthesized by placing noise clusters at random angles along a straight path to simulate the path of charged particles. Different shot parameters are randomly sampled from the empirical distribution and embedded into clean hotspot images. Shots are randomly added to the clean hotspot images (in this embodiment, specifically, 30,000 to 100,000 shots are added to every 512 images × 512 images) to generate diverse impulse noise. S16. Combining steps S14 and S15, a synthesized noise with both shot noise and impulse noise is obtained. Adding the synthesized noise to the clean signal yields a synthesized X-ray image.

[0019] S2. Real images of X-rays acquired through inertial confinement fusion (ICF) experiments (the real images in this application also contain noise); such as Figure 1 As shown, the real image acquisition module consists of a pinhole array, a filter, an imaging tube, a scintillator, a charge-coupled device (CCD), and a gas cell. The pinhole array is configured as an 8 × 7 grid with a row spacing of 0.5 mm and a column spacing of 0.45 mm, located in front of a pointer 112 mm from the black cavity target. Each experiment produces 56 pinhole images. Each pinhole has a diameter of approximately 15 μm. Integrated into the imaging path is a filter assembly consisting of a 200 μm thick beryllium layer and a 50 μm aluminum layer, placed inside the imaging tube. After the tube, a scintillator converts X-rays into visible light, which is then connected to the CCD, located 1186 mm from the pinhole array, to capture the generated visible photons. In actual experiments, due to the high cost, subsequent steps in this application involve generating a synthetic image that fits the real image, and also performing data augmentation on the acquired real image to ensure sufficient real data is used to bridge the domain gap between the synthetic and real domains.

[0020] S3. Construct synthetic image datasets and real image datasets for use as input to subsequent models; In a preferred embodiment, step S3 specifically includes: S31. Construct a synthetic image dataset consisting of three sets of paired images. Each set of paired images is an image pair of clean signal and synthetic noise. The first set uses only a small number of Legendre moments to represent hotspot asymmetry. In this embodiment, the first set uses four sets of Legendre moments, with peak intensity ranges set to 2,000-4,000, shot noise levels of 40 and 400, and 30,000–100,000 impulse noise spots, generating 3,360 pairs of paired images. The second group uses Legendre moments with a greater number of peak intensities (4,000–24,000) and the same noise configuration than the first group to enhance diversity, producing 15,120 paired images; During training, it was observed that using only Legendre-based synthetic data led to overfitting of the model to the specific signal patterns described by the Legendre function, resulting in distorted denoising output. Therefore, to improve generalization and prevent this bias, this application uses, in the third group, data such as... Figure 2 The natural texture images shown are natural texture images of five rice varieties (Arborio, Basmati, Ipsala, Jasmine, Karadag) and various clothing. After processing, using the same noise-adding method as the second group, 9,183 additional paired samples were generated. Finally, the synthetic dataset consists of 27,663 paired images, which takes into account both the richness and diversity of the samples. S32. Perform data augmentation on the real images to expand the number of samples and obtain a real image dataset; that is, apply data augmentation techniques, including rotation and flipping, to expand the real images to 1,144 samples.

[0021] Furthermore, both the synthetic and real image datasets were preprocessed, with all images standardized to a single-channel 16-bit format, resized to 512 × 512 pixels, and normalized to the range [0, 1]. The datasets were divided into training, validation, and test sets in a 7:2:1 ratio. The network was implemented in PyTorch and trained on an NVIDIA GeForce RTX 3090 GPU using a batch size of 4.

[0022] S4. Construct an image denoising model. The image enhancement model includes an enhancement module, a denoising module, and three discriminators. The synthetic image dataset and the real image dataset are used as inputs to the image denoising model. They are passed through the enhancement module and the denoising module in sequence to obtain their respective denoising results. During the enhancement and denoising process, joint adversarial learning is performed through the three discriminators in sequence to align the synthetic image and the real image in terms of image, features, and results. In actual experiments, a domain gap can be observed between the synthetic image and the real image, especially in texture and background, which limits the denoising generalization.

[0023] To alleviate this situation, this application first introduces an enhancement module to optimize the synthesized noise to better match the distribution of the real world. The enhancement module consists of a Conv convolutional layer, multiple dense blocks (three dense blocks in this embodiment, the number of dense blocks can be adjusted according to actual needs), and a fusion module (used to fuse all features and perform dimensionality reduction processing, so as to serve as the input of subsequent networks); In this embodiment, Conv convolutions and dense blocks are connected via residual jump connections, such as... Figure 3 As shown, the output of the Conv convolution is subtracted from the output of the second dense block and then input into the third dense block (the low-level features specific to the source domain can be suppressed through the subtraction residual). The output of the first dense block is then added to the output of the third dense block to obtain the residual (the addition residual is used to preserve the high-level semantic features shared across domains). Each dense block consists of multiple convolutions, instantiation, and activation functions. The enhancement module takes synthetic noise as input. For an input image tensor of size 16 in a single channel, the preprocessed tensor shape is [4,3,512,512]. The output shape after inputting into the enhancement module is also [4,3,512,512]. High-level features are extracted from multiple dense blocks to enhance the synthetic noise. In this application, as shown... Figure 3 As shown, after enhancement, the enhanced synthetic noise and real noise are trained on image adversarial training through the first discriminator to align the enhanced synthetic noise and real noise from the image perspective; that is, to reduce the inter-domain difference between the enhanced synthetic noise and real noise, so that the synthesized image can be more realistic. After enhancement, the process proceeds to the denoising module. In this application, the denoising module consists of two U-Net-like networks with shared weights, each with an encoder, a latent layer, and a decoder. The denoising process is as follows: The enhanced synthetic noise is added to the synthesized clean signal to obtain the enhanced synthetic image. The real noise is added to the reconstructed real clean signal (obtained from the real image through ordinary U-Net denoising) to obtain the real noisy image. The enhanced synthetic image and the real noisy image are respectively used as inputs to a U-Net-like network. They are first passed through convolutional layers and downsampling to extract high-level features in the encoder, and then through upsampling layers and fusion layers for decoding and denoising. After convolutional processing by the convolutional layers, the shape of the image tensor becomes [4, 64, 512, 5].

[12] Then, after multiple downsampling, the shape changes from [4,64,512,512] to [4,64,32,32]. After multiple upsampling, the shape is restored to [4,64,512,512]. Finally, after the effect of the fusion layer, the shape of the tensor changes back to [4,3,512,512]. The two U-Net-like networks output the denoising tensor results of the enhanced synthetic image and the real noisy image, respectively. After inverse normalization of the denoising tensor results, they are saved as images of the same format to obtain the final visualized denoising result.

[0024] During the denoising process, the enhanced synthetic image and the real noisy image are subjected to feature adversarial training by the second discriminator in the latent layer of the denoising module; the enhanced synthetic image and the real noisy image are aligned in terms of features; after the two U-Net meshes output the denoising tensor results, the results are subjected to another adversarial training by the third discriminator to align the denoising tensor results; at the same time, the loss function is calculated for each discriminator, and the calculated loss function is included in the hybrid loss function in step S5.

[0025] Meanwhile, during the denoising process, since the differences between the synthetic image and the real image usually lie in style attributes, such as differences in lighting and color distribution, in order to adjust these differences while preserving semantic content, the denoising process itself performs feature adversarial training on the enhanced synthetic image and the real noisy image through the second discriminator in the latent layer of the denoising module; aligns the enhanced synthetic image and the real noisy image in terms of features; after the two U-Net meshes each output the denoising tensor results, the third discriminator performs another result adversarial training to align the denoising tensor results, so as to ensure that the denoised synthetic image has visual realism similar to the denoised real image; Finally, loss functions are calculated for the first, second, and third discriminators, and the calculated loss functions are included in the mixed loss function in step S5.

[0026] Through the structural design of the above model, this application uses an enhancement module, a denoising module, and three discriminators to perform joint adversarial learning during training. These modules align the images, features, and results, respectively. The three alignment components can promote each other, thus combining to better construct cross-domain invariance and bridge the inter-domain gap between synthetic and real images.

[0027] S5. Based on the denoising and alignment results in step S4, design a hybrid loss function to guide the image denoising model to bridge the domain gap between the synthesized image and the real image, thereby improving the denoising quality while maintaining structural fidelity. In a preferred embodiment, the hybrid loss function in step S5 is specifically expressed as follows: ; in, Represents the mixed loss function; This represents the image adversarial loss function set for the first discriminator; This indicates a loss of content retention for enhanced task settings; This represents the task loss set for the denoising task; This represents the feature adversarial loss set for the second discriminator; This represents the output adversarial loss set for the third discriminator; , , , , These are the weighting coefficients that control the contribution of each loss function.

[0028] The hybrid loss function designed in this application balances pixel-level accuracy, semantic consistency, task supervision, style alignment, and output realism. By leveraging the contributions of each loss term through weight coefficients, the network is guided to narrow the interdomain gaps while maintaining structural fidelity, thereby improving denoising quality.

[0029] More specifically, the loss functions are as follows: Image adversarial loss The synthesized image, enhanced by the image denoising model, is closer to the real image. The formula is expressed as follows: ; in, Represents real noise. This represents the enhanced synthesized noise. express , Interpolation between It is the gradient penalty coefficient; Indicates the first discriminator; Represents the L2 norm; Is the discriminator in the sample gradient at; This represents the mean; Using a full-signal image with added noise during enhancement may unintentionally alter the signal, causing deviations from the fundamental facts. To avoid this, this application first performs preliminary denoising to extract synthetic noise. As a target, the entire denoising framework is then iteratively improved, gradually increasing its effectiveness. This improves image quality and achieves more accurate domain alignment. Simultaneously, a first discriminator, D1, is introduced, which uses Wasserstein GAN with Gradient Penalty (WGAN-GP) and, according to image adversarial loss, performs [the following] on the [the following]: and To make distinctions.

[0030] Loss of content retention The enhancement module is encouraged to retain high-level semantic results while reducing low-level adaptations, as expressed by the following formula: ; in, Indicates synthesized noise; It is the L1 norm. This indicates that features are extracted from the k-th layer of a pre-trained VGG19 network. For VGG19 network layers, The corresponding weights; The two loss functions mentioned above are used in the enhancement module; while the denoising module consists of two U-Net-like networks with shared weights. Weight sharing encourages the synthetic image branch and the real image branch to learn domain-invariant representations, focusing on common denoising principles such as noise statistics and structural features. By combining the enhancement noise rating from the previous stage, the denoising model adapts to both domains, solving the common problem that synthetic data has strong performance but poor performance in generating real images. Then, the following three loss functions guide the denoising module.

[0031] Mission loss As the primary loss function of the denoising framework, it guides the denoising module in restoring visual fidelity and semantic structure, and its formula is expressed as: ; in, y s The clean signal representing the synthesized image. y st Indicates the enhanced synthetic image x st The denoising results x st Depend on , y s Add them together to get; This is the set of network layers; weights β and γ control the contributions of image loss and perceptual loss, respectively. In this application, to achieve effective supervision of synthetic clean data, the task loss combines pixel-level accuracy and perceptual consistency.

[0032] Feature-based adversarial loss The guiding image denoising model adjusts the stylistic attribute differences between synthetic and real images while preserving semantic content; its formula is expressed as: ; in, This indicates the shared encoder of the noise reduction module. It is a linear interpolation between the latent features of the enhanced synthetic image and the real noisy image. λ 2 represents the gradient penalty coefficient; This is a real, noisy image; This represents the second discriminator; Output to counter the loss To ensure that the synthesized image after denoising has visual realism close to that of the original denoised image, the formula is expressed as: ; in, This indicates the third discriminator. λ 3 represents the gradient penalty coefficient; This is the denoising result of the real image.

[0033] Thus, the image denoising model and hybrid loss function proposed in this application achieve comprehensive alignment between the synthesized image and the real image through a three-level discriminator: the enhancement module with the first discriminator ensures image-level alignment, while the denoising modules with the second and third discriminators ensure feature-level alignment and output-level alignment, respectively. Ultimately, under the unified goal of bridging the domain gaps, they jointly achieve optimized denoising of X-ray images.

[0034] This embodiment also provides the results of actual experiments using the method proposed in this application, verifying the performance and superiority of the method. Schematic diagrams of denoising results for different models and comparisons of denoising effects and pixel value curves for real images are shown below. Figure 4 and Figure 5 As shown: like Figure 4 As shown, the denoising framework proposed in this application is compared with three representative baselines: DnCNN, Concat DnCNN, and U-Net. To ensure a fair comparison, all models were trained and validated on the same dataset using the same preprocessing methods and training schedule. Evaluation was performed on the validation set, and the evaluation metrics used were Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The figure shows that this invention has better domain adaptive enhancement capabilities, and its performance significantly outperforms all baselines.

[0035] To evaluate the model's performance on real experimental data, the denoising model proposed in this invention was further applied to all available X-ray images exhibiting severe noise pollution, such as... Figure 5 Some representative inspection results are presented. The leftmost two columns of images represent the real image and the denoised image, respectively. The remaining four columns are pixel value comparison charts drawn along straight lines at different angles from the image center, where blue represents the pixel curve of the real image and orange represents the pixel curve of the denoised image. It can be seen that the radial intensity profiles along multiple angles further demonstrate that the method proposed in this invention effectively suppresses shot noise and impulse noise while preserving the underlying spatial distribution, showing strong consistency with physical experiments.

[0036] In summary, this invention proposes a hybrid denoising framework that narrows the domain gap and generates realistic, statistically consistent synthetic data, achieving robust domain-invariant learning. Experiments show that this framework outperforms traditional and learning-based baselines, highlighting its effectiveness in data-scarce scientific imaging tasks.

[0037] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0038] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0039] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An X-ray image denoising inter-domain gap filling method based on physical information and data synthesis, characterized in that, Specifically comprising the following steps: S1, acquire a synthetic image of the X-ray image through physical data, the synthetic image comprising a clean signal and synthetic noise; S2, acquire a real image of the X-ray image through an inertial confinement fusion (ICF) experiment; S3, construct a synthetic image dataset and a real image dataset; S4, construct an image denoising model, the image denoising model comprising an enhancement module, a denoising module and three discriminators; input the synthetic image dataset and the real image dataset into the image denoising model, and sequentially pass through the enhancement module and the denoising module to obtain respective denoising results; in the process of enhancement and denoising, sequentially pass through the three discriminators for joint adversarial learning, and align the synthetic image and the real image in terms of image, feature and result; S5, design a hybrid loss function according to the denoising result and the alignment result in step S4, guide the image denoising model to bridge the domain gap between the synthetic image and the real image, and improve the denoising quality while maintaining the structural fidelity.

2. The method of claim 1, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. Step S1 specifically comprises: S11, first model the synthetic image as: ； wherein represents a synthetic image; is a clean signal of the synthetic image, K s and K n represent a system blur kernel and a scintillator blur kernel, respectively, P (.) represents a shot noise of a Poisson distribution, represents a pulse noise generated by a neutron interaction, represents a convolution operation; S12, the clean signal of the synthetic image is generated by using the Legendre method, and a 10th order Legendre polynomial is used to expand the clean signal hotspot to model the radial boundary, capture the asymmetric and offset conditions of the clean signal hotspot and the geometric center of the clean signal, and the formula of the radial boundary is: ； wherein, are Legendre polynomials; c i is a tracking shape coefficient, c 0 represents the overall size, c 1 represents the degree of center offset, and the remaining higher order terms represent the degree of asymmetry; is a radial boundary related to the angle . then, through random sampling by tracking the shape coefficient, a group of clean signal hotspots with different degrees of asymmetric measurement is synthesized, and the signal intensity distribution of the pixels in the hotspot range is modeled as an exponential decay starting from the clean signal hotspot, which is modulated by the radial boundary, and the formula of the signal intensity of the pixels in the hotspot range is: ； wherein, s 0 is the peak intensity, S13, after generating a clean signal dataset through step S12, the clean information is enhanced by introducing additional texture diversity using natural images; is the radial distance from the hotspot, is S15, by analyzing the real image obtained by the ICF experiment, synthesize the speckle of the pulse noise, and use a random walk algorithm to copy the irregular shape and intensity decay of the speckle; then, by placing noise clusters along straight paths at random angles, linear stripes of pulse noise are synthesized to simulate the paths of charged particles; then, different shot parameters are randomly sampled from an empirical distribution and embedded in the clean signal to randomly add shot to the clean signal, thereby generating diversified pulse noise; the signal strength at the point. S16, combining step S14 and step S15, the synthetic noise with both shot noise and pulse noise is obtained, and the synthetic noise is added to the clean signal to obtain the synthetic image of the X-ray image. S14, the shot noise is modeled as a Poisson process with variance proportional to the expected signal intensity, the shot noise texture is modeled as a convolution of white noise and a point spread function (PSF), the shot noise texture is modeled by the equation: ； wherein N Step S13 specifically comprises: is a white Gaussian noise map, is a point spread function; the shot noise level is controlled by a scale factor a, which is estimated from image patches with uniform luminosity; meanwhile, the clean signal is the pixel expectation, resulting in a shot noise is formulated as:​ ； first, convert the texture of the natural image into a gray image, then reformat and adjust the natural texture to match the data structure of the clean signal, and normalize the gray value, set the background intensity and peak intensity range to be consistent with the dynamic range observed in the ICF experiment. Step S3 specifically comprises:

3. The method of claim 2, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. ​ ​ 4. The method of claim 1, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. ​ S31, construct a synthetic image dataset composed of three groups of paired images, each group of paired images is an image pair of clean signal and synthetic noise, wherein: the first group only uses a small number of Legendre moments to represent hotspot asymmetry, the second group uses more Legendre moments than the first group, and the peak intensity is higher and the noise configuration is the same; the third group uses natural texture images and uses the same noise adding method as the second group; S32, data augmentation is performed on the real image to expand the sample quantity to obtain a real image dataset.

5. The method of claim 1, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. The enhancement module in step S4 is composed of a layer of Conv convolution, multiple dense blocks and a fusion module fusion, the Conv convolution, the dense blocks are connected through residual jump connection, and each dense block is composed of multiple convolution, instantiation normalization layer and activation function; the enhancement module takes the synthetic noise as input, and extracts high-level features through multiple dense blocks to enhance the synthetic noise.

6. The method of claim 5, wherein the method is based on physical information and data synthesis to fill the gap in the X-ray image denoising domain. The denoising module in step S4 is composed of two weight-shared U-Net-like networks, each network has an encoder, a latent layer and a decoder, and the denoising process is as follows: The enhanced synthetic noise and the synthetic clean signal are added as the enhanced synthetic image, and the real noise and the reconstructed real clean signal are added to the real noisy image; the enhanced synthetic image and the real noisy image are respectively input into a U-Net-like network, and are sequentially input into the convolution layer and the down-sampling layer in the encoder to extract high-level features, and are input into the up-sampling layer and the fusion layer for decoding and denoising. The two U-Net-like networks respectively output the denoising tensor results of the enhanced synthetic image and the real noisy image, and after the denoising tensor results are de-normalized, the same format pictures are saved to obtain the final visual denoising results.

7. The method of claim 1, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. In step S4, the three discriminators are sequentially used for joint adversarial learning in the enhancement and denoising process, and the synthetic image and the real image are aligned in terms of image, feature and result, which is as follows: After enhancement, the first discriminator is used for image adversarial training of the enhanced synthetic noise and the real noise to align the enhanced synthetic noise and the real noise in terms of image; In the denoising process, the second discriminator in the latent layer of the denoising module is used for feature adversarial training of the enhanced synthetic image and the real noisy image; The enhanced synthetic image and the real noisy image are aligned in terms of feature; after the two U-Net networks respectively output the denoising tensor results, the third discriminator is used for result adversarial training to align the denoising tensor results; meanwhile, the loss functions of the discriminators are calculated, and the calculated loss functions are included in the hybrid loss function in step S5.

8. The method of claim 1, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. The hybrid loss function in step S5 is specifically represented as: ； wherein, represents a hybrid loss function; represents an image adversarial loss function set for the first discriminator; represents a content preservation loss set for the enhancement task; represents a task loss set for the denoising task; represents a feature adversarial loss set for the second discriminator; represents an output adversarial loss set for the third discriminator; , , , , are weight coefficients that control the contribution of each loss function.

9. The method of claim 8, wherein the method is based on physical information and data synthesis for X-ray image denoising gap filling. The loss functions in step S5 are specifically as follows: Image adversarial loss The processed synthetic image is closer to the real image by using the image denoising model, and the formula is represented as: ； wherein, denotes the real noise, denotes the enhanced synthesized noise, denotes , an interpolation between is a gradient penalty coefficient; denotes the first discriminator; denotes the L2 norm; is the gradient of the discriminator at the sample ; denotes the mean; content preservation loss The encouragement enhancement module encourages the preservation of high-level semantic results while reducing low-level adaptation, which is expressed in the formula as: ； wherein, represents a synthetic noise; is an L1 norm, represents extracting features from the k-th layer of the pre-trained VGG19 network, is a set of VGG19 network layers, is a corresponding weight; task loss The guidance denoising module recovers visual fidelity and semantic structure, and its formula is expressed as: ； wherein, y s represents a clean signal of the synthesized image, y st represents an enhanced synthesized image x st de-noising result, x st is obtained by , y s adding; is a set of network layers; weights β and γ control the contribution of image loss and perceptual loss, respectively. feature adversarial loss The guided image denoising model adjusts the difference in style attributes between the synthetic image and the real image while preserving the semantic content, which is expressed in a formula as follows: ； wherein, denotes a shared encoder of the denoising module, is a linear interpolation between the latent features of the input enhanced synthetic image and the real noisy image, λ 2 is a gradient penalty coefficient; is a real noisy image; denotes a second discriminator; output adversarial loss To ensure that the denoised synthetic image has visual fidelity close to the denoised real image, the formula is expressed as: ； wherein, denotes a third discriminator, λ 3 is a gradient penalty coefficient; is a denoising result of the real image.

Images