Image denoising method based on multi-scale feature enhancement and local pixel disturbance

Abstract

The invention discloses an image denoising method based on multi-scale feature enhancement and local pixel disturbance. According to the method, features of different scales are extracted by introducing a multi-scale convolution module and cavity convolution, so that the perception capability of complex noise is enhanced. Meanwhile, flat areas are calculated through the local variance of the pseudo clean image, and pixel disturbance is carried out in the areas so as to weaken noise correlation. According to the method, channel attention and space attention mechanisms are combined, and learning of important features is enhanced. In the aspect of loss function design, pixel loss, multi-scale feature loss and local weighting loss are fused. In practical application, a cosine annealing learning rate scheduling and exponential moving average strategy is adopted, and the convergence speed and stability of the model are remarkably improved. According to the method, image details can be effectively reserved, and the method is suitable for denoising tasks of various complex scenes such as natural images and medical images.

Description

Technical Field

[0001] This invention relates to image processing technology, and in particular to an image denoising method based on deep learning, which is applicable to various complex scenarios such as natural images and medical images. Background Technology

[0002] Image denoising is a fundamental task in computer vision, aiming to remove image noise while preserving as much texture and detail as possible in the original image. Currently popular methods include denoising algorithms based on traditional models (such as wavelet transform) and deep learning-based methods. For example, a deep learning-based image denoising method, as disclosed in patent application number 202010748359.2, involves inputting a noisy image and its noise level into an image processing model for processing; the model then performs multi-level scaling and offset feature transformation adjustments on the noisy image denoising process. This denoising method addresses the low flexibility and controllability issues of current image denoising methods by performing multi-level scaling and offset feature adjustment within the image processing model, thus improving the flexibility and controllability of image denoising.

[0003] However, existing deep learning denoising methods suffer from the following problems: difficulty in striking a balance between denoising intensity and detail preservation; significant performance degradation for highly correlated noise; and a lack of importance assessment of local features, potentially leading to the loss of local information. This patent proposes an image denoising method combining multi-scale feature extraction, local pixel perturbation, and attention mechanisms. This method demonstrates significant denoising effectiveness against highly correlated noise, better preserves image texture and edge information, and improves the model's robustness and generalization ability under various noise types, thus possessing practical value. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an image denoising method that can effectively handle complex noise while preserving details. By combining multi-scale feature extraction, local pixel perturbation and attention mechanism, the model's ability to capture detailed features is improved, resulting in better denoising effect.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an image denoising method based on multi-scale feature enhancement and local pixel perturbation, including building an image denoising model, which includes input data preprocessing, multi-scale feature extraction, a local perturbation module, an attention mechanism module, and a loss calculation module.

[0006] The input data preprocessing includes calculating the local variance using a pseudo-clean image and generating a flat region mask;

[0007] Multi-scale feature extraction includes extracting features at different scales through multi-scale convolutional modules;

[0008] The local perturbation module includes noise characteristics for flat areas, which reduces noise correlation through pixel perturbation to effectively separate noise from the image signal;

[0009] The model incorporates channel attention and spatial attention modules to dynamically weight features and obtain a weighted feature map for output.

[0010] The loss calculation module uses a multi-scale loss function that combines pixel loss, feature loss, and local weighted loss to optimize model performance.

[0011] The flat region mask generated by calculating the local variance of the pseudo-clean image is used to locate low-texture regions in the image, providing a basis for local pixel perturbation; Gaussian filtering is used to perform preliminary processing on the original image to generate an approximate pseudo-clean image;

[0012] The local variance of a pseudo-clean image is calculated to reflect the degree of fluctuation of pixel values ​​in the image. A Gaussian filter is used to calculate the local variance within a sliding window.

[0013] The local variance is normalized, and a threshold is set to mark flat and non-flat regions.

[0014] Multi-scale convolutional modules enhance the model's ability to perceive noise and details of different sizes, thereby adapting to complex noise patterns. These modules include:

[0015] (1) Multi-branch convolutional structure:

[0016] A multi-branch convolution module is used, with each branch employing a convolution kernel of a different size to capture features of different receptive field ranges; the convolution results are integrated by concatenation or weighted fusion to form a multi-scale feature representation.

[0017] (2) Enhanced dilated convolution:

[0018] By incorporating dilated convolutions into the convolutional branches, the receptive field is expanded while reducing computational complexity. Combining dilated convolutional branches with different dilation rates further enhances the ability to capture multi-scale features.

[0019] The perturbation method of the local perturbation module includes: selecting the target region using a flat region mask, and performing perturbation on the region covered by the mask within a k×k sub-block range.

[0020] The perturbation method of the local perturbation module also includes: in flat areas, randomly adjusting the pixel arrangement, reordering the pixel blocks by generating random indices to ensure that the perturbed pixel distribution conforms to the noise decorrelation target, while maintaining the integrity of the local structure; adjusting the perturbation intensity according to the local variance value, setting a higher perturbation intensity in flat areas, and keeping the high variance areas unchanged.

[0021] The channel attention module in the attention mechanism module uses global average pooling to calculate the global features of each channel, and generates the attention weight W_channel of each channel through a multilayer perceptron (MLP). The weight values ​​are used to dynamically adjust the importance of the channels.

[0022] The spatial attention module within the attention mechanism calculates the channel average and maximum values ​​of the feature map, fuses them to generate a spatial attention mask, and normalizes the spatial attention map to generate a pixel-level weighted mask. ;

[0023] Channel attention is used for cross-channel feature selection, while spatial attention is used for highlighting key regions. Channel attention and spatial attention modules can be used in series or in parallel.

[0024] A multi-scale loss function is used to calculate the error and optimize the model parameters. The multi-scale loss function is calculated as follows:

[0025] + +

[0026] Pixel loss weights Feature loss weights Local weighted loss weight ; For pixel-level loss, L1 loss is used to measure the difference between the denoised image and the real image at the pixel level;

[0027] in, This represents the total number of pixels in the image. This represents the true value of the i-th pixel in the real image. The predicted value of the i-th pixel in the denoised image;

[0028] For multi-scale feature loss, the feature differences between the denoised image and the real image are calculated on feature maps at different levels. Perceptual loss is used to measure the similarity of high-level semantic features.

[0029] in, This indicates the number of layers for multi-scale features (the number of feature layers extracted by the model). For the real image in the first Representation of the layer feature space, To denoise the image at the first Representation of the layer feature space, It is the square of the L2 norm (Euclidean distance), used to calculate the difference between two eigenvectors.

[0030] For locally weighted loss, a flat region mask is used to calculate the locally weighted loss, assigning higher weights to pixel differences in flat regions to enhance the denoising effect in low-texture areas:

[0031]

[0032] The weight is a local weight, which is related to the local variance of the pseudo-clean image. It is used to reflect the weight of the i-th pixel. The higher the weight value, the more important the pixel region is.

[0033] When training the image denoising model, a cosine annealing learning rate scheduling strategy and EMA optimization are used to reduce possible oscillations during training.

[0034] The training process for the constructed image denoising model includes:

[0035] (1) Training data generation: Input data includes noisy images and target clean image To enhance the generalization ability of the model, random cropping, horizontal flipping, rotation and other enhancement operations are used to normalize the pixel values ​​to [0,1] to ensure the consistency of the model input;

[0036] (2) Model initialization: The weights are initialized using the He initialization method to ensure the stability of the deep network;

[0037] (3) Dynamic learning rate scheduling strategy: The initial learning rate value is lr=0.0004, and the minimum value is lrmin=0.00001; after each T0 round, the learning rate decays and restarts, and the learning rate is gradually adjusted; after each round, the length of the round doubles, and the scheduling formula is as follows:

[0038]

[0039] Where t is the current epoch. This represents the length of the current learning rate period.

[0040] (4) Optimization and model update: Forward propagation: Input noisy image, generate denoised image and intermediate features;

[0041] Loss calculation: The error is calculated using a multi-scale loss function;

[0042] Backpropagation: Use loss.backward() to calculate the gradient, complete the backpropagation, and update the model parameters.

[0043] (5) Model validation and saving: PSNR and SSIM are used to evaluate the denoising effect. The metrics are calculated on the validation set and the results are saved for each epoch. If the current model performs better than the previously saved best model on the validation set, the model is updated and the file is saved.

[0044] The advantages of this invention are: an image denoising method that can effectively handle complex noise while preserving details, by combining multi-scale feature extraction, local pixel perturbation, and attention mechanisms, enhances the model's ability to capture detailed features, achieving superior denoising results. It has the following beneficial effects:

[0045] 1. Significantly improves noise suppression capabilities, with particularly noticeable effects on highly correlated noise;

[0046] 2. Preserve image texture and edge information in complex scenes;

[0047] 3. Improve the model's generalization performance, making it suitable for various scenarios. Attached Figure Description

[0048] The following is a brief explanation of the contents of each of the accompanying drawings and the markings in the drawings:

[0049] Figure 1 This is a diagram illustrating the overall architecture of the noise reduction method of the present invention.

[0050] Figure 2 This is a flowchart of the local disturbance handling process;

[0051] Figure 3 This is a schematic diagram of a multi-scale module;

[0052] Figure 4 This is a schematic diagram of the attention mechanism module;

[0053] Figure 5 This is a comparison chart of the experimental results. Detailed Implementation

[0054] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the description of the preferred embodiments.

[0055] This invention discloses an image denoising method based on multi-scale feature enhancement and local pixel perturbation. This method introduces multi-scale convolutional modules and dilated convolutions to extract features at different scales, thereby enhancing the ability to perceive complex noise. Simultaneously, it calculates flat regions using the local variance of pseudo-clean images and performs pixel perturbation within these regions to reduce noise correlation. This method combines channel attention and spatial attention mechanisms to strengthen the learning of important features. The loss function design integrates pixel loss, multi-scale feature loss, and local weighted loss. In practical applications, cosine annealing learning rate scheduling and exponential moving average strategies are employed to significantly improve the model's convergence speed and stability. This invention effectively preserves image details and is suitable for denoising tasks in various complex scenes, including natural images and medical images.

[0056] This embodiment provides an image denoising method that can effectively handle complex noise while preserving details. By combining multi-scale feature extraction, local pixel perturbation, and an attention mechanism, it enhances the model's ability to capture detailed features, achieving superior denoising results. This image denoising method based on multi-scale feature enhancement and local pixel perturbation includes building an image denoising model, which comprises input data preprocessing, multi-scale feature extraction, a local perturbation module, an attention mechanism module, and a loss calculation module.

[0057] The input data preprocessing includes calculating the local variance using a pseudo-clean image and generating a flat region mask;

[0058] Multi-scale feature extraction includes extracting features at different scales through multi-scale convolutional modules;

[0059] The local perturbation module includes noise characteristics for flat areas, which reduces noise correlation through pixel perturbation to effectively separate noise from the image signal;

[0060] The model incorporates channel attention and spatial attention modules to dynamically weight features and obtain a weighted feature map for output.

[0061] The loss calculation module uses a multi-scale loss function that combines pixel loss, feature loss, and local weighted loss to optimize model performance.

[0062] like Figure 1-4 As shown, an image denoising method based on multi-scale feature enhancement and local pixel perturbation includes the following steps:

[0063] Step 1: Calculate the local variance using a pseudo-clean image to generate a flat region mask. The flat region mask can accurately locate low-texture regions (such as the sky and background) in the image, providing a basis for local pixel perturbation. See the flowchart below. Figure 1 The specific implementation is as follows:

[0064] (1) Generation of pseudo-clean images:

[0065] The original image is pre-processed using Gaussian filtering to generate an approximate pseudo-clean image, where the Gaussian filter kernel size is 13×13.

[0066] (2) Calculation of local variance:

[0067] The local variance of a pseudo-clean image is calculated to reflect the degree of fluctuation in pixel values. A Gaussian filter is used to calculate the local variance within a sliding window (15×15 window).

[0068] (3) Generate a flat region mask:

[0069] The local variance is normalized and a threshold is set (0.5 in this invention) to mark flat regions (e.g., regions with variance values ​​below the set threshold) and non-flat regions.

[0070] Step 2: Extract features at different scales using a multi-scale convolution module. The multi-scale convolution module is designed to enhance the model's ability to perceive noise and details of different sizes, thereby adapting to complex noise patterns. This design enables the model of this invention to extract global and local features in the early stages, effectively enhancing its sensitivity to complex noise. See the flowchart below. Figure 2 .

[0071] The specific implementation is as follows:

[0072] (1) Multi-branch convolutional structure:

[0073] A multi-branch convolutional module is used, with each branch employing a different kernel size (3×3, 5×5, 7×7) to capture features across different receptive field ranges. The convolutional results are then integrated through concatenation or weighted fusion to form a multi-scale feature representation.

[0074] (2) Enhanced dilated convolution:

[0075] Adding dilated convolutions to the convolutional branches expands the receptive field while reducing computational complexity. Combining dilated convolutional branches with different dilation rates further enhances the ability to capture multi-scale features.

[0076] Step 3: For the noise characteristics of flat regions, pixel perturbation is used to reduce noise correlation, which helps subsequent models more effectively separate noise from the image signal. The steps are as follows:

[0077] (1) Application of perturbation mask

[0078] The target region is selected using a flat region mask. Within a k×k sub-block (4×4 in this invention), perturbation is performed on the region covered by the mask.

[0079] (2) Pixel scrambling strategy:

[0080] Within flat regions, the pixel arrangement is randomly adjusted, for example, by reordering pixel blocks using random indices. This ensures that the perturbed pixel distribution conforms to the noise decorrelation objective while maintaining the integrity of the local structure.

[0081] (3) Adaptive perturbation strength:

[0082] The perturbation intensity is adjusted based on the local variance value. A higher perturbation intensity is set for flat areas (high mask value), while the high variance areas (low mask value) remain unchanged.

[0083] Step 4: Introduce channel attention and spatial attention modules into the model to dynamically weight features. This can improve the model's attention to key regions and important details, thus enhancing denoising performance. To improve the model's ability to focus on key regions and important features, the following attention module is designed:

[0084] (1) Channel attention module:

[0085] Global average pooling is used to compute the global features for each channel. Attention weights for each channel are generated using a multilayer perceptron (MLP). The weight values ​​are used to dynamically adjust the importance of channels.

[0086] (2) Spatial attention module:

[0087] Calculate the channel average and maximum values ​​of the feature map and fuse them to generate a spatial attention mask. Normalize the spatial attention map to generate a pixel-level weighted mask. .

[0088] (3) Module integration:

[0089] Channel attention is used for cross-channel feature selection, while spatial attention is used to enhance key regions. Channel and spatial attention modules can be used in series or parallel. The attention results are used to reweight the feature maps, strengthening the model's ability to learn important features.

[0090] Step 5: Optimize model performance using a multi-scale loss function that combines pixel loss, feature loss, and local weighted loss. The multi-scale loss function is key to improving the model's denoising performance, and its specific components are as follows:

[0091] (1) Pixel-level loss: L1 loss is used to measure the difference between the denoised image and the real image at the pixel level.

[0092]

[0093] in, This represents the total number of pixels in the image. This represents the true value of the i-th pixel in the real image. The predicted value of the i-th pixel in the denoised image.

[0094] (2) Multi-scale feature loss: Calculate the feature differences between the denoised image and the real image on feature maps at different levels. Use perceptual loss to measure the similarity of high-level semantic features.

[0095]

[0096] in, This indicates the number of layers for multi-scale features (the number of feature layers extracted by the model). For the real image in the first Representation of the layer feature space, To denoise the image at the first Representation of the layer feature space. It is the square of the L2 norm (Euclidean distance), used to calculate the difference between two eigenvectors.

[0097] (3) Local weighted loss: The local weighted loss is calculated using a flat region mask, and higher weights are given to pixel differences in flat regions to enhance the denoising effect in low-texture regions.

[0098]

[0099] The weights are local weights, related to the local variance of the pseudo-clean image, and are used to reflect the weight of the i-th pixel. A higher weight value indicates that the pixel region is more important. The meanings of the remaining symbols are the same as in the pixel-level loss formula.

[0100] (4) Overall loss:

[0101] The weights of each component of the multi-scale loss can be adjusted according to task requirements (e.g., pixel loss weights). Feature loss weights Local weighted loss weight ), weighted combination of multiple losses:

[0102] + +

[0103] This loss function can balance denoising intensity and detail preservation, effectively optimizing model performance.

[0104] Step 6: Training strategy: Use cosine annealing learning rate scheduling strategy and EMA to optimize model parameters to reduce possible oscillations during training.

[0105] ① Training data generation: Input data includes noisy images and target clean image To enhance the model's generalization ability, augmentation operations such as random cropping, horizontal flipping, and rotation are used. Pixel values ​​are normalized to [0,1] to ensure consistency of model input.

[0106] ② Model initialization: The weights are initialized using the He initialization method to ensure the stability of the deep network. The AdamW optimizer is used with a learning rate of lr=0.0004.

[0107] ③ Dynamic learning rate scheduling strategy: The initial learning rate is lr = 0.0004, and the minimum learning rate is lrmin = 0.00001. After each T0 round, the learning rate decays and restarts, gradually adjusting the learning rate. During scheduling, the period length doubles after each completed period. This further stabilizes the later stages of training. The scheduling formula is as follows:

[0108]

[0109] Where t is the current epoch. This represents the length of the current learning rate period.

[0110] ④ Optimization and Model Update: Forward Propagation: Input a noisy image and generate a denoised image and intermediate features.

[0111] Loss calculation: The error is calculated using a multi-scale loss function.

[0112] Backpropagation: Use loss.backward() to calculate the gradient.

[0113] ⑤ Model Validation and Saving: The denoising effect is evaluated using PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity). The metrics are calculated and saved on the validation set for each epoch. If the current model outperforms the previously saved best model on the validation set, the saved file is updated.

[0114] like Figure 5 As shown, the denoising performance of the method of this invention is compared with other existing single-image-based denoising methods on the SIDD validation and FMDD datasets, especially its advantages in texture preservation and noise suppression. Figure 5The paper presents three existing denoising methods, the denoising method of this invention, and a comparison of image texture and noise corresponding to the original image. The leftmost image is the original image, the three middle images are the three existing denoising methods: DIP[1], Self2Self[2], and Blind2Unblind[3], and the rightmost image is the denoising method of this invention. The three existing denoising methods are the denoising algorithms disclosed in the paper, and their specific sources are as follows:

[0115] DIP[1] denoising method: Andrea Vedaldi Ulyanov, Dmitry and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018.

[0116] Self2Self[2] denoising method: Yuhui Quan, Mingqin Chen, Tongyao Pang, and HuiJi.Self2self with dropout: Learning self-supervised denoising from singleimage. In Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pages 1890–1898, 2020.

[0117] Blind2Unblind[3] denoising method: Wang Z, Liu J, Li G, et al. Blind2Unblind: Self-Supervised Image Denoising with Visible Blind Spots[J]. 2022.DOI:10.48550 / arXiv.2203.06967.

[0118] pass Figure 5 The comparison shows that the denoising method in this scheme can remove noise and preserve image texture and edge information better than the three existing denoising methods.

[0119] Obviously, the specific implementation of this invention is not limited to the above-described methods. Any non-substantial improvements made using the inventive concept and technical solution of this invention are within the protection scope of this invention.

Claims

1. An image denoising method based on multi-scale feature enhancement and local pixel perturbation, characterized in that: This includes building an image denoising model, which includes input data preprocessing, multi-scale feature extraction, local perturbation module, attention mechanism module, and loss calculation module. The input data preprocessing includes calculating the local variance using a pseudo-clean image and generating a flat region mask; Multi-scale feature extraction includes extracting features at different scales through multi-scale convolutional modules; The local perturbation module includes noise characteristics for flat areas, which reduces noise correlation through pixel perturbation to effectively separate noise from the image signal; The model incorporates channel attention and spatial attention modules to dynamically weight features and obtain a weighted feature map for output. The loss calculation module uses a multi-scale loss function that combines pixel loss, feature loss, and local weighted loss to optimize model performance.

2. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: The flat region mask generated by calculating the local variance of the pseudo-clean image is used to locate low-texture regions in the image, providing a basis for local pixel perturbation; Gaussian filtering is used to perform preliminary processing on the original image to generate an approximate pseudo-clean image; The local variance of a pseudo-clean image is calculated to reflect the degree of fluctuation of pixel values ​​in the image. A Gaussian filter is used to calculate the local variance within a sliding window. The local variance is normalized, and a threshold is set to mark flat and non-flat regions.

3. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: Multi-scale convolutional modules enhance the model's ability to perceive noise and details of different sizes, thereby adapting to complex noise patterns. These modules include: (1) Multi-branch convolutional structure: A multi-branch convolution module is used, with each branch employing a convolution kernel of a different size to capture features of different receptive field ranges; the convolution results are integrated by concatenation or weighted fusion to form a multi-scale feature representation. (2) Enhanced dilated convolution: By incorporating dilated convolutions into the convolutional branches, the receptive field is expanded while reducing computational complexity. Combining dilated convolutional branches with different dilation rates further enhances the ability to capture multi-scale features.

4. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: The perturbation method of the local perturbation module includes: selecting the target region using a flat region mask, and performing perturbation on the region covered by the mask within a k×k sub-block range.

5. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 4, characterized in that: The perturbation method of the local perturbation module also includes: in flat areas, randomly adjusting the pixel arrangement, reordering the pixel blocks by generating random indices to ensure that the perturbed pixel distribution conforms to the noise decorrelation target, while maintaining the integrity of the local structure; adjusting the perturbation intensity according to the local variance value, setting a higher perturbation intensity in flat areas, and keeping the high variance areas unchanged.

6. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: The channel attention module in the attention mechanism module uses global average pooling to calculate the global features of each channel, and generates the attention weight W_channel of each channel through a multilayer perceptron (MLP). The weight values ​​are used to dynamically adjust the importance of the channels. The spatial attention module within the attention mechanism calculates the channel average and maximum values ​​of the feature map, fuses them to generate a spatial attention mask, and normalizes the spatial attention map to generate a pixel-level weighted mask. ; Channel attention is used for cross-channel feature selection, while spatial attention is used for highlighting key regions. Channel attention and spatial attention modules can be used in series or in parallel.

7. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: A multi-scale loss function is used to calculate the error and optimize the model parameters. The multi-scale loss function is calculated as follows: + + ； Pixel loss weights Feature loss weights Local weighted loss weight ; For pixel-level loss, L1 loss is used to measure the difference between the denoised image and the real image at the pixel level; ; in, This represents the total number of pixels in the image. This represents the true value of the i-th pixel in the real image. The predicted value of the i-th pixel in the denoised image; For multi-scale feature loss, the feature differences between the denoised image and the real image are calculated on feature maps at different levels. Perceptual loss is used to measure the similarity of high-level semantic features. ; in, This indicates the number of layers for multi-scale features (the number of feature layers extracted by the model). For the real image in the first Representation of the layer feature space, To denoise the image at the first Representation of the layer feature space, The square of the L2 norm (Euclidean distance) is used to calculate the difference between two eigenvectors; For locally weighted loss, a flat region mask is used to calculate the locally weighted loss, assigning higher weights to pixel differences in flat regions to enhance the denoising effect in low-texture areas: ； The weight is a local weight, which is related to the local variance of the pseudo-clean image. It is used to reflect the weight of the i-th pixel. The higher the weight value, the more important the pixel region is.

8. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 1, characterized in that: When training the image denoising model, a cosine annealing learning rate scheduling strategy and EMA optimization are used to reduce possible oscillations during training.

9. The image denoising method based on multi-scale feature enhancement and local pixel perturbation as described in claim 8, characterized in that: The training process for the constructed image denoising model includes: (1) Training data generation: Input data includes noisy images and target clean image To enhance the generalization ability of the model, random cropping, horizontal flipping, rotation and other enhancement operations are used to normalize the pixel values ​​to [0,1] to ensure the consistency of the model input; (2) Model initialization: The weights are initialized using the He initialization method to ensure the stability of the deep network; (3) Dynamic learning rate scheduling strategy: The initial learning rate value is lr=0.0004, and the minimum value is lrmin=0.00001; after each T0 round, the learning rate decays and restarts, and the learning rate is gradually adjusted; after each round, the length of the round doubles, and the scheduling formula is as follows: ； Where t is the current epoch. The length of the current learning rate period; (4) Optimization and model update: Forward propagation: Input noisy image, generate denoised image and intermediate features; Loss calculation: The error is calculated using a multi-scale loss function; Backpropagation: Use loss.backward() to calculate the gradient, complete the backpropagation, and update the model parameters; (5) Model validation and saving: PSNR and SSIM are used to evaluate the denoising effect. The metrics are calculated on the validation set and the results are saved for each epoch. If the current model performs better than the previously saved best model on the validation set, the model is updated and the file is saved.

Images