Adaptive underwater image enhancement algorithm

Abstract

The invention relates to the technical field of image processing, and discloses an adaptive underwater image enhancement algorithm, and the algorithm specifically comprises the following steps: S10, constructing a network architecture: designing a conditional diffusion model as a main frame of image enhancement, the model comprises a forward noise adding process and a reverse denoising process, and in the denoising process, the forward noise adding process and the reverse denoising process are carried out; a denoising network composed of a Transform block and an improved residual state space block RSSB is introduced for encoding and refining image features, the RSSB is combined with a discrete state space equation of Mamba, a long dependency relationship in an image is modeled in a recursive form, channel redundancy is avoided, a visual state space module VSSM is introduced, important pixel information in the image is captured in a self-adaptive mode, and the image quality is improved. And the modeling capability of spatial details is enhanced. According to the adaptive underwater image enhancement algorithm, the influence of underwater environment complexity on image quality is overcome, and the image contrast, the color rendition degree and the detail retention capability are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an adaptive underwater image enhancement algorithm. Background Technology

[0002] Underwater image enhancement, as a cutting-edge exploration in computer vision and underwater robotics, faces numerous challenges. The unique characteristics of the underwater environment, such as the strong scattering and absorption of light, lead to decreased image contrast and blurred details. At the same time, the differences in light absorption by water at different wavelengths cause color distortion, resulting in unnatural blue-green tones in underwater images. These problems greatly limit the application potential of underwater images in fields such as marine monitoring, underwater archaeology, and biological research. Traditional image enhancement methods, such as histogram equalization and filtering denoising, can improve image quality to some extent, but they often fall short in the face of the complexity and variability of the underwater environment. These methods struggle to accurately restore the true colors of the image, improve contrast, and retain detail information, thus failing to meet the demand for high-quality underwater images in practical applications.

[0003] In recent years, with the rapid development of deep learning technology, deep learning-based image enhancement methods, especially conditional diffusion models, have attracted widespread attention in the field of image generation and enhancement due to their flexibility and powerful modeling capabilities. Conditional diffusion models simulate the gradual addition of noise to image data during the diffusion process until it is completely destroyed, and then learn the inverse process to remove noise and restore the image, thus achieving effective editing and enhancement of image content. However, when the original diffusion model is directly applied to underwater image enhancement tasks, the following defects are found: the underwater environment is complex and variable, and phenomena such as light scattering, absorption, and color distortion have a serious impact on image quality. When processing such highly degraded images, the original diffusion model often fails to accurately capture and restore the key information in the image. During the reverse diffusion process, due to the lack of prior knowledge about the characteristics of the underwater environment, the model is prone to generating multiple possible image solutions, which increases the uncertainty of the enhancement results. This uncertainty not only affects the visual effect of the enhanced image, but also limits its reliability in subsequent analysis and processing. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] The purpose of this invention is to overcome the impact of the complexity of the underwater environment on image quality and improve image contrast, color reproduction and detail preservation capabilities, and to propose an adaptive underwater image enhancement algorithm.

[0006] (II) Technical Solution

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0008] An adaptive underwater image enhancement algorithm specifically includes the following steps:

[0009] S10. Constructing the network architecture: Design a conditional diffusion model as the main framework for image enhancement. This model includes a forward noise addition process and a reverse denoising process. In the denoising process, a denoising network composed of Transformer blocks and improved residual state space blocks RSSB is introduced to encode and refine image features. RSSB combines Mamba's discrete state space equations to recursively model long dependencies in the image, avoiding channel redundancy. A visual state space module VSSM is introduced to adaptively capture important pixel information in the image and enhance the ability to model spatial details.

[0010] S20. Model Training: Prepare a large-scale paired dataset containing the original underwater images and their corresponding enhanced images for model training and testing. Modify the number of image feature channels to adapt to the input requirements of the residual state space module to ensure that the model can be trained normally. Use a hybrid loss function to optimize the network and provide variance learning signals to improve the model training effect and image enhancement quality. During the training process, use conditional images such as reference clear images or simulated underwater images as auxiliary inputs to guide the model to generate specific enhanced images corresponding to the input images.

[0011] S30. Image Enhancement Process: Input a low-quality underwater image and a conditional image, set a time step, and gradually add Gaussian noise to the image through the forward process of the conditional diffusion model until the image becomes pure noise. In the reverse process, the trained denoising network is used to gradually remove noise and generate a clear enhanced image. The Transformer block in the denoising network uses an attention mechanism to capture global dependencies, the RSSB block further refines and encodes image features, and the VSSM module adaptively captures key pixel information.

[0012] Based on the above technical solution, the present invention can be further improved as follows.

[0013] Furthermore, in the forward noise addition process in step S10, clear image data is gradually injected with noise until the data becomes pure noise. Each step of noise injection can be regarded as a small Wiener process, in which the addition of noise is carried out step by step to ensure a smooth transition of data. The independent incremental characteristic of the Wiener process ensures that the noise added in each step is independent, and the past noise state will not affect the future noise state. The increment from time t to t+Δt follows a normal distribution with a center of 0 and a standard deviation of Δt, which ensures the randomness and controllability of noise.

[0014] Furthermore, in the reverse denoising process, the denoising process starts from the noisy state and uses a denoising network composed of Transformer blocks and an improved RSSB to gradually restore the original state of the data. The Transformer block is used to capture global dependencies in the image and encode image features through a self-attention mechanism. The improved RSSB combines the discrete state-space equation of Mamba to recursively model long dependencies in the image, avoiding channel redundancy. The RSSB enhances the capabilities of the original Mamba through local spatial priors and channel interactions, generating feature representations specific to the restoration, adaptively capturing important pixel information in the image, and enhancing the ability to model spatial details. By capturing long-term spatial dependencies, the denoising network helps the denoising network to more accurately restore image details. In the reverse process, the denoising network gradually reduces noise in the data through iteration. At each step, the trained model is used to predict the best method for removing noise until clear data is restored.

[0015] Furthermore, the trained model minimizes the difference between the generated data and the original data, uses mean squared error (MSE) or log-likelihood loss as the training objective, selects an appropriate loss function based on the specific characteristics of the data and the noise model to ensure that the model can accurately recover image details, employs optimization algorithms such as gradient descent to train the model, and updates the network parameters through backpropagation to improve the model's denoising ability.

[0016] Furthermore, in step S20, a large-scale paired dataset containing the original underwater images and their corresponding enhanced images is prepared. These images should cover different underwater environments, lighting conditions, object types, etc., to ensure the diversity and representativeness of the dataset. The data is preprocessed to remove low-quality images such as noise, blur, and duplication. If the dataset requires annotation for tasks such as object detection and segmentation, the corresponding annotation work is carried out. According to the input requirements of the Residual State Space Module (RSSB), the number of image feature channels is modified. This usually involves converting the image from the common RGB three-channel to other channel numbers, or generating input features suitable for RSSB through specific preprocessing steps such as color space conversion and feature extraction.

[0017] Furthermore, the model training is based on the Conditional Diffusion Model framework, constructing a denoising network containing Transformer blocks, an improved RSSB, and a Visual State Space Module (VSSM). This ensures the network structure can handle the adjusted number of image feature channels. A hybrid loss function is used to optimize the network, which can include multiple components, such as pixel-level losses like L1 or L2 losses to measure the pixel differences between the generated and real images, perceptual losses like feature matching losses to measure the similarity between the generated and real images in the high-level feature space, and adversarial losses like the discriminator loss in GANs to improve the realism of the generated images. These loss functions collectively provide a variance learning signal to improve model training performance and image enhancement quality. A suitable optimizer, such as Adam or SGD, is selected to update the network parameters. The choice of optimizer should consider training speed, convergence stability, and hyperparameters. To mitigate factors such as number sensitivity, the dataset is divided into training, validation, and test sets during training. The training set is used for model training, the validation set is used for adjusting hyperparameters and monitoring the training process, and the test set is used for final model performance evaluation. During training, conditional images, such as reference clear images or simulated underwater images, are used as auxiliary inputs. These conditional images can provide additional information or constraints to the model, guiding it to generate specific augmented images corresponding to the input images. The network parameters are optimized through multiple iterations of training. In each iteration, a batch of images is randomly selected from the training set to compare the original and augmented images. The loss value is calculated through forward propagation, and then the network parameters are updated through backpropagation. The model's performance on the validation set is evaluated periodically during training. If the model's performance on the validation set begins to decline, indicating overfitting, appropriate measures are taken, such as stopping training early, reducing the learning rate, and increasing regularization, to prevent overfitting.

[0018] Furthermore, the performance of the trained model is evaluated using a test set. Evaluation metrics may include objective metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), as well as subjective visual quality evaluation. The model is adjusted based on the evaluation results. If the model performance does not meet the requirements, adjustments may be made to the network structure, loss function, optimizer, or training strategy.

[0019] Furthermore, the process involves using low-quality underwater images to be enhanced as initial inputs. These images may be affected by factors such as insufficient light, water scattering, and color distortion, leading to a decline in image quality. Simultaneously, conditional images, such as reference clear images or simulated underwater images, are input. These conditional images provide additional information or constraints to the model, helping to generate enhanced images that better meet expectations. An appropriate time step is set according to requirements. The time step determines the rate and extent of adding Gaussian noise to the image during the forward process. Through the forward process of the conditional diffusion model, Gaussian noise is gradually added to the image, starting from the initial low-quality underwater image. As the time step increases, useful information in the image is gradually masked by noise until the image becomes pure noise.

[0020] Furthermore, in the reverse process, a pre-trained denoising network is used to gradually remove noise from the image. By learning the mapping relationship between low-quality and high-quality images, the denoising network can accurately predict and remove noise, restoring clear image content. The Transformer block in the denoising network uses an attention mechanism to capture global dependencies. The attention mechanism enables the model to focus on key regions in the image, thereby more accurately removing noise and restoring image details. The Residual State Space Block (RSSB) further refines and encodes image features. Through its unique structural design, the RSSB can more effectively extract and represent high-level features in the image, providing stronger support for subsequent image reconstruction. The Visual State Space Module (VSSM) adaptively captures key pixel information. The VSSM can identify key pixels in the image and perform special processing on these pixels to ensure that important image information is preserved during image enhancement.

[0021] Furthermore, during the reverse process, the denoising network gradually removes noise from the image while preserving key information and details. As the denoising process progresses, the image clarity gradually improves, ultimately generating a clear enhanced image. This enhanced image is then used as the final output. These images should be visually significantly better than the original low-quality underwater image, making them easier for subsequent analysis and processing.

[0022] (III) Beneficial Effects

[0023] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0024] First, in terms of network architecture, this algorithm innovatively designs a denoising network combining Transformer blocks and an improved Residual State Space (RSSB) block. This design fully leverages the Transformer's advantage in capturing global dependencies, enabling the model to better understand and recover long-range information lost in underwater images due to light scattering and absorption. Simultaneously, the RSSB block, combined with Mamba's discrete state space equations, recursively models long dependencies in the image, effectively avoiding channel redundancy issues in traditional models and improving the model's efficiency and accuracy in processing image features. Furthermore, the introduced Visual State Space (VSSM) module adaptively captures important pixel information in the image, enhancing the model's ability to model spatial details, thus better preserving and recovering image details. Second, during model training, the algorithm provides rich learning samples by preparing a large-scale paired dataset containing original underwater images and their corresponding enhanced images, and modifies the number of image feature channels to adapt to the input requirements of the RSSB module, ensuring the model can train normally and efficiently. Effective training is achieved by employing a hybrid loss function to optimize the network. This not only provides variance learning signals to stabilize the training process but also improves model training performance and image enhancement quality. Crucially, by using conditional images, such as reference clear images or simulated underwater images, as auxiliary inputs, the model is guided to generate specific enhanced images corresponding to the input images. This strategy significantly enhances the model's adaptability and flexibility, enabling it to more accurately handle images from different underwater environments. Finally, in the image enhancement process, the algorithm progressively adds Gaussian noise to the image through the forward process of the conditional diffusion model, simulating the underwater image degradation process. In the reverse process, a trained denoising network is used to progressively remove noise, generating a clear enhanced image. The collaborative work of the Transformer block, RSSB block, and VSSM module in the denoising network allows the model to simultaneously optimize image quality at both global and local levels, restoring color and contrast while preserving rich detail information. This comprehensive enhancement strategy effectively overcomes the negative impact of the underwater environment on image quality, significantly enhancing the application potential of underwater images. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the algorithm structure of an adaptive underwater image enhancement algorithm according to the present invention;

[0026] Figure 2 This is a diagram illustrating the effect of an adaptive underwater image enhancement algorithm according to the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0028] Combination Figures 1-2 The present invention provides an adaptive underwater image enhancement algorithm, which specifically includes the following steps:

[0029] S10. Constructing the network architecture: Design a conditional diffusion model as the main framework for image enhancement. This model includes a forward noise addition process and a reverse denoising process. In the denoising process, a denoising network composed of Transformer blocks and improved residual state space blocks RSSB is introduced to encode and refine image features. RSSB combines Mamba's discrete state space equations to recursively model long dependencies in the image, avoiding channel redundancy. A visual state space module VSSM is introduced to adaptively capture important pixel information in the image and enhance the ability to model spatial details.

[0030] The specific operational steps in step S10 include:

[0031] S101. Determine the model framework:

[0032] First, the basic structure of the conditional diffusion model is clarified, including the forward noise addition process and the reverse noise reduction process. This is the core framework of the entire algorithm, used to simulate the entire process of an image from sharpness to degradation and then to enhancement.

[0033] S102, Design a noise reduction network:

[0034] A denoising network consisting of Transformer blocks and an improved Residual State Space Block (RSSB) is introduced. This design aims to combine the global dependency capture capability of the Transformer with the recursive modeling capability of the RSSB to process image features more finely.

[0035] The Transformer block is responsible for capturing global dependencies in an image and understanding the overall structure of the image through a self-attention mechanism.

[0036] The RSSB block combines Mamba's discrete state-space equations to recursively model long dependencies in the image. This recursive modeling approach helps avoid channel redundancy and improves the model's efficiency in utilizing image features.

[0037] S103, Integrated VSSM Module:

[0038] The Visual State Space Module (VSSM) is introduced, which can adaptively capture important pixel information in the image, enhancing the model's ability to model spatial details. The addition of VSSM makes the model more outstanding in recovering image details.

[0039] S104. Optimize network structure:

[0040] Based on specific task requirements and network performance, parameter tuning and structural optimization are performed on each module in the denoising network. This includes adjusting the number, size, and connection method of Transformer blocks and RSSB blocks to ensure that the entire network achieves the best balance between performance and efficiency.

[0041] S105, Implementation details:

[0042] During implementation, attention needs to be paid to the details of network architecture, such as the way data flows are transmitted, the interface definition between modules, and the implementation of parallel computing. These details are crucial to ensuring the stability and efficiency of the network architecture.

[0043] S106. Verification and Testing:

[0044] After the network architecture is designed and implemented, it needs to be fully verified and tested. This includes preliminary testing with a small dataset to check the feasibility and effectiveness of the network architecture, followed by comprehensive testing with a large dataset to evaluate the performance and stability of the network architecture.

[0045] S20. Model Training: Prepare a large-scale paired dataset containing the original underwater images and their corresponding enhanced images for model training and testing. Modify the number of image feature channels to adapt to the input requirements of the residual state space module to ensure that the model can be trained normally. Use a hybrid loss function to optimize the network and provide variance learning signals to improve the model training effect and image enhancement quality. During the training process, use conditional images such as reference clear images or simulated underwater images as auxiliary inputs to guide the model to generate specific enhanced images corresponding to the input images.

[0046] Step S20 specifically includes the following operational steps:

[0047] S201. Prepare the dataset:

[0048] Collect and organize large-scale paired datasets containing raw underwater images and their corresponding enhanced images. These datasets should cover a variety of underwater environments, lighting conditions, and object types to ensure the model's generalization ability.

[0049] The dataset is preprocessed, including resizing images and normalizing pixel values, so that the model can better process image data;

[0050] S202. Modify the number of image feature channels:

[0051] Based on the input requirements of the Residual State Space Module (RSSB), the number of feature channels of the image data is modified. This typically involves adjusting the dimensions of the image data to ensure that it matches the input interface of the RSSB module.

[0052] S203. Design a hybrid loss function:

[0053] Design a hybrid loss function that combines multiple loss terms, such as pixel loss, structural loss, and perceptual loss, to comprehensively evaluate the difference between the augmented image generated by the model and the real augmented image.

[0054] A variance learning signal is introduced into the hybrid loss function to encourage the model to explore a wider solution space during training, while maintaining the stability and consistency of the generated enhanced images;

[0055] S204. Set training parameters:

[0056] Determine key parameters for model training, such as the initial learning rate, batch size, and number of training epochs. These parameters directly affect the model's training performance and convergence speed.

[0057] Based on the specific structure of the model and the characteristics of the dataset, the parameters are tuned to find the optimal combination of training parameters.

[0058] S205. Training using conditional images:

[0059] During training, in addition to using paired raw and enhanced underwater images, conditional images are introduced as auxiliary inputs. These conditional images can be reference sharp images or simulated underwater images, used to guide the model in generating specific enhanced images corresponding to the input images.

[0060] The conditional image and the original underwater image are input into the model together. The model is trained through the forward and backward processes of the conditional diffusion model. In the backward process, the additional information provided by the conditional image is used to optimize the parameters of the denoising network to improve the quality of the generated enhanced image.

[0061] S206. Monitor the training process:

[0062] During training, monitor the model's performance metrics in real time, such as loss function value, accuracy, and recall. These metrics reflect the model's learning progress and generalization ability during training.

[0063] Adjust training parameters or optimize model structure in a timely manner based on monitoring results to ensure that the model can be trained in the expected direction;

[0064] S207. Verification and Testing:

[0065] Use separate validation and test sets to validate and test the model. These datasets should be separate from the training set to ensure the objectivity and reliability of the test results.

[0066] Evaluate the model's performance metrics on the validation and test sets, such as the visual effect of enhanced images, contrast, and color restoration. Based on the evaluation results, further optimize or adjust the model.

[0067] S30. Image Enhancement Process: Input a low-quality underwater image and a conditional image. Set a time step. Through the forward process of the conditional diffusion model, Gaussian noise is gradually added to the image until the image becomes pure noise. In the reverse process, a trained denoising network is used to gradually remove noise and generate a clear enhanced image. The Transformer block in the denoising network uses an attention mechanism to capture global dependencies, the RSSB block further refines and encodes image features, and the VSSM module adaptively captures key pixel information.

[0068] Step S30 specifically includes the following operational steps.

[0069] S301, Input Preparation

[0070] Input low-quality underwater image: Use the low-quality underwater image to be enhanced as input to the model.

[0071] Input conditional image: Input conditional images, such as reference clear images or simulated underwater images, to guide the model to generate enhanced images of a specific style;

[0072] S301, Set time step

[0073] Determine the time step of the forward process: Based on the model design and actual needs, set the time step (or diffusion step) for gradually adding Gaussian noise in the forward process. This time step determines the speed and extent of noise addition.

[0074] S301, Forward Process: Gradually Add Noise

[0075] Initialization: Start with low-quality underwater images.

[0076] Noise addition: Gaussian noise is gradually added to the image according to the set time step. Each step uses a predefined method and noise level to increase the noise content of the image until the image finally becomes pure noise. This process simulates the transformation of data from structured to unstructured.

[0077] S301, Reverse Process: Gradually Remove Noise

[0078] Initialize the noisy image: Use the final noisy image (pure noise image) of the forward process as the starting point of the reverse process.

[0079] The role of a denoising network: A trained denoising network gradually removes noise from an image. This network typically contains multiple modules, such as Transformer blocks, RSSB blocks, and VSSM blocks.

[0080] The Transformer block uses an attention mechanism to capture global dependencies in an image, helping the model understand the overall structure and content of the image.

[0081] RSSB block: Further refines and encodes image features, extracting higher-level image representations.

[0082] VSSM module: Adaptively captures key pixel information, ensuring that important image details are preserved while removing noise.

[0083] Progressive denoising: Through multiple steps in the reverse process, noise in the image is removed layer by layer. Each step uses a denoising network to predict and remove the noise at the current noise level, gradually restoring the clarity and details of the original image.

[0084] S301, Output Enhanced Image

[0085] Generate an enhanced image: After the progressive denoising process in reverse, a clear enhanced image is finally generated. This image retains the important information of the original image while removing low-quality factors such as noise and blur, presenting higher clarity and contrast.

[0086] S301. Post-processing: As needed, the generated enhanced image can be post-processed and optimized, such as color correction and contrast enhancement, to further improve image quality.

[0087] according to Figure 2 Input represents the original underwater image, Ours represents the enhanced image obtained using this invention's algorithm, and Ground Truths represents the enhanced image paired with the original image.

[0088] In a preferred embodiment, the present invention can be further configured as follows: during the forward noise addition process in step S10, clear image data is gradually injected with noise until the data becomes pure noise. Each step of noise injection can be regarded as a small Wiener process, wherein the noise addition is performed stepwise to ensure a smooth transition of the data. The independent incremental characteristic of the Wiener process ensures that the noise added at each step is independent, and the past noise state will not affect the future noise state. The increment from time t to t+Δt follows a normal distribution with a center of 0 and a standard deviation of Δt, which ensures the randomness and controllability of the noise. The noise addition is performed stepwise, avoiding the sharp decline in image quality caused by adding a large amount of noise at once in traditional methods. The stepwise addition of noise ensures a smooth transition of image data in the noiseification process, which is beneficial for better recovery of image details in the subsequent reverse denoising process. The independent incremental characteristic of the Wiener process is the core advantage of this design, which ensures that the noise added at each step is independent. The added noise is independent, meaning that past noise states do not affect future noise states. This independence makes the addition of noise more random and uniform, avoiding local noise aggregation in the image, thereby improving the uniformity and naturalness of the enhanced image. The noise increment from time t to t+Δt follows a normal distribution with a center of 0 and a standard deviation of Δt. This setting ensures both the randomness of the noise and the controllability of the noise addition process. By adjusting the value of Δt, the speed and degree of noise addition can be flexibly controlled to adapt to different image qualities and enhancement needs. The gradually and independently added noise provides richer information for the subsequent reverse denoising process. In the reverse process, the denoising network can use this noise information to more accurately restore the original details and structure of the image, thereby achieving better enhancement results. In addition, because the randomness and controllability of the noise are guaranteed, the denoising network is also more likely to learn effective denoising strategies and feature representations during training.

[0089] In a preferred embodiment, the present invention can be further configured as follows: During the reverse denoising process, the denoising process starts from a noisy state and uses a denoising network composed of Transformer blocks and an improved RSSB to gradually restore the original state of the data. The Transformer block is used to capture global dependencies in the image and encodes image features through a self-attention mechanism. The improved RSSB combines the discrete state-space equation of Mamba to recursively model long dependencies in the image, avoiding channel redundancy. The RSSB enhances the capabilities of the original Mamba through local spatial priors and channel interactions, generating feature representations specific to the restoration and adaptively capturing the image. By capturing key pixel information and enhancing the ability to model spatial details, VSSM helps the denoising network more accurately recover image details by capturing long-term spatial dependencies. In the reverse process, the denoising network iteratively reduces noise in the data, using a trained model to predict the best method for noise removal at each step until clear data is recovered. The introduction of the Transformer block greatly enhances the denoising network's ability to capture global dependencies in the image. Through a self-attention mechanism, the Transformer block can effectively model the interactions between different regions in the image, thus better understanding and recovering the overall structure and content of the image during denoising. This global nature... The improved RSSB approach helps reduce local inconsistencies and artifacts that may occur during denoising, improving the visual coherence and naturalness of the enhanced image. Secondly, the improved RSSB incorporates Mamba's discrete state-space equations to recursively model long dependencies in the image. This design not only avoids channel redundancy issues that may arise in traditional methods but also enhances the capabilities of the original Mamba through local spatial priors and channel interactions. RSSB can generate recovery-specific feature representations, adaptively capturing important pixel information in the image, thus better preserving and restoring image details while achieving noise removal. This fine-grained processing is crucial for improving the quality of the enhanced image. Importantly, the introduction of VSSM further enhances the ability of denoising networks to capture long-term spatial dependencies. Through its unique structure and mechanism, VSSM helps denoising networks recover detailed information in images more accurately, especially when processing images with complex textures and structures. In the reverse denoising process, VSSM works in conjunction with the Transformer block and the improved RSSB to gradually reduce noise in the data through an iterative approach. At each step, the trained model is used to predict the best method for removing noise. This iterative optimization not only improves the efficiency of the denoising process but also ensures that the final recovered image has clear visual effects and rich details.

[0090] In a preferred embodiment, the present invention can be further configured as follows: the trained model minimizes the difference between generated data and original data, using mean squared error (MSE) or log-likelihood loss as the training objective, selecting an appropriate loss function based on the specific characteristics of the data and the noise model to ensure that the model can accurately recover image details, employing optimization algorithms such as gradient descent to train the model, updating network parameters through backpropagation to improve the model's denoising ability. By explicitly setting the training objective as minimizing the difference between generated data and original data, a clear and quantifiable optimization direction is set for the model training. Using mean squared error (MSE) or log-likelihood loss as training objectives can accurately measure the gap between the model output and the real data, thereby guiding the model to continuously optimize its parameters during the learning process to more accurately restore the original state of the image. This quantified training objective not only improves the efficiency and stability of model training but also ensures that the model can output high-quality enhanced images during the testing phase. Secondly, selecting an appropriate loss function based on the specific characteristics of the data and the noise model is one of the important innovative features of this technology. Different image data and noise models may have different characteristics. Different statistical characteristics and distribution patterns mean that choosing an appropriate loss function is crucial for improving the model's adaptability and generalization ability. For example, for image data with obvious Gaussian noise characteristics, MSE loss may be a good choice; while for more complex or sparse noisy models, it may be necessary to consider using a more flexible loss function to capture subtle differences in the data. By carefully selecting the loss function, this technique can further improve the model's performance in specific application scenarios. Finally, using optimization algorithms such as gradient descent to train the model and updating the network parameters through backpropagation is the key to achieving efficient training. Gradient descent is an optimization method widely used in neural network training. It can update the model's parameters based on the gradient information of the loss function, thereby gradually reducing the value of the loss function. Backpropagation is one of the key technologies for implementing gradient descent. It can propagate the gradient information of the loss function from the output layer back to the input layer layer by layer, so as to update the corresponding parameters in each layer. Through this iterative optimization method, this technique can continuously improve the model's denoising ability and detail recovery accuracy, ultimately outputting high-quality enhanced images.

[0091] In a preferred embodiment, the present invention can be further configured as follows: In step S20, a large-scale paired dataset containing original underwater images and their corresponding enhanced images is prepared. These images should cover different underwater environments, lighting conditions, object types, etc., to ensure the diversity and representativeness of the dataset. The data is preprocessed to remove low-quality images such as noise, blur, and duplication. If the dataset requires annotation for tasks such as object detection or segmentation, corresponding annotation work is performed. The number of image feature channels is modified according to the input requirements of the Residual State Space Module (RSSB). This typically involves converting the image from the common RGB three-channel format to other channel counts, or generating suitable input features for RSSB through specific preprocessing steps such as color space conversion and feature extraction. By preparing a large-scale paired dataset containing original underwater images and their corresponding enhanced images, and ensuring that these images cover different underwater environments, lighting conditions, object types, etc., this technical feature greatly improves the diversity and representativeness of the dataset. This diverse dataset helps the model learn more comprehensive and robust feature representations, thereby improving the model's generalization ability in different scenarios. Furthermore, the dataset is further processed... Preprocessing to remove low-quality images such as noise, blur, and duplicates is a crucial step in ensuring the quality of model training. This step significantly reduces the interference of noise on model training, improving the model's stability and accuracy. Removing duplicate images also helps reduce data redundancy and improve training efficiency. Furthermore, labeling the dataset is essential. This annotation not only provides the model with richer supervision signals but also helps the model perform more accurately and efficiently in advanced vision tasks. For example, in object detection and segmentation, accurate annotations can guide the model to learn the precise boundaries and features of objects, thereby improving detection and segmentation accuracy. Finally, modifying the number of image feature channels according to the input requirements of the Residual State Space Module (RSSB) is an important step in ensuring the model can fully utilize the advantages of RSSB. As an advanced network structure, RSSB typically has specific requirements for input features. By modifying the number of image feature channels or performing preprocessing steps such as color space conversion and feature extraction, suitable input features for RSSB can be generated, thus fully leveraging RSSB's advantages in capturing long-term dependencies and detailed information in images.

[0092] In a preferred embodiment, the present invention can be further configured as follows: the model training is based on the conditional diffusion model framework, constructing a denoising network including a Transformer block, an improved RSSB, and a visual state space module VSSM, ensuring that the network structure can handle the adjusted number of image feature channels. A hybrid loss function is used to optimize the network. The hybrid loss function can include multiple components, such as pixel-level losses like L1 or L2 losses to measure the pixel difference between the generated image and the real image, perceptual losses like feature matching losses to measure the similarity between the generated image and the real image in the high-level feature space, and adversarial losses like the discriminator loss in GANs to improve the realism of the generated image. These loss functions together provide variance. Learning signals are used to improve model training performance and image enhancement quality. Appropriate optimizers, such as Adam and SGD, are selected to update network parameters. The choice of optimizer should consider factors such as training speed, convergence stability, and hyperparameter sensitivity. During training, the dataset is divided into training, validation, and test sets. The training set is used for model training, the validation set is used to adjust hyperparameters and monitor the training process, and the test set is used to finally evaluate model performance. Conditional images, such as reference clear images or simulated underwater images, are used as auxiliary inputs during training. These conditional images can provide the model with additional information or constraints, guiding the model to generate specific enhanced images corresponding to the input images. Network parameters are optimized through multiple iterations of training. In each iteration, the training set is used to refine the network parameters. A batch of images is randomly selected for comparison with the original and enhanced images. Loss values ​​are calculated through forward propagation, and network parameters are updated through backpropagation. The model's performance on the validation set is periodically evaluated during training. If the model's performance on the validation set begins to decline, indicating overfitting, appropriate measures are taken to prevent overfitting, such as early stopping training, reducing the learning rate, and increasing regularization. A conditional diffusion model framework is used as the basis for model training. This framework has powerful generative capabilities and flexibility, capable of capturing complex distributions in the data and generating high-quality images. A denoising network is constructed by combining Transformer blocks, an improved RSSB, and the Visual State Space Module (VSSM). Each of these modules has its unique advantages. The Transformer block captures global dependencies in an image, the improved RSSB models long-term dependencies and avoids channel redundancy, and VSSM captures long-term spatial dependencies to recover image details. This network structure not only handles adjusted image feature channel counts but also captures key image information at different levels, thereby improving denoising and enhancement effects. Secondly, the use of hybrid loss functions to optimize the network is another important aspect of this technique. These hybrid loss functions include pixel-level loss, perceptual loss, and adversarial loss, which together provide the model with rich learning signals. Pixel-level loss ensures the pixel-level similarity between the generated image and the real image.Perceptual loss focuses on similarity in the high-level feature space, while adversarial loss further improves the realism of the generated images by introducing a discriminator. This multi-layered loss function design helps the model gradually optimize its performance during training, ultimately generating high-quality enhanced images. The selection of optimizers is also carefully considered; suitable optimizers such as Adam or SGD can balance factors such as training speed, convergence stability, and hyperparameter sensitivity, ensuring that the model can steadily improve its performance during training. Furthermore, by dividing the dataset into training, validation, and test sets, this technique provides a scientific basis for model training and evaluation. The training set is used for model training, the validation set is used to adjust hyperparameters and monitor the training process to prevent overfitting, and the test set is used to finally evaluate the model's performance. In addition, using conditional images as auxiliary input is another feature of this technique. Another highlight of this feature is that conditional images, such as reference clear images or simulated underwater images, can provide the model with additional information or constraints, thereby guiding the model to generate specific augmented images corresponding to the input images. This conditional guidance method helps the model better understand the correspondence between images during training and generate augmented results that better meet expectations. Finally, through multiple iterations of training and optimization of network parameters, this technique ensures that the model can gradually improve its performance during training. In each iteration, a batch of image pairs is randomly selected from the training set for forward and backward propagation operations to calculate the loss value and update the network parameters. Simultaneously, the model's performance on the validation set is periodically evaluated during training, and corresponding measures are taken based on the evaluation results to prevent overfitting. This scientific training process ensures that the model performs well in different scenarios and has high generalization ability.

[0093] In a preferred embodiment, this invention can be further configured as follows: The performance of the trained model is evaluated using a test set. Evaluation metrics may include objective metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), as well as subjective visual quality assessment. The model is adjusted based on the evaluation results. If the model performance does not meet the requirements, adjustments can be made to the network structure, loss function, optimizer, or training strategy. Using a test set to evaluate the performance of the trained model reflects an emphasis on the model's practical application capabilities. The evaluation metrics not only include objective metrics such as PSNR and SSIM, which can quantitatively measure the similarity between the generated image and the real image at the pixel and structural levels, thus objectively reflecting the model's denoising and restoration capabilities; but also emphasize the importance of subjective visual quality assessment. This evaluation method can… This technology takes into account the subjective perception of image quality by the human eye, including multiple aspects such as color, sharpness, and contrast, providing a powerful supplement to the comprehensive evaluation of model performance. Based on the evaluation results, this technology also proposes strategies for adjusting the model. If the model performance does not meet the requirements, adjustments can be made by comprehensively considering multiple aspects, including but not limited to network structure, loss function, optimizer, or training strategy. This systematic adjustment strategy helps to address the problems existing in the model in a targeted manner, thereby improving its performance in practical applications. For example, adjusting the network structure can introduce more powerful feature extraction capabilities, optimizing the loss function can better balance the contribution of different aspects of loss, changing the optimizer can speed up training or improve convergence stability, and adjusting the training strategy, such as adding data augmentation or using pre-trained models, can further improve the model's generalization ability and robustness.

[0094] In a preferred embodiment, the present invention can be further configured as follows: During input, a low-quality underwater image to be enhanced is used as the initial input. These images may be affected by factors such as insufficient light, water scattering, and color distortion, leading to a decrease in image quality. Simultaneously, conditional images, such as a reference clear image or a simulated underwater image, are input. These conditional images provide additional information or constraints to the model, helping to generate an enhanced image that better meets expectations. An appropriate time step is set according to requirements. The time step determines the rate and extent of adding Gaussian noise to the image during the forward process. Through the forward process of the conditional diffusion model, Gaussian noise is gradually added to the image, starting from the initial low-quality underwater image. As the time step increases, useful information in the image is gradually masked by noise until the image becomes pure noise. The importance of simultaneously providing both the low-quality underwater image to be enhanced and the conditional images during the input stage is highlighted. The quality of the low-quality underwater image as the initial input is limited by various natural factors, while the conditional images (such as a reference clear image or a simulated underwater image) provide additional information or constraints to the model. This dual-input approach allows the model to learn the mapping between low-quality and high-quality images during training, and to use useful information from the conditional image to guide the enhancement process, thereby generating more desirable enhanced images. This design not only improves the model's flexibility and adaptability but also enhances its ability to handle complex underwater environments. Secondly, the forward process of the conditional diffusion model is introduced as a key step in image enhancement. By setting an appropriate time step, the model can control the rate and extent of adding Gaussian noise to the image during the forward process. This process simulates the natural process of image quality deterioration, allowing the model to learn the feature representation of the image at different noise levels. As the time step increases, useful information in the image is gradually masked by noise until the image becomes pure noise. However, in the reverse process (i.e., during training), the model attempts to recover the original high-quality image from pure noise. This progressively challenging training method helps the model better learn the key information and structural features in the image, thereby improving the quality and realism of the enhanced image.

[0095] In a preferred embodiment, the present invention can be further configured as follows: during the reverse process, a pre-trained denoising network is used to gradually remove noise from the image. The denoising network learns the mapping relationship between low-quality and high-quality images, enabling it to accurately predict and remove noise, thus restoring clear image content. The Transformer block in the denoising network utilizes an attention mechanism to capture global dependencies. This attention mechanism allows the model to focus on key regions in the image, thereby more accurately removing noise and restoring image details. The Residual State Space Block (RSSB) further refines and encodes image features. Through its unique structural design, the RSSB can more effectively extract and represent high-level features in the image, providing a basis for subsequent image processing. To provide stronger support for image reconstruction, the Visual State Space Module (VSSM) adaptively captures key pixel information. VSSM can identify key pixels in the image and perform special processing on these pixels to ensure that important image information is preserved during image enhancement. The denoising network learns the mapping relationship between low-quality and high-quality images, accurately predicting and removing noise in the image to restore clear image content. This process not only solves the problem of traditional methods performing poorly when dealing with complex noise, but also significantly improves the efficiency and accuracy of image enhancement. The design of the denoising network allows the model to automatically adapt to different noise types and levels, thus maintaining stability in various complex underwater environments. Secondly, regarding performance, the Transformer block in the denoising network utilizes an attention mechanism to capture global dependencies. This feature allows the model to focus on key regions in the image while removing noise. By dynamically adjusting the model's attention to different parts of the image, the attention mechanism enables the model to more accurately identify and remove noise that affects image quality, while preserving important information and details. This targeted approach not only improves the image enhancement effect but also makes the enhanced image more consistent with human visual perception. Furthermore, the introduction of the Residual State Space Block (RSSB) further refines and encodes image features. Through its unique structural design, the RSSB can more effectively extract and represent features in the image. High-level features not only contain the basic contour and texture information of the image, but also contain deeper semantic information. This information provides stronger support for subsequent image reconstruction, enabling the enhanced image to maintain clarity while better restoring the scene and atmosphere of the original image. Finally, the addition of the Visual State Space Module (VSSM) allows the model to adaptively capture key pixel information in the image. By identifying key pixels in the image and performing special processing on these pixels, VSSM ensures that important image information is preserved during the image enhancement process. This targeted processing method not only improves the quality of image enhancement, but also makes the enhanced image more in line with the needs of practical applications.

[0096] In a preferred embodiment, the present invention can be further configured as follows: during the reverse process, the denoising network gradually removes noise from the image while preserving key information and details. As the denoising process progresses, the image clarity gradually improves, ultimately generating a clear enhanced image. This enhanced image is then used as the final output. These images should be visually significantly superior to the original low-quality underwater image, facilitating subsequent analysis and processing. This emphasizes the importance of the denoising process in preserving key information and details. Under the action of the denoising network, noise in the image is gradually removed, while key information such as edges, textures, and colors are effectively preserved. This targeted denoising method ensures that the enhanced image is both clear and natural, avoiding excessive noise. Firstly, the image enhancement process avoids distortion or blurring caused by excessive denoising. Secondly, as the denoising process progresses, the image clarity gradually improves. This process not only visually demonstrates the effect of image enhancement but also provides users with a better visual experience. The resulting clear enhanced image is visually superior to the original low-quality underwater image, making the information in the image easier to identify and analyze. Furthermore, the clear enhanced image provides a more reliable data foundation for subsequent image analysis and processing. In fields such as machine vision, marine monitoring, and underwater archaeology, high-quality images are crucial to ensuring the accuracy of analysis results. The enhanced image generated through this technology can significantly improve the efficiency and accuracy of subsequent image processing, providing strong support for research and applications in related fields.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive underwater image enhancement algorithm, It is characterized in that The specific steps include: S10. Build network architecture: Design a conditional diffusion model as the main framework of image enhancement. The model includes a forward noise addition process and a reverse denoising process. In the denoising process, a denoising network composed of a Transformer block and an improved residual state space block RSSB is introduced to encode and refine image features. RSSB combines Mamba's discrete state space equation to recursively model long dependencies in the image to avoid channel redundancy. The visual state space module VSSM is introduced to adaptively capture important pixel information in the image and enhance the modeling ability of spatial details. S20, model training: prepare a large-scale paired dataset containing underwater original images and their corresponding enhanced images for model training and testing, modify the number of image feature channels to adapt to the input requirements of the residual state space module, ensure that the model can be trained normally, use a mixed loss function to optimize the network, and provide variance learning signals to improve the model training effect and image enhancement quality. During the training process, use conditional images such as reference clear images or simulated underwater images as auxiliary inputs to guide the model to generate specific enhanced images corresponding to the input images; S30, image enhancement process: input low-quality underwater images and conditional images, set the time step, and gradually add Gaussian noise to the image through the forward process of the conditional diffusion model until the image becomes pure noise. In the reverse process, the trained denoising network is used to gradually remove the noise to generate a clear enhanced image. The Transformer block in the denoising network uses the attention mechanism to capture global dependencies, the RSSB block further refines and encodes image features, and the VSSM module adaptively captures key pixel information.

2. The adaptive underwater image enhancement algorithm according to claim 1, It is characterized in that In the forward noise adding process in step S10, the clear image data is gradually injected with noise until the data becomes pure noise. Each step of noise injection can be regarded as a small Wiener process, in which the noise is added step by step to ensure a smooth transition of the data. The independent incremental characteristic of the Wiener process ensures that the noise added at each step is independent, and the past noise state will not affect the future noise state. The increment from time t to t+Δt follows a normal distribution with a center of 0 and a standard deviation of Δt, which ensures the randomness and controllability of the noise.

3. The adaptive underwater image enhancement algorithm according to claim 2, It is characterized in that In the reverse denoising process, the denoising process starts from the noisy state, and uses a denoising network composed of a Transformer block and an improved RSSB to gradually restore the original state of the data. The Transformer block is used to capture the global dependencies in the image and encode the image features through the self-attention mechanism. The improved RSSB combines the discrete state space equations of Mamba to model the long dependencies in the image in a recursive form to avoid channel redundancy. RSSB enhances the capabilities of the original Mamba through local spatial priors and channel interactions, generates feature representations specific to restoration, adaptively captures important pixel information in the image, and enhances the modeling capabilities of spatial details. VSSM helps the denoising network to more accurately restore image details by capturing long-term spatial dependencies. In the reverse process, the denoising network gradually reduces the noise in the data in an iterative manner. At each step, the trained model is used to predict the best method to remove noise until clear data is restored.

4. The adaptive underwater image enhancement algorithm according to claim 3, It is characterized in that The trained model minimizes the difference between the generated data and the original data, uses the mean square error MSE or the log-likelihood loss as the training target, selects a suitable loss function according to the specific characteristics of the data and the noise model to ensure that the model can accurately restore the image details, and trains the model using optimization algorithms such as gradient descent. The network parameters are updated through the back propagation algorithm to improve the denoising ability of the model.

5. The adaptive underwater image enhancement algorithm according to claim 1, It is characterized in that In step S20, a large-scale paired data set including underwater original images and their corresponding enhanced images is prepared. These images should cover different underwater environments, lighting conditions, object types, etc. to ensure the diversity and representativeness of the data set, and these data are preprocessed to remove low-quality images such as noise, blur, and duplication in the data set. If the data set needs to be labeled for tasks such as object detection and segmentation, the corresponding labeling work is performed. According to the input requirements of the residual state space module RSSB, the number of image feature channels is modified, which usually involves converting the image from the common RGB three channels to other channel numbers, or generating input features suitable for RSSB through specific preprocessing steps such as color space conversion and feature extraction.

6. The adaptive underwater image enhancement algorithm according to claim 1, It is characterized in that The model training is based on the conditional diffusion model framework, and a denoising network including a Transformer block, an improved RSSB and a visual state space module VSSM is constructed to ensure that the network structure can handle the adjusted number of image feature channels. A mixed loss function is used to optimize the network. The mixed loss function may include multiple components, such as pixel-level losses such as L1 or L2 losses for measuring the pixel difference between the generated image and the real image, perceptual losses such as feature matching losses for measuring the similarity between the generated image and the real image in the high-level feature space, and adversarial losses such as the discriminator loss in GAN for improving the realism of the generated image. These loss functions jointly provide variance learning signals to improve the model training effect and image enhancement quality. Appropriate optimizers such as Adam, SGD, etc. are selected to update the network parameters. The selection of the optimizer should take into account the training speed, convergence stability and hyperparameter sensitivity. The training set is used for model training, the validation set is used to adjust hyperparameters and monitor the training process, and the test set is used to finally evaluate the model performance. During the training process, conditional images such as reference clear images or simulated underwater images are used as auxiliary inputs. These conditional images can provide additional information or constraints for the model, guiding the model to generate specific enhanced images corresponding to the input images, and optimizing the network parameters through multiple iterations of training. In each iteration, a batch of image pairs of original images and enhanced images are randomly selected from the training set, and the loss value is calculated through forward propagation. Then, the network parameters are updated through back propagation. The performance of the model on the validation set is regularly evaluated during the training process. If the performance of the model on the validation set begins to decline, that is, overfitting occurs, appropriate measures are taken to prevent overfitting, such as stopping training early, reducing the learning rate, and increasing regularization.

7. The adaptive underwater image enhancement algorithm according to claim 6, It is characterized in that The test set is used to evaluate the performance of the trained model. The evaluation indicators may include objective indicators such as peak signal-to-noise ratio PSNR, structural similarity index SSIM, and subjective visual quality evaluation. The model is adjusted according to the evaluation results. If the model performance does not meet the requirements, the network structure, loss function, optimizer or training strategy can be adjusted.

8. The adaptive underwater image enhancement algorithm according to claim 1, It is characterized in that The low-quality underwater images to be enhanced are used as initial inputs. These images may be affected by factors such as insufficient light, water scattering, and color distortion, resulting in reduced image quality. At the same time, conditional images are input, such as reference clear images or simulated underwater images. These conditional images provide additional information or constraints for the model, which helps to generate enhanced images that are more in line with expectations. A suitable time step is set according to requirements. The time step determines the rate and degree of adding Gaussian noise to the image in the forward process. Through the forward process of the conditional diffusion model, starting from the initial low-quality underwater image, Gaussian noise is gradually added to the image. As the time step increases, the useful information in the image is gradually masked by the noise until the image completely becomes pure noise.

9. The adaptive underwater image enhancement algorithm according to claim 1, It is characterized in that In the reverse process, the trained denoising network is used to gradually remove the noise in the image. The denoising network can accurately predict and remove noise and restore clear image content by learning the mapping relationship between low-quality images and high-quality images. The Transformer block in the denoising network uses the attention mechanism to capture global dependencies. The attention mechanism enables the model to focus on key areas in the image, thereby more accurately removing noise and restoring image details. The residual state space block RSSB further refines and encodes image features. RSSB, through its unique structural design, can more effectively extract and represent high-level features in the image, providing stronger support for subsequent image reconstruction. The visual state space module VSSM adaptively captures key pixel information. VSSM can identify key pixel points in the image and perform special processing on these pixel points to ensure that important image information can be retained during the image enhancement process.

10. An adaptive underwater image enhancement algorithm according to claim 9, It is characterized in that In the reverse process, the denoising network gradually removes the noise in the image while maintaining the key information and details of the image. As the denoising process proceeds, the clarity of the image gradually improves, and finally a clear enhanced image is generated. The generated enhanced image is used as the final output. These images should be significantly better than the original low-quality underwater images in visual effects and are more convenient for subsequent analysis and processing.

Images