Non-uniform illumination underwater image recovery method based on discrete entropy gain

Abstract

The invention discloses a non-uniform illumination underwater image reply method based on discrete entropy gain, which is applied to the technical field of computer vision and image processing, and aims to solve the problem that an existing method based on image enhancement or a physical model is difficult to give consideration to global brightness balance and local detail reservation. Particularly, the problems of light spot overexposure and gradual change shadow caused by an artificial light source cannot be effectively inhibited are solved; according to the method, an underwater image under non-uniform illumination is modeled by using a vignetting component, a central overexposure region in the modeled vignetting component is constructed by using an exponential term, and an edge shadow region in the modeled vignetting component is constructed by using a polynomial term. Dynamic modeling of exponential term construction is realized through brightness distribution characteristics in the centroid region; optimal modeling of polynomial construction is achieved through discrete entropy minimization, and image attenuation caused by the influence of an extra light source of an underwater medium is effectively recovered through a gain function.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision and image processing, and particularly relates to an underwater image restoration technology. BACKGROUND

[0002] Underwater images have important application value in the fields of marine exploration and underwater engineering, but the image quality is significantly degraded due to the influence of water absorption, scattering and secondary interference (such as local overexposure, shadow shielding and uneven illumination distribution) introduced by artificial light sources. The traditional underwater image restoration method is based on the assumption of uniform illumination, and does not consider the vignetting effect caused by the spatial heterogeneity of light sources, resulting in insufficient adaptability in non-uniform illumination scenes. The existing methods based on image enhancement or physical models cannot balance global brightness equalization and local detail preservation, especially cannot effectively suppress the spot overexposure and gradual shadow caused by artificial light sources. Therefore, there is an urgent need for an underwater image restoration method that can dynamically compensate for non-uniform illumination attenuation and adaptively correct vignetting effect. SUMMARY

[0003] To solve the above technical problems, the present application proposes a non-uniform illumination underwater image restoration method based on discrete entropy gain, which realizes adaptive correction of non-uniform illumination caused by additional light sources by constructing a non-linear gain function model, combining image centroid statistical features and parameter optimization strategy.

[0004] The technical scheme adopted by the present application is: a non-uniform illumination underwater image restoration method based on discrete entropy gain, comprising:

[0005] S1, modeling the underwater image under non-uniform illumination using vignetting component;

[0006] The vignetting effect component is introduced into the traditional underwater imaging model, and the model is expanded as:

[0007] I(x)=J0(x)·t(x)+A·(1-t(x))

[0008] J0(x)=J(x)·E V

[0009] Where I(x) represents the underwater image under non-uniform illumination, J(x) represents the real underwater scene, J0(x) represents the underwater scene after non-uniform illumination, t(x) represents the attenuation transmittance, A represents the background light, E V To represent the vignetting component of the spatial distribution of illumination, its expression is:

[0010]

[0011] r is the normalized distance of the current pixel to the image centroid, a, b, c, k, γ are dynamic fitting parameters;

[0012] S2, constructing the center overexposed area in the vignetting component using an exponential term;

[0013] S3, constructing the edge shadow area in the vignetting component using a polynomial term;

[0014] S4, based on the discrete entropy minimization correction framework, the optimal modeling of the polynomial construction is realized by discrete entropy minimization, so as to obtain the optimal gain function;

[0015] S5, based on the optimal gain function, the underwater image under non-uniform illumination is recovered.

[0016] The method of the present application uses the vignetting component to model the underwater image under non-uniform illumination, uses an exponential term to model the center overexposed area in the vignetting component, and uses a polynomial term to model the edge shadow area in the vignetting component. The dynamic modeling of the exponential term construction is realized by the brightness distribution characteristics in the centroid area; the optimal modeling of the polynomial construction is realized by discrete entropy minimization, and the image attenuation caused by the influence of the additional light source of the underwater medium is effectively recovered by the gain function, which has the advantages of high reliability and strong stability, and can provide convenience for subsequent computer vision tasks. The method of the present application has the following advantages:

[0017] 1. Physical reliability: the present application models the spatial decay law of non-uniform illumination through the vignetting effect component, overcoming the limitations of the traditional uniform illumination assumption;

[0018] 2. Adaptive correction: the present application combines centroid statistical characteristics and discrete entropy minimization to dynamically suppress overexposure, shadow and brightness gradient abnormalities;

[0019] 3. Strong compatibility: the present application can be seamlessly integrated into existing underwater recovery network algorithms to improve the recovery effect of these methods for non-uniformly illuminated images. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the flowchart of the method of the present application;

[0021] Figure 2 is the vignetting component and gain function effect visualization diagram;

[0022] Figure 3 is the visual recovery effect comparison of the present application and the existing method. DETAILED DESCRIPTION

[0023] As shown in Figure 1 the implementation process of the method of the present application includes the following steps:

[0024] S1. Non-uniform illumination modeling and vignetting effect introduction. Introduce the vignetting effect component in the traditional underwater imaging model, and expand the model as follows:

[0025] I(x) = J0(x) t(x) + A (1-t(x))

[0026] J0(x) = J(x) E V

[0027] where E V is the vignetting component representing the spatial distribution of illumination, and its expression is:

[0028]

[0029] r is the normalized distance of the current pixel to the image centroid, and a, b, c, k, γ are dynamic fitting parameters.

[0030] S2. Dynamic optimization of the parameters of the vignetting component. Specifically, it includes:

[0031] S21, use the exponential term to construct the modeling of the central overexposure region in the vignetting component, the exponential term includes k and γ.

[0032] Exponential term k construction: according to the average brightness μ center of the centroid region over , and the proportion of overexposed pixels p

[0033]

[0034] where T l , T h , k min , k max , and α are hyperparameters, and the specific value is explained as follows: for example, for the OceanDark dataset, through experiments, it is determined that T L = 180, T H = 220; k min , k max and α are set to 0.3, 0.6 and 0.4 respectively.

[0035] Define the centroid neighborhood Ω as a circular region with the image centroid (x cm , y cm ) as the center and a radius of R. The brightness distribution characteristics of the pixels in this region are counted:

[0036]

[0037] where indicates that the value is 1 when the internal condition is true, and 0 otherwise.

[0038] Exponential term γ construction, based on the dynamic γ value calculation model of distance variance:

[0039]

[0040] Where, γ max and γ min is a parameter, in this embodiment, γ min ,γ max are set to 0.1, 1.0, σ d It represents the standard deviation of brightness within Ω, and is calculated as:

[0041]

[0042] S22. Use polynomial terms to construct the edge shadow area in the modeling vignetting component:

[0043] By constraining the parameters a, b, and c to ensure that the gain function is monotonically decreasing, and combining the discrete entropy minimization objective function for iterative optimization, overexposure is suppressed and brightness distribution is balanced. Three cases are discussed according to the value of c, and the polynomial constructor is defined as E + :

[0044]

[0045] When c=0: If a≤0 and a+2b≤0, and a and b are not 0 at the same time, then E + Monotonically decreasing.

[0046] When c>0: the derivative function opens upward, which needs to satisfy and That is, a≤0 and a+2b+3c≤0. It is E V Polynomial constructor for ; yes The first derivative of ; and Respectively The values ​​are between 0 and 1.

[0047] When c<0, the derivative function opens downward, which can be discussed in two sub-cases:

[0048] If b 2 ≤3ac, There is no solution, then E + Monotonically decreasing.

[0049] If b 2 >3ac, The stationary point value is

[0050]

[0051] If you want to make Constant establishment, need to meet q + ≤0 or q - ≥1.

[0052] S3. Optimal modeling of polynomial construction by discrete entropy minimization

[0053] Discrete entropy minimization correction framework. Define the improved discrete entropy objective function:

[0054]

[0055] H represents the discrete entropy objective function, n represents the gray level of the image, p i represents the probability of gray level i, λ represents the penalty coefficient, which is determined to be 10 according to the experiment; N over is the number of overexposed pixels, that is, the number of pixels whose gray value exceeds the maximum value of the gray level; N total is the total number of pixels in the image.

[0056] The optimal parameter combination is searched by gradient descent method, and the specific steps include:

[0057] (1) Initialize the polynomial parameters a, b, c and the step size δ.

[0058] (2) For the current parameters (a, b, c), generate six adjacent parameter combinations:

[0059] (a+δ,b,c),(a-δ,b,c),(a,b+δ,c),(a,b-δ,c),(a,b,c+δ) and (a,b,c-δ).

[0060] (3) For each adjacent parameter combination, check whether it meets the condition that the vignetting component E V is monotonically decreasing in the interval 0~1. If it meets the condition, calculate its corresponding gain function, and calculate the discrete entropy of the image after gain.

[0061] (4) If the discrete entropy of a certain adjacent parameter combination is less than the current minimum discrete entropy, update the minimum discrete entropy and the parameters.

[0062] The calculation process of discrete entropy is:

[0063] Adjacent parameter combinations are used to calculate the vignetting component E V based on step S1, and then the underwater scene J0(x) after non-uniform illumination is calculated based on the formula in step S1. Then map the adjusted underwater scene J0(x) brightness value to the corresponding gray level interval, calculate the probability pi of each gray level interval, and calculate the discrete entropy H according to the formula in step S3.

[0064] The process of updating the parameters is:

[0065] When the discrete entropy of a certain adjacent parameter combination is less than the current minimum discrete entropy, the parameters are updated, the discrete entropy of the current adjacent parameter combination is used as the new minimum discrete entropy, and the current adjacent parameter combination is saved as the current discrete entropy parameter (a, b, c).

[0066] (5) If the current parameters (a, b, c) are the parameters that minimize the discrete entropy, the step size δ is halved.

[0067] (6) Repeat steps (2)-(5) until the step size δ is less than the termination condition ∈. The current parameter combination is used as the optimal parameter combination, and the optimal gain function is calculated based on the optimal parameter combination. Output the optimal gain function g(r) = ζ / E V (r), ζ is the weight of the gain function, that is, the overall magnification and reduction weight for the vignetting component, which is 1 in the experiment. Figure 2 The figure shows a visualization of the vignetting component and the gain function effect. The vignetting component is separated and the gain function is used to achieve local restoration of the image.

[0068] S4. The gain function is used to effectively restore the image attenuation caused by the influence of additional light sources in the underwater medium.

[0069]

[0070] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The specific process is as follows (floating point numbers are used for calculation in the following calculation process):

[0071] 1. Data preparation and parameter setting. This example uses the OceanDark dataset, the input image is denoted as I(x); set the low / high brightness threshold T L =180, T H =220; low / high brightness k value k min , k max The high brightness penalty weights α are set to 0.3, 0.6, and 0.4 respectively; the radius of the centroid region R = 0.2; the upper and lower bounds of γ are γ min ,γ max Set to 0.1, 1.0; initial parameters a = b = c = 0, step size δ = 2, penalty coefficient λ = 10, termination condition ∈ = 1 / 256.

[0072] 2. Calculation and optimization of vignetting components. Calculate the image centroid coordinate x cm ,y cm , normalize the pixel distance r; dynamically adjust k and γ according to the brightness distribution of the centroid area; and determine the optimal polynomial parameters a, b, and c through iterative search.

[0073] 3. Image restoration and evaluation, the corrected image is input into the underwater image restoration method, and the restoration result is output, such asFigure 3 As shown, 7 groups of images are given, and the processing results of different methods are shown. From the processing results of the first six groups of images, it can be seen that the present application can effectively remove the light imbalance, overexposure and shadow caused by vignetting component. The prior art cannot consider this point for the restoration of underwater images. From the processing result of the last group of images, it can be seen that the present application can significantly improve the details of underwater objects under insufficient illumination, and is more conducive to detecting targets.

[0074] Those skilled in the art will realize that the embodiments described herein are for the purpose of illustration and should not be construed as limiting the scope of the present application. The present application can be modified and varied in various ways, and any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for underwater image restoration under non-uniform illumination based on discrete entropy gain, characterized in that: include: A1. Modeling underwater images under non-uniform illumination using vignetting components; The vignetting effect component is introduced into the traditional underwater imaging model, and the extended model is: I(x)=J0(x)·t(x)+A·(1-t(x)) J0(x)=J(x)·E V Among them, I(x) represents the underwater image under non-uniform illumination, J(x) represents the real underwater scene, J0(x) represents the underwater scene after non-uniform illumination, t(x) represents the attenuation transmittance, A represents the background light, and E V To characterize the vignetting component of the spatial distribution of illumination, its expression is: r is the normalized distance from the current pixel to the image centroid, a, b, c, k, γ are dynamic fitting parameters; A2. Use exponential terms to construct and model the central overexposed area in the vignetting component; A3. Use polynomial terms to construct and model the edge shadow area in the vignetting component; A4. Based on the discrete entropy minimization correction framework, the optimal modeling of polynomial construction is achieved through discrete entropy minimization, thereby obtaining the optimal gain function g(r) = ζ / E V (r), ζ is the gain function weight; A5. Restoring underwater images under non-uniform illumination based on the optimal gain function.

2. The method for underwater image restoration under non-uniform illumination based on discrete entropy gain according to claim 1, characterized in that: The parameters of the exponential term in step A2 include k and γ. The specific implementation process of step A2 is: A21, according to the mean brightness μ of the centroid area center And the proportion of overexposed pixels over , dynamically adjust the attenuation strength parameter k: Among them, T l , T h , k min , k max , α is a hyperparameter; A22. Dynamic γ value calculation model based on distance variance: Where, γ max and γ min is a parameter, σ d represents the standard deviation of brightness within Ω.

3. The method for underwater image restoration under non-uniform illumination based on discrete entropy gain according to claim 2, characterized in that: The parameters of the polynomial in step A3 include a, b, and c. The specific implementation process of step A3 is as follows: Define the polynomial constructor as E + : When c=0: If a≤0 and a+2b≤0, and a and b are not 0 at the same time, then E + Monotonically decreasing; When c>0: the derivative function opens upward, which needs to satisfy and That is, a≤0 and a+2b+3c≤0; yes The first derivative of ; and Respectively The values ​​at 0 and 1; When c<0, the derivative function opens downward, which can be discussed in two sub-cases: If b 2 ≤3ac, There is no solution, then E + Monotonically decreasing; If b 2 >3ac, The stationary point value is If you want to make Constant establishment, need to meet q + ≤0 or q - ≥1.

4. The method for underwater image restoration under non-uniform illumination based on discrete entropy gain according to claim 3, characterized in that: Step A4 is specifically as follows: A41. Define the improved discrete entropy objective function as: Among them, H represents the discrete entropy objective function, n represents the gray level of the image, and p i represents the probability of gray level i, λ represents the penalty coefficient, N over Indicates the number of overexposed pixels, N total Indicates the total number of pixels in the image; A42. Search for the optimal parameter combination using the gradient descent method; specifically, the following steps are included: A421, initialize parameters a, b, c and step size δ; A422. For the current parameters (a, b, c), generate six adjacent parameter combinations: (a+δ, b, c), (a-δ, b, c), (a, b+δ, c), (a, b-δ, c), (a, b, c+δ), and (a, b, c-δ); A423. For each adjacent parameter combination, check whether it satisfies the vignetting component E V The monotonically decreasing condition is satisfied between 0 and 1. If so, the corresponding gain function is calculated and the discrete entropy corresponding to the image after gain is obtained. Otherwise, the calculation of the current adjacent parameter combination is terminated and the calculation of the next adjacent parameter combination is performed. A424. If the discrete entropy of a certain adjacent parameter combination is less than the current minimum discrete entropy, update the minimum discrete entropy and parameters; A425. If the current parameters (a, b, c) are the parameters that minimize the discrete entropy, then the step size δ is halved; A43. Repeat steps A422-A425 until the step size δ is less than the termination condition ∈; output the optimal gain function.

5. The method for underwater image restoration under non-uniform illumination based on discrete entropy gain according to claim 4, characterized in that: The calculation formula corresponding to step A5 is: Among them, I0(x) represents the restored underwater image.

Images