A multi-teacher supervised underwater image enhancement method based on diffusion recovery
By employing a multi-teacher supervision framework and a diffusion model based on underwater physical imaging priors, combined with a depth unfolding architecture, the problems of structural consistency and data dependency in underwater image enhancement are solved, achieving efficient and physically reasonable image enhancement results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA SHENZHEN INTERNATIONAL GRADUATE SCHOOL
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-29
AI Technical Summary
Existing underwater image enhancement methods based on diffusion models are insufficient in maintaining image structural consistency and physical interpretability, and rely too heavily on high-quality paired data, which limits their application in real-world scenarios.
We employ a multi-teacher supervision framework and a diffusion model based on underwater physical imaging priors, combined with a learnable deep unfolding architecture. By iteratively optimizing physical parameters, we reduce dependence on paired data, preserve image structural information, and improve physical plausibility.
It significantly improves the physical consistency and structure preservation of underwater image enhancement, reduces the dependence on real data, and enhances the model's generalization ability and enhancement effect in different underwater environments.
Smart Images
Figure CN122115243A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a multi-teacher supervised underwater image enhancement method based on diffusion restoration. Background Technology
[0002] With the increasing frequency of activities such as marine resource exploration, underwater engineering operations, and ecological monitoring, underwater vision systems play a crucial role. However, due to the absorption and scattering effects of water on light, underwater images generally suffer from degradation phenomena such as color distortion, low contrast, and blurred details, which severely restricts the accuracy of automated analysis and decision-making based on underwater images.
[0003] In recent years, methods based on the Transformer architecture and diffusion models have made significant progress in underwater image enhancement tasks, improving image quality to a certain extent. However, existing methods still face several key technical bottlenecks, restricting their deployment and application in real-world scenarios:
[0004] 1. Difficulty in maintaining structural consistency: Current diffusion-based methods typically employ Gaussian noise addition strategies, which can easily damage the structural information of the original image during denoising and reconstruction, leading to inconsistencies in content or semantics between the images before and after enhancement, affecting visual coherence and the reliability of subsequent tasks.
[0005] 2. Lack of physical interpretability: Most existing models are "black box" architectures, which make it difficult to effectively integrate the physical prior knowledge of underwater imaging (such as light attenuation, scattering effects, etc.), which may cause the generated results to violate the real underwater optical laws and reduce the physical rationality and credibility of the enhanced images.
[0006] 3. High dependence on high-quality paired data: Mainstream methods mostly adopt a fully supervised learning paradigm, which heavily relies on a large number of realistic and accurately aligned underwater-clear image pairs. However, such realistic paired data is extremely difficult to obtain, and existing studies often rely on synthetic datasets. However, there is a significant domain shift between synthetic data and the real underwater environment, which limits the generalization ability and practicality of the model.
[0007] The aforementioned problems collectively restrict the performance improvement and practical application of underwater image enhancement algorithms. Therefore, there is an urgent need in this field to develop a new generation of enhancement methods that are more physically consistent, more data efficient, and have stronger structure preservation capabilities.
[0008] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0009] The top-level technical problem to be solved in this application is: how to provide an underwater image enhancement method with higher physical consistency and stronger structure preservation capability; specifically, it can be reflected in: how to solve the problem that existing diffusion model-based methods lose image structure information and lack physical authenticity in the restoration results due to the mismatch between the forward noise process and the underwater physical degradation process, and how to solve the problem that existing deep learning methods rely too much on a large number of paired training data.
[0010] The technical solution adopted in this application to solve the above-mentioned technical problems is as follows.
[0011] A multi-teacher supervised underwater image enhancement method based on diffusion restoration includes a model training phase and an image enhancement phase; The model training phase employs a multi-teacher supervision framework to train the student model. This framework includes an explorer teacher model, a stabilizer teacher model, and a memory teacher model. The pseudo-labels generated by these teacher models collectively guide the training of the student model, reducing reliance on paired data. The image enhancement phase includes the following steps: Input the underwater image to be enhanced, and construct a diffusion model based on underwater physical imaging priors. The diffusion model is used to simulate the degradation of underwater images in the underwater environment and preserve the structural information of the underwater image during degradation. Based on a diffusion model, an inverse restoration of underwater images is performed using a learnable deep unfolding architecture to obtain an enhanced, clear image. In reverse recovery, the deep unfolding architecture optimizes multiple physical parameters in an iterative and alternating manner, where the value of each physical parameter is estimated by a corresponding parameter estimation function; Output a clear image after inverse restoration and enhancement.
[0012] In some embodiments, underwater physical imaging priors are based on the attenuation and scattering effects of light underwater, and multiple physical parameters include a clear image. J Transmission diagram, which characterizes the ability of light to transmit. T and the backscattering intensity characterizing the background scattered light. N .
[0013] In some embodiments, the diffusion model is defined by the formula: where P This is an underwater image, representing a matrix of all 1s.
[0014] In some embodiments, the inverse recovery process is performed by a depth unfolding architecture iteratively estimating a clear image. J Transmission image T and backscattering intensity N This is achieved through iterative estimation, where the objective of the estimation is to minimize the underwater image. PThe difference between the reconstructed image and the currently estimated sharp image is... J Transmission image T and backscattering intensity N Calculated using a diffusion model.
[0015] In some embodiments, the iterative and alternating manner includes: in each iteration, the depth-unfolded architecture sequentially updates the sharpened image. J Transmission image T and backscattering intensity N The estimated value is such that when updating one parameter, the current estimated values of the other parameters are fixed.
[0016] In some embodiments, update clear image J The parameter estimation function used requires at least the underwater image and the currently estimated transmission map as input. T and the currently estimated backscattering intensity N .
[0017] In some embodiments, the transmission map is updated. T The parameter estimation function used requires at least an underwater image and a currently estimated sharp image as input. J and the currently estimated backscattering intensity N .
[0018] In some embodiments, the backscattering intensity is updated. N The parameter estimation function used requires at least an underwater image and a currently estimated sharp image as input. J and the currently estimated transmission map T .
[0019] In some embodiments, the parameter estimation function is implemented by a learnable neural network module.
[0020] In some embodiments, an underwater image enhancement apparatus is also provided, comprising: a processor having a computer program stored thereon; wherein, when the computer program is executed by the processor, the underwater image enhancement method of this application is implemented.
[0021] In some embodiments, an underwater robot vision system is also provided, comprising: an image acquisition module for acquiring raw underwater images; an image processing module, including the underwater image enhancement device of this application, for receiving and enhancing the raw underwater images; and a control module for making navigation or target recognition decisions based on the enhanced images.
[0022] In some embodiments, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, which, when executed, implements the underwater image enhancement method of this application.
[0023] The present invention has the following beneficial effects: This application effectively solves the fundamental technical problem of "how to provide an underwater image enhancement method with higher physical consistency and stronger structure preservation capability" by adopting a technical solution of "constructing a diffusion model based on underwater physical imaging priors" and "performing inverse restoration of underwater images through a learnable deep unfolding architecture". Specifically, the diffusion model constructed in this application directly simulates the underwater physical degradation process, abandoning the strategy of adding Gaussian noise, which makes the inverse restoration process essentially match the physical mechanism of image degradation. Through iterative optimization within this physically matched framework, the structural information of the original image can be preserved and restored to the greatest extent, avoiding information loss due to model mismatch, thereby significantly enhancing the physical realism and structural integrity of the restoration results.
[0024] Furthermore, this application introduces a multi-teacher supervision framework for training, utilizing explorer, stabilizer, and memory teacher models to generate pseudo-labels to jointly guide the student model. This significantly reduces the reliance on a large amount of accurately paired real underwater-clear image data, i.e., paired data, thereby improving the model's data utilization efficiency and generalization ability in real-world scenarios.
[0025] Furthermore, this application explicitly introduces light attenuation and scattering effects (clear image J, transmission image) T Backscattering intensity N As a physical prior, the diffusion model is constructed as The form of this process gives the entire enhancement process a clear physical meaning and interpretability, ensuring that the enhancement results conform to the laws of underwater optics.
[0026] Furthermore, this application employs a deep unfolding architecture and optimizes physical parameters in an iterative and alternating manner. Each parameter update is implemented through a dedicated parameterized function whose input includes other current estimates. This enables the model to learn the complex coupling relationships between various physical parameters in a data-driven manner, thereby achieving more accurate and stable joint optimization.
[0027] Furthermore, the parameter estimation function of this application is implemented by a learnable neural network module, which endows the model with powerful nonlinear fitting capabilities and enables it to adaptively learn complex underwater degradation patterns and recovery strategies, thereby further improving the performance of underwater image enhancement.
[0028] Furthermore, this application adopts a multi-teacher supervision framework, which integrates the advantages of different teacher models (such as exploring new knowledge, stabilizing training, and memorizing historical knowledge). It can provide high-quality and diverse supervision signals for student models in the absence of a large amount of real paired data, effectively alleviating the problem of over-reliance on paired data.
[0029] In summary, the technical solution adopted in this application forms an organic whole. The diffusion model based on physical priors provides the correct optimization direction and physical constraints for the entire enhancement process, ensuring the rationality of the restoration results. The deep unfolding architecture and its iterative alternating optimization mechanism decompose the complex inverse restoration problem into multiple learnable sub-problems, achieving refined estimation of physical parameters. Dedicated neural network parameterization functions provide powerful fitting tools for solving each sub-problem. Meanwhile, the multi-teacher supervision framework provides effective learning signals for the entire model during the training phase, reducing data requirements. These technical features work together synergistically to achieve the overall inventive objective of significantly improving the physical consistency and structure preservation capabilities of underwater image enhancement while reducing data dependence.
[0030] Other beneficial effects of the present invention will be further described below. Attached Figure Description
[0031] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a diagram showing the overall architecture and module composition of the diffusion model in this application; Figure 2 This is a visual representation of the supervision methods within a multi-teacher supervision framework. The left side represents existing supervision methods, while the right side represents the supervision method proposed in this invention. Detailed Implementation
[0032] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.
[0033] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0034] This invention proposes a multi-teacher supervised underwater image enhancement method based on diffusion restoration. By combining a diffusion model with prior underwater physical imaging knowledge with a learnable deep unfolding architecture, it effectively compensates for the shortcomings of image enhancement methods that do not conform to physical reality. Furthermore, the introduction of a multi-teacher supervised framework effectively alleviates the excessive reliance on underwater data. The core structure and workflow of this invention are described in detail below.
[0035] In some embodiments, the overall architecture and module composition of the diffusion model are as follows: Figure 1As shown, the diffusion process is described as follows: This process introduces an underwater physical imaging model into the diffusion process, replacing the previous noise-adding process that used Gaussian noise. This enables image degradation while maintaining the image's consistent structure (maintaining the consistent structure is based on the fact that this application does not add Gaussian noise; by selecting a physical mode of degradation, the essence of the image is not destroyed; it can be understood as simply darkening the image). The underwater imaging formula is as follows:
[0036] (1) in Representing underwater images, A clear image is a clean picture without underwater interference. Represents a matrix of all ones. Represents a transmission diagram, while This represents the backscattering intensity. Based on the above formula (1), the formula for promoting the forward diffusion process in this invention is defined as follows:
[0037] (2) at this time, Adding noise to an image means that the images share the same dimensions. and (Right now , is a commonly used symbol in diffusion model theory, representing the original, undegraded, clear image. and The transmission diagrams and backscattering intensities represent different stages, and T1 represents the decay factor at different times t. The following formula is further derived:
[0038] (3) In some embodiments, for Figure 1 The reverse process shown in this invention employs a deep unfolding architecture, such as a deep unfolding network, for the reconstruction process. Compared to the traditional UNeT architecture's method of predicting noise, the deep unfolding network better reflects the integration with underwater physical imaging models, resulting in images that are more consistent with physical reality.
[0039] In some embodiments, the degradation model and energy form involved in the reverse process are as follows: In order to physically separate underwater image formation from the diffusion process, the problem is modeled in an optimization framework, and a cooperative energy function is derived based on the imaging model in Equation (1), which naturally incorporates the factors of image degradation.
[0040] (4) in , and These represent the regularization terms for each item, preventing overfitting. This invention uses three variables. , and To approximate the true value separately , and This leads to the following minimization problem:
[0041]
[0042] In some embodiments, the physical parameters involved in the reverse process are optimized as follows: This invention introduces three physical parameters at different stages. , and The exact optimization strategy is as follows:
[0043] optimization Given the current clear image J The transmission map of the (k-1)th iteration and backscattering intensity ,variable J You can update it in the following way: (6) This is a clear image. J The objective function is to find an optimal one. J This makes the current estimate and In order to obtain an image as close as possible to the observed image. P Next, we use a strong mapping module to build... and An approximate mapping between them is named J-GFMM and denoted as . The mapping capability of the strong mapping module is based on the stacking capability of Transformer blocks, as shown in the following expression:
[0044] (7) Auxiliary variables It can be updated as follows: (8) at this time, It is a learnable parameter that can be adaptively adjusted and optimized.
[0045] optimization : Give J k N k-1 And T, variable T can be updated to: (9) Similar to equation (6), we utilize the T-GFMM module and denote it as... The expression is as follows: (10) Auxiliary variables It can be updated as follows: (11) at this time, It is a learnable parameter that can be adaptively adjusted and optimized.
[0046] optimization : Give J k T k And N, the update of this variable N can be as follows: (12) Similar to equation (6), we utilize the N-GFMM module and denote it as : (13) Our auxiliary variable can be calculated as follows: (14) At this time, v k It is a learnable parameter.
[0047] In some embodiments, the multi-teacher supervision framework is specifically as follows: Most current underwater image enhancement methods heavily rely on paired datasets; however, obtaining high-quality paired data in real-world underwater scenarios is extremely difficult. To alleviate this over-reliance on paired data, this paper proposes a multi-teacher collaborative supervision mechanism based on dynamic pseudo-label generation. This method constructs three teacher models with different responsibilities—Explorer, Stabilizer, and Memorizer—responsible for exploring new features, maintaining training stability, and preserving historical knowledge, respectively. These three models work collaboratively (e.g., using NR-IQA to select the model with the highest quality score for supervision) to jointly guide the learning process of the student model, thereby achieving more robust and efficient underwater image enhancement even without paired real data.
[0048] like Figure 2 The diagram shows a visual representation of the supervision method, with the left side representing existing technology supervision methods and the right side representing the supervision method proposed in this invention.
[0049] During the training process, the model uses both labeled and unlabeled images. Labeled images are directly supervised, while unlabeled images are supervised by generating pseudo-labels through a multi-teacher network and then fed back to the student model.
[0050] Specifically, the Explorer: Among all teachers, the Explorer is the only one who needs to self-train to update the model. The training dataset is consistent with the student model source, but has undergone data augmentation to become more difficult samples. The update form can be represented as:
[0051] (15) in It is the raw data, and This is the enhanced data. The model parameters represent the explorer. Represents the loss function. This represents the explorer model function.
[0052] The memory recorder stores the best past performance of the student model, effectively ensuring the model's stability. The update mechanism is as follows:
[0053] (16) in This represents the parameters of the student model. The representative evaluation metric (which can be updated based on the highest PSNR+SSIM performance, and the memory set is updated if the performance of the validation set improves), and This represents the parameters of the memory model.
[0054] Stables: The stables model is updated using the EMA update mechanism, as shown in the following expression: (17) Here, This represents the model parameters of the stabilizer, and It is a momentum parameter.
[0055] In some embodiments, an underwater image enhancement apparatus is also provided, comprising: a processor having a computer program stored thereon; wherein, when the computer program is executed by the processor, the underwater image enhancement method as described in this application is implemented.
[0056] In some embodiments, an underwater robot vision system is also provided, comprising: an image acquisition module for acquiring raw underwater images; an image processing module, including the underwater image enhancement device of this application, for receiving and enhancing the raw underwater images; and a control module for making navigation or target recognition decisions based on the enhanced images.
[0057] In some embodiments, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, which, when executed, implements the underwater image enhancement method of this application.
[0058] The following will provide some specific application examples of the technical solutions in this application.
[0059] Example 1: Vision Enhancement System for Underwater Robots In marine exploration or underwater inspection missions, cameras mounted on autonomous underwater vehicles (AUVs) often acquire low-quality images due to factors such as turbid water and reduced light intensity. This invention can be integrated into the real-time image processing module of an AUV: first, it uses an underwater physics prior model to perform structure-preserving degradation simulation on the input image, and then performs high-quality reconstruction using a deep unfolding network; simultaneously, a multi-teacher supervised framework continuously optimizes the model without paired real data, improving its generalization ability to different water bodies (such as freshwater lakes, nearshore areas, and deep seas), thereby providing clear and reliable visual input for navigation, target recognition, and environmental mapping.
[0060] Example 2: Underwater archaeology and ecological monitoring image restoration platform: Images captured by researchers during underwater archaeology or coral reef ecological monitoring often suffer from color distortion and low contrast, and it is difficult to obtain "clean-degraded" paired samples of the same scene. This invention can be deployed as an offline / online image enhancement platform: after users upload original underwater photos, the system generates reasonable degradation paths based on physical priors and performs inverse enhancement through a student model; a multi-teacher mechanism (explorers discover new degradation patterns, stabilizers suppress artifacts, and memoryers retain historical knowledge) ensures that the enhancement results are both realistic and rich in detail, significantly improving the accuracy of subsequent manual interpretation or AI analysis.
[0061] Example 3: Underwater video live streaming quality enhancement plugin: For real-time video applications such as live underwater streaming and aquaculture monitoring, this invention can be packaged as a lightweight video enhancement plugin. This plugin runs on edge devices (such as underwater cameras or shore-based servers): it replaces the traditional end-to-end model that relies on large amounts of paired data with a physically guided degradation-recovery process, significantly reducing training costs; simultaneously, the dynamic pseudo-label mechanism allows the model to continuously learn from newly acquired unlabeled videos after deployment, adapting to imaging changes under different time, depth, and water quality conditions, achieving long-term stable high-definition underwater video output.
[0062] It is understood that the core innovations of this application include, but are not limited to, the following three aspects: 1. Structure preservation degradation modeling based on underwater physical priors: This application abandons the Gaussian noise addition method that destroys the image structure in the traditional diffusion model, and introduces a diffusion model based on underwater physical imaging priors that conforms to the real underwater imaging mechanism to perform controllable degradation of the image. While simulating the underwater degradation process, it effectively preserves the structural and semantic information of the original image.
[0063] 2. Deep Deployment-Driven Reverse High-Quality Recovery Architecture: A physically interpretable deep unfolded network is constructed. By deriving the mathematical relationships between key variables in the degradation process and designing a parameter estimation function to approximate the solution, accurate and stable inverse reconstruction from degraded images to high-quality underwater images is achieved.
[0064] 3. A dynamic pseudo-labeling supervision mechanism involving multiple teachers: We propose a multi-teacher supervision framework that does not require paired data. We design three types of teacher models with complementary roles (explorer, stabilizer, and memorizer) to jointly guide student model training by dynamically generating reliable pseudo-labels. This significantly reduces the dependence on scarce real underwater paired data and improves the generalization ability and robustness of the model in complex real-world scenarios.
[0065] These three technologies are organically integrated to jointly solve key problems in current underwater image enhancement methods, such as structural distortion, reliance on paired data, and weak generalization ability, providing an efficient, realistic, and deployable solution for practical underwater vision applications.
[0066] In summary, the underwater image enhancement method based on diffusion restoration proposed in this invention combines a diffusion model with a learnable depth unfolding architecture based on prior underwater physical imaging, and employs a multi-teacher supervision framework to achieve accurate underwater enhancement. Compared with existing technologies, it has the following significant technical effects and advantages: 1. Traditional diffusion models typically rely on adding Gaussian white noise layer by layer to construct the forward degradation process. However, this noise addition method lacks a physical characterization of underwater imaging characteristics and easily damages key structural information such as image edges and textures, leading to distortion or blurring of subsequent reconstruction results. This invention, starting from the physical laws of underwater light propagation, constructs a degradation model incorporating core factors such as wavelength-selective attenuation, backscattering, and forward scattering. This model accurately simulates phenomena such as color shift, contrast reduction, and detail blurring in real underwater environments. This degradation process not only more closely resembles the actual imaging mechanism but also strictly preserves the geometric structure and semantic layout of the original image during degradation, laying a solid foundation for subsequent high-quality reconstruction.
[0067] 2. This invention designs a deep unfolding architecture based on the idea of optimization iteration, transforming the underwater image restoration problem into a series of learnable approximate solution steps. By systematically deriving the coupling relationships between key variables such as illumination, transmission map, and background light in the degradation model, a solution formula with clear physical meaning for the inverse problem is established. Based on this, lightweight yet efficient fitting modules (such as adaptive residual correction units and cross-scale feature fusion modules) are designed to perform end-to-end approximation of complex nonlinear relationships. This architecture not only inherits the interpretability and stability of traditional model-driven methods but also fully utilizes the expressive power of deep learning, achieving coordinated restoration of color, contrast, and detail without relying on large amounts of labeled data, significantly improving the visual quality and realism of enhanced images.
[0068] 3. To address the practical bottleneck of obtaining "clean-degraded" image pairs in real underwater scenarios, this invention proposes a self-supervised learning paradigm that does not require paired data. This framework comprises three complementary teacher models: an explorer model responsible for discovering potential degradation patterns and novel features in unlabeled data; a stabilizer model focused on suppressing noise interference and artifact generation during training to ensure output consistency; and a memory model that retains valid priors accumulated during historical training through a knowledge distillation mechanism. These three models are dynamically weighted and fused to generate high-quality pseudo-labels, jointly supervising the learning process of the student model. This mechanism not only significantly reduces the dependence on expensive paired datasets but also enhances the model's generalization ability and robustness under different water conditions, lighting conditions, and turbidity levels.
[0069] In summary, this invention constructs an efficient, realistic, and practical underwater image enhancement system through a three-pronged approach: physically guided degradation modeling, deep unrolling inverse solving, and multi-teacher self-supervised learning. This method not only fundamentally solves the structural distortion problem caused by the use of Gaussian noise in traditional methods but also overcomes the high dependence of existing deep learning models on large-scale paired data. Simultaneously, it ensures that the enhancement results highly match the real underwater scene in terms of color, detail, and overall appearance. Therefore, this invention has broad application value and engineering potential in various practical scenarios such as marine exploration, underwater robot vision, ecological monitoring, underwater archaeology, and underwater live streaming.
[0070] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0071] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0074] The background section of this invention may include background information about the problems or environment in which the invention is being developed, and is not necessarily a description of prior art. Therefore, the content included in the background section does not constitute an admission of prior art by the applicant.
[0075] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.
Claims
1. A multi-teacher supervised underwater image enhancement method based on diffusion restoration, characterized in that, This includes the model training phase and the image enhancement phase; The model training phase employs a multi-teacher supervision framework to train the student model. This framework includes an explorer teacher model, a stabilizer teacher model, and a memory teacher model. The pseudo-labels generated by these teacher models collectively guide the training of the student model, reducing reliance on paired data. The image enhancement phase includes the following steps: Input the underwater image to be enhanced, and construct a diffusion model based on underwater physical imaging priors. The diffusion model is used to simulate the degradation of the underwater image in the underwater environment and to preserve the structural information of the underwater image during degradation. Based on the diffusion model, an inverse reconstruction of the underwater image is performed using a learnable deep unfolding architecture to obtain an enhanced, clear image. In the reverse recovery, the deep unfolding architecture optimizes multiple physical parameters in an iterative and alternating manner, wherein the value of each physical parameter is estimated by a corresponding parameter estimation function; Output a clear image after inverse restoration and enhancement.
2. The underwater image enhancement method according to claim 1, characterized in that, The underwater physical imaging prior is based on the attenuation and scattering effects of light underwater, and multiple physical parameters including clear images. J Transmission diagram, which characterizes the ability of light to transmit. T and the backscattering intensity characterizing the background scattered light. N .
3. The underwater image enhancement method according to claim 2, characterized in that, The diffusion model is given by the formula: Definition, where P The underwater image, It represents a matrix of all 1s.
4. The underwater image enhancement method according to claim 3, characterized in that, The reverse recovery process is performed by the depth unfolding architecture iteratively estimating the sharpened image. J The transmission map T and the backscattering intensity N To achieve this; wherein, the objective of the iterative estimation is to minimize the underwater image. P The difference between the reconstructed image and the currently estimated sharp image. J The transmission map T and the backscattering intensity N Calculated using the diffusion model.
5. The underwater image enhancement method according to claim 2, characterized in that, The iteration and alternation method includes: in each iteration, the depth unfolding architecture sequentially updates the sharpened image. J The transmission map T and the backscattering intensity N The estimated value is such that when updating one parameter, the current estimated values of the other parameters are fixed.
6. The underwater image enhancement method according to claim 5, characterized in that, Update the clear image J The parameter estimation function used takes as input at least the underwater image and the currently estimated transmission map. T and the currently estimated backscattering intensity N .
7. The underwater image enhancement method according to claim 5, characterized in that, Update the transmission map T The parameter estimation function used takes as input at least the underwater image and the currently estimated sharp image. J and the currently estimated backscattering intensity N .
8. The underwater image enhancement method according to claim 5, characterized in that, Update the backscattering intensity N The parameter estimation function used takes as input at least the underwater image and the currently estimated sharp image. J and the currently estimated transmission map T .
9. The underwater image enhancement method according to claim 1, characterized in that, The parameter estimation function is implemented by a learnable neural network module.
10. An underwater image enhancement device, characterized in that, include: A processor having a computer program stored thereon; wherein, when the computer program is executed by the processor, the underwater image enhancement method as described in any one of claims 1 to 9 is implemented.
11. An underwater robot vision system, characterized in that, include: The image acquisition module is used to acquire raw underwater images; The image processing module includes the underwater image enhancement device as described in claim 10, for receiving and enhancing the original underwater image; the control module makes navigation or target recognition decisions based on the enhanced image.
12. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed, it implements the underwater image enhancement method as described in any one of claims 1 to 9.