Mask-free ocean snow removal method based on residual degradation self-guided network

Abstract

The application relates to the technical field of image processing, and particularly discloses a mask-free marine snow removal method based on a residual degradation self-guided network, which proposes a residual degradation self-guided network composed of a coarse recovery module, a residual degradation prior generation module and a residual-guided refinement module. The coarse recovery module provides a coarse recovery image, the residual degradation prior generation module captures residual marine snow artifacts and guides a refinement process to generate a prior map. The residual-guided refinement module extracts coarse features of the coarse recovery image, and after scale alignment and dimension alignment of the prior map and the coarse features, selectively enhances the part of the corresponding degradation area in the coarse features, and then decodes and reconstructs to obtain a final recovery image. The network can automatically extract degradation perception guidance from the residual difference between the input marine snow image and the coarse recovery result, adaptively focus on the residual degradation area, and improve the recovery quality without mask supervision.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a maskless ocean snow removal method based on a residual degradation self-guided network. Background Technology

[0002] Underwater images are frequently affected by marine snow degradation, which is primarily caused by suspended organic and inorganic particles in the water. These particles manifest in acquired images as bright spots, flocculent occlusion, translucent veil-like artifacts, and motion-like streaks. Marine snow not only reduces image visibility and contrast but also obscures important scene structures, further impacting downstream underwater visual tasks such as target detection, tracking, navigation, and environmental monitoring. Therefore, removing marine snow from underwater images is a crucial issue in underwater image restoration.

[0003] In reality, the morphology and concentration of marine snow vary dramatically with factors such as water depth, ocean currents, and biological activity: small-scale particles may appear as isolated bright spots, while large-scale flocculents can cover large areas and form a translucent haze, even producing trailing streaks in dynamic water bodies. This heterogeneity results in high uncertainty in degradation patterns in both the spatial and intensity domains, posing a significant challenge to traditional filtering or statistical modeling methods. Furthermore, the different scattering and absorption characteristics of marine snow at different wavelengths lead to uneven distribution of degradation across the red, green, and blue channels, further increasing the difficulty of recovery. Therefore, designing a removal method that can adaptively handle diverse degradation patterns has become a key requirement for underwater image preprocessing.

[0004] However, most existing learning-based methods still rely on strong supervised information, such as paired clean images or explicit ocean snow masks. While mask-guided methods can provide useful spatial priors for identifying ocean snow particles, accurate ocean snow masks are difficult and costly to obtain in real-world underwater scenarios. Ocean snow typically has diverse appearances, including small particles, large flocculent areas, and semi-transparent blurs, making manual annotation highly ambiguous and extremely time-consuming. Furthermore, inaccurate masks can mislead the recovery network, leading to residual artifacts or over-smoothing of background details.

[0005] Specifically, annotating ocean snow masks requires specialized underwater imagery knowledge. Annotators must distinguish, at the pixel level, which areas belong to degraded particles, which belong to background texture, or which belong to organisms. This process is not only highly subjective but also suffers from low consistency among annotators. Even with semi-automatic annotation tools, the blurred transition between ocean snow and background edges makes boundary delineation exceptionally difficult. More importantly, when ocean snow density is high or overlaps with background structures, even manual assessment of the degree of occlusion is challenging. Consequently, annotated masks often contain omissions or mislabeling. Such noisy masks force the network to learn incorrect attention distributions during training, ultimately impairing reconstruction quality. Considering that obtaining a large number of high-quality paired masks is virtually impossible in real-world underwater operations, developing a self-supervised or self-guided method driven solely by information from the degraded image itself, without relying on external masks, offers significant advantages for practical deployment. Summary of the Invention

[0006] This invention provides a maskless ocean snow removal method based on a residual degenerate self-guided network. The technical problem it solves is: how to achieve adaptive removal of heterogeneous ocean snow while maintaining a clear background structure without the need for manual labeling of ocean snow masks.

[0007] To address the above technical problems, this invention provides a maskless marine snow removal method based on a residual degraded self-guided network, comprising:

[0008] A residual degradation self-guided network is constructed and trained. The residual degradation self-guided network consists of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module cascaded together. The coarse restoration module performs preliminary artifact removal processing on the original ocean snow image to obtain a coarse restored image. The residual degradation prior generation module utilizes the pixel-by-pixel difference between the original ocean snow image and the coarse restored image to perform feature encoding and normalization mapping on the difference, generating a prior map representing the spatial distribution of residual degradation. The residual-guided refinement module extracts coarse features from the coarse restored image, selectively enhances the corresponding degradation regions in the coarse features based on the prior map, and then decodes and reconstructs to obtain the final restored image.

[0009] The image of ocean snow to be removed is input into the trained residual degradation self-guided network. The coarse restoration image is obtained by the coarse restoration module, the prior image is obtained by the residual degradation prior generation module, and the final restored image is reconstructed and output by the residual-guided refinement module.

[0010] Furthermore, the coarse restoration module adopts an encoder-decoder structure. The encoder sequentially performs multi-level convolution and downsampling on the original ocean snow image to output multi-scale semantic features. The decoder sequentially upsamples and convolves the multi-scale semantic features, and fuses the outputs of each level of the encoder with the corresponding level features of the decoder through skip connections to output the coarse restored image.

[0011] Furthermore, the residual degradation prior generation module is composed of a difference calculation unit and a lightweight feature encoding unit connected in series. The difference calculation unit performs pixel-by-pixel absolute difference calculation on the original ocean snow image and the coarse restoration image, and outputs a residual map. The high response regions in the residual map correspond to the locations of residual ocean snow artifacts that the coarse restoration module failed to eliminate, and the low response regions correspond to regions where the background structure is preserved. The lightweight feature encoding unit performs deep spatial feature extraction and cross-pixel correlation modeling on the residual map, and outputs the prior map after normalizing the extracted features.

[0012] Furthermore, the lightweight feature encoding unit is composed of multiple stacked convolutional layers, each followed by a batch normalization layer and a nonlinear activation layer; the last convolutional layer of the lightweight feature encoding unit compresses the number of channels to one, and normalizes the value of each pixel to the interval through an activation function to obtain the prior image.

[0013] Furthermore, the residual-guided refinement module is composed of a feature extraction unit, a feature modulation unit, and a decoding and reconstruction unit cascaded together. The feature extraction unit performs multi-layer convolutional transformation on the coarsely restored image to obtain the coarse features. The feature modulation unit first performs scale alignment and dimension alignment between the prior image and the coarse features. Then, it performs element-wise collaborative modulation between the scale-aligned and dimension-aligned prior image and the coarse features to output refined features. The decoding and reconstruction unit performs progressive upsampling and multi-level feature fusion on the refined features and maps them back to the image space to obtain the final restored image.

[0014] Furthermore, the feature modulation unit upsamples the prior image to the same spatial size as the coarse feature to obtain a scale-aligned prior image; after expanding the scale-aligned prior image along the channel dimension to be consistent with the coarse feature, it multiplies it with the modulation weight and adds 1 to generate element-wise modulation coefficients, and then multiplies the element-wise modulation coefficients with the coarse feature element-wise to obtain refined features.

[0015] Furthermore, when training the residual degradation self-guided network, paired ocean snow degradation images and corresponding clean underwater ground truth images are used as training samples, and a joint loss function is constructed using the reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image.

[0016] Furthermore, the joint loss function is constructed as a weighted sum of the reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image, wherein the reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image adopt L1 loss.

[0017] This invention provides a maskless ocean snow removal method based on a residual degradation self-guided network (RDSNet). RDSNet consists of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module. The coarse restoration module provides a coarsely restored image, while the residual degradation prior generation module captures residual ocean snow artifacts and guides the refinement process, generating a prior map representing the spatial distribution of residual degradation. The residual-guided refinement module extracts coarse features from the coarsely restored image, performs scale and dimension alignment between the prior map and the coarse features, and selectively enhances the corresponding degradation regions within the coarse features (while retaining the original features for clean regions). The final restored image is then obtained through decoding and reconstruction. Unlike existing methods that rely on explicit ocean snow masks or additional degradation annotations, RDSNet automatically extracts degradation-aware guidance from the residual differences between the input ocean snow image and the coarse restoration result. With this self-generated residual degradation prior, the proposed framework can adaptively focus on the remaining degradation regions and improve restoration quality without mask supervision. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the residual degradation self-guiding network provided in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are given for illustrative purposes only and should not be construed as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0020] The maskless marine snow removal method based on a residual degenerate self-guided network provided in this invention includes:

[0021] A Residual Degradation Self-Guided Network (RDSNet) was constructed and trained. The Residual Degradation Self-Guided Network consists of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module cascaded together. The coarse restoration module performs preliminary artifact removal processing on the original ocean snow image to obtain a coarse restored image. The residual degradation prior generation module utilizes the pixel-by-pixel difference between the original ocean snow image and the coarse restored image to perform feature encoding and normalization mapping on the difference, generating a prior map representing the spatial distribution of residual degradation. The residual-guided refinement module extracts coarse features from the coarse restored image, selectively enhances the corresponding degradation regions in the coarse features based on the prior map, and then decodes and reconstructs to obtain the final restored image.

[0022] The image of ocean snow to be removed is input into the trained residual degradation self-guided network. The coarse restoration image is obtained by the coarse restoration module, the prior image is obtained by the residual degradation prior generation module, and the final restored image is reconstructed and output by the residual-guided refinement module.

[0023] The structure of the residual degenerate self-guiding network is as follows: Figure 1 As shown, the network consists of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module cascaded together. The entire network takes the original ocean snow image as its sole input and the final restored image as its sole output, without relying on externally labeled ocean snow masks for supervision.

[0024] 1. Coarse recovery module

[0025] The coarse restoration module employs a deep convolutional encoder-decoder structure. The encoder uses multi-level downsampling operations, each level containing multiple stacked convolutional layers, batch normalization layers, and ReLU activation layers to progressively extract global semantic features from the input image. The decoder restores spatial resolution through upsampling operations at the corresponding level (using transposed convolution or bilinear interpolation followed by convolutional layers) and establishes skip connections between features at the same level in the encoder and decoder to compensate for the detail lost during downsampling. The coarse restoration module performs the following steps on the original ocean snow image: ( This represents the number of pixels in the height direction of the image. (This represents the number of pixels in the image width direction; 3 represents the red, green, and blue color channels.) Preliminary artifact removal processing is performed, roughly suppressing high-brightness particles, flocculent occlusions, and semi-transparent veil-like degradations globally, while preserving the main structure and color distribution of the underwater scene as much as possible. Finally, a coarsely restored image is output. ,in This indicates the operation of the coarse restoration module. (Coarse restored image) Its spatial dimensions are related to the input image. Maintain consistency.

[0026] 2. Residual Degeneracy Prior Generation Module

[0027] The residual degradation prior generation module consists of a difference calculation unit and a lightweight feature encoding unit connected in series. Its output prior map maintains the spatial correspondence with the input image and is scale-aligned according to the resolution of the coarse features in the subsequent feature modulation stage.

[0028] The difference calculation unit is used to solve the original ocean snow image pixel by pixel. With coarse restoration image The absolute difference between them is calculated, and a residual map reflecting the degree of difference between the two pixels is output. :

[0029] ,

[0030] in, This indicates the operation of taking the absolute value pixel by pixel, that is, for and The absolute difference is calculated for the values ​​at every spatial location and for every color channel. Because... and All The residual map obtained after performing pixel-by-pixel absolute difference operations on the tensor. Its spatial dimensions are still The number of channels is 3. In the residual map, the high response areas correspond to the locations of residual ocean snow artifacts that were not effectively eliminated in the coarse recovery stage, while the low response areas correspond to areas where the background structure is well preserved.

[0031] Lightweight feature coding unit It consists of multiple stacked convolutional layers, each followed by a batch normalization layer and a nonlinear activation layer, used to process the residual map. Deep spatial feature extraction and cross-pixel correlation modeling are performed. The final convolutional layer of the lightweight feature encoding unit compresses the number of channels from the middle feature channels to 1, and then applies the sigmoid activation function. After normalizing the value of each pixel to the (0,1) interval, the prior image is obtained. :

[0032] .

[0033] Its spatial dimensions are It has 1 channel and is a two-dimensional continuous-value image with the same spatial size as the input image. The value of each pixel is... This characterizes the relative severity of residual degradation at a given location—a value closer to 1 indicates more significant residual degradation, while a value closer to 0 indicates that the location has been largely restored—thus providing spatial distribution guidance for subsequent refinement operations. The generated prior map... Instead of being considered an explicit degradation mask, it serves as a soft degradation-aware guiding map, providing a spatial degradation prior for the subsequent residual-guided refinement process. Upon entering the feature modulation stage, due to the prior map... The prior image may have a different spatial resolution than the features to be modulated. Therefore, it is necessary to first perform a scale alignment operation on the prior image to make it spatially consistent with the features to be modulated, and then participate in element-wise feature modulation.

[0034] 3. Residual-guided refinement module

[0035] The residual-guided refinement module consists of a feature extraction unit, a feature modulation unit, and a decoding and reconstruction unit connected in sequence.

[0036] The feature extraction unit employs a pre-constructed shallow convolutional backbone network to coarsely reconstruct the image. Perform multi-layer convolutional transformations, followed by batch normalization layers and ReLU activation layers after each convolutional layer, to obtain coarse features containing spatial structure and texture details. :

[0037] ,

[0038] in, The mapping function represents the feature extraction unit.

[0039] The feature modulation unit receives coarse features. and prior diagram Both are subjected to degradation-sensory feature modulation. Due to the prior image Generated in the image space by the residual degradation prior generation module, while coarse features Features extracted through convolution typically have different spatial resolutions; therefore, prior images need to be processed before element-wise modulation. Perform scale alignment and dimension alignment.

[0040] Specifically, bilinear interpolation, nearest neighbor interpolation, transposed convolution, or other learnable / non-learnable scaling transformations are used to modify the prior image. Upsampling or scaling is performed to adjust its spatial dimensions and coarse features. To maintain consistency, we obtain a scale-aligned prior image. .

[0041] Due to residual degenerate prior maps It is a single-channel, two-dimensional continuous value plot, with coarse features. It contains multiple feature channels, therefore further processing along the channel dimension Perform replication expansion to adjust its channel count and coarse features. To maintain consistency, we obtain a prior image that is both scale-aligned and dimension-aligned. .

[0042] Prior graphs with scale and dimension alignment For coarse features Element-by-element modulation yields refined features:

[0043] ,

[0044] in, This indicates element-wise multiplication. The prior modulation weights control the degree of influence of residual degradation priors on feature enhancement. This allows the network to enhance the feature representation of regions with strong residual degradation responses while preserving the original feature information of the recovered background regions. Through this modulation process, the feature modulation unit outputs refined features that have undergone degradation-selective enhancement. Its spatial dimensions and number of channels are both similar to Maintain consistency.

[0045] The decoding and reconstruction unit employs a lightweight decoder for refined features. Perform progressive upsampling and multi-level feature fusion operations to map it back to image space, obtaining the final reconstructed image:

[0046] ,

[0047] in, This indicates the operation of the decoding and reconstruction unit to restore the image. The space size is restored to The number of channels is restored to 3, which is consistent with the input image. The dimensions are completely consistent.

[0048] The decoding and reconstruction unit does not introduce any external masks or additional priors; image reconstruction is entirely driven by refined feature maps.

[0049] 4. Loss Function

[0050] During the training phase, the residual degradation self-guided network uses paired images of ocean snow degradation. and the corresponding clean underwater truth image These are the training samples. To ultimately reconstruct the image. With the truth image Reconstruction loss and coarse restored image With the truth image The reconstruction loss between the two is used to construct a joint loss function.

[0051] Restore image With the truth image The reconstruction loss between them is defined as:

[0052] ,

[0053] in, L1 norm is represented by summing the absolute values ​​of all elements of the tensor. Using L1 loss helps to preserve the edge sharpness and structural integrity of the recovered image.

[0054] To maintain consistency between the coarse restoration result and the final restoration result, the coarse restoration image... With the truth image The reconstruction loss between them is defined as:

[0055] .

[0056] The joint loss function is then constructed as follows:

[0057] ,

[0058] in, The weighting coefficient is used to balance the two objectives, and its value ranges from 0.1 to 1.0.

[0059] 5. Training Setup

[0060] The backpropagation algorithm simultaneously updates all trainable parameters of the coarse recovery module, the residual degradation prior generation module, and the residual-guided refinement module end-to-end. This includes the kernel weights and biases of each convolutional layer, and the scaling and translation factors of the batch normalization layer. The prior map output by the residual degradation prior generation module is used during training. It is generated automatically by the network forward computation and does not involve any manually labeled ocean snow mask in loss calculation or gradient update.

[0061] This embodiment trains and evaluates the proposed RDSNet on two publicly available ocean snow removal datasets: MSRB and Snowy-VAROS. MSRB is a widely used ocean snow removal benchmark that provides paired ocean snow images and their corresponding clean underwater images for supervised restoration. Snowy-VAROS contains more challenging underwater scenes with diverse ocean snow distributions, including different particle scales, densities, and spatial patterns, making it suitable for evaluating the robustness of restoration methods under complex degradation conditions. To ensure a fair comparison, we follow existing work and use the official training and test sets of the two datasets.

[0062] During training, all images were randomly cropped and resized. Pixel, i.e. , Data augmentation strategies such as random horizontal flipping and random rotation were employed to increase the diversity of training samples. During evaluation, the original test images were resized to the same resolution to ensure consistent comparison. Recovery performance was evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Higher PSNR and SSIM indicate better recovery quality.

[0063] All experiments were performed on a single NVIDIA RTX 5090 GPU with mixed-precision training enabled to reduce memory consumption and speed up training. The Adam optimizer was used, and the initial learning rate was set to [value missing]. The weights are gradually reduced using a cosine decay strategy. The batch size is set to 8, and the model is trained for 200 epochs. The weight coefficients in the training objective... Determined through experience.

[0064] During training, optimization is performed using only pairs of ocean snow images and clean reference images, without requiring additional ocean snow masks or degradation annotations. During inference, RDSNet takes only degraded ocean snow images as input and directly outputs the restored results, thus demonstrating its maskless restoration capability.

[0065] 6. Comparison with state-of-the-art methods

[0066] This embodiment compares RDSNet with representative methods, including underwater image enhancement methods, image denoising methods, and ocean snow removal methods. The comparison results are shown in Table 1. The methods used for comparison specifically include:

[0067] UGAN stands for Generative Adversarial Network (GAN)-based underwater image enhancement method, which improves the color and sharpness of underwater images through adversarial training between the generator and the discriminator.

[0068] NU2Net, an underwater image enhancement network proposed in the AAAI 2023 Oral paper, has only 3.15M parameters. It improves image clarity and color balance while maintaining high efficiency, but there is still room for improvement in color control accuracy.

[0069] DnCNN, a deep convolutional neural network image denoising model, uses residual learning strategy and batch normalization technique to accelerate training and improve denoising performance. Its core idea is to predict noise components rather than directly output clean images. A single model can uniformly handle multiple tasks such as Gaussian denoising, super-resolution and JPEG deblocking.

[0070] Restormer, an efficient Transformer image restoration model proposed at CVPR 2022, captures long-distance pixel interactions while maintaining computational feasibility through key designs of multi-head attention and feedforward networks. It is suitable for various restoration tasks such as image deraining, deblurring, and denoising.

[0071] HDCW-Net, a single-image snow removal algorithm proposed at ICCV 2021, uses hierarchical dual-tree complex wavelet transform (DTCWT) to better capture snow morphology in different directions and introduces contradictory channel loss as a feature to judge residual snow. It cuts large snow blocks into sub-bands for processing through hierarchical decomposition.

[0072] TransWeather, a Transformer-based end-to-end weather degradation removal model proposed at CVPR 2022, employs a single encoder-single decoder architecture. It enhances local attention through Transformer blocks within patches to remove small-scale degradation and introduces learnable weather type embeddings to adapt to different degradation types.

[0073] HiTrans, a hierarchical Transformer integration for ocean snow removal, leverages the Transformer's strengths in capturing global representations to remove ocean snow artifacts.

[0074] Of the methods mentioned above, UGAN and NU2Net are for underwater image enhancement, DnCNN is for general image denoising, Restormer and TransWeather are general image restoration models, HDCW-Net is for image desnowing, and HiTrans is specifically designed for ocean snow removal tasks.

[0075] As shown in Table 1, RDSNet achieves competitive performance on both the MSRB and Snowy-VAROS datasets. Compared to existing methods, RDSNet consistently demonstrates improved performance, proving the effectiveness of the proposed residual degradation self-guided mechanism. Unlike mask-guided methods that require explicit ocean snow localization, RDSNet automatically extracts degradation-aware cues from the residual differences between the input image and the coarsely reconstructed image. Although some methods achieve similar PSNR values, RDSNet performs better in structure restoration, especially in challenging scenarios with heterogeneous ocean snow distributions. This indicates that residual degradation priors can provide effective guidance for restoring degraded regions while preserving background structure.

[0076] Table 1 Quantitative comparisons on the MSRB and Snowy-VAROS datasets

[0077]

[0078] 7. Ablation test

[0079] To verify the effectiveness of each component, an ablation experiment was conducted on MSRB in this embodiment. The results are shown in Table 2, where the baseline model directly predicts the restored image without using residual guidance and residual degradation priors.

[0080] Table 2 Ablation experiments of RDSNet on MSRB

[0081]

[0082] As shown in Table 2, the recovery performance was improved after introducing the residual-guided refinement module, indicating that the coarse-to-fine recovery strategy is effective. Further improvement was achieved by adding a residual degradation prior generation module, demonstrating that the self-generated degradation cues can provide effective guidance for recovering residual ocean snow artifacts.

[0083] In summary, this invention proposes a maskless ocean snow removal method based on a residual degradation self-guided network (RDSNet). RDSNet consists of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module. The coarse restoration module provides a coarse restored image, while the residual degradation prior generation module captures residual ocean snow artifacts and guides the refinement process, generating a prior map characterizing the spatial distribution of residual degradation. The residual-guided refinement module extracts coarse features from the coarse restored image, aligns these features with the row scale and dimension of the prior map, and uses them for degradation-aware feature modulation. The final restored image is then obtained through decoding and reconstruction. Unlike existing methods that rely on explicit ocean snow masks or additional degradation annotations, RDSNet automatically extracts degradation-aware guidance from the residual differences between the input ocean snow image and the coarse restoration result. With this self-generated residual degradation prior, the proposed framework can adaptively focus on the remaining degradation region and improve restoration quality without mask supervision. Extensive experiments on publicly available ocean snow removal datasets demonstrate that RDSNet can effectively remove heterogeneous ocean snow degradation while preserving background structure and image details.

[0084] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A maskless marine snow removal method based on residual degenerate self-guided networks, characterized in that, include: Construct and train a residual degenerate self-guided network; The residual degradation self-guided network is composed of a coarse restoration module, a residual degradation prior generation module, and a residual-guided refinement module cascaded together. The coarse restoration module performs preliminary artifact removal processing on the original ocean snow image to obtain a coarse restored image. The residual degradation prior generation module utilizes the pixel-by-pixel difference between the original ocean snow image and the coarse restored image to perform feature encoding and normalization mapping on the difference, generating a prior map representing the spatial distribution of residual degradation. The residual-guided refinement module extracts coarse features from the coarse restored image, selectively enhances the corresponding degradation regions in the coarse features based on the prior map, and then decodes and reconstructs to obtain the final restored image. The image of ocean snow to be removed is input into the trained residual degradation self-guided network. The coarse restoration image is obtained by the coarse restoration module, the prior image is obtained by the residual degradation prior generation module, and the final restored image is reconstructed and output by the residual-guided refinement module.

2. The maskless marine snow removal method based on a residual degenerate self-guided network according to claim 1, characterized in that: The coarse restoration module adopts an encoder-decoder structure. The encoder sequentially performs multi-level convolution and downsampling on the original ocean snow image to output multi-scale semantic features. The decoder sequentially upsamples and convolves the multi-scale semantic features, and fuses the outputs of each level of the encoder with the corresponding level features of the decoder through skip connections to output the coarse restored image.

3. The maskless marine snow removal method based on a residual degenerate self-guiding network according to claim 1, characterized in that: The residual degradation prior generation module is composed of a difference calculation unit and a lightweight feature encoding unit connected in series. The difference calculation unit performs pixel-by-pixel absolute difference calculation on the original ocean snow image and the coarse restoration image, and outputs a residual map. The high response regions in the residual map correspond to the locations of residual ocean snow artifacts that the coarse restoration module failed to eliminate, and the low response regions correspond to regions where the background structure is preserved. The lightweight feature encoding unit performs deep spatial feature extraction and cross-pixel correlation modeling on the residual map, and outputs the prior map after normalizing the extracted features.

4. The maskless marine snow removal method based on a residual degenerate self-guiding network according to claim 3, characterized in that: The lightweight feature encoding unit consists of multiple stacked convolutional layers, each followed by a batch normalization layer and a non-linear activation layer. The last convolutional layer of the lightweight feature encoding unit compresses the number of channels to one and normalizes the value of each pixel to a range using an activation function to obtain the prior image.

5. The maskless marine snow removal method based on a residual degenerate self-guided network according to claim 1, characterized in that: The residual-guided refinement module is composed of a feature extraction unit, a feature modulation unit, and a decoding and reconstruction unit cascaded together; the feature extraction unit performs multi-layer convolutional transformation on the coarsely restored image to obtain the coarse features; The feature modulation unit first performs scale alignment and dimension alignment between the prior image and the coarse feature; then it performs element-wise collaborative modulation between the scale-aligned and dimension-aligned prior image and the coarse feature to output refined features. The decoding and reconstruction unit performs progressive upsampling and multi-level feature fusion on the refined features, and maps them back to the image space to obtain the final restored image.

6. The maskless marine snow removal method based on a residual degenerate self-guided network according to claim 5, characterized in that: The feature modulation unit upsamples the prior image to the same spatial size as the coarse feature to obtain a scale-aligned prior image; after expanding the scale-aligned prior image along the channel dimension to match the coarse feature, it multiplies it with the modulation weight and adds 1 to generate element-wise modulation coefficients; then, the element-wise modulation coefficients are multiplied element-wise with the coarse feature to obtain refined features.

7. The maskless marine snow removal method based on a residual degenerate self-guided network according to claim 6, characterized in that: When training the residual degradation self-guided network, paired ocean snow degradation images and corresponding clean underwater ground truth images are used as training samples, and a joint loss function is constructed using the reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image.

8. The maskless marine snow removal method based on a residual degenerate self-guided network according to claim 7, characterized in that: The joint loss function is constructed as a weighted sum of the reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image. The reconstruction loss between the final restored image and the ground truth image and the reconstruction loss between the coarse restored image and the ground truth image are both constructed using L1 loss.

Images