An underwater light halo image correction method based on radial gradient distribution characteristics

Abstract

The application provides an underwater light halo image correction method based on radial gradient distribution characteristics, which comprises the following steps: constructing an imaging model of an underwater light halo image; solving the imaging model by using a Bayesian framework, introducing radial gradient distribution and hyper-Laplacian distribution to constrain the optimization process, and obtaining a target equation by using maximum posterior probability in combination with probability distribution of different image layers; then, using an alternating iterative updating method to decompose the target function into a plurality of sub-problems for solving; and performing brightness adjustment, color correction, contrast enhancement and other post-processing on the estimated reflection layer by using an adaptive brightness scaling factor and a multi-scale gradient domain contrast stretching to obtain a final clear underwater image after light halo correction. The application firstly considers a real complex light scene during underwater operation of a UUV, can effectively remove the light halo effect caused by an artificial strong light source, reduces serious interference caused by the light halo effect on subsequent computer vision tasks, and significantly improves underwater image quality.

Description

Technical Field

[0001] This invention relates to the fields of marine environmental sensing, digital image processing, and underwater image sharpening, and more particularly to an underwater halo image correction method based on radial gradient distribution characteristics. Background Technology

[0002] Autonomous underwater robots (AUVs) have become crucial for acquiring seabed information during seabed observation, and seabed imaging technology has developed rapidly. However, due to the strong absorption and scattering of light by water, the underwater environment suffers from severe low-light conditions. Underwater robots typically require artificial lighting devices to enhance their perception of the surrounding environment. Because of the large amount of suspended matter in the water, strong artificial light sources cause significant backscattering, creating intense light spots in the camera foreground or on the seabed. This results in non-uniform lighting in the captured underwater images, significantly impacting subsequent visual tasks such as feature extraction and edge detection. While existing underwater image enhancement technologies have advanced rapidly, few have considered the severe impact of artificial light sources in real-world underwater environments, resulting in limited practicality. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, this invention proposes an underwater halo image correction method based on radial gradient distribution characteristics. This method has excellent effect and robustness in the sharpening of underwater images with halos introduced by artificial strong light sources.

[0004] The technical means employed in this invention are as follows:

[0005] An underwater halo image correction method based on radial gradient distribution characteristics includes the following steps:

[0006] Step 1: Obtain the underwater halo image S to be processed;

[0007] Step 2: Construct an underwater halo image imaging model to describe the physical process of underwater image imaging in artificial light source-assisted lighting scenarios. The underwater halo image imaging model is as follows:

[0008]

[0009] Where S represents the underwater halo image to be processed, R is the reflectance map, I is the illuminance map, V is the halo map, and the operators are... This indicates element-wise multiplication;

[0010] Step 3: Solve the underwater halo image imaging model using a Bayesian framework, introduce radial gradient distribution and hyperLaplace distribution to constrain the optimization results, and obtain the objective function based on the maximum a posteriori probability and the probability distribution model of different image layers;

[0011] Step 4: Decompose the objective function into several sub - problems for solution by using the alternating iteration update method, and calculate the halo layer V, the illumination layer I, and the reflection layer R. The non - convex optimization sub - problems involved in the sub - problems are solved by the iteratively reweighted least squares method;

[0012] Step 5: Post - process the reflection layer R. Adjust the brightness of the reflection layer by an adaptive brightness scaling factor, and combine multi - scale gradient domain processing to perform color correction and detail enhancement on the reflection layer R, so as to obtain the final clear underwater image Z after halo correction.

[0013] Furthermore, a Bayesian framework is used to solve the underwater halo image imaging model, including:

[0014] Step 3.1: Respectively obtain the reflection map R, the illumination map I, and the halo map V by maximizing the posterior probability. The posterior probability is solved according to the following formula:

[0015]

[0016] where, P(R, I, V|S) is the posterior probability distribution, P(S|R, I, V) is the likelihood function, P(R), P(I), and P(V) are the probability distributions of the reflection map R, the illumination map I, and the halo map V respectively, and P(S) is the probability distribution of the input underwater halo image S.

[0017] Furthermore, introduce the hyper - Laplacian distribution and the radial gradient distribution to constrain the optimization result. According to the maximum posterior probability and the probability distribution models of different image layers, the objective function is obtained, including:

[0018] Step 3.2: Introduce the image hyper - Laplacian distribution and the radial gradient distribution to constrain the optimization result, and we have:

[0019]

[0020] where, is the smoothing constraint term, is the first estimated value obtained by using , is the L p norm, 0 < p < 1 is the hyper - Laplacian distribution constraint, D i (i = 1, 2) is the matrix corresponding to the difference operator, and P(I) is the probability distribution of the illumination map I;

[0021]

[0022] where, P(R) is the probability distribution of the reflection map R;

[0023]

[0024] in, The radial gradient smoothing constraint term is represented by the following: Considering that the halo layer has radial symmetry with the optical center as the center and is smooth, a discrete, uniform sampling method is adopted for the halo map V: (V(r) t ),r t ∈S r ), where S r ={r0,r1,...,r n-1 The image is divided into several sectors along these discrete radii, i.e., r m Let m be the inner diameter of sector m, and δr represent the sector width. Then:

[0025]

[0026] in, This represents the data fidelity term, used to constrain the proposed model. The difference between the input image S and the input image S;

[0027] Substituting equations (3), (4), (5), and (6) into equation (2), we obtain the objective equation:

[0028]

[0029] Among them, λ1, λ2, λ3 and λ4 are constant coefficients that control the importance of different terms.

[0030] Furthermore, the objective function is decomposed into several sub-problems by using an alternating iterative update method, and the halo layer V, illuminance layer I, and reflection layer R are calculated, including:

[0031] Step 4.1: Initialize the halo parameter V 0 =255, weight parameters And the number of iterations k = 0;

[0032] Step 4.2: Treat R and I as a whole {R,I} for estimation, and decompose the objective equation (7) to obtain formulas (8) and (9):

[0033]

[0034]

[0035] Step 4.3, for a given V k (k=0,1,2,...), calculate subproblem (8), and optimize using the iterative reweighted least squares method. In each iteration, problem (8) is transformed into a quadratic problem as shown in formula (10), and solved using the gradient descent method (10). The calculated {R,I} (k+1) Substitute into (11) to calculate the weight parameters

[0036]

[0037]

[0038] Step 4.4: Based on the calculated {R,I} (k+1) The updated V is obtained by solving equation (9) using the gradient descent method. (k +1) ;

[0039] Step 4.5: Update the iteration count k = k + 1; alternately execute steps 4.3 and 4.4 until ||{R,I} (k+1) -{R,I} k If ||<ε, the final halo map V and {R,I} are obtained;

[0040] Step 4.6: Based on the calculated {R,I} and objective equation (7), in order to obtain the two unknowns R and I, the alternating iteration method is used to obtain two sub-problems as shown in formulas (12) and (13):

[0041]

[0042]

[0043] Step 4.7: Estimate the initial illuminance map And initialize the illuminance map The initial weight parameter is W. i 0 =1, iteration number k=0;

[0044] Step 4.8, for a given I k (k=0,1,2,...), using the iterative reweighted least squares method, the non-convex subproblem (12) is computed to obtain R. (k+1) ;

[0045] Step 4.9: Based on the calculated R (k+1) Using the iterative reweighted least squares method, the nonconvex subproblem (13) is computed to obtain I. (k+1) ;

[0046] Step 4.10: Update the iteration count, i.e., k = k + 1; alternately execute steps 4.8 and 4.9 until ||R (k+1) -R k ||<ε, thus obtaining the final reflection map R and illuminance map I.

[0047] Furthermore, post-processing is performed on the reflective layer R. The brightness of the reflective layer is adjusted using an adaptive brightness scaling factor, and color correction and detail enhancement are performed on the reflective layer R using multi-scale gradient domain processing, including:

[0048] Step 5.1: Extract the V channel luminance information of the reflection map R and the product map {R,I} in the HSV color space to obtain R. val and {R,I} val ;

[0049] Step 5.2: Based on the brightness information, design an adaptive brightness scaling factor f, and calculate the adaptive brightness map R′ using formula (14). val ;

[0050]

[0051] The adaptive brightness scaling factor f is defined as:

[0052]

[0053] Among them, GF {R,I} This represents a guided filter, where β is a constant parameter controlling brightness.

[0054] Step 5.3: Convert the adaptive luminance map from HSV space to RGB space to obtain an enhanced reflectance map R′. Perform WLS-based multi-scale image decomposition on R′ to obtain a base layer representing low-frequency color information and two detail layers d representing texture information at different scales. i (i = 1, 2);

[0055] Step 5.4: Perform image color correction on the base layer using a Gaussian distribution-based linear stretching model.

[0056] First, the maximum and minimum values ​​of the base layer in each color channel are obtained using formulas (16) and (17):

[0057]

[0058]

[0059] Where, c∈{r,g,b}, and These represent the mean and standard deviation of the base layer in channel c, respectively, and μ is a parameter controlling the dynamic range of the image.

[0060] Color correction of the base layer is performed using equation (18):

[0061]

[0062] Step 5.5, in two detail layers d i Gradient domain contrast stretching of different magnitudes is performed on (i=1,2), specifically:

[0063] First, divide the two detail layers d i (i = 1, 2) are all transformed to the gradient domain to obtain

[0064] Secondly, multi-scale gradient domain contrast stretching is performed according to equation (19):

[0065]

[0066] Where, ω i To control the non-negativity of the enhancement amplitude, The average value of the transmittance t, ω1≤ω2,

[0067] Finally, by solving the Poisson equation, the detail layer d after contrast stretching is obtained. i (i = 1, 2);

[0068] Step 5.6: Through the above steps, the final clear underwater image Z with halo correction can be obtained as follows:

[0069] Z = base cr +d′1+d′2 (20).

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] This invention proposes a physical imaging model capable of describing underwater halo images under artificial light source-assisted illumination scenarios, and based on this model, proposes an underwater halo image correction method based on radial gradient distribution characteristics. This method effectively avoids the common defects of existing underwater image sharpening methods, such as overexposure in the luminous area and under-enhancement of the dark background area, when faced with non-uniform illumination. While removing halos, it significantly improves the quality of underwater images, demonstrating excellent effectiveness and robustness, thus laying a theoretical and technical foundation for subsequent visual tasks such as panoramic seabed observation. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart of the underwater halo image correction method based on radial gradient distribution characteristics according to the present invention.

[0074] Figure 2 The image shown is an underwater halo image input in the example.

[0075] Figure 3 The halo pattern obtained by the method of the present invention in the embodiment is shown.

[0076] Figure 4 The illuminance map is obtained by the method of the present invention in the embodiment.

[0077] Figure 5 The reflection map obtained by the method of the present invention in the embodiment is shown.

[0078] Figure 6 The image shown is a brightness-adjusted reflection map obtained by the method of the present invention in this embodiment.

[0079] Figure 7 The diagram shows the final output of the method of the present invention in the embodiment. Detailed Implementation

[0080] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0081] like Figure 1 As shown, this invention provides an underwater halo image correction method based on radial gradient distribution characteristics, comprising the following steps:

[0082] Step 1: Obtain the underwater halo image S to be processed, such as... Figure 2 The image shown is an underwater halo image input in an embodiment of the present invention.

[0083] Step 2: Construct an underwater halo image imaging model that describes the physical process of underwater image imaging under artificial light source-assisted lighting scenarios. Define the underwater halo image as a multiplicative combination of the illuminance layer, reflection layer, and halo layer. Specifically, this includes:

[0084] Step 2.1: Based on the similarity in brightness distribution characteristics between vignetting images and halo images, and drawing on the formation principle of vignetting images, an imaging model is constructed to describe the physical process of underwater halo image imaging in artificial light source-assisted lighting scenarios. The underwater halo image is defined as a multiplicative combination of the reflection layer R, the illuminance layer I, and the halo layer V, i.e.

[0085]

[0086] Among them, S represents the underwater halo image to be processed, and the operator represents element-wise multiplication. In this formula, only S is the known quantity, and R, I, and V are all unknown quantities. The goal of the algorithm is to obtain R as accurate as possible, that is, the clear image after halo correction, and this process is an ill-posed estimation problem.

[0087] Step 2.2: Estimate the light source center point (x0, y0) of the underwater halo image. First, use a Gaussian filter to remove the texture in the original image and extract the brightness pattern of the image; then calculate the classical centroid of the image as the light source center point.

[0088] Step 3: Solve the imaging model using the Bayesian framework, innovatively introduce the radial gradient distribution and the hyper-Laplacian distribution to constrain the optimization process, and use the maximum a posteriori probability (MAP) to combine the probability distributions of different image layers to obtain the target equation. Specifically, it includes:

[0089] Step 3.1: According to formula (1), when the underwater halo image is known, the reflection layer R, the illumination layer I, and the halo layer V are obtained respectively by maximizing the posterior probability. According to the Bayesian formula, we can get:

[0090]

[0091] Since the underwater halo image S is known and P(S) is a constant, we can get: P(R, I, V|S) ∝ P(S|R, I, V)P(R)P(I)P(V). According to the distribution characteristics of different image layers, the image hyper-Laplacian distribution and the radial gradient distribution are innovatively introduced to constrain the optimization result, and we have:

[0092]

[0093] Among them, is the smoothing constraint term, is the first estimated value obtained by using . This term is used to reduce the difference between the estimated illumination layer I and the initial illumination layer to avoid the scale scaling problem in the process of estimating the illumination layer I; is the L p norm, 0 < p < 1, that is, the hyper-Laplacian distribution constraint, which reflects the prior of the gradient sparsity of the image and is used to retain the structural information in the illumination layer. In this invention, p is set to 0.8 according to experience; D i (i = 1, 2) are the matrices corresponding to the first-order and second-order difference operators.

[0094]

[0095] The meaning is the same as above. This is to better preserve details and edge information in the reflectance image, considering that natural images follow a super-Laplacian distribution.

[0096]

[0097] This represents the radial gradient smoothing constraint term. This design takes into account the characteristic that the halo layer has radial gradient symmetry with the optical center as the origin, and that it is smooth.

[0098]

[0099] This represents the data fidelity term. This term is used to constrain the proposed model. The difference between the observed image S and the image S.

[0100] Substituting equations (3), (4), (5), and (6) into equation (2), we obtain the objective equation:

[0101]

[0102] Among them, λ1, λ2, λ3 and λ4 are constant coefficients that control the importance of different terms and can be manually adjusted according to different situations of the actual image.

[0103] Step 4: Decompose the objective function into several subproblems using an alternating iterative update method. The non-convex optimization subproblems are solved using the Iterative Reweighted Least Squares (IRLS) method. Specifically, this includes:

[0104] Step 4.1: Initialize parameters, halo parameter V 0 =255, weight parameter The number of iterations k = 0.

[0105] Step 4.2: The objective equation can be solved using an alternating iterative update method. To estimate R, I, and V, R and I can first be considered as a whole {R,I} for estimation. Then the objective equation (7) is divided into two subproblems as follows:

[0106]

[0107]

[0108] Step 4.3, for a given V k(k = 0, 1, 2, ...), calculate subproblem (8). Since the subproblem contains the p-norm, it is a non-convex optimization problem, so the iterative reweighted least squares (IRLS) method is used for optimization. In each iteration, problem (8) can be transformed into the following quadratic problem (10), which is solved using the gradient descent method (10), and the calculated {R, I} (k+1) Substitute into (11) to calculate the weight parameter W i (k+1) :

[0109]

[0110]

[0111] Step 4.4: Based on the calculated {R,I} (k+1) The updated V is obtained by solving equation (9) using the gradient descent method. (k+1) .

[0112] Step 4.5: Update the iteration count, i.e., k = k + 1; alternately execute steps 4.3 and 4.4 until ||{R,I} (k+1) -{R,I} k ||<ε, thus obtaining the final halo map V and {R,I}.

[0113] Step 4.6: Based on the calculated {R,I} and objective equation (7), in order to obtain the two unknowns R and I, the alternating iteration method is used to obtain two sub-problems as follows:

[0114]

[0115]

[0116] Step 4.7: Estimate the initial illuminance layer And initialize the illumination layer The initial weight parameters involved are The number of iterations k = 0.

[0117] Step 4.8, for a given I k (k=0,1,2,...), using the IRLS method, the non-convex subproblem (12) is calculated to obtain R. (k+1) .

[0118] Step 4.9: Based on the calculated R (k+1) Using the IRLS method, the non-convex subproblem (13) is calculated to obtain I. (k+1) .

[0119] Step 4.10: Update the iteration count, i.e., k = k + 1; alternately execute steps 4.8 and 4.9 until ||R (k+1)-R k ||<ε, thus obtaining the final reflective layer R and illuminance layer I. For example... Figure 3-5 The images shown are the halo diagram, illuminance diagram, and reflection diagram obtained from the embodiments of the present invention.

[0120] Step 5: Based on the adaptive brightness scaling factor, an enhanced reflection layer is obtained. Post-processing, including brightness adjustment, color correction, and contrast enhancement, is performed on the estimated reflection layer using the adaptive brightness scaling factor and multi-scale gradient domain contrast stretching to obtain a clear underwater image with halo correction. Specifically, this includes:

[0121] Step 5.1: Extract the V channel luminance information of the reflection map R and the product map {R,I} in the HSV color space to obtain R. val and {R,I} val ;

[0122] Step 5.2: Based on the brightness information, design an adaptive brightness scaling factor f, and calculate the adaptive brightness map R′ using formula (14). val ;

[0123]

[0124] The adaptive brightness scaling factor f is defined as:

[0125]

[0126] Among them, GF {R,I} This represents the guided filter, where β is a constant parameter controlling brightness.

[0127] Step 5.3: Convert the adaptive brightness map from HSV space to RGB space to obtain an enhanced reflectance map R′. Perform multi-scale image decomposition on R′ based on WLS (Weighted Least Squares) to obtain a base layer representing low-frequency color information and two detail layers d representing texture information at different scales. i (i = 1, 2);

[0128] Step 5.4: Perform image color correction on the base layer using a Gaussian distribution-based linear stretching model.

[0129] First, the maximum and minimum values ​​of the base layer in each color channel are obtained using formulas (16) and (17):

[0130]

[0131]

[0132] Where, c∈{r,g,b}, and These represent the mean and standard deviation of the base layer in the c channel, respectively, and μ is a parameter that controls the dynamic range of the image, set to 2.3.

[0133] Color correction of the base layer is performed using equation (18):

[0134]

[0135] Step 5.5, in two detail layers d i Gradient domain contrast stretching of different magnitudes is performed on (i=1,2). Specifically:

[0136] First, divide the two detail layers d i (i = 1, 2) are all transformed to the gradient domain to obtain

[0137] Secondly, multi-scale gradient domain contrast stretching is performed according to equation (19):

[0138]

[0139] Where, ω i To control the non-negativity of the enhancement amplitude, Let ω1 be the average transmittance t. Since detail layer d1 typically contains minute details (including noise), a smaller enhancement is desired. Detail layer d2, on the other hand, typically contains important contours and edge information, requiring a larger enhancement. Therefore, we set ω1 ≤ ω2.

[0140] Finally, by solving the Poisson equation, the detail layer d after contrast stretching is obtained. i (i = 1, 2);

[0141] Step 5.6: Through the above steps, the final clear underwater image Z with halo correction can be obtained as follows:

[0142] Z = base cr +d′1+d′2 (20)

[0143] like Figure 6 The image shown is a brightness-adjusted reflection diagram output by an embodiment of the present invention. Figure 7 The image shown is a clear underwater image output by an embodiment of the present invention.

[0144] The advantage of this invention is that it takes into account the real and complex lighting scenarios during UUV underwater operations, effectively removes the halo effect caused by artificial strong light sources, reduces its serious interference to subsequent computer vision tasks, and significantly improves the quality of underwater images.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An underwater light halo image correction method based on radial gradient distribution characteristics, characterized in that, The method comprises the following steps: Step 1, obtaining an underwater halo image S to be processed; Step 2, constructing an underwater halo image imaging model for describing an imaging physical process of an underwater image under an artificial light source auxiliary lighting scene, wherein the underwater halo image imaging model is as follows: where S denotes the underwater halo image to be processed, R is the reflectance map, I is the intensity map, V is the halo map, and the operator denotes element-wise multiplication. Step 3, solving the underwater halo image imaging model by using a Bayesian framework, introducing a radial gradient distribution and a super Laplace distribution to constrain the optimization result, and obtaining a target function according to a maximum posterior probability and a probability distribution model of different image layers; the method comprises the following steps: Step 3.1, obtaining a reflection map R, an illumination map I and a halo map V respectively by maximizing a posterior probability, wherein the posterior probability is solved according to the following formula: Wherein, P(R, I, V|S) is a posterior probability distribution, P(S|R, I, V) is a likelihood function, P(R), P(I) and P(V) are respectively a probability distribution of the reflection map R, the illumination map I and the halo map V, and P(S) is a probability distribution of the input underwater halo image S, Step 3.2, introducing an image super Laplace distribution and a radial gradient distribution to constrain the optimization result, and having: wherein is a smoothing constraint term, is obtained using is a first estimate, is a L p norm, 0 < p < 1 is a super-Laplacian distribution constraint, D i are difference operator matrices corresponding to the difference operators, P(I) is the probability distribution of the intensity map I; Wherein, P(R) is a probability distribution of the reflection map R; wherein, represents a radial gradient smoothing constraint term, taking into account the radial symmetry of the halo layer with respect to the optical center and the smoothness thereof, and is applied to the halo map V using a discrete, uniform sampling approach: (V(r t ), r t ∈ S r ), where S r = {r0, r1, …, r n-1}, dividing the image into sectors along these discrete radii, i.e. r m is the inner radius of sector m and δr represents the sector width, having: wherein, denotes a data fidelity term constraining the proposed model to the difference between the input image S; Substituting the formula (3), (4), (5) and (6) into the formula (2), the target equation can be obtained: Wherein, λ1, λ2, λ3 and λ4 are constant coefficients for controlling the importance of different terms; Step 4, decomposing the target function into several sub-problems by using an alternating iterative updating method to solve the target function, and calculating a halo layer V, an illumination layer I and a reflection layer R, wherein a non-convex optimization sub-problem in the sub-problems is solved by using an iterative re-weighted least square method; Step 5, post-processing the reflection layer R, adjusting the brightness of the reflection layer by using an adaptive brightness scaling factor, and combining a multi-scale gradient domain processing to perform color correction and detail enhancement on the reflection layer R, so as to obtain a clear underwater image Z after halo correction. 2.The underwater light halo image correction method based on radial gradient distribution characteristics according to claim 1, characterized in that, The target function is decomposed into several sub-problems by using an alternating iterative updating method to solve the target function, and a halo layer V, an illumination layer I and a reflection layer R are calculated, comprising: Step 4.1, initialize halo parameters V 0 = 255, weight parameter W i 0 = 1 and iteration number k = 0; Step 4.2, taking R and I as a whole {R, I} to estimate, and decomposing the target equation (7) to obtain the formula (8) and (9): Step 4.

3. For a given V k (k = 0, 1, 2,...), compute subproblem (8) using the iteratively reweighted least squares method, at each iteration, reduce problem (8) to a quadratic problem as shown in equation (10), solve (10) using gradient descent, and compute {R, I} (k+1) Substitute into (11) to compute weight parameter W i (k+1) : Q1 = |D i {R,I} k+1 | Q2 = p(Q1) p-1 (11) Step 4.4, update V according to the calculated {R, I} (k+1) ; using gradient descent method to solve equation (9) to get the updated V (k+1) ; Step 4.5, update iteration number k = k + 1; alternately execute step 4.3 and step 4.4 until ||{R, I} (k+1) -{R,I} k ||<ε, get the final halo map V and {R, I}; Step 4.6, according to the calculated {R, I} and the target equation (7), to obtain two unknown quantities R and I, using an alternating iterative method to obtain two sub-problems as shown in the formula (12) and (13): Step 4.7, Estimate initial illumination map and initialize the illumination map Involving initial weight parameter W i 0 = 1, iteration number k = 0; Step 4.8, for a given I k (k = 0, 1, 2,...), using the iteratively reweighted least squares method, to obtain R (k+1) ; Step 4.9, based on the calculated R (k+1) , the non-convex subproblem (13) is computed using an iteratively reweighted least squares method to obtain I (k+1) ; Step 4.

10. Update the iteration number, i.e., k = k + 1; alternately perform Step 4.8 and Step 4.9 until ||R (k+1) -R k || < ε, the final reflectance map R and the illumination map I are obtained. 3.The underwater light halo image correction method based on radial gradient distribution characteristics according to claim 2, characterized in that, The reflection layer R is post-processed, the brightness of the reflection layer is adjusted by using an adaptive brightness scaling factor, and the reflection layer R is color corrected and detail enhanced by combining a multi-scale gradient domain processing, comprising: Step 5.1, extract the reflectance map R and the product map {R, I} in the V channel luminance information of the HSV color space, respectively, to obtain R val and {R, I} val ; Step 5.2, according to the brightness information, design adaptive brightness scaling factor f, through the formula (14) to calculate the adaptive brightness map R' val ; The adaptive brightness scaling factor f is defined as: where GF {R,I} denotes a guided filter, and β is a constant parameter controlling the brightness. Step 5.

3. Transforming the adaptive intensity map from HSV space to RGB space gives the enhanced reflectance map R', and performing WLS-based multi-scale image decomposition on R' gives the base layer base representing low-frequency color information, and two detail layers d representing texture information at different scales i (i = 1, 2); Step 5.4, performing image color correction on the base layer by using a linear stretching model based on a Gaussian distribution: Firstly, the maximum value and the minimum value of the base layer in each color channel are obtained by the formula (16) and (17): where c e {r, g, b}, and denote the mean and standard deviation of the base layer in the c channel, respectively, and μ is a parameter that controls the dynamic range of the image. The base layer is color corrected by the formula (18): Step 5.5, on the two detail layers d i (i = 1, 2) different amplitude gradient field contrast stretching is performed, specifically: First, the two detail layers d i (i = 1, 2) are converted to the gradient domain, resulting in i , Secondly, the multi-scale gradient domain contrast stretching is performed according to the formula (19): where ω i is a non-negative parameter controlling the amplitude of enhancement, is the average value of the transmittance t, ω1≤ ω2, Finally, by solving the Poisson equation, the detail layer d after contrast stretching is obtained i (i = 1, 2); Step 5.6, through the above steps, the clear underwater image Z after halo correction can be obtained as: Z = base cr + d1' + d2' (20).

Images