A method for underwater image quality enhancement based on coupled degradation processing

Abstract

The present invention discloses a method for enhancing underwater image quality based on coupled degradation processing, comprising: obtaining an underwater image dataset; constructing a joint degradation processing model, inputting the underwater image dataset into the joint degradation processing model for training, and obtaining a trained joint degradation processing model; obtaining a target underwater image, inputting the trained joint degradation processing model into the model for underwater image enhancement, and obtaining a target underwater quality-enhanced image. This method addresses the problem of traditional image processing methods being unable to effectively restore image quality when faced with complex coupled degradation problems, enabling underwater robots to efficiently and accurately perceive their environment, thereby improving their operational safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater image enhancement, and in particular relates to an underwater image quality enhancement method based on coupled degradation processing. Background Art

[0002] Underwater image enhancement is a key technology in image processing. Its effectiveness directly impacts the underwater robot's ability to perceive the environment, and thus its operational safety. During underwater operations, the complex, variable, and unpredictable nature of the marine environment—such as light absorption and scattering by water, the concentration of suspended particulate matter, the presence of currents and plankton, and the agitation of sediment during operations—inevitably leads to various degradation phenomena in underwater images. These degradation phenomena are not independent but rather coupled, significantly reducing the quality of underwater images. While many existing methods perform well in specific degraded environments, in real-world scenarios, the complexity of degradation characteristics and the presence of multiple couplings make it difficult for a single method to achieve ideal restoration results. Furthermore, the unique characteristics of the underwater environment make it difficult to effectively improve image quality through manual intervention, which can easily lead to erroneous perception and abnormal judgments of the underwater robot's environment. This not only severely impacts the accuracy of underwater vision but also adversely affects subsequent image analysis and understanding tasks, ultimately threatening the operational safety of the underwater robot. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes an underwater image quality enhancement method based on coupled degradation processing, which enables the underwater robot to perceive the environment efficiently and accurately, thereby improving the operation safety of the underwater robot.

[0004] To achieve the above object, the present invention provides an underwater image quality enhancement method based on coupled degradation processing, comprising: acquiring an underwater image dataset;

[0005] Constructing a joint degradation processing model, inputting the underwater image dataset into the joint degradation processing model for training, and obtaining a trained joint degradation processing model;

[0006] An underwater image to be tested is acquired, and the trained joint degradation processing model is inputted to perform underwater image enhancement to obtain a target underwater quality enhanced image.

[0007] Optionally, the joint degradation processing model includes: a joint feature extraction module, a probability guidance module, a feature post-processing module and a water balance loss function;

[0008] The joint feature extraction module is used to mine the potential information of underwater images and extract coupled potential feature maps through a multi-stage cascade design;

[0009] The probability guidance module is used to adjust the statistical properties of the coupled potential feature map, match the training and weight feature distributions, and output features;

[0010] The feature post-processing module is used to optimize the features by adopting grayscale correction and boundary enhancement strategies;

[0011] The water balance loss function is used to constrain the entire training process of the joint degradation processing model.

[0012] Optionally, the joint feature extraction module includes: a feature extraction unit, a first exponential linear unit-based compression and excitation residual unit, and an upsampling unit;

[0013] The feature extraction unit is used to perform nonlinear processing on the input underwater image to extract initial underwater image features;

[0014] The first exponential linear unit-based compression and excitation residual unit is used to enhance the coupled latent feature representation through dynamic weight allocation and smooth nonlinear transformation according to the initial underwater image features;

[0015] The upsampling unit is used to restore the spatial scale of the coupled potential feature map.

[0016] Optionally, the probability guidance module includes: a probability generation unit, an adaptive instance normalization unit, and a second compression and excitation residual unit based on an exponential linear unit;

[0017] The probability generation unit is used to extract the required mean and standard deviation distribution from the coupled potential feature map after restoration of the scale;

[0018] The adaptive instance normalization unit is used to adjust the feature map of the content image using the mean and variance of the target style image;

[0019] The second compression and excitation residual unit based on the exponential linear unit is used to optimize and dynamically adjust the required feature distribution.

[0020] Optionally, the feature post-processing module includes a grayscale correction component and a boundary enhancement mask strategy;

[0021] The grayscale correction component is used to correct the model's over-adjustment of the image, adjust the image's color balance, and eliminate unnatural color casts;

[0022] The boundary enhancement mask strategy is used to enhance detail information through Gaussian adjustment, reduce artifacts, and eliminate losses caused by boundary deviations.

[0023] Optionally, the water area balance loss function is calculated as:

[0024] AnquBalanceLoss=|ABL out -ABL in +λ imp | 2

[0025] Among them, ABL out Indicates the value of the output image after relevant calculations during the training process; ABL in Represents the value of the input image after similar calculations during the training process; λ imp Indicates the amount of deviation.

[0026] Optionally, the feature extraction unit consists of a first convolutional layer, a first activation function, a random dropout layer, and a pooling layer;

[0027] The first convolutional layer extracts local features of the input underwater image to obtain a primary feature map, the first activation function is used to introduce nonlinear characteristics, the random inactivation layer is used to prevent overfitting, and the primary feature map is further downsampled through the pooling layer to extract initial underwater image features.

[0028] Optionally, the first exponential linear unit-based compression and excitation residual unit includes reflection padding, a second convolutional layer, a second activation function, an ESE layer and a residual connection, specifically including:

[0029] The first exponential linear unit-based compression and excitation residual unit uses reflection filling to mirror the edge of the feature map to ensure that the convolution operation does not introduce boundary artifacts;

[0030] The second convolutional layer and the second activation function are used to extract features, and channel attention adjustment is performed through the ESE layer;

[0031] The ESE layer extracts global channel features through adaptive average pooling and performs nonlinear transformation using two of the second convolutional layers and the second activation function;

[0032] Channel weighting is achieved through Sigmoid function;

[0033] The channel attention mechanism dynamically weights the feature contributions of different channels according to the global information of each channel, and further strengthens the feature expression through the residual connection.

[0034] Optionally, the upsampling unit includes a third convolutional layer, a third activation function and an upsampling layer;

[0035] The feature map is processed by the third convolutional layer and the third activation function, and the spatial size of the feature map is increased by using the upsampling layer to restore the feature space resolution.

[0036] Technical Effects of the Invention: This invention discloses a method for underwater image quality enhancement based on coupled degradation processing. This method utilizes a joint feature extraction module to mine potential information from various coupled degradations, capturing the complexity of underwater images. A probability guidance module adjusts the overall statistical properties of underwater images to match the distribution of training features. A feature post-processing module optimizes and enhances the framework's output features, reducing information loss in edge regions and boundary deviation during inference, thereby enhancing boundary clarity and detail while maintaining overall image quality. A water balance loss is implemented and, based on the supervision of this loss function, a model is trained on the publicly available underwater dataset, UIEB, ensuring that the model effectively learns the features of various coupled degradations and achieves a balance between color, clarity, and contrast. The model's performance is tested on various underwater scene datasets to demonstrate its image enhancement effects. This method achieves underwater image quality enhancement in complex coupled degradation scenarios. This invention effectively handles coupled degradation and exhibits good generalization performance, enabling underwater robots to achieve efficient and accurate environmental perception, thereby improving their operational safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0038] Figure 1 This is a flow chart of a method for enhancing underwater image quality based on coupled degradation processing according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the structure of the joint degradation processing model according to an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of a joint feature extraction module according to an embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram showing the effectiveness visualization results of the joint degradation processing model according to an embodiment of the present invention on the U45 dataset;

[0042] Figure 5 This is a schematic diagram of the visual comparison results of the joint degradation processing model of the present invention and other SOTA methods on the UIEB dataset. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] like Figure 1 As shown, this embodiment provides an underwater image quality enhancement method based on coupled degradation processing, including: acquiring an underwater image dataset;

[0046] Construct a joint degradation processing model, input the underwater image dataset into the joint degradation processing model for training, and obtain a trained joint degradation processing model;

[0047] Obtain the underwater image to be tested, input the trained joint degradation processing model to perform underwater image enhancement, and obtain the target underwater quality enhanced image.

[0048] Furthermore, if Figure 2 The joint degradation processing model shown includes: a joint feature extraction module, a probability guidance module, a feature post-processing module and an aquabalance loss function (AquaBalanceLoss);

[0049] Joint feature extraction module, which is used to mine the potential information of underwater images and extract coupled latent feature maps through a multi-stage cascade design;

[0050] The probability guidance module is used to adjust the statistical properties of the coupled latent feature map, match the training feature distribution, and output features;

[0051] Feature post-processing module, used to optimize features using grayscale correction and boundary enhancement strategies;

[0052] AquaBalanceLoss is a function that is used to jointly handle the constraints of the entire training process of the degradation processing model.

[0053] Specifically, the input underwater image is first processed through a multi-layer feature extraction module to extract preliminary multi-scale features. After feature extraction, the features are dynamically weighted and enhanced using an exponential linear unit-based compression and excitation residual unit (ESE-ResBlock). Following joint feature extraction, the processed features are further subjected to probability estimation and distribution adjustment to ensure stable image quality. Adaptive normalization is used to enhance style and content adaptability. Finally, the image is post-processed using a grayscale correction block and a boundary enhancement mask to optimize the image's grayscale and edge information, reduce artifacts, and enhance edge detail, ultimately generating high-quality underwater images.

[0054] First, the present invention introduces a multi-stage cascade design in the Joint Feature Extraction Module (JFEModule) to mine potential information from various coupled degradations. This encoding module accounts for the influence of factors such as color, illumination, and scattering, capturing the complexity of underwater images. These feature maps are then passed through a Probabilistic Bootstrap Module (PB Module), whose statistical properties are adjusted to match the distribution of the training features. To further optimize and enhance the output features of the framework, the present invention designs a Feature Post-Processing Module (FPP Module). The Feature Post-Processing Module employs grayscale correction and a Border Enhancement Mask (BEM) strategy to reduce information loss in edge regions, address boundary deviation during inference, and enhance boundary clarity and detail while maintaining overall image quality. Finally, to balance color, clarity, and contrast and ensure that the model can effectively learn features from various coupled degradations, the present invention proposes a novel AquaBalanceLoss loss function. This loss function supervises the model's learning process and enhances its feature inference capabilities.

[0055] Furthermore, if Figure 3 The joint feature extraction module includes a feature extraction module, a first exponential linear unit based compression and excitation residual unit (ESE-ResBlock) and an upsampling unit (UpBlock);

[0056] A feature extraction module, used for introducing nonlinearity into the input underwater image to extract initial underwater image features;

[0057] The first one is an exponential linear unit based compression and excitation residual unit (ESE-ResBlock), which is used to extract the coupled latent feature map according to the initial underwater image features through dynamic weight allocation and smooth nonlinear transformation;

[0058] Upsampling unit, used to restore the scale of the coupled latent feature map.

[0059] Furthermore, the feature extraction unit includes a first convolutional layer (Conv), a first activation function (ELU), a random dropout layer (Dropout) and a maximum pooling layer (MaxPooling).

[0060] The input underwater image first passes through the first convolutional layer for feature extraction, obtaining a preliminary feature representation. The first activation function then introduces nonlinearity, enabling the model to capture complex nonlinear relationships. A dropout layer prevents overfitting and enhances the model's generalization when processing complex images. A second convolution layer and the first activation function extract further feature representations. Finally, a max pooling layer reduces the spatial size of the feature map while retaining the most representative features for subsequent processing.

[0061] Furthermore, the first exponential linear unit-based compression and excitation residual unit (ESE-ResBlock) includes a reflection padding (ReflectionPad), a second convolutional layer (Conv), a second activation function (ELU), an ESE layer and a residual connection. First, the first exponential linear unit-based compression and excitation residual unit (ESE-ResBlock) structure uses reflection padding to mirror the edges of the feature map to ensure that the convolution operation does not introduce boundary artifacts. Then, the second convolutional layer and the second activation function (ELU) are used to extract features, and channel attention adjustment is performed through the ESE layer. The ESE layer extracts global channel features through adaptive average pooling, and uses two second convolutional layers and the second activation function (ELU) for nonlinear transformation, and finally implements channel weighting through the Sigmoid function. The channel attention mechanism dynamically weights the feature contributions of different channels according to the global information of each channel, and further enhances the feature expression through residual connections.

[0062] Furthermore, the upsampling unit includes a third convolutional layer (Conv), a third activation function (ELU) and an upsampling layer (Upsampling).

[0063] The upsampling unit restores the spatial resolution of the feature map through upsampling. First, the feature map is processed through the third convolutional layer and the third activation function (ELU). The upsampling layer increases the spatial size of the feature map and restores the spatial resolution of the feature map to support subsequent modules.

[0064] Specifically, the present invention proposes a joint feature extraction module, which adopts an encoder-decoder structure and consists of a feature extraction module (FE Module) and a first exponential linear unit based compression and excitation residual unit (ESE-ResBlock), aiming to extract potential features from the input image. Figure 3As shown in Figure 2, by introducing nonlinearity and overfitting prevention measures, the model's learning and generalization capabilities are enhanced. This lays the foundation for the JFE module's ability to effectively extract and abstract input data features. In order to more efficiently utilize model parameters, enhance the model's adaptive feature recalibration capability, and achieve the extraction of key coupled degradation features and effective suppression of irrelevant features, as shown in Figure 2. Figure 2 As shown, the present invention proposes a first compression and excitation residual unit (ESE-ResBlock) based on an exponential linear unit. This module improves the representation ability of the model by modeling channel dependencies. This design enhances the model's focus on key features by recalibrating the convolution response and suppressing irrelevant features, guiding the model to prioritize restoring important features of multiple coupling degradations. The first compression and excitation residual unit (ESE-ResBlock) based on an exponential linear unit enhances the model's ability to model nonlinear distortions of underwater images by introducing nonlinearity, making the capture of nonlinear features more effective. In addition, it will also further enhance the model's adaptability and learning ability to image features under various environmental conditions. Finally, the first compression and excitation residual unit (ESE-ResBlock) based on an exponential linear unit realizes dynamic weight allocation and smooth nonlinear transformation, which improves the adaptability of the model, enabling it to handle multiple coupling degradations and correct nonlinear phenomena in underwater images, ultimately improving the generalization performance of the model in diverse underwater environments. The present invention combines the first exponential linear unit-based compression and excitation residual unit (ESE-ResBlock) with the FE module to ensure the smooth flow of gradients, alleviate the problem of gradient vanishing during training, and improve the learning speed and stability of the model.

[0065] Furthermore, if Figure 3 The probability guidance module shown includes a probability generation unit (PG Module), an adaptive instance normalization unit (AdaIN), and a second exponential linear unit-based compression and excitation residual unit (ESE-ResBlock);

[0066] A probability generation unit is used to extract the required mean and standard deviation distribution according to the latent feature map after the scale is restored;

[0067] Adaptive Instance Normalization (AdaIN) unit, which is used to construct the required Gaussian distribution based on the desired mean and standard deviation distribution;

[0068] The second exponential linear unit based compression and excitation residual unit (ESE-ResBlock) is used to realign the required Gaussian distribution.

[0069] Specifically, the probability-guided module adaptively optimizes the image's color, contrast, and macrostructure through style adjustment and feature recalibration, while preserving the scene structure of the original content. The present invention combines a second exponential linear unit-based compression and excitation residual unit (ESE-ResBlock) with an adaptive instance normalization (AdaIN) unit to enhance the model's flexibility in style adjustment and content preservation in multiple coupled degradation environments.

[0070] In the probability guidance module proposed in this paper, in order to further optimize the input features, a Probability Generation Module (PG Module) is designed to construct the required mean and standard deviation distribution. The feature map output by the joint feature extraction module (JFE Module) is set as the data matrix F1. Where B, C, H, and W represent the batch size, number of channels, height, and width, respectively.

[0071] First, calculate the mean μ of each channel f from the obtained feature map data matrix F1 f and standard deviation σ f ,get:

[0072]

[0073] and

[0074]

[0075] Among them, f ij Represents the pixel value at a specific position in channel f.

[0076] Afterwards, the invention uses a 1×1 convolution kernel to extract features from the mean vector, and the result is:

[0077]

[0078] and

[0079]

[0080] a represents the feature extracted from the mean vector through the convolution layer, and m represents the feature extracted from the standard deviation vector through the convolution layer.

[0081] The present invention also uses a 1×1 convolution kernel to extract features from the standard deviation vector to obtain the mean and standard deviation

[0082] Next, the present invention extends the application of Gaussian distribution to the input optimization and adjustment of the Adaptive Instance Normalization (AdaIN) unit. The present invention uses the extracted mean a and standard deviation m parameters to construct an n-dimensional Gaussian distribution Expressed as:

[0083]

[0084] Used to adjust the mean of the data.

[0085] Similarly, the present invention uses b and n to construct another n-dimensional Gaussian distribution Expressed as

[0086]

[0087] The standard deviation used to adjust the data.

[0088] Finally, the present invention starts from N mean and N std The mean and standard deviation are extracted from the Gaussian distribution and reused as the optimized unique mean and standard deviation to form the input mean μ of AdaIN optimal :

[0089] μ optimal =ExtractMean(N mean ) (7)

[0090] and standard deviation σ optimal :

[0091] σ optimal =ExtractStd(N std ) (8)

[0092] ExtractMean() and ExtractStd() respectively represent the operations of extracting the mean and standard deviation from the Gaussian distribution.

[0093] After global adjustment, the integrated second exponential linear unit-based compression and excitation residual unit (ESE-ResBlock) continues to guide the model to focus on key features, enhance the consistency of the adaptive instance normalization (AdaIN) effect, and improve the model's adaptability to different underwater degradation conditions.

[0094] Furthermore, the feature post-processing module includes a grayscale correction component and a boundary enhancement mask strategy;

[0095] Grayscale correction component, used to correct the model's over-adjustment of the image;

[0096] A boundary enhancement mask strategy is used to enhance detail information through Gaussian adjustment and eliminate artifacts and boundary deviation losses.

[0097] Specifically, such as Figure 2 As shown in the figure, to reduce information loss in edge regions, lower boundary deviation during inference, and enhance texture details, a Feature Post-Processing Module (FPP Module) was invented. This module further enhances feature representation based on the existing feature maps. This module consists of a grayscale correction component and a boundary enhancement mask (BEM) strategy. The grayscale correction module aims to correct the problem of over-scaling the image in the first few layers of the network, while the boundary enhancement mask method enhances detail information through Gaussian adjustment, addressing the artifacts and boundary deviation loss encountered during model inference.

[0098] Assume that the feature map obtained by the probability generation unit is F2, which enters the feature post-processing module and is processed by white balance to obtain F3. A Gaussian filter is applied to F3 to extract its low-frequency component, which is recorded as F low :

[0099] F low =G(F3,ω) (9)

[0100] Where G represents the Gaussian filter, and ω represents the balance parameter of the Gaussian filter.

[0101] The boundary enhancement mask is denoted as F BEM :

[0102] F BEM =F3-F low +λ (10)

[0103] Among them, λ is used to maintain the strength of the boundary enhancement mask BEM. Finally, the boundary enhancement mask BEM is merged with F3 to obtain the adjusted feature map F P :

[0104]

[0105] After convolution adjustment, the final module output is obtained as the final output image of the model, that is, the enhanced underwater image. At this point, the model inference process is completed.

[0106] Furthermore, the water balance loss function (AquaBalanceLoss) is calculated as:

[0107] AnquBalanceLoss=|ABL out -ABL in +λ imp | 2 (12)

[0108] Among them, ABL out Indicates the value of the output image after relevant calculation, ABL inRepresents the value of the input image after similar calculations. imp Indicates the deviation. ABL consists of three parts: color index (L coi ), clarity index (L si ) and contrast index (L cti The basic expression of ABL is as follows:

[0109] ABL=c1×L coi +c2×L si +c3×L cti (13)

[0110] Where c1, c2, and c3 represent the weights assigned to each attribute, c1 = 0.029, c2 = 0.295, and c3 = 3.550.

[0111] For L coi , the formula is as follows:

[0112] L coi =-0.027×l+0.159×r (14)

[0113] Among them, the mean correlation l and variance correlation r can be expressed as:

[0114]

[0115] and

[0116]

[0117] Among them, μ α (·) represents the alpha trimmed mean of pixel values, s α (·) represents the variance of the pixel value after alpha clipping is applied, that is, the sum of squared deviations from the alpha clipping mean divided by the number of pixels. G And the difference between the fourth and third color components Y B It can be obtained by the following formula:

[0118] R G =RG (17)

[0119] and

[0120]

[0121] Among them, R, G, and B represent the first color, second color, and third color channel values of the image respectively.

[0122] For L si , which evaluates the clarity of an image by calculating the weighted sum of the edge maps obtained after applying Sobel edge detection to the image. The formula is as follows:

[0123]

[0124] Among them EME i Represents the result of applying Sobel edge detection to a channel and computing the channel edge map (EME) strength. i Indicates the weighting coefficient of the channel. c indicates the number of channels.

[0125] For L cti , which evaluates the contrast of underwater images by calculating the contrast within image blocks. The formula is:

[0126]

[0127] Where α is a parameter used to adjust the entropy scale. Increasing α can enhance the randomness of the evaluation. top is the difference between the maximum and minimum values of each image block, indicating the contrast range of the image block. It is calculated as:

[0128] top=max-min (21)

[0129] bot is the sum of the maximum and minimum values of each image block, which is used to normalize the contrast range. It is calculated as:

[0130] bot=max+min (22)

[0131] Among them, max represents the maximum value in each image block, and min represents the minimum value of the image block.

[0132] The contrast weight coefficient ω is defined as:

[0133]

[0134] Used to perform weighted averaging of the contrast scores of each block. k1 and k2 are the number of blocks after the image is divided by the window in the height and width directions, respectively.

[0135] Among them, H loss Indicates the height of the tensor when calculating the loss, W loss Indicates the width of the tensor when calculating the loss, and win_size is the partition window size parameter.

[0136] Combined with other loss functions, the total loss function of the model is:

[0137] Loss = λ1 × loss vgg +λ2×loss kl +λ3×loss re +λ4×AnquBalanceLoss (24)

[0138] Among them, λ1, λ2, λ3, and λ4 are weight coefficients, which are 0.025, 1, 0.1, and 0.1, respectively. vgg is the vgg loss function, loss kl is the divergence loss, loss re For reconstruction loss, AnquBalanceLoss is the water balance loss function proposed in this invention.

[0139] loss vgg The expression is:

[0140]

[0141] Among them, φ l (·) represents the feature map of the lth layer. N l Represents the size of the feature map of the lth layer, I(output) represents the output image generated by the model, I(target) represents the target image, and L is the total number of feature map layers.

[0142] loss kl The expression is:

[0143]

[0144] Among them, q(z|x) is the posterior distribution, which represents the distribution of the latent variable z under the condition of a given input x, and p(z) is the prior distribution, which is set to the standard normal distribution. represents the difference measure between q(z|x) and p(z), and ∫(·)dz represents the integration of the latent variable z.

[0145] loss re The expression is:

[0146]

[0147] like Figure 4 The figure below demonstrates the effectiveness of our model in a coupled degradation scenario. After color correction, our model effectively handles forward scattering, avoiding a hazy appearance. Similarly, the image below, after color correction, further enhances brightness and texture detail.

[0148] like Figure 5Figure 2 shows a visual comparison of our method with other state-of-the-art methods on the UIEB dataset. The visual comparison of different methods includes, from left to right, HS2CM2A, Shallow-UWnet, PUIE (MC, MP), FiveAPlus, CWR, Spectroformer, NU2Net, and our method, JDPNet. In the first and fifth rows, our method not only corrects color but also successfully adjusts brightness and eliminates scattering effects. Other methods correct for color cast but underperform on scattering effects. In the second and third rows, our method achieves excellent descattering and performs exceptionally well, with processed images displaying improved clarity and saturation, surpassing GT. In the last row, after color correction, other methods introduce varying degrees of distortion in biological details, while our method effectively preserves these details. In the sixth row, after brightness enhancement, our method eliminates underwater scattering effects and corrects for secondary color hues, resulting in improved visual distance. This demonstrates the model's ability to effectively handle environments with multiple coupled degradations. In addition, the proposed method has good detail post-processing and shows excellent detail preservation ability in all images.

[0149] Table 1 shows a comparison of JDPNet with other state-of-the-art methods on the UIEB dataset. We selected commonly used quantitative metrics to evaluate image quality. These metrics include peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), underwater image quality measure (UIQM), and underwater color image quality evaluation (UCIQE). PSNR and SSIM assess overall image quality and visual similarity, while UIQM and UCIQE specifically address the unique quality issues of underwater images.

[0150] Table 1

[0151]

[0152]

[0153] comprehensive Figure 5 From the qualitative visual comparison, it can be seen that the proposed method is the best performing model, demonstrating its superiority in different coupling degradation scenarios.

[0154] The method of the present invention shows the best performance on multiple datasets, including UIEB, EUVP, U45, C60 and UCCS. On the UIEB dataset, the method of the present invention achieved the highest performance in PSNR, SSIM, UIQM and UCIQE indicators, which preliminarily demonstrated the superiority of the method of the present invention. Due to the poor quality of some GT data, the method of the present invention does not strictly pursue consistency with the true value, and therefore fails to achieve overwhelmingly high scores in PSNR and SSIM. On the large-scale EUVP dataset, the method of the present invention achieved the highest scores in PSNR, UIQM and UCIQE indicators, and the second highest score in SSIM, demonstrating its strong generalization ability. On the other three datasets, the method of the present invention is significantly ahead in the UIQM score, indicating that an excellent balance is achieved between color, contrast and sharpness. In the highly challenging C60 dataset, the method of the present invention performed excellently, demonstrating its outstanding ability to handle complex coupled degradations and difficult scenes. This success is attributed to the design of the overall framework of the present invention and the effectiveness of the novel loss function.

[0155] This invention discloses a method for underwater image quality enhancement based on coupled degradation processing. It utilizes a joint feature extraction module to mine potential information from various coupled degradations, capturing the complexity of underwater images. A probability guidance module adjusts the overall statistical properties of underwater images to match the distribution of training features. A feature post-processing module optimizes and enhances the framework's output features, reducing information loss in edge regions and boundary deviation during inference, enhancing boundary clarity and detail while maintaining overall image quality. A water balance loss is implemented and, supervised by this loss function, the model is trained on the publicly available underwater dataset, UIEB, to ensure that the model effectively learns the features of various coupled degradations and achieves a balance between color, clarity, and contrast. The model's performance is tested on various underwater scene datasets to demonstrate its image enhancement effects. This method achieves underwater image quality enhancement in complex coupled degradation scenarios. This invention effectively handles coupled degradation and exhibits good generalization performance, enabling underwater robots to achieve efficient and accurate environmental perception, thereby improving their operational safety.

[0156] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for underwater image quality enhancement based on coupled degradation processing, characterized in that: include: Obtain underwater image dataset; Constructing a joint degradation processing model, inputting the underwater image dataset into the joint degradation processing model for training, and obtaining a trained joint degradation processing model; Acquire an underwater image to be tested, input the trained joint degradation processing model to perform underwater image enhancement, and obtain a target underwater quality-enhanced image; The joint degradation processing model includes: a joint feature extraction module, a probability guidance module, a feature post-processing module and a water balance loss function; The joint feature extraction module is used to mine the potential information of underwater images and extract coupled potential feature maps through a multi-stage cascade design; The probability guidance module is used to adjust the statistical properties of the coupled potential feature map, match the training and weight feature distributions, and output features; The feature post-processing module is used to optimize the features by adopting grayscale correction and boundary enhancement strategies; The water balance loss function is used to constrain the entire training process of the joint degradation processing model; The joint feature extraction module includes: a feature extraction unit, a first compression and excitation residual unit based on an exponential linear unit, and an upsampling unit; The feature extraction unit is used to perform nonlinear processing on the input underwater image to extract initial underwater image features; The first exponential linear unit-based compression and excitation residual unit is used to enhance the coupled latent feature representation through dynamic weight allocation and smooth nonlinear transformation according to the initial underwater image features; The upsampling unit is used to restore the spatial scale of the coupled latent feature map; The probability guidance module includes: a probability generation unit, an adaptive instance normalization unit and a second compression and excitation residual unit based on an exponential linear unit; The probability generation unit is used to extract the required mean and standard deviation distribution from the coupled potential feature map after restoration of the scale; The adaptive instance normalization unit is used to adjust the feature map of the content image using the mean and variance of the target style image; The second compression and excitation residual unit based on the exponential linear unit is used to optimize and dynamically adjust the required feature distribution; The water balance loss function is calculated as: AnquBalanceLoss=|ABL out -ABL in +λ imp | 2 Among them, ABL out Indicates the value of the output image after relevant calculations during the training process; ABL in Represents the value of the input image after similar calculations during the training process; λ imp Indicates the deviation; ABL indicates the value of the image; ABL consists of color index, clarity index and contrast index.

2. The underwater image quality enhancement method based on coupled degradation processing according to claim 1, characterized in that: The feature post-processing module includes a grayscale correction component and a boundary enhancement mask strategy; The grayscale correction component is used to correct the model's over-adjustment of the image, adjust the image's color balance, and eliminate unnatural color casts; The boundary enhancement mask strategy is used to enhance detail information through Gaussian adjustment, reduce artifacts, and eliminate losses caused by boundary deviations.

3. The underwater image quality enhancement method based on coupled degradation processing according to claim 1, characterized in that: The feature extraction unit consists of a first convolutional layer, a first activation function, a random dropout layer and a pooling layer; The first convolutional layer extracts local features of the input underwater image to obtain a primary feature map, the first activation function is used to introduce nonlinear characteristics, the random inactivation layer is used to prevent overfitting, and the primary feature map is further downsampled through the pooling layer to extract initial underwater image features.

4. The underwater image quality enhancement method based on coupled degradation processing according to claim 1, characterized in that: The first exponential linear unit-based compression and excitation residual unit includes reflection padding, a second convolutional layer, a second activation function, an ESE layer and a residual connection, specifically including: The first exponential linear unit-based compression and excitation residual unit uses reflection filling to mirror the edge of the feature map to ensure that the convolution operation does not introduce boundary artifacts; The second convolutional layer and the second activation function are used to extract features, and channel attention adjustment is performed through the ESE layer; The ESE layer extracts global channel features through adaptive average pooling and performs nonlinear transformation using two of the second convolutional layers and the second activation function; Channel weighting is achieved through Sigmoid function; The channel attention mechanism dynamically weights the feature contributions of different channels according to the global information of each channel, and further strengthens the feature expression through the residual connection.

5. The underwater image quality enhancement method based on coupled degradation processing according to claim 1, characterized in that: The upsampling unit includes a third convolutional layer, a third activation function and an upsampling layer; The feature map is processed by the third convolutional layer and the third activation function, and the spatial size of the feature map is increased by using the upsampling layer. Restore feature space resolution.

Images