An underwater image denoising method and system of a multi-order dynamic wavelet gated aggregation network

Abstract

The application provides an underwater image denoising method and system of a multi-order dynamic wavelet gated aggregation network. The method comprises: obtaining a to-be-processed underwater degraded image and performing preprocessing; constructing an expansion context branch to perform multi-scale feature extraction, to obtain an expansion context feature map; constructing an adaptive reuse branch to perform cross-layer feature reuse and dynamic weighting processing, to obtain an adaptive reuse feature map; constructing a wavelet gated transform branch to perform frequency domain noise separation and adaptive weighting processing, to obtain a wavelet gated feature map; constructing a dynamic hierarchical gated aggregation module to perform adaptive weighted aggregation on the three-path features, to obtain aggregated optimized features; and constructing a feature fusion layer and a reconstruction layer to output a denoised target image. Through the architecture design of three-path cooperation and frequency domain-spatial domain fusion, the application effectively solves the problems in the prior art, such as weak noise and texture distinguishing ability, poor adaptability to underwater coupling degradation, and difficulty in balancing noise suppression and detail preservation.

Description

Technical Field

[0001] This invention relates to the field of underwater image processing technology, and more particularly to an underwater image denoising method and system using a multi-order dynamic wavelet gated aggregation network. Background Technology

[0002] Underwater imagery plays an indispensable role in various fields, including marine resource exploration, underwater archaeology, ecological monitoring, and navigation of autonomous underwater vehicles. However, due to the inherent absorption and scattering characteristics of light in water, underwater images often suffer from complex degradation phenomena. These phenomena include low contrast, severe color distortion, and most notably, significant noise interference caused by suspended particles (sea snow) and low-light sensor limitations. These degradation problems severely affect visual quality and the accuracy of subsequent downstream tasks.

[0003] Compared to traditional land images, the noise mechanisms in underwater scenes are far more complex. First, wavelength-dependent attenuation leads to a noticeable blue-green cast and low-light environments, necessitating high ISO settings, which introduce significant sensor noise. Second, forward scattering from plankton and suspended particles causes image blurring, while backscattering produces a haze-like effect. Crucially, these noise components are often coupled with color casts and haze, making standard denoising algorithms ineffective. Furthermore, factors such as uneven lighting and water current disturbances exacerbate these problems. The coupling of these degrading factors makes underwater image denoising an extremely challenging ill-posed problem.

[0004] Early research focused on physical models, with researchers proposing the Dark Channel Prior (DCP) theory. This was subsequently adapted for underwater environments, forming the Underwater Dark Channel Prior (UDCP) method. Galdran et al. further proposed a red channel prior method, providing a new technical approach for underwater image haze removal. While these methods achieve haze removal by constructing physical models to estimate transmission maps and demonstrate some effectiveness in haze suppression, they generally suffer from design flaws—failing to consider noise components in underwater images. This leads to a significant amplification of potential sensor noise within the image during inverse image restoration, ultimately resulting in noticeable artifacts in the restored image and failing to meet the image quality requirements of high-precision applications.

[0005] To overcome the limitations of the aforementioned physical model-based methods, image enhancement techniques have been widely studied, aiming to improve the visibility of underwater images by enhancing image contrast and detail. For example, Hisam et al. employed histogram equalization, and Fu et al. developed an enhancement algorithm based on Retinex theory. While these methods offer high computational efficiency, they suffer from a core technical weakness: they misclassify noise as high-frequency detail components. Therefore, while enhancing image texture information, they inevitably over-enhance noise particles, resulting in a noticeable grainy visual defect in the processed image. This fails to achieve true noise suppression and yields ineffective denoising.

[0006] In recent years, deep learning technology has made groundbreaking progress in the field of image processing, providing a new technical approach for underwater image restoration. However, existing research exhibits a clear domain bias: a large body of literature focuses on underwater color correction and haze removal (such as typical solutions like WaterNet and UW-CycleGAN), while targeted research specifically for underwater image denoising is relatively scarce. To address this research gap, the current common technical strategy is to directly transfer state-of-the-art (SOTA) architectures in general image denoising to underwater scenarios, but this strategy suffers from significant adaptability limitations.

[0007] In general image denoising deep learning architectures, the Denoising Convolutional Neural Network (DnCNN) pioneered the introduction of residual learning mechanisms, achieving blind Gaussian denoising and laying the technological foundation for the field of general image denoising. Building upon this, researchers have designed various improved architectures for complex noise features: DudeNet employs a dual-branch feature extraction structure to enhance the processing capability for complex noise; Attention-Guided Denoising Network (ADNet) and Enhanced Convolutional Neural Denoising Network (ECNDNet) introduce attention mechanisms to attempt to effectively distinguish noise from texture; and Multi-Level Wavelet Denoising CNN (MWDCNN) integrates wavelet transform technology to separate high-frequency noise through frequency domain decomposition.

[0008] Compared to the rapid development of general image denoising models, dedicated networks specifically designed for underwater image restoration are extremely rare. UDRN is a representative attempt in this direction, adapting to underwater optical characteristics through a customized residual learning mechanism. However, this approach and existing similar underwater-specific technologies still have significant shortcomings: Firstly, the aforementioned general denoising models (such as ADNet and MWDCNN) lack inductive biases to address the coupling degradation problem of turbidity and light attenuation unique to underwater environments, making them unable to effectively cope with complex multi-factor degradation scenarios underwater. Secondly, early underwater-specific models (such as UDRN) failed to fully integrate current advanced feature extraction techniques, resulting in limited feature representation capabilities. Therefore, existing technical solutions struggle to achieve a balance between effective noise suppression and accurate preservation of fine marine details in highly turbid underwater environments such as the deep sea, failing to meet the demands of high-end applications such as deep-sea exploration.

[0009] Besides convolutional neural networks (CNNs), the Transformer architecture has also shown potential application value in the field of underwater image processing in recent years. U-shaped Transformers, WaterFormers, and other schemes significantly improve the global correlation of feature representations by introducing global attention mechanisms. However, these Transformer architectures have inherent technical limitations: they lack the local inductive bias found in convolutional neural networks, resulting in poor performance in capturing and removing fine-grained, high-frequency noise patterns unique to the underwater environment, and making it difficult to adapt to the complex distribution characteristics of underwater noise.

[0010] In summary, existing underwater image denoising techniques, whether traditional or deep learning-based, all have their own shortcomings and limitations, specifically: weak noise and texture differentiation capabilities, poor adaptability to underwater coupled degradation problems, difficulty in balancing noise suppression and detail preservation, and poor performance in removing fine-grained noise. These technical bottlenecks not only prevent existing solutions from effectively handling image denoising in complex underwater environments but also restrict the application of underwater imaging technology in high-end practical scenarios. Therefore, there is an urgent need to propose a novel underwater image denoising technique that can adapt to complex underwater degradation characteristics and balance noise suppression and detail preservation to address the deficiencies of existing technologies. Summary of the Invention

[0011] To address the shortcomings of existing underwater image denoising techniques, such as weak noise and texture differentiation, poor adaptability to underwater coupling degradation, difficulty in balancing noise suppression and detail preservation, and inadequate fine-grained noise removal, this invention provides a multi-level dynamic wavelet-gated aggregation network-based underwater image denoising method. This invention primarily utilizes an extended context branch to capture multi-scale global semantic information, an adaptive reuse branch to achieve cross-layer feature reuse and dynamic weighting through a dynamically hierarchical gated aggregation block, a wavelet-gated transform branch to separate noise and effective signals in the frequency domain through multi-level wavelet decomposition and gating mechanisms, and a dynamically hierarchical gated aggregation module to achieve adaptive weighted aggregation of three-path features. This effectively suppresses complex noises such as sensor noise and suspended particle noise in underwater images while preserving fine details such as marine biological textures and reef edges. It achieves a balance between noise suppression and detail preservation in the high-turbidity environment of the deep sea, meeting the requirements for high-precision underwater image analysis.

[0012] The technical means employed in this invention are as follows:

[0013] A method for underwater image denoising using a multi-order dynamic wavelet-gated aggregation network includes: S1. Obtain the underwater degradation image to be processed, and preprocess the underwater degradation image to obtain the preprocessed underwater degradation image; S2. Construct an extended context branch. Input the preprocessed underwater degradation image into the extended context branch for multi-scale feature extraction. The extended context branch captures global semantic information and local details of different receptive fields through multiple parallel extended residual blocks to obtain an extended context feature map. S3. Construct an adaptive reuse branch. Input the expanded context feature map into the adaptive reuse branch for cross-layer feature reuse and dynamic weighting. The adaptive reuse branch achieves effective fusion of shallow detail features and deep semantic features through dense connections and dynamic hierarchical gating aggregation mechanism to obtain an adaptive reuse feature map. S4. Construct a wavelet-gated transform branch, input the adaptive reuse feature map into the wavelet-gated transform branch for frequency domain noise separation and adaptive weighting processing. The wavelet-gated transform branch separates noise and effective signal in the frequency domain through multi-level wavelet decomposition and gating mechanism to obtain the wavelet-gated feature map. S5. Construct a dynamic hierarchical gated aggregation module. Input the expanded context feature map, adaptive reuse feature map and wavelet gated feature map into the dynamic hierarchical gated aggregation module for adaptive weighted aggregation and feature reuse, suppress redundant information and enhance effective features to obtain aggregated optimized features. S6. Construct a feature fusion layer and a reconstruction layer, input the aggregated optimized features into the feature fusion layer and the reconstruction layer, and output the denoised target image.

[0014] Further, step S1 includes: S11. Perform sliding window cropping on the acquired underwater degraded image; S12. Normalize the cropped image patch to the [0,1] value range; S13. Perform data augmentation transformation on the cropped image patches to expand the training dataset.

[0015] Further, step S2 includes: S21. Construct dilated residual blocks as basic processing units and replace standard convolutions with dilated convolutions; S22. Set up four parallel extended residual blocks, configure different expansion rates for each extended residual block, and the expansion rates are distributed in an increasing manner. S23. The preprocessed underwater degradation image is processed in parallel at multiple scales through the four parallel extended residual blocks to obtain feature representations of different receptive fields, and to perceive local details and global structural background at the same time. S24. Aggregate the features output by the four parallel extended residual blocks to form an extended context feature map.

[0016] Furthermore, in step S22, the dilation rate is distributed in an increasing manner to ensure that the receptive fields are densely interleaved rather than sparsely sampled, effectively mitigating the grid effect introduced by dilated convolution, and enabling the network to gradually aggregate contextual information at continuous scales.

[0017] Further, step S3 includes: S31. Construct dense connection blocks to achieve effective fusion of shallow detailed features and deep semantic features through cross-layer connections; S32. Input the expanded context feature map into the dense connection block to obtain the dense connection features; S33. Construct a dynamic hierarchical gated aggregation block based on the dense connection block, and use a dynamic channel segmentation mechanism to replace the fixed channel segmentation strategy, and adaptively allocate channel resources according to the specific characteristics of the input features. S34. Input the densely connected features into the dynamic hierarchical gating aggregation block, dynamically adjust the weight allocation of each channel, use the dynamic aggregation attention mechanism to adaptively suppress redundant information, improve noise extraction and feature expression capabilities, and output an adaptive reuse feature map.

[0018] Further, step S34 includes: S341. The dynamic hierarchical gated aggregation block performs global average pooling on the densely connected features of the input to obtain channel-level global information. S342. The channel-level global information is mapped to channel attention weights through a fully connected layer, and the channel attention weights are adaptively adjusted according to the degradation characteristics of the input features. S343. Based on the channel attention weights, the channels of the densely connected features are dynamically segmented and weighted to achieve adaptive filtering and aggregation of features. S344. Perform residual connections between the weighted aggregated features and the densely connected features to output the optimized adaptive reuse features.

[0019] Further, step S4 includes: S41. Construct a wavelet transform module, input the adaptive reuse feature map into the wavelet transform module for multi-level wavelet decomposition, decompose the image into low-frequency sub-band and high-frequency sub-band, the low-frequency sub-band contains color and brightness information, the high-frequency sub-band contains detail and noise information; S42. Construct a low-frequency processing unit, input the low-frequency sub-band into the low-frequency processing unit, and perform color correction and overall brightness adjustment processing. S43. Construct a high-frequency processing unit, input the high-frequency sub-band into the high-frequency processing unit, and perform noise suppression and detail preservation processing; S44. Construct a gating mechanism module to dynamically adjust the weights of each frequency sub-band based on the noise characteristics of the input image; S45. Input the processed low-frequency sub-band and high-frequency sub-band into the gating mechanism module for adaptive weighted fusion; S46. Perform wavelet reconstruction on each frequency sub-band after weighted fusion and output wavelet-gated feature maps.

[0020] Further, step S5 includes: S51. Construct a feature receiving unit to receive the extended context feature map, the adaptive reuse feature map, and the wavelet-gated feature map; S52. Construct a channel splicing unit to splice the feature maps output by the three paths along the channel dimension to form a comprehensive feature representation; S53. Construct a dynamic weighting unit and adaptively weight the spliced ​​features through a dynamic hierarchical gating aggregation mechanism to learn the importance weights of each path feature. S54. Construct a feature fusion unit, and perform weighted fusion of the three-path features based on the learned weights to achieve cross-path feature reuse; S55. Construct an attention optimization unit, suppress redundant information through the attention mechanism, strengthen features that are effective for denoising, and output aggregated optimized features.

[0021] Further, step S6 includes: S61. Input the aggregated optimized features into the feature fusion layer to perform deep fusion of multi-scale features; S62. Input the fused features into the reconstruction layer, and perform image reconstruction through convolution operation to restore the clear image after denoising. S63. Output the denoised target image, which retains fine ocean details while sufficiently suppressing noise.

[0022] This invention also provides an underwater image denoising system based on the above-mentioned underwater image denoising method using a multi-order dynamic wavelet gated aggregation network, comprising: The preprocessing module is used to acquire the underwater degradation image to be processed and to preprocess the underwater degradation image to obtain the preprocessed underwater degradation image. The extended context branch module is used to extract multi-scale features from the preprocessed underwater degraded image input. It captures global semantic information and local details of different receptive fields through multiple parallel extended residual blocks to obtain an extended context feature map. The adaptive reuse branch module is used to perform cross-layer feature reuse and dynamic weighting processing on the input of the expanded context feature map. Through dense connections and dynamic hierarchical gating aggregation blocks, it achieves effective fusion of shallow detail features and deep semantic features to obtain an adaptive reuse feature map. The wavelet-gated transform branch module is used to perform frequency domain noise separation and adaptive weighting processing on the adaptive reuse feature map input. Through multi-level wavelet decomposition and gating mechanism, noise and effective signal are separated in the frequency domain to obtain the wavelet-gated feature map. The dynamic hierarchical gating aggregation module is used to adaptively weight and aggregate the expanded context feature map, adaptive reuse feature map and wavelet gating feature map into the input, suppress redundant information and enhance effective features to obtain aggregated optimized features. The image reconstruction module is used to fuse and reconstruct the aggregated optimized feature inputs and output the denoised target image.

[0023] Compared with the prior art, the present invention has the following advantages: 1. The extended context branch provided by this invention achieves dense interleaving of multi-scale receptive fields by setting four parallel extended residual blocks and configuring increasing dilation rates (1, 2, 4, 8), effectively alleviating the grid effect, and simultaneously perceiving local details and global structural background, thereby enhancing the model's ability to perceive objects of different scales and improving training stability.

[0024] 2. The adaptive reuse branch provided by this invention combines dense connection blocks with dynamic hierarchical gating aggregation blocks, and adopts a dynamic channel segmentation mechanism to replace the fixed allocation strategy, thereby achieving effective fusion of shallow detail features and deep semantic features. It adaptively allocates channel resources according to the degradation characteristics of input features, suppresses redundant information, and improves noise extraction and feature expression capabilities.

[0025] 3. The wavelet gated transform branch provided by this invention, through the combination of multi-level wavelet decomposition and adaptive gating mechanism, realizes the separation of noise and effective signal in the frequency domain, processes different frequency sub-bands separately (low-frequency color correction, high-frequency noise suppression), and dynamically adjusts the weight of each frequency sub-band according to underwater environmental conditions, effectively coping with complex underwater multi-factor degradation scenarios.

[0026] 4. The dynamic hierarchical gating aggregation module provided by this invention achieves collaborative optimization of three-path features by adaptively weighting and reusing the extended context feature map, adaptive reuse feature map, and wavelet gating feature map, thereby suppressing redundant information and strengthening effective features, and achieving a balance between effective noise suppression and accurate preservation of fine ocean details.

[0027] In summary, the technical solution of this invention addresses the problems in existing technologies, such as the lack of inductive bias in general denoising models that address the coupling degradation of turbidity and light attenuation specific to underwater environments, the limited feature representation capabilities of early underwater-specific models, and the poor fine-grained noise removal effect caused by the lack of local inductive bias in the Transformer architecture. Through a three-path collaborative and frequency-spatial domain fusion architecture design, it achieves precise processing of coupling degradation such as light attenuation, turbidity interference, and high-frequency noise in underwater images. Therefore, the technical solution of this invention solves the problems of weak noise and texture differentiation capabilities, poor adaptability to underwater coupling degradation problems, difficulty in balancing noise suppression and detail preservation, and poor fine-grained noise removal effect in existing technologies.

[0028] Based on the above reasons, this invention can be widely applied in fields such as marine resource exploration, underwater archaeology, ecological monitoring, autonomous underwater vehicle navigation, deep-sea exploration, and underwater vision systems. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the overall architecture of the multi-order dynamic wavelet gated aggregation network of the present invention.

[0031] Figure 2 This is a schematic diagram of the extended context branch structure of the present invention.

[0032] Figure 3 This is a schematic diagram of the adaptive branch reuse structure of the present invention.

[0033] Figure 4This is a schematic diagram of the wavelet-gated transform branch of the present invention.

[0034] Figure 5 This invention provides a comparison of the performance metrics of different models on an underwater image test set.

[0035] Figure 6 A visual comparison of underwater image denoising using different methods provided in embodiments of the present invention at a noise level of 25.

[0036] Figure 7 A visual comparison of underwater image denoising using different methods provided in embodiments of the present invention at a noise level of 25.

[0037] Figure 8 A visual comparison of different methods provided in the embodiments of the present invention for denoising land images at a noise level of 25. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0039] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product or device.

[0040] like Figure 1 As shown, this invention provides an underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network, comprising: S1. Obtain the underwater degradation image to be processed, and preprocess the underwater degradation image to obtain the preprocessed underwater degradation image; S2. Construct an extended context branch. Input the preprocessed underwater degradation image into the extended context branch for multi-scale feature extraction. The extended context branch captures global semantic information and local details of different receptive fields through multiple parallel extended residual blocks to obtain an extended context feature map. S3. Construct an adaptive reuse branch. Input the expanded context feature map into the adaptive reuse branch for cross-layer feature reuse and dynamic weighting. The adaptive reuse branch achieves effective fusion of shallow detail features and deep semantic features through dense connections and dynamic hierarchical gating aggregation mechanism to obtain an adaptive reuse feature map. S4. Construct a wavelet-gated transform branch, input the adaptive reuse feature map into the wavelet-gated transform branch for frequency domain noise separation and adaptive weighting processing. The wavelet-gated transform branch separates noise and effective signal in the frequency domain through multi-level wavelet decomposition and gating mechanism to obtain the wavelet-gated feature map. S5. Construct a dynamic hierarchical gated aggregation module. Input the expanded context feature map, adaptive reuse feature map and wavelet gated feature map into the dynamic hierarchical gated aggregation module for adaptive weighted aggregation and feature reuse, suppress redundant information and enhance effective features to obtain aggregated optimized features. S6. Construct a feature fusion layer and a reconstruction layer, input the aggregated optimized features into the feature fusion layer and the reconstruction layer, and output the denoised target image.

[0041] In this embodiment, the present invention achieves precise processing of coupling degradation such as light attenuation, turbidity interference, and high-frequency noise in underwater images by organically integrating wavelet-gated residual networks, dynamic multi-level gated aggregation (DHGB) attention mechanisms, and three-path collaborative architecture, and finally outputs clear images with sufficient noise suppression and high detail restoration.

[0042] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. Select 40 underwater images from the dataset, of which 20 underwater images are used for pre-training and 20 underwater images are used for testing. Perform sliding window cropping on the acquired degraded underwater images. In this embodiment, the preferred cropping window size is 41×41 with a step size of 10. S12. Normalize the cropped image patch to the [0,1] value range; S13. Data augmentation transformations are performed on the cropped image patches to expand the training dataset. These transformations include: horizontal flipping, 90° counter-clockwise rotation, 180° counter-clockwise rotation, 270° counter-clockwise rotation, 90° counter-clockwise rotation followed by horizontal flipping, 180° counter-clockwise rotation followed by horizontal flipping, and 270° counter-clockwise rotation followed by horizontal flipping. Ultimately, 3872 41×41 underwater images are obtained and used as the real training dataset. All methods are performed on the same training and test sets with a consistent number of training iterations to ensure experimental fairness.

[0043] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Construct dilated residual blocks as basic processing units, replacing standard convolutions with dilated convolutions; for example... Figure 2 As shown, the basic building block of this branch is the dilated residual block (DRB), in which the standard convolution is replaced by a dilated convolution (a block with a dilation rate of 1 is shown in the figure for reference). S22. Four parallel dilated residual blocks are configured, each with a different dilation rate, and the dilation rates are distributed in an increasing manner. In this embodiment, underwater images exhibit various degradation characteristics at different scales, including fine suspended particles, global color shifts, and uneven illumination. Standard convolutions with fixed small receptive fields struggle to model these complex patterns simultaneously. To address this issue, we construct this branch using four parallel dilated residual blocks, configuring each block with a different dilation rate. By changing the dilation factor, the network obtains diverse receptive fields, enabling it to simultaneously perceive local details and global structural background.

[0044] S23. The preprocessed underwater degradation image is processed in parallel at multiple scales through the four parallel extended residual blocks to obtain feature representations of different receptive fields, and to perceive local details and global structural background at the same time. S24. Aggregate the features output by the four parallel extended residual blocks to form an extended context feature map.

[0045] In a preferred embodiment of the invention, in step S22, the dilation rate is progressively increased, set sequentially to 1, 2, 4, and 8. This ensures dense interleaving of the receptive field rather than sparse sampling, effectively mitigating the grid effect introduced by dilated convolution and enabling the network to gradually aggregate contextual information at continuous scales. In this embodiment, although dilated convolution can effectively expand the receptive field without increasing computational cost, it inevitably introduces the "grid effect," leading to the loss of local information continuity. By assigning progressively increasing dilation rates to parallel blocks, dense interleaving of the receptive field is ensured, rather than sparse sampling. Therefore, the network can gradually aggregate contextual information at continuous scales, effectively smoothing grid artifacts while maintaining strong feature extraction capabilities.

[0046] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. Construct dense connection blocks to achieve effective fusion of shallow detailed features and deep semantic features through cross-layer connections; S32. Input the expanded context feature map into the dense connection block to obtain the dense connection features; S33. A dynamic hierarchical gated aggregation block is constructed based on the densely connected block. A dynamic channel segmentation mechanism is used to replace the fixed channel segmentation strategy, and channel resources are adaptively allocated according to the specific characteristics of the input features. In this embodiment, inspired by the MOGA design, the dynamic hierarchical gated aggregation block proposed in this invention adopts a dynamic channel segmentation mechanism, which enables the module to adaptively allocate channel resources according to the specific characteristics of the input features, overcoming the limitation that the fixed allocation mechanism cannot adaptively capture complex and spatially changing degradation patterns in underwater images.

[0047] S34. Input the densely connected features into the dynamic hierarchical gating aggregation block, dynamically adjust the weight allocation of each channel, use the dynamic aggregation attention mechanism to adaptively suppress redundant information, improve noise extraction and feature expression capabilities, and output an adaptive reuse feature map.

[0048] In this embodiment, as Figure 3 As shown, adaptive reuse of branch fusion dense connection blocks and dynamic hierarchical gating aggregation blocks, with cross-layer feature reuse as the core, achieves effective fusion of shallow detailed features and deep semantic features.

[0049] In a specific implementation, as a preferred embodiment of the present invention, step S34 includes: S341. The dynamic hierarchical gated aggregation block performs global average pooling on the densely connected features of the input to obtain channel-level global information. S342. The channel-level global information is mapped to channel attention weights through a fully connected layer, and the channel attention weights are adaptively adjusted according to the degradation characteristics of the input features. S343. Based on the channel attention weights, the channels of the densely connected features are dynamically segmented and weighted to achieve adaptive filtering and aggregation of features. S344. Perform residual connections between the weighted aggregated features and the densely connected features to output the optimized adaptive reuse features.

[0050] In a specific implementation, as a preferred embodiment of the present invention, step S4 includes: S41. Construct a wavelet transform module, input the adaptive reuse feature map into the wavelet transform module for multi-level wavelet decomposition, decompose the image into low-frequency sub-band and high-frequency sub-band, the low-frequency sub-band contains color and brightness information, the high-frequency sub-band contains detail and noise information; S42. Construct a low-frequency processing unit, input the low-frequency sub-band into the low-frequency processing unit, and perform color correction and overall brightness adjustment processing. S43. Construct a high-frequency processing unit, input the high-frequency sub-band into the high-frequency processing unit, and perform noise suppression and detail preservation processing; S44. Construct a gating mechanism module to dynamically adjust the weights of each frequency sub-band based on the noise characteristics of the input image; S45. Input the processed low-frequency sub-band and high-frequency sub-band into the gating mechanism module for adaptive weighted fusion; S46. Perform wavelet reconstruction on each frequency sub-band after weighted fusion and output wavelet-gated feature maps.

[0051] In this embodiment, the constructed wavelet-gated transform branch is as follows: Figure 4 As shown, due to the optical properties of the water medium, underwater images not only contain traditional Gaussian and salt-and-pepper noise, but also structured noise caused by scattering from suspended particles and color channel imbalance degradation caused by wavelength-selective absorption. This complex noise distribution makes it difficult for denoising methods relying solely on spatial convolution to effectively distinguish useful signals from noise components. To address these limitations, wavelet-gated convolution combines the multi-resolution analysis capabilities of wavelet transform with an adaptive gating mechanism. Through multi-level wavelet decomposition, the model can process noise at different frequency levels: for low-frequency sub-bands, it focuses on color correction and overall brightness adjustment; for high-frequency sub-bands, it emphasizes noise suppression and detail preservation. The introduced gating module gives the model adaptive selection capabilities, dynamically adjusting the weights of each frequency sub-band based on the noise characteristics of the input image. In underwater environments, image noise distribution varies significantly under different depths and water quality conditions. The gating mechanism allows the model to adapt to these changes without manual parameter tuning, automatically learning when to rely more on low-frequency information (e.g., scenes with severe color cast) or high-frequency information (e.g., scenes with rich detail).

[0052] In a specific implementation, as a preferred embodiment of the present invention, step S5 includes: S51. Construct a feature receiving unit to receive the extended context feature map, the adaptive reuse feature map, and the wavelet-gated feature map; S52. Construct a channel splicing unit to splice the feature maps output by the three paths along the channel dimension to form a comprehensive feature representation; S53. Construct a dynamic weighting unit and adaptively weight the spliced ​​features through a dynamic hierarchical gating aggregation mechanism to learn the importance weights of each path feature. S54. Construct a feature fusion unit, and perform weighted fusion of the three-path features based on the learned weights to achieve cross-path feature reuse; S55. Construct an attention optimization unit, suppress redundant information through the attention mechanism, strengthen features that are effective for denoising, and output aggregated optimized features.

[0053] In a specific implementation, as a preferred embodiment of the present invention, step S6 includes: S61. Input the aggregated optimized features into the feature fusion layer to perform deep fusion of multi-scale features; S62. Input the fused features into the reconstruction layer, and perform image reconstruction through convolution operation to restore the clear image after denoising. S63. Output the denoised target image, which retains fine ocean details while sufficiently suppressing noise.

[0054] By using the denoising model of the present invention, the following effects can be achieved: 1. Objective indicators: WMAPNet of this invention improves PSNR by 1.8~2.5 dB, SSIM by 0.03~0.05, and UIQM by 0.06~0.09 on the test set, which is significantly better than general denoising models and existing underwater-specific models.

[0055] 2. Subjective effect: In deep-sea high turbidity scenes, WMAPNet can effectively suppress high-frequency noise particles and artifacts, while preserving details such as marine life textures and reef edges; while the comparison models (such as ADNet and WaterFormer) have noise residue or blurred details.

[0056] 3. Core Advantage Verification: Through ablation experiments (removing the wavelet gating module, DHGB attention mechanism, and three-path architecture respectively), the necessity of each core module is verified. After removing any module, the PSNR decreases by 0.8~1.2 dB, proving the synergistic effect of the architecture design of this invention.

[0057] 4. The model adaptively separates noise and effective signals in the frequency domain using a wavelet-gated residual network, solving the problems of difficulty in distinguishing noise from texture and incomplete removal of high-frequency noise in existing technologies. Addressing the unique sensor noise and fine-grained noise caused by turbid particles in underwater environments, the noise residue in the denoised image is reduced by more than 40% compared to existing UDRN models, with no noticeable graininess in subjective visual perception, and an artifact removal rate improved by 50%, meeting the needs of high-precision underwater image analysis.

[0058] 5. Superior performance on multiple benchmark datasets: Comprehensive qualitative and quantitative experimental results show that WMAPNet exhibits good robustness and image restoration ability under different noise levels.

[0059] This invention also provides an underwater image denoising system based on the above-mentioned underwater image denoising method using a multi-order dynamic wavelet gated aggregation network, comprising: The preprocessing module is used to acquire the underwater degradation image to be processed and to preprocess the underwater degradation image to obtain the preprocessed underwater degradation image. Expansion context branch module, which is used to input the preprocessed underwater degraded image for multi-scale feature extraction, capture the global semantic information and local details of different receptive fields through multiple parallel dilated residual blocks, and obtain the expansion context feature map; Adaptive reuse branch module, which is used to input the expansion context feature map for cross-layer feature reuse and dynamic weighting processing, and effectively fuse the shallow detail features and deep semantic features through dense connection and dynamic hierarchical gating aggregation block to obtain the adaptive reuse feature map; Wavelet gating transform branch module, which is used to input the adaptive reuse feature map for frequency domain noise separation and adaptive weighting processing, and separate the noise and effective signals in the frequency domain through multi-level wavelet decomposition and gating mechanism to obtain the wavelet gating feature map; Dynamic hierarchical gating aggregation module, which is used to input the expansion context feature map, adaptive reuse feature map and wavelet gating feature map for adaptive weighted aggregation and feature reuse, suppress redundant information and strengthen effective features, and obtain the aggregated and optimized features; Image reconstruction module, which is used to input the aggregated and optimized features for fusion and reconstruction, and output the denoised target image.

[0060] For the embodiments of the present invention, since it corresponds to the above embodiments, the description is relatively simple. For the relevant similarities, please refer to the description in the above embodiments, and details are not described here.

[0061] Embodiment In this embodiment, the Adam algorithm is used to train and optimize the network. The network is trained based on the Pytorch framework. After multiple experiments, the hyperparameters controlling the decay rates of the first-order and second-order momentum estimates are both 0.9. During the training process, the learning rate is dynamically adjusted according to different stages. Specifically, when epoch is 60, the learning rate is set to . When 40 < epoch < 60, the learning rate is set to . When epoch is 40, the learning rate is set to .

[0062] This embodiment uses Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Normalized Mean Squared Error (NMSE), and Learned Perceptual Image Patch Similarity (LPIPS) for quantitative evaluation, and quantitatively compares the evaluation values ​​of different methods. Table 1 lists the value range and evaluation effect of each objective evaluation index.

[0063] Table 1. Comparison of different model results on underwater image test set metrics.

[0064] The following table lists ECNDNet, DnCNN, DudeNet, UDRN, MWDCNN, and ADNet as denoising models based on convolutional neural networks, and WMAPNet as the convolutional neural network denoising model proposed in this invention. Red data represents the best performance, and blue data represents the next best performance.

[0065] Table 2 Comparison of performance metrics for different models on the underwater image test set.

[0066] Table 3 Comparison of performance metrics for different models on the underwater image test set.

[0067] Table 4 Comparison of different models on the land image test set

[0068] Table 5 Comparison of different models on the land image test set

[0069] Table 6. Running time and parameters of different models in a single training session.

[0070] like Figure 6As shown, this paper compares the visual performance of the proposed WMAPNet model with existing methods (such as ECNDNet, DnCNN, DudeNet, UDRN, MWDCNN, ADNet, etc.) in underwater image denoising under a noise level of 25. The figure demonstrates that WMAPNet effectively suppresses high-frequency noise particles and artifacts while preserving details such as marine life textures and reef edges, while the contrasting methods exhibit varying degrees of noise residue or detail blurring. Figure 7 As shown, another set of underwater image denoising visualization results under a noise level of 25 is provided, further validating the denoising performance of WMAPNet in different underwater scenarios. The figures demonstrate that WMAPNet achieves a better balance between noise suppression and detail preservation when dealing with complex underwater degradation (such as structured noise and color distortion caused by suspended particles), resulting in significantly better visual quality output images than existing methods. Figure 8 As shown, the denoising performance of each method on a standard land image test set is compared under a noise level of 25. The results demonstrate that WMAPNet not only performs excellently in dedicated underwater image denoising tasks but also exhibits good generalization ability in general land image denoising scenarios. It can effectively remove Gaussian noise while preserving image edge and texture details, proving the multi-scenario applicability of the method of this invention.

[0071] In summary, the underwater image denoising model WMAPNet proposed in this invention achieves excellent denoising performance on multiple benchmark datasets through wavelet gating and attention mechanisms, demonstrating good high-frequency detail preservation capabilities and underwater application potential.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for underwater image denoising using a multi-order dynamic wavelet-gated aggregation network, characterized in that, include: S1. Obtain the underwater degradation image to be processed, and preprocess the underwater degradation image to obtain the preprocessed underwater degradation image; S2. Construct an extended context branch. Input the preprocessed underwater degradation image into the extended context branch for multi-scale feature extraction. The extended context branch captures global semantic information and local details of different receptive fields through multiple parallel extended residual blocks to obtain an extended context feature map. S3. Construct an adaptive reuse branch. Input the expanded context feature map into the adaptive reuse branch for cross-layer feature reuse and dynamic weighting. The adaptive reuse branch achieves effective fusion of shallow detail features and deep semantic features through dense connections and dynamic hierarchical gating aggregation mechanism to obtain an adaptive reuse feature map. S4. Construct a wavelet-gated transform branch, input the adaptive reuse feature map into the wavelet-gated transform branch for frequency domain noise separation and adaptive weighting processing. The wavelet-gated transform branch separates noise and effective signal in the frequency domain through multi-level wavelet decomposition and gating mechanism to obtain the wavelet-gated feature map. S5. Construct a dynamic hierarchical gated aggregation module. Input the expanded context feature map, adaptive reuse feature map and wavelet gated feature map into the dynamic hierarchical gated aggregation module for adaptive weighted aggregation and feature reuse, suppress redundant information and enhance effective features to obtain aggregated optimized features. S6. Construct a feature fusion layer and a reconstruction layer, input the aggregated optimized features into the feature fusion layer and the reconstruction layer, and output the denoised target image.

2. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S1 includes: S11. Perform sliding window cropping on the acquired underwater degraded image; S12. Normalize the cropped image patch to the [0,1] value range; S13. Perform data augmentation transformation on the cropped image patches to expand the training dataset.

3. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S2 includes: S21. Construct dilated residual blocks as basic processing units and replace standard convolutions with dilated convolutions; S22. Set up four parallel extended residual blocks, configure different expansion rates for each extended residual block, and the expansion rates are distributed in an increasing manner. S23. The preprocessed underwater degradation image is processed in parallel at multiple scales through the four parallel extended residual blocks to obtain feature representations of different receptive fields, and to perceive local details and global structural background at the same time. S24. Aggregate the features output by the four parallel extended residual blocks to form an extended context feature map.

4. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, In step S22, the dilation rate is distributed in an increasing manner to ensure that the receptive fields are densely interlaced rather than sparsely sampled, effectively mitigating the grid effect introduced by dilated convolution and enabling the network to gradually aggregate contextual information at continuous scales.

5. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S3 includes: S31. Construct dense connection blocks to achieve effective fusion of shallow detailed features and deep semantic features through cross-layer connections; S32. Input the expanded context feature map into the dense connection block to obtain the dense connection features; S33. Construct a dynamic hierarchical gated aggregation block based on the dense connection block, and use a dynamic channel segmentation mechanism to replace the fixed channel segmentation strategy, and adaptively allocate channel resources according to the specific characteristics of the input features. S34. Input the densely connected features into the dynamic hierarchical gating aggregation block, dynamically adjust the weight allocation of each channel, use the dynamic aggregation attention mechanism to adaptively suppress redundant information, improve noise extraction and feature expression capabilities, and output an adaptive reuse feature map.

6. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 5, characterized in that, Step S34 includes: S341. The dynamic hierarchical gated aggregation block performs global average pooling on the densely connected features of the input to obtain channel-level global information. S342. The channel-level global information is mapped to channel attention weights through a fully connected layer, and the channel attention weights are adaptively adjusted according to the degradation characteristics of the input features. S343. Based on the channel attention weights, the channels of the densely connected features are dynamically segmented and weighted to achieve adaptive filtering and aggregation of features. S344. Perform residual connections between the weighted aggregated features and the densely connected features to output the optimized adaptive reuse features.

7. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S4 includes: S41. Construct a wavelet transform module, input the adaptive reuse feature map into the wavelet transform module for multi-level wavelet decomposition, decompose the image into low-frequency sub-band and high-frequency sub-band, the low-frequency sub-band contains color and brightness information, the high-frequency sub-band contains detail and noise information; S42. Construct a low-frequency processing unit, input the low-frequency sub-band into the low-frequency processing unit, and perform color correction and overall brightness adjustment processing. S43. Construct a high-frequency processing unit, input the high-frequency sub-band into the high-frequency processing unit, and perform noise suppression and detail preservation processing; S44. Construct a gating mechanism module to dynamically adjust the weights of each frequency sub-band based on the noise characteristics of the input image; S45. Input the processed low-frequency sub-band and high-frequency sub-band into the gating mechanism module for adaptive weighted fusion; S46. Perform wavelet reconstruction on each frequency sub-band after weighted fusion and output wavelet-gated feature maps.

8. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S5 includes: S51. Construct a feature receiving unit to receive the extended context feature map, the adaptive reuse feature map, and the wavelet-gated feature map; S52. Construct a channel splicing unit to splice the feature maps output by the three paths along the channel dimension to form a comprehensive feature representation; S53. Construct a dynamic weighting unit, and adaptively weight the spliced ​​features through a dynamic hierarchical gating aggregation mechanism to learn the importance weights of each path feature. S54. Construct a feature fusion unit, and perform weighted fusion of the three-path features based on the learned weights to achieve cross-path feature reuse; S55. Construct an attention optimization unit, suppress redundant information through the attention mechanism, strengthen features that are effective for denoising, and output aggregated optimized features.

9. The underwater image denoising method using a multi-order dynamic wavelet-gated aggregation network according to claim 1, characterized in that, Step S6 includes: S61. Input the aggregated optimized features into the feature fusion layer to perform deep fusion of multi-scale features; S62. Input the fused features into the reconstruction layer, and perform image reconstruction through convolution operation to restore the clear image after denoising. S63. Output the denoised target image, which retains fine ocean details while sufficiently suppressing noise.

10. An underwater image denoising system based on the underwater image denoising method of the multi-order dynamic wavelet gated aggregation network according to any one of claims 1-9, characterized in that, include: The preprocessing module is used to acquire the underwater degradation image to be processed and to preprocess the underwater degradation image to obtain the preprocessed underwater degradation image; The extended context branch module is used to extract multi-scale features from the preprocessed underwater degraded image input. It captures global semantic information and local details of different receptive fields through multiple parallel extended residual blocks to obtain an extended context feature map. The adaptive reuse branch module is used to perform cross-layer feature reuse and dynamic weighting processing on the input of the expanded context feature map. Through dense connections and dynamic hierarchical gating aggregation blocks, it achieves effective fusion of shallow detail features and deep semantic features to obtain an adaptive reuse feature map. The wavelet-gated transform branch module is used to perform frequency domain noise separation and adaptive weighting processing on the adaptive reuse feature map input. Through multi-level wavelet decomposition and gating mechanism, noise and effective signal are separated in the frequency domain to obtain the wavelet-gated feature map. The dynamic hierarchical gating aggregation module is used to adaptively weight and aggregate the expanded context feature map, adaptive reuse feature map and wavelet gating feature map into the input, suppress redundant information and enhance effective features to obtain aggregated optimized features. The image reconstruction module is used to fuse and reconstruct the aggregated optimized feature inputs and output the denoised target image.

Images