Single image snow removal method and system based on frequency domain enhanced multi-scale feature extraction, and medium

Abstract

The invention relates to a single image snow removal method and system based on frequency domain enhancement multi-scale feature extraction and a medium, and the method comprises the following steps: obtaining an input image degraded by snow, inputting the input image into a pre-trained frequency domain enhancement-based snow removal model for processing, outputting a high-fidelity snow removal image, and carrying out the processing of the high-fidelity snow removal image; the frequency domain enhancement-based snow removal model comprises a resampling module and a plurality of channels, wherein each channel comprises an overlapped image block embedding module, a multi-scale feature extraction module, a Transform module and a snowless image reconstruction module which are connected in sequence. Compared with the prior art, attention can be dynamically adjusted so as to effectively cope with snowflake noise of different scales, and therefore the snow removing effect and the like are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image restoration technology, and in particular to a single-image desnowing method based on multi-scale feature extraction with frequency domain enhancement. Background Technology

[0002] In the field of computer vision, snowfall presents a significant challenge. Snow not only blurs important details in images but also introduces high-frequency noise, leading to a severe loss of visual information. This quality degradation poses a major challenge to computer algorithms that rely on high-quality visual input. For example, in autonomous driving scenarios, vehicle perception systems depend on clear images to accurately identify roads, pedestrians, and obstacles. Therefore, effectively removing snowflakes and restoring a clear background from a single input image has always been an important research direction in image restoration.

[0003] Early image desnow removal methods primarily relied on manually defined physical models and prior assumptions. Representative methods typically used sparse coding or low-rank decomposition to separate the foreground snow layer from the background image, based on characteristics such as sparsity, high brightness, and random distribution of snow particles. Some studies employed a hierarchical modeling approach, dividing the image into background and snow layers and estimating them separately. However, these methods often rested on overly idealistic assumptions, such as the assumption that snow particles and background were completely separable in the frequency domain, or that the distribution of snow particles in the image was fixed. In real-world scenes, snow has diverse and complex forms, ranging from fine particles to blurred snow fog, and is often affected by factors such as lighting, wind speed, and camera exposure. Due to the limited expressive power of traditional models, it is difficult to balance desnow removal accuracy and robustness, thus often resulting in poor performance in complex natural scenes.

[0004] With the rapid development of deep learning, convolutional neural networks (CNNs) have become a core technology for image desnowing. Researchers construct end-to-end network models to directly learn the mapping relationship from snow-laden images to clear images. These methods no longer rely on manual feature design but utilize a large amount of labeled data for training, thereby obtaining stronger feature representation and modeling capabilities. Compared with traditional methods, deep learning-based desnowing methods have achieved significant performance improvements on synthetic snow and some real snow datasets. However, standard convolutional structures have a fixed receptive field, mainly focusing on local features, and have limited ability to model large-scale snow fog or snow particles spreading across regions. Therefore, convolutional neural networks are prone to problems such as blurred background details, texture loss, or residual snow marks during the restoration process.

[0005] In recent years, the Transformer has demonstrated strong potential in image restoration. Its core self-attention mechanism can effectively capture long-range dependencies, compensating for the insufficient receptive field of convolution. In single-image snow removal tasks, the Transformer structure exhibits superior characteristics, simultaneously focusing on global structure and local details, thus more accurately distinguishing snow particles from the background. However, the computational complexity of Transformer attention increases quadratically with image resolution, leading to a sharp increase in computational and storage costs in high-resolution image processing, thereby limiting its widespread adoption in practical applications. Overall, whether based on convolution-based local modeling or attention-based global modeling, existing methods still have significant limitations when dealing with complex snow scenes.

[0006] In summary, although existing methods have made some progress in performance, they still face two key challenges when processing complex snow scene images: (1) How to effectively deal with multi-scale degradation caused by snow is the foundation for processing complex snow scenes, especially when removing large-scale snowflakes; (2) Global contextual information is crucial for accurately reconstructing a clear scene, but previous methods either ignored this or could only generate blurry results. The fundamental reason is that the noise of large snowflakes and small snowflakes has a significant difference in frequency domain distribution. The noise of large snowflakes is mainly distributed in the high-frequency domain, while the noise of small snowflakes is mainly concentrated in the low-frequency domain.

[0007] Furthermore, degraded snowscape images often contain various snow effects (such as snowflakes, snow tracks, and masking effects), and different types of degradation exhibit unique scale characteristics. Relying solely on a single desnowing strategy or a single attention mechanism makes it difficult to fully capture these complex degradation patterns and global interaction information, thus affecting the overall desnowing effect. Summary of the Invention

[0008] The purpose of this invention is to provide a single-image desnowing method based on frequency domain enhancement and multi-scale feature extraction to improve image desnowing performance.

[0009] The objective of this invention can be achieved through the following technical solutions: A single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction includes the following steps: A snow-degraded input image is acquired and processed by a pre-trained frequency-domain enhancement-based snow removal model to output a high-fidelity snow-removed image. The frequency-domain enhancement-based snow removal model includes a resampling module and multiple channels. Each channel includes a sequentially connected overlapping image patch embedding module, a multi-scale feature extraction module, a Transformer module, and a snow-free image reconstruction module. The steps for outputting the high-fidelity snow-removed image include: The input image is resampled using a resampling module to obtain input images of different sizes; Input the images at each scale into the corresponding channels and perform the following steps: Based on the input images at each scale, the corresponding initial features are obtained through the overlapping image patch embedding module. Based on the corresponding initial features, the coarse grid features and fine network features are extracted first using the multi-scale feature extraction module, and then the channel weighting of Fourier transform and dual attention mechanism is combined to obtain the fused features. The fused features are first input into the Transformer module, and then spatial modulation and feature aggregation are performed using the snowless image reconstruction module to finally output a high-fidelity desnowed image.

[0010] Furthermore, the coarse mesh feature and the fine mesh feature are respectively represented as follows: , , In the formula, It is a coarse grid feature. , These are the radius and the angle, respectively. For low frequency channels, The initial features are the input. For fine network characteristics, Indicates a high-frequency channel.

[0011] Furthermore, the step of obtaining the fusion feature includes: The coarse mesh features and the fine network features are respectively generated into feature maps by the encoder. Then, the coarse mesh features and the fine network features are fused into the feature maps using bilinear interpolation to obtain an initial fused feature map, represented as follows: , In the formula, The fused feature map Indicates the interpolation weights; Based on the initial fused feature map, it is decomposed into multi-scale frequency sub-bands using the Fourier transform method; A dual attention mechanism is used to adjust the weights of frequency sub-bands at each scale, and the frequency sub-bands at each scale are weighted and fused to obtain the final fused features.

[0012] Furthermore, the Fourier transform method decomposes the initial fused feature map into multi-scale frequency sub-bands using a wavelet filter, wherein the operational expression for obtaining the multi-scale frequency sub-bands is: , In the formula, These are frequency sub-bands of different scales. It is a low-pass filter. , , All are high-pass filters. This represents the convolution operation. This represents the initial fused feature map of the input.

[0013] Furthermore, the steps for the final fusion feature include: Using the multi-scale frequency sub-bands as input, corresponding frequency domain attention weights and spatial attention weights are generated using frequency domain attention mechanism and spatial attention mechanism, respectively. Based on the frequency domain attention weights and spatial attention weights, a hybrid attention weight vector is generated, represented as follows: , In the formula, This represents the final generated hybrid attention weight vector. Represents the frequency domain attention weights. Spatial attention weights; The multi-scale frequency sub-bands are weighted and fused using the hybrid attention weight vector to obtain the final fused feature, which is expressed as: , In the formula, The fused features are obtained by weighting multi-scale frequency subbands. These are frequency sub-bands of different scales. and These represent the weighting coefficients and their corresponding frequency sub-bands at different scales. These are low-frequency components and high-frequency components, respectively.

[0014] Furthermore, the operational expressions for the frequency domain attention weights and spatial attention weights are as follows: , , In the formula, For frequency domain attention weights, Indicates input features, This represents the Sigmoid activation function. This represents the convolution operation used to generate attention weights. This represents the Fourier magnitude of the input feature. Spatial attention weights, and These represent the average pooling and max pooling operations performed on the input features, respectively.

[0015] Furthermore, the step of acquiring the high-fidelity desnow image includes: Based on the fused features, encoding and decoding are first performed using the Transformer module to establish global dependencies between features and achieve information recombination. Then, a multilayer perceptron is used as the reconstruction function for image reconstruction, and its reconstruction operation expression is as follows: , In the formula, For images after the snow has been removed, For reconstruction function, and These represent the weighting coefficients and their corresponding frequency sub-bands at different scales. These are low-frequency components and high-frequency components, respectively. Then, using the snowless image reconstruction module, spatial modulation and feature aggregation are performed at different scales to obtain features at different scales, represented as follows: , In the formula, These are features at different scales. This is a channel partitioning operation used to decompose the input desnowed feature map into multiple sub-feature maps to support subsequent multi-scale feature processing. The features at different scales are concatenated and then convolved to obtain intermediate features, which are represented as follows: , In the formula, Indicates intermediate features, For convolution operations, For connection operation; Based on the intermediate features and features at different scales, the modulation features are obtained, represented as follows: , In the formula, Indicates modulation characteristics, It is a nonlinear mapping function. For features at different scales, representing ; For the modulation features Decode the image to obtain a high-fidelity desnowed image.

[0016] Furthermore, the loss function used in the training process of the frequency domain-enhanced snow removal model is a composite multi-scale loss function, expressed as: , In the formula, For composite multi-scale loss functions, , As weight, For MSE loss, For Charbonnier loss, This is the edge loss.

[0017] The present invention also provides a single-image desnowing system based on frequency domain enhanced multi-scale feature extraction, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the single-image desnowing method based on frequency domain enhanced multi-scale feature extraction as described above.

[0018] The present invention also provides a storage medium storing a program thereon, which, when executed, implements the single-image desnowing method based on frequency domain enhancement and multi-scale feature extraction as described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention introduces a multi-scale feature extraction module for each channel and integrates a dual attention mechanism therein. By extracting features hierarchically and dynamically weighting them at different scales, it achieves efficient removal of multi-scale snow noise. At the same time, Fourier transform is introduced during the feature extraction process to enhance the frequency domain feature modeling capability, thereby capturing the degradation information in the snow scene image more fully. Finally, combined with the snowless image reconstruction module, the snow removal result is gradually refined and the image details and brightness levels are restored. This invention significantly improves the image clarity and overall visual quality by taking into account the fusion of spatial and frequency domain information in the structural design and reconstruction strategy.

[0020] (2) The present invention combines frequency domain and spatial attention in a multi-scale feature extraction module. This innovative module can dynamically adjust attention to effectively deal with snow noise at different scales, thereby significantly improving the snow removal effect.

[0021] (3) This invention proposes a snowless image reconstruction method that can independently calculate multi-scale features and perform dynamic aggregation, thereby achieving better results in clear image reconstruction.

[0022] (4) This invention overcomes the shortcomings of existing methods in terms of insufficient multi-scale degradation modeling, limited global context capture capability, and insufficient utilization of frequency domain features, thereby improving the efficiency and accuracy of complex snowy day image restoration.

[0023] (5) The present invention designs a composite multi-scale loss function, which simultaneously constrains pixel accuracy, structural consistency and edge detail fidelity during training, ensuring that the model can still obtain high-quality snowless images under complex snow scenes.

[0024] (6) This invention has not only achieved excellent experimental results on multiple public datasets, but also has good generalization ability and practical application value, and is suitable for deployment in autonomous driving, security monitoring and other intelligent vision systems that need to cope with complex weather conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram illustrating the overall framework of the frequency-domain-enhanced snow removal model of the present invention. Figure 3 This is a structural diagram of the snowless image reconstruction module of the present invention; Figure 4 This is a comparison of the snow removal effect of the present invention on commonly used synthetic snowy day datasets; Figure 5 This is a comparison chart of the snow removal effect of the present invention on a commonly used real-world snowy day dataset. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] Example 1 This embodiment provides a single-image desnowing method based on frequency domain enhancement and multi-scale feature extraction, aiming to solve the problems of insufficient multi-scale degradation modeling and limited global context capture ability of traditional convolutional neural networks. This method utilizes, for example... Figure 2 The frequency-domain enhanced snow removal model shown achieves efficient removal of multi-scale snow noise by hierarchically extracting and dynamically weighting features at different scales. During feature extraction, Fourier transform is introduced to enhance frequency domain feature modeling capabilities, while simultaneously expanding the convolutional receptive field to more fully capture degradation information in snow scene images. Finally, combined with a snow-free image reconstruction module, the snow removal results are gradually refined, image details and brightness levels are restored, and a clearer, more realistic snow-free image is generated. This invention integrates spatial and frequency domain information in its structural design and reconstruction strategy, significantly improving image clarity and overall visual quality. Specifically, as... Figure 1 As shown, the method includes the following steps: S1. Obtain the snow-degraded input image and perform initial feature extraction.

[0028] like Figure 2The snow removal model based on frequency domain enhancement shown includes a resampling module and multiple channels. Each channel includes an overlapping image patch embedding module, a multi-scale feature extraction module, a Transformer module, and a snow-free image reconstruction module connected in sequence.

[0029] This step serves as the initial feature extraction stage. To simultaneously capture degradation patterns at different spatial frequencies, this invention first constructs a three-channel architecture. For input images degraded by snow... Resampling was performed to obtain images at half-scale and quarter-scale. and Images at each scale are embedded by an overlapping image embedding module to obtain corresponding initial features.

[0030] S2, deep feature extraction and cross-domain interaction fusion.

[0031] This step utilizes multi-scale feature extraction modules for each channel to perform deep feature extraction and cross-domain interactive fusion. First, feature extraction is performed using pixel coordinates to obtain coarse grid features and fine network features: , , in, It is a coarse grid feature. For fine network characteristics, and These represent the radius and the angle, respectively. Indicates the low-frequency channel, only retains Less than the specified threshold The proportion; on the contrary, Indicates high-frequency channels, only reserved Greater than the specified threshold The amount.

[0032] Coarse grids are primarily used to extract low-resolution, blurry regions, while fine grids focus on high-resolution, sharper areas, thus enabling more effective processing of different regions in the image. Both inputs are fed separately into an encoder to generate feature maps. Subsequently, bilinear interpolation is used to fuse the coarse and fine grid features in the feature maps, as follows: , in, This is the initial fused feature map. This indicates the interpolation weights. This interpolation method preserves image detail while utilizing low-frequency components from the coarse grid to support the overall structure.

[0033] To further optimize image feature representation, this invention designs a dual attention mechanism—a frequency domain and spatial attention module—to adaptively weight the extracted multi-scale features. First, the frequency domain attention mechanism dynamically adjusts the importance of different frequency components (high-frequency and low-frequency signals), making the network focus more on task-relevant key frequency information. Its output calculation formula is as follows: , in, Represents the frequency domain attention weights. Indicates input features, This represents the convolution operation used to generate attention weights. The Fourier amplitude spectrum representing the input features. This represents the Sigmoid activation function.

[0034] Subsequently, a spatial attention mechanism is used to enhance the representation of key regions in the feature map, making the features of important regions more prominent. Its calculation method is as follows: , in, Spatial attention weights, This represents the Sigmoid activation function. and These represent the average pooling and max pooling operations performed on the input features, respectively. This dual attention mechanism enables the network to extract and process multi-scale features more accurately, thereby effectively improving image desnowing performance.

[0035] Furthermore, this embodiment combines Fourier transform with frequency domain and spatial attention modules to more effectively capture multi-scale features. The Fourier transform utilizes discrete wavelet transform to decompose the input feature map into multiple sub-bands, which contain high-frequency and low-frequency components, including... , , , .in, The remaining portion acts as a low-pass filter, while the rest acts as a high-pass filter to extract different frequency sub-bands from the input feature map, enabling the network to effectively capture multi-scale frequency domain information. Then, a wavelet filter decomposes the input image into multi-scale frequency sub-bands, thereby simultaneously capturing both low-frequency and high-frequency components. , in, These are frequency sub-bands of different scales. This represents the initial fused feature map of the input. For each input channel, the convolution operation generates a corresponding output.

[0036] Subsequently, these four frequency sub-bands at different scales are input into the frequency domain and spatial attention mechanism module, where spatial attention weights are generated through global pooling. This weight vector adaptively adjusts the weights of sub-band features across channels, and is calculated as follows: , in, This represents the final generated hybrid attention weight vector. This represents the frequency domain attention weights. The fused features are obtained by weighting frequency sub-bands at different scales. The calculation formula is as follows: , By combining the multi-scale feature decomposition of Fourier transform with the channel weighting mechanism of frequency domain and spatial attention mechanisms, more refined removal of snow noise can be achieved. , and These represent the weighting coefficients and their corresponding frequency sub-bands, respectively.

[0037] S3, Snowless Image Reconstruction Subsequently, regarding fusion characteristics First, the multi-scale fused features are encoded and decoded using the Transformer module to establish global dependencies between features and reconstruct information. Then, a multilayer perceptron is used as the reconstruction function. Reconstructed image: .

[0038] In the formula, This is an image after the snow has been removed.

[0039] Again Figure 3 The snowless image reconstruction module shown performs spatial modulation and feature aggregation at different scales, enhancing the representation of both local and non-local information in the image and obtaining features at different scales: , Subsequently, this embodiment first stitches together features at different scales, and then processes the stitched result... Convolution is used to complete the final processing: , , in, Indicates intermediate features, This represents the modulated features, an intermediate result of the control section, used to dynamically adjust the input features. ( The weight of ) This is a non-linear mapping function; in this embodiment, the Sigmoid activation function is used. Finally, The decoder generates a clear, snow-free, high-fidelity desnow image.

[0040] In this embodiment, the frequency domain-enhanced desnowing model adopts an end-to-end composite multi-scale loss function during the training phase. The model is jointly optimized by the constraints of MSE loss, Charbonnier loss and edge loss, thereby comprehensively improving the desnowing effect and generalization performance in terms of pixel accuracy, structural consistency and edge detail fidelity.

[0041] The overall loss function is a weighted sum of the three factors: .

[0042] This invention effectively removes snow noise at different scales by constructing a frequency domain and spatial attention mechanism. It utilizes the synergistic effect of a multi-scale feature extraction module and a snow-free image reconstruction module to achieve dynamic aggregation of multi-scale features and high-fidelity image reconstruction. Simultaneously, it combines Fourier transform to enhance global receptive field modeling capabilities and improve computational efficiency, visualizing the snow removal effect as shown below. Figure 4 , Figure 5 As shown. The embodiments of the present invention adopt a lightweight computational structure composed of convolution and state space, which is easy to deploy on embedded platforms and can run efficiently on edge devices and vehicle platforms, thereby meeting the real-time and stability requirements of applications such as autonomous driving, security monitoring and mobile vision.

[0043] In summary, this invention employs a multi-scale feature extraction and frequency-spatial fusion structure. By constructing a multi-scale feature extraction module, it achieves hierarchical modeling and processing of snow noise at different scales. Spatial and frequency domain feature extraction pathways are established at each scale. Simultaneously, a visual Transformer module is introduced to encode and decode the multi-scale fused features, utilizing a self-attention mechanism to model the global dependencies between features, thereby improving the completeness of feature representation. Finally, a snow-free image reconstruction module is used to progressively refine the snow removal results and restore image details and brightness levels. To further improve the realism and clarity of image reconstruction, this invention designs a comprehensive loss function that simultaneously constrains pixel accuracy, structural consistency, and edge detail fidelity during training, ensuring that the model can still obtain high-quality snow-free images even in complex snow scenes.

[0044] Example 2 This embodiment provides a single-image snow removal system based on frequency domain enhancement and multi-scale feature extraction, including a memory, a processor, and a program stored in the memory. The feature is that when the processor executes the program, it implements the single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction as described in Embodiment 1 above.

[0045] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A single-image snow removal method based on multi-scale feature extraction with frequency domain enhancement, characterized in that, Includes the following steps: A snow-degraded input image is acquired and processed by a pre-trained frequency-domain enhancement-based snow removal model to output a high-fidelity snow-removed image. The frequency-domain enhancement-based snow removal model includes a resampling module and multiple channels. Each channel includes a sequentially connected overlapping image patch embedding module, a multi-scale feature extraction module, a Transformer module, and a snow-free image reconstruction module. The steps for outputting the high-fidelity snow-removed image include: The input image is resampled using a resampling module to obtain input images of different sizes; Input the images at each scale into the corresponding channels and perform the following steps: Based on the input images at each scale, the corresponding initial features are obtained through the overlapping image patch embedding module. Based on the corresponding initial features, the coarse grid features and fine network features are extracted first using the multi-scale feature extraction module, and then the channel weighting of Fourier transform and dual attention mechanism is combined to obtain the fused features. The fused features are first input into the Transformer module, and then spatial modulation and feature aggregation are performed using the snowless image reconstruction module to finally output a high-fidelity desnowed image.

2. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 1, characterized in that, The coarse mesh feature and the fine mesh feature are respectively represented as follows: ， ， In the formula, It is a coarse grid feature. , These are the radius and the angle, respectively. For low frequency channels, The initial features are the input. For fine network characteristics, Indicates a high-frequency channel.

3. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 1, characterized in that, The steps for obtaining the fusion features include: The coarse mesh features and fine network features are respectively generated into feature maps by an encoder. Then, the coarse mesh features and fine network features are fused into the feature maps using bilinear interpolation to obtain an initial fused feature map, represented as follows: ， In the formula, This is the initial fused feature map. Indicates the interpolation weights; Based on the initial fused feature map, it is decomposed into multi-scale frequency sub-bands using the Fourier transform method; A dual attention mechanism is used to adjust the weights of frequency sub-bands at each scale, and the frequency sub-bands at each scale are weighted and fused to obtain the final fused features.

4. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 3, characterized in that, The Fourier transform method decomposes the initial fused feature map into multi-scale frequency sub-bands using a wavelet filter, wherein the operational expression for obtaining the multi-scale frequency sub-bands is: ， In the formula, These are frequency sub-bands of different scales. It is a low-pass filter. , , All are high-pass filters. This represents the convolution operation. This represents the initial fused feature map of the input.

5. A single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 3, characterized in that, The steps for the final fusion feature include: Using the multi-scale frequency sub-bands as input, corresponding frequency domain attention weights and spatial attention weights are generated using frequency domain attention mechanism and spatial attention mechanism, respectively. Based on the frequency domain attention weights and spatial attention weights, a hybrid attention weight vector is generated, represented as follows: ， In the formula, This represents the final generated hybrid attention weight vector. Represents the frequency domain attention weights. Spatial attention weights; The multi-scale frequency sub-bands are weighted and fused using the hybrid attention weight vector to obtain the final fused feature, which is expressed as: ， In the formula, The fused features are obtained by weighting multi-scale frequency subbands. These are frequency sub-bands of different scales. and These represent the weighting coefficients and their corresponding frequency sub-bands at different scales. These are low-frequency components and high-frequency components, respectively.

6. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 5, characterized in that, The computational expressions for the frequency domain attention weight and the spatial attention weight are as follows: ， ， In the formula, For frequency domain attention weights, Indicates input features, This represents the Sigmoid activation function. This represents the convolution operation used to generate attention weights. This represents the Fourier magnitude of the input feature. Spatial attention weights, and These represent the average pooling and max pooling operations performed on the input features, respectively.

7. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 1, characterized in that, The steps for acquiring the high-fidelity desnow image include: Based on the fused features, encoding and decoding are first performed using the Transformer module to establish global dependencies between features and achieve information recombination. Then, a multilayer perceptron is used as the reconstruction function for image reconstruction, and its reconstruction operation expression is as follows: ， In the formula, Images after the snow has been removed. For reconstruction function, and These represent the weighting coefficients and their corresponding frequency sub-bands at different scales. These are low-frequency components and high-frequency components, respectively. Then, using the snowless image reconstruction module, spatial modulation and feature aggregation are performed at different scales to obtain features at different scales, represented as follows: ， In the formula, These are features at different scales. This is a channel partitioning operation used to decompose the input desnowed feature map into multiple sub-feature maps to support subsequent multi-scale feature processing. The features at different scales are concatenated and then convolved to obtain intermediate features, which are represented as follows: ， In the formula, Indicates intermediate features. For convolution operations, For connection operation; Based on the intermediate features and features at different scales, the modulation features are obtained, represented as follows: ， In the formula, Indicates modulation characteristics, It is a nonlinear mapping function. For features at different scales, representing ; For the modulation features Decode the image to obtain a high-fidelity desnowed image.

8. The single-image snow removal method based on frequency domain enhancement and multi-scale feature extraction according to claim 1, characterized in that, The frequency-domain-enhanced snow removal model uses a composite multi-scale loss function during training, expressed as: ， In the formula, For composite multi-scale loss functions, , As weight, For MSE loss, For Charbonnier loss, This is the edge loss.

9. A single-image snow removal system based on frequency domain enhancement and multi-scale feature extraction, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the single-image desnowing method based on frequency domain enhancement and multi-scale feature extraction as described in any one of claims 1-8.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the single-image desnowing method based on frequency domain enhancement and multi-scale feature extraction as described in any one of claims 1-8.

Images