Zero sample image restoration method based on underwater polarization robot

Abstract

The invention relates to a zero sample image restoration method and system based on an underwater polarization robot. Image restoration is realized by acquiring a transmission image, a reflectivity component and an illumination component. Constructing a zero sample underwater polarization image restoration model, and constructing an optimization objective function of the zero sample underwater polarization image restoration model under the non-uniform illumination condition based on a regular item and a control item; solving the optimization objective function through an alternating direction multiplier algorithm, and iteratively optimizing the zero sample underwater polarization image restoration model based on the optimization objective function to obtain an optimized zero sample underwater polarization image restoration model; the mosaic images of multiple polarization angles are input into the zero-sample underwater polarization image restoration model, and the objective function is solved and optimized through the alternating direction multiplier algorithm, so that the performance of the underwater imaging system can be improved, and the application prospect of the underwater vision technology in multiple industries can be expanded.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of underwater image restoration technology, and in particular to a zero-sample image restoration method and system based on an underwater polarization robot. Background Art

[0002] Underwater imaging technology plays a vital role in ocean exploration, environmental monitoring, underwater robotics, and other fields. With the growing demand for underwater detection, traditional underwater imaging technology faces numerous challenges. Due to water turbidity, light scattering, and absorption, the performance of traditional optical imaging systems is significantly reduced in underwater environments, often severely impacting image quality, resulting in blurred images, loss of detail, and difficulty identifying targets. This is particularly true in complex underwater environments, where interference from light variations, turbid water, and sensor noise is even more pronounced, posing significant obstacles to tasks such as underwater detection and robotic visual perception.

[0003] To overcome these challenges, polarization imaging technology has gradually become an important research direction in the field of underwater imaging. Underwater polarization imaging can effectively identify and extract polarization information caused by factors such as surface reflection and light scattering, thereby enhancing the contrast and clarity of underwater images.

[0004] However, existing underwater polarization imaging systems still face several technical bottlenecks in practical applications, particularly in image clarity and detail recovery. Traditional underwater image restoration methods rely on various optical correction techniques, but these methods often fail to fully account for the complexities of non-uniform illumination and underwater optical media. Summary of the Invention

[0005] In view of this, the embodiments of the present application propose a zero-sample image restoration method and system based on an underwater polarization robot, aiming to solve the problems of poor underwater image quality and missing image details.

[0006] To achieve the above-mentioned purpose, an embodiment of the present application provides a zero-sample sample image restoration method based on an underwater polarization robot, comprising: obtaining a transmission map, a reflectivity component, and an illumination component of an underwater imaging image; constructing a zero-sample underwater polarization image restoration model, wherein the zero-sample underwater polarization image restoration model is constructed based on the matrix dot product of the transmission map, the reflectivity component, and the illumination component and the sum of the backscattered light; a regularization term is constructed based on the difference between the theoretical value calculated by the zero-sample underwater polarization image restoration model and the actual light intensity image; and a weighted internal prior learning factor, a first external prior learning factor, and a first external prior learning factor are used to restore the image. An empirical learning factor and a second external prior learning factor are used to construct a control term for controlling the importance of the regularization term. Based on the regularization term and the control term, an optimization objective function of a zero-shot underwater polarization image restoration model under non-uniform illumination conditions is constructed to jointly optimize the transmission map, reflectance component and illumination component. Mosaic images of multiple polarization angles are input into the zero-shot underwater polarization image restoration model, and the optimization objective function is solved by the alternating direction multiplier algorithm. The optimized transmission map, reflectance component and illumination component are output, and the underwater image is restored based on the optimized transmission map, reflectance component and illumination component.

[0007] Optionally, before obtaining the transmission map, reflectivity component and illumination component of the underwater polarization angle independent image, the method also includes: obtaining mosaic images of four polarization angles of 0°, 45°, 90° and 135° through a polarization camera carried by an underwater robot, demosaicing the mosaic images of the four polarization angles to generate a first polarization angle independent image and a second polarization angle independent image, wherein the first polarization angle independent image and the second polarization angle independent image respectively represent polarization images captured at polarization angles corresponding to maximum intensity and minimum intensity of background scattered light; calculating the sum of the first polarization angle independent image and the second polarization angle independent image to obtain an actual light intensity image; and decomposing the mosaic images of the four polarization angles to obtain the corresponding transmission map, reflectivity component and illumination component.

[0008] Optionally, before constructing the control term for controlling the importance of the regularization term based on the weighted internal prior learning factor, the first external prior learning factor and the second external prior learning factor, the method further includes: constructing a weight matrix based on the local variance weight matrix and the gradient feature weight matrix of the illumination component, and constructing the internal prior learning factor based on the square of the L2 norm of the matrix dot product of the weight matrix and the first-order derivative of the illumination component.

[0009] Optionally, constructing a weight matrix based on the local variance weight matrix and the gradient feature weight matrix of the illumination component includes: constructing a weight matrix based on the matrix dot product of the local variance weight matrix and the gradient feature weight matrix of the illumination component, wherein the local variance weight matrix is ​​used to distinguish between uniform and complex illumination, and the gradient feature weight matrix is ​​used to maintain the details of the illumination boundary.

[0010] Optionally, the expression of the local variance weight matrix of the illumination component is:

[0011]

[0012] Among them, g r,σ (·) is a low-pass Gaussian filter, r is the filter kernel size, the values ​​of the parameters {r,σ} are set to {5,2}, and i represents the illumination component;

[0013] W2(i)=1-max(P h (i),P v (i))

[0014] in, and Represent the weights of the lighting component on the horizontal and vertical gradients respectively.

[0015] Optionally, the expression of the optimization objective function is:

[0016]

[0017] Where I represents the illumination component, R represents the reflectivity component, and t represents the transmission map. represents the matrix dot product, ψ1(I), ψ2(R) and ψ3(t) represent the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively. α, β and λ represent the weights of the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively.

[0018] Optionally, before solving the optimization objective function by using the alternating direction multiplier algorithm, the method further comprises: constructing an augmented Lagrangian function based on the optimization objective function introducing the first auxiliary variable and the second auxiliary variable.

[0019] Optionally, the augmented Lagrangian function is expressed as:

[0020]

[0021] in, To include T1 and T2, T1 represents the first auxiliary variable, T2 represents the second auxiliary variable, It includes M1 and M2, where M1 represents the first Lagrange multiplier, M2 represents the second Lagrange multiplier, and μ represents the weighting factor.

[0022] Optionally, the method of solving the optimization objective function by the alternating direction multiplier algorithm includes: solving the matrix equation by the preconditioned conjugate gradient method to obtain the illumination component; minimizing the objective function by a closed-form solution to update the reflectance component; detecting the first noise level in real time by the image noise estimation algorithm, and denoising the first noise level of the reflectance component by using the pre-trained denoising network FFDNet to obtain a first auxiliary variable; obtaining a transmission map by solving a group of linear equations; detecting the first noise level in real time by the image noise estimation algorithm, and denoising the transmission map and the second noise level by using the pre-trained denoising network FFDNet to obtain a second auxiliary variable; iteratively executing the above steps to update the first Lagrange multiplier and the second Lagrange multiplier, and terminating the iteration if the optimization objective function converges or the number of iterations exceeds the maximum number of iterations.

[0023] To achieve the above-mentioned objectives, the present application also provides a zero-sample image restoration system based on an underwater polarization robot, comprising an underwater robot, a polarization image acquisition module and an image processing module: the polarization image acquisition module is arranged on the underwater robot, and the image processing module is communicatively connected to the polarization image acquisition module; wherein the polarization image acquisition module is used to obtain a polarization mosaic image based on the acquired underwater imaging image; and the image processing module is used to execute the zero-sample image restoration method based on the underwater polarization robot provided above.

[0024] The embodiment of the present application proposes a zero-sample image restoration method and system based on an underwater polarization robot, which obtains the transmission map, reflectivity component and illumination component of the underwater imaging image; constructs a zero-sample underwater polarization image restoration model, and constructs a regularization term based on the difference between the theoretical value calculated by the zero-sample underwater polarization image restoration model and the actual light intensity image. Based on the weighted internal prior learning factor, the first external prior learning factor and the second external prior learning factor, a control term for controlling the importance of the regularization term is constructed, and an optimization objective function of the zero-sample underwater polarization image restoration model under non-uniform lighting conditions is constructed based on the regularization term and the control term, wherein the zero-sample underwater polarization image restoration model is based on the matrix dot product and backward multiplication of the transmission map, the reflectivity component and the illumination component. The invention relates to an underwater polarization image restoration model based on the zero-sample learning method, which is constructed by combining the sum of scattered light and internal and external prior learning to effectively deal with the scattering effect and non-uniform illumination problem in underwater images; the optimization objective function is solved by the alternating direction multiplier algorithm, and the zero-sample underwater polarization image restoration model is iteratively optimized based on the optimization objective function to obtain the optimized zero-sample underwater polarization image restoration model. This application combines the zero-sample learning method to realize automatic underwater image sharpening processing for the case of unlabeled data or lack of training samples; mosaic images of multiple polarization angles are input into the zero-sample underwater polarization image restoration model, and the optimization objective function is solved by the alternating direction multiplier algorithm. This application can not only improve the performance of the underwater imaging system, but also expand the application prospects of underwater vision technology in multiple industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flowchart of a zero-sample image restoration method based on an underwater polarization robot provided in one embodiment of the present application;

[0026] Figure 2 is a schematic diagram of a zero-sample image restoration method based on an underwater polarization robot provided in one embodiment of the present application;

[0027] Figure 3 Schematic diagram of an underwater polarization imaging system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, each embodiment of the present application will be described in detail below with reference to the accompanying drawings. However, it will be understood by those skilled in the art that in each embodiment of the present application, many technical details are proposed to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present application can be implemented. The division of the following embodiments is for convenience of description and should not constitute any limitation on the specific implementation of the present application. The various embodiments can be combined and referenced with each other under the premise of no contradiction.

[0029] One embodiment of the present application proposes a zero-shot image restoration method based on an underwater polarization robot, which is applied to an electronic device, where the electronic device can be a terminal or a server. This embodiment and the following embodiments are described using a server as an example. The implementation details of the zero-shot image restoration method based on an underwater polarization robot proposed in this embodiment are described below. The following content is provided for ease of understanding only and is not required for the implementation of this solution.

[0030] The main problem we investigate is how to recover accurate illumination, reflectance, and transmission maps from underwater images, while effectively handling the light scattering and non-uniform illumination problems of underwater images. To achieve this goal, we formulate an optimization problem where the objective is to minimize the error between the reconstructed image and the observed image, while introducing regularization terms to control the smoothness of each component.

[0031] The zero-shot image sharpening method proposed in this application relies on an improved polarization imaging model that not only considers the effects of underwater scattering media but also introduces a non-uniform illumination model to enhance the quality of underwater images. Specifically, we use an optimization method that combines internal and external prior learning to effectively handle scattering effects and non-uniform illumination problems in underwater images.

[0032] Underwater robots equipped with polarization cameras can efficiently perform tasks in complex waters such as deep seas and lakes, providing clear and accurate image information, and providing powerful visual support for underwater exploration, environmental monitoring, underwater robot navigation and control, and other fields. The technology proposed in this patent can not only improve the performance of underwater imaging systems, but also expand the application prospects of underwater vision technology in multiple industries.

[0033] The underwater polarization imaging system built by this patent is based on an underwater robot with a volume of 208x204x358 mm. It can dive to a depth of 100 meters and has 5 degrees of freedom. The system is equipped with two micro-polarization array polarization cameras LUCIDPHX050S-PC and a control system. It can achieve real-time control and image transmission through cables, and can obtain 4K resolution polarization mosaic images or videos with four polarization angles in real time. System diagram Figure 3 shown.

[0034] The specific process of the zero-sample image restoration method based on the underwater polarization robot proposed in this embodiment can be as follows: Figure 1 Shown, including:

[0035] S101, obtaining a transmission map, a reflectivity component, and an illumination component of an underwater imaging image;

[0036] Specifically, before obtaining the transmission map, reflectivity component, and illumination component of the underwater polarization angle independent image, the method further includes:

[0037] Acquire mosaic images at four polarization angles of 0°, 45°, 90°, and 135° using a polarization camera carried by an underwater robot, perform demosaicing on the mosaic images at the four polarization angles, and generate a first polarization angle-independent image and a second polarization angle-independent image, wherein the first polarization angle-independent image and the second polarization angle-independent image represent polarization images captured at polarization angles corresponding to maximum and minimum intensities of background scattered light, respectively;

[0038] calculating the sum of the first polarization angle independent image and the second polarization angle independent image to obtain an actual light intensity image;

[0039] The mosaic images of the four polarization angles are decomposed to obtain the corresponding transmission maps, reflectance components and illumination components.

[0040] For example, the present application first uses the constructed underwater polarization imaging system to obtain a polarization mosaic image containing polarization information at four angles of 0°, 45°, 90°, and 135°, performs demosaicing processing to obtain images with different polarization angles, and then uses the underwater imaging model to extract preliminary transmission maps, reflectivity, and illumination components.

[0041] The underwater imaging process can be described by the Jaffe-McGlamery model, also known as the underwater imaging model. The image consists of three components: target information light, forward scattered light, and backscattered light. In real-world scenarios, forward scattering has a minor impact and can usually be ignored. Therefore, the underwater imaging model can be simplified to:

[0042] S=D+B

[0043]

[0044] Among them, S represents the image acquired by the sensor, D is the target information light, B is the backscattered light; J is the original radiance of the scene, t is the transmission map, A ∞ represents the infinite background light intensity, and the symbol represents element-by-element multiplication.

[0045] Backscattered light usually has a strong polarization effect, while the polarization degree of the target information light is close to zero. By collecting two orthogonal polarization images of the same scene, denoted as S || and S ⊥ We can get:

[0046]

[0047] Among them, S || and S⊥ Respectively represent the polarization images obtained at the polarization angles that make the backscattered light intensity reach the maximum and minimum; B || and B ⊥ represents the backscattered light components in these polarization states.

[0048] Therefore, the degree of linear polarization (DOLP) of the backscattered light is denoted as P B , which can be expressed as:

[0049]

[0050] The polarization degree of the entire scene image can be expressed as:

[0051]

[0052] Combining the first three formulas, we can solve the expression of the transmission graph t:

[0053]

[0054] A background region can be selected in the image to estimate P B and A ∞ , and then combining the first and fifth formulas, the original scene radiance J can be restored:

[0055]

[0056] However, the model described by the first formula does not take into account the non-uniform illumination degradation effect that is common in underwater imaging. To address this problem, the Retinex model is introduced and the underwater imaging model is expanded to:

[0057]

[0058] Among them, I and R represent the illumination component and reflectivity component of the scene respectively.

[0059] S102, constructing a zero-sample underwater polarization image restoration model, constructing a regularization term based on the difference between a theoretical value calculated by the zero-sample underwater polarization image restoration model and an actual light intensity image, constructing a control term for controlling the importance of the regularization term based on a weighted internal prior learning factor, a first external prior learning factor, and a second external prior learning factor, constructing an optimization objective function for the zero-sample underwater polarization image restoration model under non-uniform illumination conditions based on the regularization term and the control term, wherein the zero-sample underwater polarization image restoration model is constructed based on the matrix dot product of the transmission map, the reflectivity component, and the illumination component, and the sum of the backscattered light;

[0060] In one embodiment of the present application, before constructing a control term for controlling the importance of the regularization term based on the weighted internal prior learning factor, the first external prior learning factor, and the second external prior learning factor, the method further includes the following execution process:

[0061] A weight matrix is ​​constructed based on the local variance weight matrix of the illumination component and the gradient feature weight matrix, and an internal prior learning factor is constructed based on the square of the L2 norm of the matrix dot product of the weight matrix and the first-order derivative of the illumination component.

[0062] Specifically, in complex underwater environments, the lighting may have local non-uniform changes (such as strong light sources or shadows). To this end, the weight matrix W is used to distinguish between uniform and complex lighting areas and adjust the smoothing intensity of different lighting areas. The internal prior learning of lighting is defined as:

[0063]

[0064] in, Represents the first-order derivative. The weight matrix W consists of the local variance weight matrix W1(·) and the gradient feature weight matrix W2(·), which adjust the processing intensity of different regions based on local variance and gradient information respectively. The weight matrix W is defined as:

[0065]

[0066] In one embodiment of the present application, a weight matrix is ​​constructed based on the local variance weight matrix of the illumination component and the gradient feature weight matrix, including:

[0067] The weight matrix is ​​constructed based on the matrix dot product of the local variance weight matrix of the illumination component and the gradient feature weight matrix of the illumination component, where the local variance weight matrix is ​​used to distinguish between uniform and complex illumination, and the gradient feature weight matrix is ​​used to preserve the details of the illumination boundary.

[0068] The local variance can be used to distinguish between uniform and complex lighting. The local variance weight matrix W1(·) is defined as:

[0069]

[0070] Among them, g r,σ (·) is a low-pass Gaussian filter, r is the filter kernel size, the values ​​of the parameters {r, σ} are set to {5, 2}, and i represents the illumination component. This ensures that the weight matrix W1(·) effectively identifies different illumination areas.

[0071] Gradient-based constraints help preserve details of lighting boundaries. The gradient feature weight matrix W2(·) is defined as:

[0072] W2(i)=1-max(P h(i),P v (i))

[0073] Among them, among them, and Represents the weights of the illumination component on the horizontal and vertical gradients. If the gradient in a certain direction is stronger, the corresponding weight value is lower.

[0074] The parameters {r, σ} are set to {5, 2}, and the weight matrix W is constrained to the interval [0, 1]. This ensures that stronger processing is applied to uniformly illuminated areas, while less processing is applied to areas with complex lighting. By combining local variance and gradient information, it effectively prevents oversmoothing, preserves detail information, and adaptively adjusts the lighting component.

[0075] It is worth noting that in the descattering processing of underwater images, the reflectivity component R represents the inherent characteristics of the scene and is not affected by lighting changes; the transmission map t represents the remaining part of the light transmitted in the water medium and is affected by turbidity. In order to improve the quality of underwater images, it is necessary to estimate the reflectivity component R and the transmission map t more accurately. This application proposes an external prior learning framework that uses a large-scale external image database for denoising training without the need to construct complex paired data. This application integrates a pre-trained denoising network (FFDNet) into the optimization process of the reflectivity component and the transmission map, thereby avoiding the solution of an explicit regularization function and improving the efficiency of data processing.

[0076] Specifically, the external prior learning framework treats the estimation of reflectance and transmittance maps as an additive white noise denoising problem and uses a deep denoiser at each iteration to constrain the reflectance and transmittance maps, thereby removing noise and preserving details. This external prior learning framework improves the generalization ability under different underwater conditions by using a denoising model trained on a large-scale dataset.

[0077] This application constructs a regularization term based on the difference between the theoretical value calculated by the zero-sample underwater polarization image restoration model and the actual light intensity image. Based on the weighted internal prior learning factor, the first external prior learning factor and the second external prior learning factor, a control term for controlling the importance of the regularization term is constructed. Based on the regularization term and the control term, an optimization objective function of the zero-sample underwater polarization image restoration model under non-uniform lighting conditions is constructed.

[0078] For example, the optimization objective function of the zero-sample underwater polarization image restoration model under non-uniform illumination conditions is:

[0079]

[0080] Where I represents the illumination component, R represents the reflectivity component, and t represents the transmission map. represents the matrix dot product, ψ1(I), ψ2(R) and ψ3(t) represent the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively, α, β and λ represent the weights of the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively

[0081] S103, inputting mosaic images of multiple polarization angles into the zero-sample underwater polarization image restoration model, solving the optimization objective function through the alternating direction multiplier algorithm, and outputting the optimized transmission map, reflectivity component, and illumination component.

[0082] In one embodiment of the present application, before solving the optimization objective function using the alternating direction multiplier algorithm, the method further includes:

[0083] An augmented Lagrangian function is constructed based on the optimization objective function introducing the first auxiliary variable and the second auxiliary variable.

[0084] Among them, the expression of the augmented Lagrangian function is:

[0085]

[0086] in, To include T1 and T2, T1 represents the first auxiliary variable, T2 represents the second auxiliary variable, Including M1 and M2, M1 represents the first Lagrange multiplier, M2 represents the second Lagrange multiplier, and μ represents the penalty parameter.

[0087] Specifically, the present application adopts the framework of alternating direction method of multipliers (ADMM), and the optimization objective function is decomposed into sub-problems, and each iteration solves a sub-problem. The method constructs the augmented Lagrangian function by introducing two variables T1 = R and T2 = t.

[0088] In one embodiment of the present application, solving the optimization objective function using the alternating direction multiplier algorithm may include the following execution process:

[0089] Solve the matrix equation by preconditioned conjugate gradient method to obtain the illumination component;

[0090] The reflectivity component is updated by minimizing the objective function through a closed-form solution;

[0091] An image noise estimation algorithm is used to detect the first noise level in real time, and a pre-trained denoising network FFDNet is used to denoise the first noise level of the reflectivity component to obtain a first auxiliary variable.

[0092] By solving the linear equations, the transmission map is obtained;

[0093] An image noise estimation algorithm is used to detect the first noise level in real time, and a pre-trained denoising network FFDNet is used to denoise the transmission image and the second noise level to obtain a second auxiliary variable.

[0094] The above steps are iteratively performed to update the first Lagrange multiplier and the second Lagrange multiplier. If the optimization objective function converges or the number of iterations exceeds the maximum number of iterations, the iteration is terminated.

[0095] Specifically, the process of solving the optimization objective function may include:

[0096] Input: Polarization image S || and S ⊥ , parameter α.

[0097] Output: illumination I, reflectance R, transmission map t and degree of linear polarization (DoLP)

[0098] 1) Initialization

[0099] 2) Perform the following iterative process:

[0100] (a) Update illumination I: By solving the matrix equation, the preconditioned conjugate gradient method (PCG) is used to improve the computational efficiency.

[0101] (b) Update the reflectivity R: Minimize the objective function through a closed-form solution to avoid directly solving complex equations.

[0102] (c) Update variable T1: Use image noise estimation method to detect noise level in real time The pre-trained deep denoiser (FFDNet) is used for denoising.

[0103] (d) Update the transmission graph t: Minimize the objective function through a closed-form solution and obtain t.

[0104] (e) Update variable T2: Use image noise estimation method to detect noise level in real time The pre-trained deep denoiser (FFDNet) is used for denoising.

[0105] (f) Update the Lagrange multiplier According to the ADMM principle, the multiplier is updated through iteration. If convergence or the number of iterations is greater than the maximum number K (K = 10), the loop ends.

[0106] The step division of the above various methods is only for the purpose of clear description. During implementation, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application; adding insignificant modifications or introducing insignificant designs to the algorithm or process without changing the core design of the algorithm and process are all within the scope of protection of this application.

[0107] Based on the above method embodiments, the present application also provides a zero-sample image restoration system based on an underwater polarization robot, including an underwater robot, a polarization image acquisition module and an image processing module: the polarization image acquisition module is arranged on the underwater robot, and the image processing module is communicatively connected to the polarization image acquisition module; wherein the polarization image acquisition module obtains a polarization mosaic image based on the acquired underwater imaging image; the image processing module is used to execute the zero-sample image restoration method based on the underwater polarization robot provided in the above method embodiments.

[0108] It is worth mentioning that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0109] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and that in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A zero-sample image restoration method based on an underwater polarization robot, characterized in that: include: Obtaining a transmission map, a reflectivity component, and an illumination component of an underwater imaging image; A zero-sample underwater polarization image restoration model is constructed, wherein the zero-sample underwater polarization image restoration model is constructed based on the matrix dot product of the transmission map, the reflectivity component, and the illumination component and the sum of the backscattered light; a regularization term is constructed based on the difference between a theoretical value calculated by the zero-sample underwater polarization image restoration model and an actual light intensity image; a control term for controlling the importance of the regularization term is constructed based on a weighted internal prior learning factor, a first external prior learning factor, and a second external prior learning factor; and an optimization objective function of the zero-sample underwater polarization image restoration model under non-uniform illumination conditions is constructed based on the regularization term and the control term to jointly optimize the transmission map, the reflectivity component, and the illumination component; Mosaic images of multiple polarization angles are input into the zero-shot underwater polarization image restoration model, and the optimization objective function is solved by the alternating direction multiplier algorithm. The optimized transmission map, reflectivity component and illumination component are output, and the underwater image is restored based on the optimized transmission map, reflectivity component and illumination component.

2. The zero-sample image restoration method based on an underwater polarization robot according to claim 1, characterized in that: Before acquiring the transmission map, reflectivity component, and illumination component of the underwater polarization angle independent image, the method further includes: Acquire mosaic images at four polarization angles of 0°, 45°, 90°, and 135° using a polarization camera carried by an underwater robot, perform demosaicing on the mosaic images at the four polarization angles, and generate a first polarization angle-independent image and a second polarization angle-independent image, wherein the first polarization angle-independent image and the second polarization angle-independent image represent polarization images captured at polarization angles corresponding to maximum and minimum intensities of background scattered light, respectively; calculating the sum of the first polarization angle independent image and the second polarization angle independent image to obtain an actual light intensity image; The mosaic images of the four polarization angles are decomposed to obtain the corresponding transmission maps, reflectance components and illumination components.

3. The zero-sample image restoration method based on an underwater polarization robot according to claim 2, characterized in that: Before constructing a control term for controlling the importance of the regularization term based on the weighted internal priori learning factor, the first external priori learning factor, and the second external priori learning factor, the method further includes: A weight matrix is ​​constructed based on the local variance weight matrix of the illumination component and the gradient feature weight matrix, and an internal prior learning factor is constructed based on the square of the L2 norm of the matrix dot product of the weight matrix and the first-order derivative of the illumination component.

4. The zero-sample image restoration method based on an underwater polarization robot according to claim 1, characterized in that: The weight matrix is ​​constructed based on the local variance weight matrix and the gradient feature weight matrix of the illumination component, including: The weight matrix is ​​constructed based on the matrix dot product of the local variance weight matrix of the illumination component and the gradient feature weight matrix of the illumination component, where the local variance weight matrix is ​​used to distinguish between uniform and complex illumination, and the gradient feature weight matrix is ​​used to preserve the details of the illumination boundary.

5. The zero-sample image restoration method based on an underwater polarization robot according to claim 4, characterized in that: The expression of the local variance weight matrix of the illumination component is: W1(i)=1-min(4500*|g r,σ (i 2 )-(g r,σ (i)) 2 |,1) Among them, g r,σ (·) is a low-pass Gaussian filter, r is the filter kernel size, the values ​​of the parameters {r,σ} are set to {5,2}, and i represents the illumination component; W2(i)=1-max(P h (i),P v (i)) in, and Represent the weights of the lighting component on the horizontal and vertical gradients respectively.

6. The zero-sample image restoration method based on an underwater polarization robot according to claim 1, characterized in that: The expression of the optimization objective function is: Where I represents the illumination component, R represents the reflectivity component, and t represents the transmission map. represents the matrix dot product, ψ1(I), ψ2(R) and ψ3(t) represent the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively. α, β and λ represent the weights of the internal prior learning factor, the first external prior learning factor and the second external prior learning factor, respectively.

7. The zero-sample image restoration method based on an underwater polarization robot according to claim 4, characterized in that: Before solving the optimization objective function by the alternating direction multiplier algorithm, the method further includes: An augmented Lagrangian function is constructed based on the optimization objective function introducing the first auxiliary variable and the second auxiliary variable.

8. The zero-sample image restoration method based on an underwater polarization robot according to claim 7, characterized in that: The expression of the augmented Lagrangian function is: in, To include T1 and T2, T1 represents the first auxiliary variable, T2 represents the second auxiliary variable, Including M1 and M2, M1 represents the first Lagrange multiplier, M2 represents the second Lagrange multiplier, and μ represents the penalty parameter.

9. The zero-sample image restoration method based on an underwater polarization robot according to claim 1, characterized in that: The method of solving the optimization objective function by using the alternating direction multiplier algorithm includes: Solve the matrix equation by preconditioned conjugate gradient method to obtain the illumination component; The reflectivity component is updated by minimizing the objective function through a closed-form solution; An image noise estimation algorithm is used to detect the first noise level in real time, and a pre-trained denoising network FFDNet is used to denoise the first noise level of the reflectivity component to obtain a first auxiliary variable. By solving the linear equations, the transmission map is obtained; An image noise estimation algorithm is used to detect the first noise level in real time, and a pre-trained denoising network FFDNet is used to denoise the transmission image and the second noise level to obtain a second auxiliary variable. The above steps are iteratively performed to update the first Lagrange multiplier and the second Lagrange multiplier of the Lagrange multiplier. If the optimization objective function converges or the number of iterations exceeds the maximum number of iterations, the iteration is terminated.

10. A zero-sample image restoration system based on an underwater polarization robot, characterized in that: Including underwater robot, polarization image acquisition module and image processing module: The polarization image acquisition module is provided on the underwater robot, and the image processing module is communicatively connected with the polarization image acquisition module; Among them, the polarization image acquisition module is used to acquire original underwater polarization mosaic images; The image processing module is used to execute the zero-sample image restoration method based on the underwater polarization robot as described in any one of claims 1-9.

Images