Image snow removal method based on multi-frequency enhanced degradation perception routing attention

Abstract

The invention provides an image snow removal method based on dual-frequency domain enhancement, which adopts a discrete wavelet transform technology to decompose high and low frequencies and fuse high-frequency enhancement features and low-frequency enhancement features after the high and low frequencies pass through a feature enhancement network so as to enhance the recovery of detail textures of a snowflake-degraded image. According to the method provided by the invention, the high-frequency characteristics of the snow degradation image are enhanced by using a differential convolutional network, and meanwhile, the serious degradation influence caused by large-scale snowflakes is reduced by using a low-frequency enhancement network, so that the performance of the snow degradation image in different frequency domains during snowflake removal and image restoration is improved. And meanwhile, global modeling is carried out on a seriously degraded area through a degradation sensing routing attention module, so that seriously degraded snowflakes are further removed more thoroughly.

Description

Technical Field

[0001] The invention relates to an image snow removal method in the field of bad weather image restoration. Background Art

[0002] With the development of image processing technology, single image snow removal technology is increasingly used in target detection, video surveillance, image restoration and other fields. Due to the diversity of snowflake shapes, sizes, reflectivity and transparency, the degradation characteristics of snowy scenes are more complex and changeable than other severe weather scenes, which greatly affects the clarity and detail presentation of the image. These degradation phenomena usually manifest as snowflakes blocking the scene, texture destruction and color shift, which brings great challenges to the true restoration of the target scene.

[0003] In the prior art, traditional image snow removal methods are usually based on physical models and image prior knowledge, and achieve snow removal by separating the foreground and background or analyzing features such as snowflake transparency and saturation. However, due to the complexity of snowflake morphology and its similarity to background texture, traditional methods are difficult to accurately distinguish snowflakes from the background, which easily leads to residual snowflakes or loss of image details.

[0004] In recent years, with the rapid development of deep learning technology, image snow removal methods based on deep neural networks have gradually become a research hotspot. For example, methods such as DesnowNet and HDCW-Net have significantly improved the effect of image snow removal by using convolutional neural networks (CNNs) to design multi-scale feature extraction networks or introducing attention mechanisms. However, this short-range modeling method based on CNNs is difficult to capture long-range spatial dependencies, resulting in poor performance when dealing with large-scale snow degradation scenes. To overcome this limitation, recent studies have adopted algorithms based on Swin Transformer, but due to window limitations, it is still difficult to capture global dependencies. At the same time, since it is difficult to distinguish snow from background information in scenes where the snowflake morphology is similar to the background, the field of image snow removal still faces the following challenges: (1) Large-scale snowflakes are not completely removed. The distribution of snowflakes in the image is usually uneven. Some areas are blocked by large-scale snowflakes, resulting in the loss of important information. Existing methods are difficult to effectively restore the scene structure and semantic information of the occluded area. (2) It is difficult to effectively distinguish snow degradation from scene information. In snow scene images, some snow degradation is often highly similar to the morphological features of the background scene. This similarity makes it difficult for traditional snow removal methods to effectively distinguish between snow degradation and scene information when processing images, resulting in the failure to completely remove snow or the loss of some scene information in the restored image, thus affecting the quality of the snow removal effect and the integrity of the restored details.

[0005] Therefore, in response to the above problems, there is an urgent need for a new image snow removal method that can effectively capture the characteristics of degraded areas and adapt to multi-scale snow degradation at the same time, so as to further improve the accuracy and effect of image snow removal. Summary of the invention

[0006] In order to solve the problem of incomplete removal of large-scale snow degradation and excessive removal leading to loss of scene information in the field of image snow removal, this paper proposes an image snow removal method based on dual-frequency enhancement and routed self-attention. The proposed image snow removal method can restore images in complex snow degradation scenes. Figure 1 As shown, the technical solution for implementing the present invention is based on multi-frequency enhanced routing attention, which is generally a U-shaped encoding and decoding structure, including two processes: network training and reasoning.

[0007] The network training process includes the following steps:

[0008] Step 1: Acquire paired clear natural images and images after snow degradation. First, use the camera to acquire a clear natural image J(x). Then, based on the snow degradation model I(x)=K(x)t(x)+A(1-t(x)), where K(x)=(1-Z(x)R(x))+C(x)Z(x)R(x), generate the corresponding snow-degraded image.

[0009] Step 2: Input the snow degradation image I(x) modeled by the snow degradation model into a 7×7 grouped convolutional feature extractor to obtain the snow degradation feature map F(x).

[0010] Step 3: Input the feature map F(x) into the multi-frequency enhancement module, enhance the high-frequency features and low-frequency features respectively, and then fuse them. The calculation steps are as follows:

[0011] a. Use discrete wavelet transform (DWT) to decompose the feature map F(x) to obtain high-pass sub-band and low-pass sub-band (HH, HL, LH, LL) = DWT(F(x));

[0012] b. Divide the low-pass subband LL into windows and perform window division on the i-th window P i Calculate the low-frequency feature LL′(x) that is enhanced by the long-range dependency information of other feature points in the window. The calculation steps are described as follows:

[0013]

[0014] LL′=MSFN(WSA(LL))+WSA(LL)

[0015] , where LL q , LL k , LL v They are obtained by linear transformation of LL, Wd It is a depth-separable convolution with a convolution kernel of 3×3. MSFN is a multi-scale feedforward neural network based on depth-separable convolution.

[0017] c. By inputting F(x) into the central difference convolution module to extract its high-frequency information, the high-frequency enhanced feature map F′(x) is obtained. The feature point p0 is selected to calculate the corresponding output. The calculation process is as follows:

[0018]

[0019] , where θ represents a weight parameter, w(p n ) is the weight parameter of the convolution kernel.

[0020] d. Decompose the high-frequency enhanced feature map F′(x) by discrete wavelet transform to obtain a high-pass subband HH, and perform inverse discrete wavelet transform (IDWT) on HH and the enhanced low-frequency feature LL′ to obtain a multi-frequency enhanced feature map;

[0021] Step 4: The multi-frequency enhanced feature map is further input into the degradation-aware routing attention module. The degradation-aware module generates a guide map of the degradation area, and then divides the guide map into windows. Softmax is used to select the feature points at the same position (u, v) in different windows. The calculation is performed to obtain the One-hot tensor P, and the tensor P is stacked and merged to obtain the mask map M.

[0022] Step 5: divide the multi-frequency enhanced feature map into windows Use the one-hot tensor P with Perform matrix multiplication to obtain the degraded region routing window F D , and then F D Perform global correlation attention calculation with the feature map to obtain the degraded region routing code F established through global dependency D ′. The calculation steps are as follows:

[0023]

[0024] Step 6: Encode the degraded region window F D ′ is copied, extended, stacked and merged, and then the mask map M is used to filter it to obtain the final degraded area coding feature, and then the coding feature is fused with the multi-frequency enhanced feature map to obtain the multi-frequency enhanced feature map with global dependency of the degraded area.

[0025] Step 7: After three stages of multi-frequency enhanced routing attention encoding and downsampling, the obtained features are input into the global context interaction module to obtain globally enhanced feature encoding.

[0026] Step 8: Upsampling the feature code. First, use a 2×2 convolution to convolve the reference frame feature, so that the number of channels becomes 2C. Then, use the PixelShuffle operation to double the width and height dimensions, and the number of channels becomes

[0027] Step nine: Use a dynamic fusion mechanism and residual connection to fuse shallow features with deep features at the same level.

[0028] Step 10: Input the fused features into the decoder that combines multi-frequency enhancement and route-related attention, and perform residual connection between the output R(x) and the input I(x) to obtain the predicted snow-free image J. p (x) The robustness of the network is enhanced by obtaining the residual between the sparser snow-free image and the snow-degraded image.

[0029] Step 11: Calculate the predicted snow-free image J p L1 loss and dynamic frequency domain loss L between (x) and J(x) FFL , and the mean absolute error regularization function L for one-hot tensors MADR .

[0030] The degradation-aware routing attention method based on multi-frequency enhancement is characterized in that the multi-frequency enhancement degradation-aware routing attention module used includes two stages. Stage one consists of a multi-frequency enhancement module, which enhances the high and low frequencies of the input feature map, and stage two consists of a degradation-aware routing attention module. The degradation-aware routing attention module includes a degradation-aware routing module, which routes the degraded area, uses the degraded area features as a query, performs a global dependency establishment operation in the global feature map, and finally fuses and enhances the globally modeled degraded area with the input feature map to further restore the severely degraded area.

[0031] The beneficial effect of this method is that the proposed snow removal method based on multi-frequency enhancement and routing-associated attention enables the model to enhance the perception of the texture details of snow-degraded images by combining discrete wavelet and differential convolution frequency enhancement and fusion methods, better capture the texture difference between snowflakes and background, and improve the recognition ability of snowflakes and scene information. The global dependency modeling of snow-degraded areas is carried out by degradation-aware routing, and its receptive field is increased to capture more long-range information to further enhance the recovery of degraded areas, so as to further solve the problem of incomplete snow removal in large-scale snow-degraded areas. Therefore, the proposed snow removal method based on multi-frequency enhancement and routing-associated attention can more thoroughly remove snowflakes in complex scenes with multi-scale snow degradation while retaining more scene information. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Attached Figure 1 Schematic diagram of the network model of the multi-frequency enhanced degradation-aware routing attention method of the present invention.

[0033] Attached Figure 2 Schematic diagram of a multi-frequency enhancement module of the multi-frequency enhancement degradation-aware routing attention method of the present invention.

[0034] Attached Figure 3 Schematic diagram of the degradation-aware routing module of the multi-frequency enhanced degradation-aware routing attention method of the present invention.

[0035] Attached Figure 4 This is a demonstration of the snow removal effect of the multi-frequency enhanced degradation-aware routing attention method of the present invention. DETAILED DESCRIPTION

[0036] A typical embodiment of a snow removal method based on multi-frequency enhanced routing attention of the present invention is described in detail below, and the method is further described in detail. It is necessary to point out here that the following embodiments are only used to further illustrate the method and cannot be understood as limiting the scope of protection of the method. Personnel skilled in the art may make some non-essential improvements and adjustments to the method based on the above content of the method, which still fall within the scope of protection of the present invention. The present invention proposes a snow removal method based on multi-frequency enhanced routing associated attention, such as Figure 1 As shown, the basic steps of this method are as follows:

[0037] Step 1: Acquire paired clear natural images and images after snow degradation. First, use the camera to acquire a clear natural image J(x). Then, based on the snow degradation model I(x)=K(x)t(x)+A(1-t(x)), where K(x)=(1-Z(x)R(x))+C(x)Z(x)R(x), generate the corresponding snow-degraded image.

[0038] Step 2: Input the snow degradation image I(x) modeled by the snow degradation model into a 7×7 grouped convolutional feature extractor to obtain the snow degradation feature map F(x).

[0039] Step 3: Input the feature map F(x) into the multi-frequency enhancement module, enhance the high-frequency features and low-frequency features respectively, and then fuse them. The calculation steps are as follows:

[0040] a. Use discrete wavelet transform (DWT) to decompose the feature map F(x) to obtain high-pass sub-band and low-pass sub-band (HH, HL, LH, LL) = DWT(F(x));

[0041] b. Divide the low-pass subband LL into windows and perform window division on the i-th window P iCalculate the low-frequency feature LL′(x) that is enhanced by the long-range dependency information of other feature points in the window. The calculation steps are described as follows:

[0042]

[0043] LL′=MSFN(WSA(LL))+WSA(LL)

[0044] , where LL q , LL k , LL v They are obtained by linear transformation of LL, W d It is a depth-separable convolution with a convolution kernel of 3×3. MSFN is a multi-scale feedforward neural network based on depth-separable convolution.

[0046] c. By inputting F(x) into the central difference convolution module to extract its high-frequency information, the high-frequency enhanced feature map F′(x) is obtained. The feature point p0 is selected to calculate the corresponding output. The calculation process is as follows:

[0047]

[0048] , where θ represents a weight parameter, w(p n ) is the weight parameter of the convolution kernel.

[0049] d. Decompose the high-frequency enhanced feature map F′(x) by discrete wavelet transform to obtain a high-pass subband HH, and perform inverse discrete wavelet transform (IDWT) on HH and the enhanced low-frequency feature LL′ to obtain a multi-frequency enhanced feature map;

[0050] Step 4: The multi-frequency enhanced feature map is further input into the degradation-aware routing attention module. The degradation-aware module generates a guide map of the degradation area, and then divides the guide map into windows. Softmax is used to select the feature points at the same position (u, v) in different windows. The calculation is performed to obtain the One-hot tensor P, and the tensor P is stacked and merged to obtain the mask map M.

[0051] Step 5: divide the multi-frequency enhanced feature map into windows Use the one-hot tensor P with Perform matrix multiplication to obtain the degraded region routing window F D , and then F D Perform global correlation attention calculation with the feature map to obtain the degraded region routing code F established through global dependency D ′. The calculation steps are as follows:

[0052]

[0053] Step 6: Encode the degraded region window F D ′ is copied, extended, stacked and merged, and then the mask map M is used to filter it to obtain the final degraded area coding feature, and then the coding feature is fused with the multi-frequency enhanced feature map to obtain the multi-frequency enhanced feature map with global dependency of the degraded area.

[0054] Step 7: After three stages of multi-frequency enhanced routing attention encoding and downsampling, the obtained features are input into the global context interaction module to obtain globally enhanced feature encoding.

[0055] Step 8: Upsampling the feature code. First, use a 2×2 convolution to convolve the reference frame feature, so that the number of channels becomes 2C. Then, use the PixelShuffle operation to double the width and height dimensions, and the number of channels becomes

[0056] Step nine: Use a dynamic fusion mechanism and residual connection to fuse shallow features with deep features at the same level.

[0057] Step 10: Input the fused features into the decoder that combines multi-frequency enhancement and route-related attention, and perform residual connection between the output R(x) and the input I(x) to obtain the predicted snow-free image J. p (x) The robustness of the network is enhanced by obtaining the residual between the sparser snow-free image and the snow-degraded image.

[0058] Step 11: Calculate the predicted snow-free image J p L1 loss and dynamic frequency domain loss L between (x) and J(x) FFL , and the mean absolute error regularization function L for one-hot tensors MADR .

[0059] The degradation-aware routing attention method based on multi-frequency enhancement is characterized in that the network structure adopts three stages. The first stage is the encoding stage, in which semantic features are obtained by encoding through the multi-frequency enhanced routing attention module and three downsamplings. The three downsamplings all use a convolution kernel size of 3×3 and a step size of 2 to perform a 2-fold downsampling operation. The second stage is the deep semantic information extraction stage. In this stage, a self-attention mechanism based on a moving window is used to perform global information modeling to obtain deep semantic information. The third stage is the decoding stage. In this stage, three upsamplings are used for scale recovery, and then jump connections are used at each level to dynamically fuse shallow and deep information. The frequency-enhanced routing attention module is used at each level for decoding operations.

[0060] The snow removal method based on multi-frequency enhanced degradation-aware routing attention is characterized in that the multi-frequency enhanced degradation-aware routing attention module used includes two stages. Stage one consists of a multi-frequency enhancement module, which enhances the high and low frequencies of the input feature map, and stage two consists of a degradation-aware routing attention module. The degradation-aware routing attention module includes a degradation-aware routing module, which routes the degraded area, uses the degraded area features as a query, performs a global dependency establishment operation in the global feature map, and finally fuses and enhances the globally modeled degraded area with the input feature map to further restore the severely degraded area.

[0061] The snow removal method based on multi-frequency enhanced degradation-aware routing attention is characterized in that the steps are divided into three stages. The first stage is composed of a Haar-type discrete wavelet transform to perform frequency division and down-sampling output on the input feature map. The second stage is composed of a parallel central difference convolution layer and a sliding window self-attention module to enhance the high and low frequency information of the input features respectively. The third stage is an inverse wavelet transform to fuse and up-sample the enhanced high and low frequency information to obtain a feature map enhanced at high and low frequencies.

[0062] The snow removal method based on multi-frequency enhanced degradation-aware routing attention is characterized in that the degradation-aware routing attention consists of three stages. First, the degradation-aware network of the degradation-aware routing attention module is used to obtain the degradation-guided graph, and the window is divided into windows. The softmax is used to select the feature points at the same position (u, v) in different windows. The calculation is performed to obtain the one-hot tensor P, and the tensor P is stacked and merged to obtain the mask map M. The second stage is to divide the multi-frequency enhanced feature map into windows to obtain Use the one-hot tensor P with Perform matrix multiplication to obtain the degraded region routing window F D , and then F D Perform global correlation attention calculation with the feature map to obtain the degraded region routing code F established through global dependency D ′. The third stage is to encode the degraded region window F D ′ is copied, extended, stacked and merged, and then the mask map M is used to filter it to obtain the final degraded area coding feature, and then the coding feature is fused with the multi-frequency enhanced feature map to obtain the multi-frequency enhanced feature map with global dependency of the degraded area.

[0063] In the example of the present invention, the datasets used for network training come from five widely used public synthetic snow removal datasets including CSD, SRRS, and Snow100K. The loss function of the model includes the pixel L1 loss between the predicted snow-free image and the clear image, and the loss function L between the frequency domain obtained after Fourier transform. FFL and a penalty entropy function to constrain the generation of one-hot tensors. In our overall network, we adopt a U-shaped architecture, in which skip connections are used between the encoder and decoder. Convolutional layers are used at the beginning and end of the network. We use multi-frequency enhanced routing attention modules in the encoder and decoder stages, and a sliding window self-attention mechanism in the high-dimensional stage. The dimensions of each layer are set to 32, 64, 128, 256, respectively. The number of blocks deployed at different stages is set to [1, 2, 3, 6, 3, 2, 1]. We use the Adam optimizer and set β1, β2, and ε to the default values, i.e., 0.9, 0.999, and 1e-8. In addition, the initial learning rate and batch size are set to 2e-4 and 8, respectively. We use a cosine annealing strategy to adjust the learning rate, which gradually decreases from the initial value to 2e-6. In terms of data augmentation, we randomly crop 256×256 image patches, rotate them by a fixed angle (i.e., 0°, 90°, 180°, or 270°), and apply vertical or horizontal flipping. During the entire training process, the model was trained for 1200 cycles. The training environment of the proposed method is the Pytorch framework of Nvidia RTX 3090GPU and Intel Core(TM)i7-7820X CPU@3.60Hz.

[0064] Table 1 Comparison of snow removal effects of different image snow removal algorithms on public datasets

[0065]

[0066] In order to verify the effectiveness and generalization ability of the proposed method, the present invention uses the trained network and multiple image snow removal algorithms to conduct sufficient comparisons on the datasets CSD, SRRS, and Snow100K. Table 1 shows the snow removal effect evaluation results of the method proposed in the present invention and other image snow removal methods. According to the results in Table 1, the snow removal performance of the proposed method is better than that of other methods. Specifically, among other comparison methods, Restormer has the highest PSNR and SSIM on all datasets. Despite this, the present invention achieves higher PSNR and SSIM than Restormer. This is mainly due to the multi-frequency enhancement module, which can capture more high-frequency detail information to reduce the loss of details in the restored image and the residue of snowflakes. In addition, the degradation-aware routing attention module uses the degradation-aware routing module to route severely degraded areas, and then uses global aggregation to globally model this part of information, further enhancing the recovery of large-scale snowflake degradation areas. Figure 4 The snow removal effect of the present invention on the CSD dataset is intuitively demonstrated. According to the graphical results, it can be seen that the snow removal algorithm of the present invention can better restore clear snow-free images. Benefiting from the enhancement characteristics of the multi-frequency enhancement module on high-frequency and low-frequency areas, the proposed network can better restore the details in the snow-degraded image and reduce the artifacts caused by incomplete removal of some snowflakes.

Claims

1. A degradation-aware routing attention method based on multi-frequency enhancement, characterized in that: The multi-frequency enhanced degradation-aware routing attention module used consists of two stages: The network training process is specifically described as follows: Step 1: Acquisition of paired clear natural images and images after snow degradation. First, use the camera to acquire a clear natural image J(x). Then, based on the snow degradation model I(x)=K(x)t(x)+A(1-t(x)), where K(x)=(1-Z(x)R(x))+C(x)Z(x)R(x), generate the corresponding snow degradation image. Step 2: Input the snow degradation image I(x) modeled by the snow degradation model into a 7×7 grouped convolutional feature extractor to obtain the snow degradation feature map F(x); Step 3: Input the feature map F(x) into the multi-frequency enhancement module, enhance the high-frequency features and low-frequency features respectively, and then fuse them. The calculation steps are as follows: a. Use discrete wavelet transform (DWT) to decompose the feature map F(x) to obtain high-pass sub-band and low-pass sub-band (HH, HL, LH, LL) = DWT(F(x)); b. Divide the low-pass subband LL into windows and perform window division on the i-th window P i Calculate the low-frequency feature LL′(x) that is enhanced by the long-range dependency information of other feature points in the window. The calculation steps are described as follows: LL′=MSFN(WSA(LL))+WSA(LL), Among them LL q , LL k , LL v They are obtained by linear transformation of LL, W d It is a depth-separable convolution with a convolution kernel of 3×3. MSFN is a multi-scale feedforward neural network based on depth-separable convolution. By inputting F(x) into the central difference convolution to extract its high-frequency information, the high-frequency enhanced feature map F′(x) is obtained, and the feature point p0 is selected to calculate the corresponding output. The calculation process is as follows: , where θ represents a weight parameter, w(p n ) is the weight parameter of the convolution kernel; c. Decompose the high-frequency enhanced feature map F′(x) by discrete wavelet transform to obtain the high-pass subband HH, and perform inverse discrete wavelet transform (IDWT) on HH and the enhanced low-frequency feature LL′ to obtain the multi-frequency enhanced feature map Step 4: The multi-frequency enhanced feature map is further input into the degradation-aware routing attention module. The degradation-aware module generates a guide map of the degradation area, and then divides the guide map into windows. Softmax is used to select the feature points at the same position (u, v) in different windows. Calculate to get the one-hot tensor P, and stack and merge the tensor P to get the mask map M; Step 5: divide the multi-frequency enhanced feature map into windows Use the one-hot tensor P with Perform matrix multiplication to obtain the degraded region routing window F D , and then F D Perform global correlation attention calculation with the feature map to obtain the degraded region routing code F established through global dependency D ′, the calculation steps are as follows: Step 6: Encode the degraded region window F D ′ is copied, expanded, stacked and merged, and then the mask map M is used to filter it to obtain the final degraded area coding feature, and then the coding feature is fused with the multi-frequency enhanced feature map to obtain the multi-frequency enhanced feature map with global dependency of the degraded area; Step 7: After three stages of multi-frequency enhanced routing attention encoding and downsampling, the obtained features are input into the global context interaction module to obtain globally enhanced feature encoding; Step 8: Upsampling the feature code. First, use a 2×2 convolution to convolve the reference frame feature, so that the number of channels becomes 2C. Then, use the PixelShuffle operation to double the width and height dimensions, and the number of channels becomes Step nine: Use a dynamic fusion mechanism and residual connection to fuse shallow features with deep features at the same level; Step 10: Input the fused features into the decoder that combines multi-frequency enhancement and route-related attention, and perform residual connection between the output R(x) and the input I(x) to obtain the predicted snow-free image J. p (x) Enhance the robustness of the network by obtaining the residual of the sparser snow-free image and the snow-degraded image; Step 11: Calculate the predicted snow-free image J p L1 loss and dynamic frequency domain loss L between (x) and J(x) FFL , and the mean absolute error regularization function L for one-hot tensors MADR .

2. The snow removal method based on multi-frequency enhanced degradation-aware routing attention according to claim 1 is characterized in that: The network structure adopts three stages. The first stage is the encoding stage, in which semantic features are encoded through a multi-frequency enhanced routing attention module and three downsamplings. The three downsamplings all use a convolution with a convolution kernel size of 3×3 and a step size of 2 to perform a 2-fold downsampling operation. The second stage is the deep semantic information extraction stage. In this stage, a self-attention mechanism based on a moving window is used to perform global information modeling to obtain deep semantic information. The third stage is the decoding stage. In this stage, three upsamplings are used for scale recovery, and then jump connections are used at each level to dynamically fuse shallow and deep information. The frequency enhanced routing attention module is used at each level for decoding operations.

3. The snow removal method based on multi-frequency enhanced degradation-aware routing attention according to claim 2 is characterized in that: The multi-frequency enhanced degradation-aware routing attention module used includes two stages. Stage one consists of a multi-frequency enhancement module, which enhances the high and low frequencies of the input feature map. Stage two consists of a degradation-aware routing attention module, which includes a degradation-aware routing module. By routing the degraded area, the degraded area features are used as queries, and global dependency establishment operations are performed in the global feature map. Finally, the globally modeled degraded area is fused and enhanced with the input feature map to further restore the severely degraded area.

4. The snow removal method based on multi-frequency enhanced degradation-aware routing attention according to claim 3 is characterized in that: The steps are divided into three stages. The first stage is composed of a haar-type discrete wavelet transform to perform frequency division and down-sampling on the input feature map, the second stage is composed of a parallel center difference convolution layer and a sliding window self-attention module to enhance the high and low frequency information of the input features respectively, and the third stage is composed of an inverse wavelet transform to fuse and up-sample the enhanced high and low frequency information to obtain a feature map enhanced at high and low frequencies.

5. The snow removal method based on multi-frequency enhanced degradation-aware routing attention according to claim 4 is characterized in that: The degradation-aware routing attention consists of three stages. First, the degradation-aware network of the degradation-aware routing attention module is used to obtain the degradation-guided graph, and the window is divided into windows. The softmax is used to select the feature points at the same position (u, v) in different windows. The calculation is performed to obtain the One-hot tensor P, and the tensor P is stacked and merged to obtain the mask map M. The second stage is to divide the multi-frequency enhanced feature map into windows to obtain Use the one-hot tensor P with Perform matrix multiplication to obtain the degraded region routing window F D , and then F D Perform global correlation attention calculation with the feature map to obtain the degraded region routing code F established through global dependency D ′, the third stage is to encode the degraded area window F D ′ is copied, extended, stacked and merged, and then the mask map M is used to filter it to obtain the final degraded area coding feature, and then the coding feature is fused with the multi-frequency enhanced feature map to obtain the multi-frequency enhanced feature map with global dependency of the degraded area.

Images