Polarization defogging method based on multi-angle polarization modeling and optimal contrast constraint

Abstract

The invention discloses a polarization defogging method based on multi-angle polarization modeling and optimal contrast constraint, which comprises the following steps: acquiring polarization images at different angles, calculating a polarization state Stokes vector of light, and further generating a multi-angle polarization degree image MDoP; calculating an initial transmissivity according to the scene depth information of the MDoP image and the total light intensity image; the initial transmissivity is optimized, and a transmissivity image with rich information and low noise is obtained; obtaining an initial defogging result through an atmospheric scattering model in combination with the optimized transmissivity and known infinite atmospheric light parameters; and taking the initial defogging result as input, optimizing a loss function through iteration until convergence, and outputting a final defogged image. According to the polarization defogging method based on multi-angle polarization modeling and optimal contrast constraint, the image detail retention and target recognition performance can be improved, the polarization difference between the image texture and the target background can be effectively enhanced, and an excellent defogging effect can be achieved even in a high-concentration haze environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a multi-angle polarization modeling and optimal contrast constraint polarization defogging method. BACKGROUND

[0002] The existing defogging algorithms are mainly divided into three categories: image enhancement, deep learning and physical model. The method based on image enhancement improves the image quality through global or local enhancement, but does not consider the physical degradation mechanism of foggy image, and the defogging effect is limited; the method based on deep learning has made progress in scattering image clarification, but it is prone to domain migration problem due to the lack of training data and the dependence on synthetic images. The defogging method based on physical model is modeled and recovered from the physical mechanism of image degradation, and has stronger theoretical basis and adaptability. It is mainly divided into two categories: based on statistical priori and based on polarization principle.

[0003] The method based on statistical priori (such as dark channel priori, non-local priori and color attenuation priori) estimates the transmittance and atmospheric light by inducing the statistical characteristics of the fog-free image, but it depends on specific scene assumptions, such as the dark channel defogging algorithm mainly applicable to non-sky area. In the sky, the transmittance estimation is easy to be distorted in the bright background such as white wall, resulting in the decline of defogging effect. The defogging method based on polarization imaging obtains the polarization characteristics of the image, introduces additional polarization information in addition to the traditional intensity information, and provides more abundant constraint conditions for transmittance estimation, effectively enhancing the ability to describe the object edge and texture, so as to achieve clearer and more detailed defogging results under high fog concentration or complex background conditions.

[0004] In the prior art, the method based on the joint constraint of polarization angle and polarization degree improves the defogging performance under high fog concentration. Then, the relationship between polarization degree and scene depth is further explored to expand the application range. Therefore, the introduction of polarization information is the key to the superiority of the polarization defogging method over the statistical priori method, and improving the utilization efficiency of polarization information is the core way to further enhance the defogging effect. Based on this, the least square estimation method is used to improve the information quantity of the polarization degree image, which improves the image quality to a certain extent. However, since it introduces multiple regularization optimizations in the transmittance estimation process, the direct expression ability of the original polarization information is weakened, resulting in limited improvement of the defogging effect. The polarization imaging method based on Mueller matrix can comprehensively describe the polarization characteristics of the object, has stronger information retention ability, and has been successfully applied to underwater image clarification. However, this method relies on active imaging system and complex measurement process, and the popularization and application in atmospheric imaging scene still face high technical threshold and implementation difficulty.

[0005] Therefore, although the introduction of polarization information makes the polarization dehazing better than the traditional statistical prior-based method in performance, the current research on improving the utilization efficiency of polarization information still faces two major challenges: first, the existing methods have limited performance in enhancing effects; second, the polarization parameters relied on are often costly to obtain, making it difficult to promote to practical polarization dehazing applications. In addition, some methods based on joint estimation of polarization angle and degree of polarization to estimate the transmittance are prone to color noise and local dehazing failure in certain scenarios, affecting the quality and stability of the restored image.

[0006] Therefore, how to efficiently and simply obtain more rich polarization information parameters and effectively integrate them into the dehazing process is a key research direction for further improving the performance of polarization dehazing. SUMMARY

[0007] The purpose of the present application is to provide a multi-angle polarization modeling and optimal contrast constraint polarization dehazing method, based on the definition of the degree of polarization in the Stokes vector, a multi-angle polarization parameter MDoP is introduced, compared with the traditional degree of polarization image, MDoP can obtain more rich scene texture and edge information without complex Mueller matrix measurement. By analyzing the correlation between MDoP and atmospheric light, a mapping relationship between the two is established and introduced into the calculation of transmittance, combined with the dual transmittance estimation method of polarization information and scene depth, a transmittance image containing more scene details is generated, and in the transmittance optimization process, an optimal contrast constraint model is introduced to enhance the accuracy of transmittance estimation and the efficiency of dehazing results.

[0008] To achieve the above purpose, the present application provides a multi-angle polarization modeling and optimal contrast constraint polarization dehazing method, comprising the following steps:

[0009] Step S1, acquiring polarization images at different angles and calculating Stokes vectors for describing the polarization state of light;

[0010] Step S2, generating multi-angle degree of polarization images MDoP based on Stokes vectors;

[0011] Step S3, calculating the initial transmittance according to the scene depth information of MDoP image and total light intensity image;

[0012] Step S4, optimizing the initial transmittance to obtain a transmittance image with rich information and small noise;

[0013] Step S5, combining the optimized transmittance and the known infinite atmospheric light parameters to obtain the initial dehazing result through the atmospheric scattering model;

[0014] Step S6, taking the initial defogging result as input, iteratively optimizing the loss function until convergence, and outputting the final defogging image.

[0015] Preferably, in step S1, polarized images of different angles are obtained, and Stokes vectors for describing the polarization state of light are calculated, the specific process being as follows:

[0016] Step S11, obtaining foggy images of 0°, 60°, and 120° polarization angles by a polarization camera, and denoting them as I0, I 60 , and I 120 , respectively.

[0017] Step S12, calculating Stokes vectors for describing the polarization state of light according to the obtained polarized images, as follows:

[0018]

[0019] wherein I represents a total light intensity image; Q and U represent two different polarization difference images; I0, I 60 , and I 120 represent polarized foggy images of polarization angles of 0°, 60°, and 120°, respectively.

[0020] Preferably, in step S2, based on the Stokes vectors, a multi-angle degree of polarization image MDoP is generated, the specific steps being as follows:

[0021] Step S21, calculating polarization images under different polarization angles according to the Stokes vectors, as follows:

[0022]

[0023] wherein θ represents the angle of the polarization image, and the value range is θ∈[0, 180];

[0024] Step S22, calculating polarization difference images by using polarized images of different angles, as follows:

[0025] I pd (θ)=I θ -I θ+90 (9);

[0026] wherein I pd (θ) represents a polarization difference image; I θ represents a polarization image corresponding to the polarization angle θ; and I θ+90 represents a polarization image corresponding to the polarization angle θ+90.

[0027] Step S23, calculating the MDoP image by combining the polarization difference image and the total light intensity image, as follows:

[0028]

[0029] wherein ε is a regulation parameter of MDoP.

[0030] Preferably, in step S3, the initial transmittance is calculated according to the scene depth information of the MDoP image and the total light intensity image, and the specific steps are as follows:

[0031] Step S31, the MDoP image is inverted to obtain a flipped multi-angle polarization image FMDoP, satisfying the following relationship:

[0032] FMDoP = 1 - MDoP (11);

[0033] Step S32, the total light intensity information represented by the Stokes vector is used to calculate a scene depth map z(x), as follows:

[0034] z(x) = 0.121779 + 0.959710v(x) - 0.780245s(x) + 0.041337 (12);

[0035] wherein v(x) is the brightness of the pixel point x; s(x) is the saturation of the pixel point x;

[0036] Step S33, the initial transmittance is calculated in combination with the FMDoP and the scene depth map, as follows:

[0037]

[0038] wherein t(x) is the initial transmittance; ε ∈ [0, 1], n ∈ [0, 1], and the initial values are both 0; FMDoP(x) is the flipped multi-angle polarization image; and z(x) is the scene depth map.

[0039] Preferably, in step S4, the initial transmittance image is optimized to obtain a transmittance image rich in information and small in noise, and the specific steps are as follows:

[0040] Step S41, the initial transmittance is divided into two parts, satisfying the following relationship:

[0041] t(x) = t1(x) + t2(x) (14);

[0042] Step S42, t1(x) and t2(x) are respectively decomposed into a base layer and a detail layer through Gaussian filtering, as follows:

[0043] t1(x) = t 1_base (x) + t 1_detail (x) = t1(x) x G σ (x) + [t1(x) - t1(x) x Gσ (x)] (15);

[0044] t2(x) = t 2_base (x) + t 2_detail (x) = t2(x) x G σ (x) + [t2(x) - t2(x) x G σ (x)] (16);

[0045] wherein, t 1_base (x), t 2_base (x) is the base layer; t 1_detail (x), t 2_detail (x) is the detail layer; G σ (x) is the Gaussian filter kernel; and σ is the Gaussian filter radius;

[0046] After substituting formula (15) and formula (16) into formula (14), the transmittance satisfies:

[0047] t(x) = t 1_base (x) + t 1_detail (x) + t 2_base (x) + t 2_detail (x) (17);

[0048] Step S43, an energy discrimination mechanism is introduced, and the local energy of the base layer and the detail layer is compared to adaptively select the final response value, as follows:

[0049] t base (x) = t 1_base (x) mask base (x) + t 2_base (x) [1-mask base (x)] (18);

[0050] t detail (x) = t 1_detail (x) mask detail (x) + t 2_detail (x) [1-mask detail (x)] (19);

[0051] wherein, mask base (x) represents the fusion proportion of the transmittance of the two base layers; and mask detail (x) represents the fusion proportion of the transmittance of the two detail layers.

[0052] Step S44, the base layer and the detail layer response value are fused to obtain the optimized transmittance, as follows:

[0053] t(x) = t base(x) + t detail (x) (20).

[0054] Step S45, guided filtering processing with a radius of 2 is performed on the optimized transmittance image to suppress noise and retain edge details.

[0055] Preferably, in step S5, the initial defogging result is obtained by combining the optimized transmittance and the known infinite atmospheric light parameter through the atmospheric scattering model, as follows:

[0056]

[0057] wherein J(x) is the defogging image; I(x) is the total light intensity image; A ∞ is the infinite atmospheric light.

[0058] Preferably, in step S6, the initial defogging result is inputted, and the loss function is iteratively optimized until convergence, and the final defogging image is outputted, with the specific steps as follows:

[0059] Step S61, the contrast loss function of the initial defogging image is calculated, as follows:

[0060]

[0061] wherein E contrast (x) is the contrast loss function; c is a certain color channel; r is the red channel; g is the green channel; b is the blue channel; p is a pixel point within the field window B; B is a window with a size of 3 by 3; J c (x) is the intensity value of the defogging image J in a certain color channel c; N B is the total number of pixel points in the field B; is the average value of the defogging image J in a certain color channel c within the field B;

[0062] Step S62, the information loss degree function of the initial defogging image is calculated, as follows:

[0063] E loss (x) = ∑ x [ρ1(J(x)-255) 2 +ρ2(0-J(x)) 2 ] (23).

[0064] wherein E loss (x) is the information loss degree function, and ρ1, ρ2 are control intensities;

[0065] Step S63, the total loss function is constructed, as follows:

[0066] E L (x) = λ1·Econtrast (x)+λ2·E loss (x) (24);

[0067] wherein, E L (x) is a total loss function; λ1, λ2 are weight parameters, respectively control the optimization deviation of contrast loss and information loss degree;

[0068] Step S64, define the parameter optimization interval ε∈[0, 1], n∈[0, 1], repeat steps S3-S6 by a fixed step, until the loss function reaches the minimum value, output the final defogging image.

[0069] Therefore, the polarized defogging method with multi-angle polarization modeling and optimal contrast constraint has the following beneficial effects:

[0070] (1) The present application can obtain more rich polarization parameters without complex imaging configuration, the multi-angle polarization degree image (MDoP) has stronger target distinguishing ability, and contains more rich scene structure information, which helps to improve image detail retention and target recognition performance.

[0071] (2) The present application emphasizes efficient utilization of polarization information, estimates the transmittance by combining the correlation between MDoP and atmospheric light, and fuses scene depth information, proposes a double transmittance estimation strategy, constructs an optimal contrast constraint model to optimize the preliminary result, effectively enhances the polarization difference between image texture and target background, and even in high concentration haze environment, excellent defogging effect can be realized.

[0072] (3) The present application realizes higher physical consistency and visual quality without increasing additional imaging cost, and is suitable for various scene types, and shows good robustness and defogging quality in natural objects (such as flowers, grass and trees) and artificial structures (such as buildings and vehicles).

[0073] The technical solutions of the present application will be further described in detail below with the aid of drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is a flow chart of the polarized defogging method with multi-angle polarization modeling and optimal contrast constraint of the present application;

[0075] Figure 2 is an embodiment block diagram of the polarized defogging method with multi-angle polarization modeling and optimal contrast constraint of the present application;

[0076] Figure 3 is a 0° polarized haze-containing image in the embodiment of the present application;

[0077] Figure 4is a 60° polarized foggy image in the embodiment of the present application;

[0078] Figure 5 is a 120° polarized foggy image in the embodiment of the present application;

[0079] Figure 6 is a polarized de-foggy image in the embodiment of the present application. DETAILED DESCRIPTION

[0080] The technical solutions of the present application are further described below through the drawings and embodiments.

[0081] EMBODIMENT

[0082] In the prior art, a mathematical model is established for the problem of low visibility in foggy weather, and it is considered that there are two main reasons for the quality decline of the foggy image: one is that the target reflected light is attenuated due to absorption and scattering; and the other is that the strong background light is formed by scattering the ambient light, which covers the target light and causes the imaging to be unclear. Under this background, a widely used polarization-based de-fogging model is proposed by those skilled in the art, as shown below:

[0083] I(x) = D(x) + A(x) (1);

[0084] In the formula, I(x) is the total light intensity received by the detector; A(x) represents the atmospheric light; and D(x) represents the attenuated target light.

[0085] Among them, the attenuated target light and the atmospheric light are represented as follows:

[0086] D(x) = J(x) t(x) (2);

[0087] A(x) = A ∞ (1-t(x)) (3);

[0088] In the formula, J(x) is the unattenuated target light, that is, the clear image; t(x) represents the transmittance; and A ∞ represents the atmospheric light at infinity.

[0089] Among them, the transmittance t(x) refers to the proportion of the light that can reach the detection system after the light is attenuated by the particles, as shown below:

[0090] t(x) = e -β·z(x) (4);

[0091] In the formula, β represents the atmospheric scattering coefficient; and z(x) represents the scene depth.

[0092] Therefore, based on the above derivation process, formula (3) can be rewritten as:

[0093] A(x) = A ∞ (1-e-β·z(x) ) (5);

[0094] Solving formula of haze-free image can be obtained by simultaneous equations (1) (2) (3):

[0095]

[0096] It can be seen from formula (6) that the haze-free image J(x) can be obtained by solving the transmittance t(x) and the atmospheric light value A ∞ at infinity.

[0097] Therefore, the key of haze removal based on atmospheric scattering model lies in the accuracy of the transmittance and the atmospheric light at infinity. Existing researches show that the average value of the top 0.1% pixels with the highest intensity value in the haze image can be used as an effective estimation of the atmospheric light at infinity. Therefore, the method is also used to determine the atmospheric light value at infinity in the present application. The estimation method of the transmittance can be mainly divided into two categories: the method based on image prior and the method based on polarization information. Among them, the dark channel prior is a representative method of the former, and the polarization difference method is one of the most representative strategies in the latter.

[0098] It can be seen from the definition of the transmittance in formula (4) that the transmittance reflects the distribution of the global fog concentration and the change degree of the scene depth in the haze image. In the estimation process of the transmittance, the traditional image prior method is widely used due to its advantage of not needing additional imaging conditions, and the dark channel prior (DCP) is a representative method. The method is based on the statistical assumption that at least one color channel intensity is extremely low in the non-sky area of the natural image, and the transmittance is calculated by local minimum value.

[0099] However, when there are strong reflection, white objects or areas close to the color of the atmospheric light in the scene, the DCP often overestimates the concentration of the fog, resulting in low estimation of the transmittance, loss of image details or appearance of artifacts. In contrast, the haze removal method based on polarization information utilizes the difference in polarization characteristics of light in the imaging process, which can effectively distinguish the reflected light of the target object from the background scattered light, so as to realize more accurate separation of the target and the background, and then realize more realistic estimation of the transmittance.

[0100] In addition, the transmittance calculated based on the polarization information can better preserve and depict the edge structure information in the scene, thereby improving the clarity of the target area in the image. Therefore, the polarization-based haze removal method exhibits significant advantages in enhancing image details, highlighting target outlines and accurately separating foreground and background.

[0101] Based on this, as shown in Figures 1-2 the present application proposes a polarization haze removal method based on multi-angle polarization modeling and optimal contrast constraint, which comprises the following steps:

[0102] As Figure 1 shown, the present application proposes a multi-angle polarization modeling and optimal contrast constraint polarization defogging method, comprising the following steps:

[0103] Step S1, acquiring polarization images at different angles, and calculating Stokes vectors for describing the polarization state of light.

[0104] Step S11, acquiring foggy images at polarization angles of 0°, 60°, and 120° by a polarization camera, denoted as I0, I 60 , and I 120 respectively.

[0105] Step S12, calculating Stokes vectors for describing the polarization state of light according to the obtained polarization images, as follows:

[0106]

[0107] wherein I represents a total light intensity image; Q and U represent two different polarization difference images; I0, I 60 , and I 120 represent polarization foggy images at polarization angles of 0°, 60°, and 120° respectively, as shown in Figure 3 , Figure 4 , Figure 5 .

[0108] Step S2, generating multi-angle polarization degree images (Multi-angle Degree of Polarization, MDoP) based on Stokes vectors.

[0109] Step S21, calculating polarization images at different polarization angles according to Stokes vectors, as follows:

[0110]

[0111] wherein θ represents the angle of the polarization image, and the value range is θ∈[0, 180].

[0112] Step S22, calculating polarization difference images by using polarization images at different angles, as follows:

[0113] I pd (θ)=I θ -I θ+90 (9);

[0114] wherein I pd (θ) represents a polarization difference image; I θ represents a polarization image corresponding to the polarization angle θ; and I θ+90The polarization image corresponding to the polarization angle θ+90 is represented.

[0115] Step S23, the MDoP image is calculated by combining the polarization difference image and the total light intensity image, as follows:

[0116]

[0117] Wherein, ε is the adjustment parameter of MDoP, the default value is 1, when the highlight scattering image and the scene transform large image, ε decreases by step.

[0118] Step S3, calculate the initial transmittance.

[0119] Step S31, the MDoP image is processed to obtain FMDoP, as follows:

[0120] FMDoP=1-MDoP (11);

[0121] Step S32, the total light intensity information represented by Stokes vector is used to calculate the scene depth map z(x), as follows:

[0122] z(x)=0.121779+0.959710v(x)-0.780245s(x)+0.041337 (12);

[0123] Wherein, v(x) is the brightness of pixel point x; s(x) is the saturation of pixel point x.

[0124] Step S33, the initial transmittance is calculated by combining FMDoP and scene depth map, as follows:

[0125]

[0126] Wherein, t(x) is the initial transmittance; ε∈[0, 1], n∈[0, 1], the initial value is set to 0, and iteration is performed according to the step; FMDoP(x) is the flipped multi-angle polarization image; z(x) is the scene depth map.

[0127] Step S4, optimize the initial transmittance to obtain a transmittance image with rich information and small noise.

[0128] Step S41, the initial transmittance is divided into two parts, satisfying the following relationship:

[0129] t(x)=t1(x)+t2(x) (14);

[0130] Step S42, t1(x) and t2(x) are respectively decomposed into a base layer and a detail layer by Gaussian filtering, as follows:

[0131] t1(x)=t1_base (x)+t 1_detail (x) = t1(x) x G σ (x)+[t1(x)-t1(x) x G σ (x)] (15);

[0132] t2(x) = t 2_base (x)+t 2_detail (x) = t2(x) x G σ (x)+[t2(x)-t2(x) x G σ (x)] (16);

[0133] wherein, t 1_base (x), t 2_base (x) is the base layer; t 1_detail (x), t 2_detail (x) is the detail layer; G σ (x) is the Gaussian filter kernel; and σ is the Gaussian filter radius, σ = 3.

[0134] After substituting formula (15) and formula (16) into formula (14), the transmittance satisfies:

[0135] t(x) = t 1_base (x)+t 1_detail (x)+t 2_base (x)+t 2_detail (x) (17);

[0136] Step S43, introducing an energy discrimination mechanism, comparing the local energy of the base layer and the detail layer, and adaptively selecting the higher energy part as the final response value, so as to realize the optimal strategy of information preservation, as follows:

[0137] t base (x) = t 1_base (x) mask base (x)+t 2_base (x) [1-mask base (x)] (18);

[0138] t detail (x) = t 1_detail (x) mask detail (x)+t 2_detail (x) [1-mask detail (x)] (19);

[0139] wherein, mask base (x) represents the fusion proportion of the two base layer transmittances;

[0140] mask detail(x) represents the fusion proportion of controlling the transmittance of two detail layers.

[0141] Step S44, add the base layer and detail layer response values to obtain the optimized transmittance, as follows:

[0142] t(x) = t base (x) + t detail (x) (20).

[0143] Step S45, perform guided filtering processing with a radius of 2 on the optimized transmittance to suppress noise and retain edge details.

[0144] Step S5, calculate the initial defogging result based on the atmospheric scattering model.

[0145] In combination with the optimized transmittance in step S4 and the known infinite atmospheric light parameter, the initial defogging result is obtained through the atmospheric scattering model, as follows:

[0146]

[0147] Wherein, J(x) is the defogging image; I(x) is the total light intensity image; A ∞ is the infinite atmospheric light, which is estimated by the mean value of 0.1% high luminance pixels in front of the fog image.

[0148] Step S6, take the initial defogging result as input, and output the final defogging image by iteratively optimizing the loss function until convergence, the specific steps are as follows:

[0149] Step S61, calculate the contrast loss function of the initial defogging image, as follows:

[0150]

[0151] Wherein, E contrast (x) is the contrast loss function; c is a certain color channel; r is the red channel; g is the green channel; b is the blue channel; p is a pixel point in the neighborhood window B; B is a window with a size of 3 by 3; J c (x) is the intensity value of the defogging image J in a certain color channel c; N B is the total number of pixel points in the neighborhood B; is the average value of the defogging image J in the color channel c within the neighborhood B.

[0152] Step S62, calculate the information loss degree function of the initial defogging image, as follows:

[0153] E loss (x) = ∑ x [ρ1(J(x)-255) 2 +ρ2(0-J(x))2 ] (23);

[0154] wherein, E loss (x) is the information loss function, and ρ1, ρ2 are control intensities;

[0155] Step S63, a total loss function is constructed as follows:

[0156] E L (x) = λ1·E contrast (x) + λ2·E loss (x) (24);

[0157] wherein, E L (x) is the total loss function; λ1, λ2 are weight parameters, respectively controlling the optimization bias of contrast loss and information loss.

[0158] Step S64, defining parameter optimization interval ε∈[0, 1], n∈[0, 1], repeating steps S3-S6 at a fixed step size until the loss function reaches a minimum value, outputting a final defogging image as shown in Figure 6 .

[0159] Therefore, the polarized defogging method of the present application has the following beneficial effects:

[0160] (1) The present application can obtain more rich polarized parameters without complex imaging configuration, the multi-angle polarization degree image (MDoP) has stronger target distinguishing ability and contains more rich scene structure information, which helps to improve the image detail retention and target recognition performance.

[0161] (2) The present application emphasizes the efficient use of polarization information, estimates the transmittance by combining the correlation between MDoP and atmospheric light, and fuses scene depth information, proposes a double transmittance estimation strategy, constructs an optimal contrast constraint model to optimize the preliminary result, effectively enhances the polarization difference between image texture and target background, and even in high concentration of fog environment, excellent defogging effect can be achieved.

[0162] (3) The present application realizes higher physical consistency and visual quality without increasing additional imaging cost, and is suitable for various scene types, and shows good robustness and defogging quality in natural objects (such as flowers, trees and plants) and artificial structures (such as buildings and vehicles).

[0163] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A polarization dehazing method with multi-angle polarization modeling and optimal contrast constraints, characterized in that, Includes the following steps: Step S1: Obtain polarization images at different angles and calculate the Stokes vector used to describe the polarization state of light; Step S2: Generate a multi-angle polarization degree image (MDoP) based on the Stokes vector; Step S3: Calculate the initial transmittance based on the scene depth information of the MDoP image and the total light intensity image; Step S4: Optimize the initial transmittance to obtain a transmittance image with rich information and low noise; Step S5: Combining the optimized transmittance and the known atmospheric light parameters at infinity, the initial defogging result is obtained through an atmospheric scattering model; Step S6: Using the initial dehazing result as input, iteratively optimize the loss function until convergence, and output the final dehazed image.

2. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 1, characterized in that, In step S1, polarization images at different angles are acquired, and the Stokes vector used to describe the polarization state of light is calculated. The specific process is as follows: Step S11: Acquire hazy images with polarization angles of 0°, 60°, and 120° using a polarization camera, denoted as I0, I1, and I2 respectively. 60 I 120 ; Step S12: Based on the obtained polarization image, calculate the Stokes vector used to describe the polarization state of light, as shown below: Where I represents the total light intensity image; Q and U represent two different sets of polarization difference images; I0, I 60 I 120 These represent polarized foggy images with polarization angles of 0°, 60°, and 120°, respectively.

3. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 1, characterized in that, In step S2, a multi-angle polarization image (MDoP) is generated based on the Stokes vector. The specific steps are as follows: Step S21: Calculate the polarization images at different polarization angles based on the Stokes vector, as shown below: Where θ represents the angle of the polarization image, and its value range is θ∈[0,180]; Step S22: Calculate the polarization difference image using polarization images at different angles, as shown below: I pd (θ)=I θ -I θ+90 (9); Among them, I pd (θ) represents the polarization difference image; I θ This represents the polarization image corresponding to the polarization angle θ; I θ+90 This represents the polarization image corresponding to the polarization angle θ+90. Step S23: By combining the polarization difference image with the total intensity image, the MDoP image is calculated, as shown below: Where ε is the adjustment parameter of MDoP.

4. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 1, characterized in that, In step S3, the initial transmittance is calculated based on the scene depth information of the MDoP image and the total light intensity image. The specific steps are as follows: Step S31: Invert the MDoP image to obtain the flipped multi-angle polarization image FMDoP, which satisfies the following relationship: FMDoP = 1 - MDoP(11); Step S32: Calculate the scene depth map z(x) using the total light intensity information represented by the Stokes vector, as shown below: z(x)=0.121779+0.959710v(x)-0.780245s(x)+0.041337(12); Where v(x) is the brightness of pixel x; s(x) is the saturation of pixel x; Step S33: Combine FMDoP and the scene depth map to calculate the initial transmittance, as shown below: Where t(x) is the initial transmittance; ε∈[0,1], n∈[0,1], and the initial values ​​are all set to 0; FMDoP(x) is the flipped multi-angle polarization image; z(x) is the scene depth map.

5. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 1, characterized in that, In step S4, the initial transmittance image is optimized to obtain an informative and low-noise transmittance image. The specific steps are as follows: Step S41: Divide the initial transmittance into two parts, satisfying the following relationship: t(x) = t1(x) + t2(x) (14); Step S42: Decompose t1(x) and t2(x) into a base layer and a detail layer respectively using Gaussian filtering, as shown below: t1(x)=t 1_base (x)+t 1_detail (x)=t1(x)×G σ (x)+[t1(x)-t1(x)×G σ (x)](15); t2(x)=t 2_base (x)+t 2_detail (x)=t2(x)×G σ (x)+[t2(x)-t2(x)×G σ (x)](16); Among them, t 1_base (x), t 2_base (x) is the basic layer; t 1_detail (x), t 2_detail (x) represents the detail layer; G σ (x) is the Gaussian filter kernel; σ is the Gaussian filter radius; Substituting formulas (15) and (16) into formula (14), the transmittance satisfies: t(x)=t 1_base (x)+t 1_detail (x)+t 2_base (x)+t 2_detail (x)(17); Step S43: Introduce an energy discrimination mechanism to compare the local energies of the base layer and the detail layer, and adaptively select the final response value, as shown below: t base (x)=t 1_base (x)·mask base (x)+t 2_base (x)·[1-mask base (x)](18); t detail (x)=t 1_detail (x)·mask detail (x)+t 2_detail (x)·[1-mask detail (x)](19); Among them, mask base (x) represents the fusion ratio that controls the transmittance of the two base layers; mask detail (x) represents the blending weight that controls the transmittance of the two detail layers; Step S44: Fuse the response values ​​of the base layer and the detail layer to obtain the optimized transmittance, as shown below: t(x)=t base (x)+t detail (x) (20); Step S45: Perform guided filtering with a radius of 2 on the optimized transmittance image to suppress noise and preserve edge details.

6. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 5, characterized in that, In step S5, combining the optimized transmittance and the known atmospheric light parameters at infinity, the initial dehazing result is obtained through an atmospheric scattering model, as shown below: Where J(x) is the dehazed image; I(x) is the total light intensity image; A ∞ It is the light of the atmosphere at infinity.

7. The polarization dehazing method with multi-angle polarization modeling and optimal contrast constraint according to claim 1, characterized in that, In step S6, the initial dehazing result is used as input, and the loss function is iteratively optimized until convergence, outputting the final dehazed image. The specific steps are as follows: Step S61: Calculate the contrast loss function of the initial dehazed image, as shown below: Among them, E contrast (x) represents the contrast loss function; c is a color channel; r is the red channel; g is the green channel; b is the blue channel; p is the pixel within the neighborhood window B; B is a window of size 3 by 3; J c (x) represents the intensity value of the dehazed image J in color channel c; N B This represents the total number of pixels in the neighborhood. The average value of the dehazed image J in the neighborhood B on the color channel c; Step S62: Calculate the information loss function of the initial dehazed image, as shown below: E loss (x)=Σ x [ρ1(J(x)-255) 2 +ρ2(0-J(x)) 2 ] (23); Among them, E loss (x) is the information loss function, and ρ1 and ρ2 are the control strengths; Step S63: Construct the total loss function, as shown below: E L (x)=λ1·E contrast (x)+λ2·E loss (x) (24); Among them, E L (x) represents the total loss function; λ1 and λ2 are weight parameters that control the optimization bias of contrast loss and information loss, respectively. Step S64: Define the parameter optimization interval ε∈[0,1], n∈[0,1], and repeat steps S3-S6 with a fixed step size until the loss function reaches its minimum value, and then output the final dehazed image.

Images