Fog map reconstruction method and electronic equipment based on atmospheric polarization orthogonal blind separation

Abstract

This invention relates to the field of image processing technology, specifically to a fog image reconstruction method and electronic device based on orthogonal blind separation of atmospheric light polarization. The method automatically estimates the depth-related atmospheric polarization degree matrix using the atmospheric polarization characteristics of the Stokes vector. In polarization space, atmospheric light at infinity is decomposed into orthogonal polarization parameters, and a multi-regularized constrained blind separation model of atmospheric scattered light is constructed to achieve accurate estimation of atmospheric scattered light. The method combines polarization degree offset image estimation of atmospheric light at infinity to reduce the influence of bright white objects. Experimental results show that this method effectively improves the visibility of target scenes at different depths and further enhances the quality of reconstructed images. It is particularly applicable to fog images taken in dense fog, solving the problems of existing methods that rely on sky regions and manually selected bias coefficients to estimate and stabilize atmospheric polarization degree, as well as the inaccurate estimation of atmospheric scattered light in dense fog scenes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a fog map reconstruction method and electronic device based on atmospheric light polarization orthogonal blind separation. Background Technology

[0002] Under severe weather conditions such as fog and haze, imaging detection systems suffer from reduced image contrast and visibility due to the scattering and absorption of a large number of suspended particles in the atmosphere, which seriously affects the image quality. Therefore, researching fog and haze image reconstruction methods to improve the clarity and contrast of detection images has important application value for target detection [1], urban traffic monitoring [2,3] and military reconnaissance.

[0003] With the in-depth research on image dehazing algorithms, a large amount of work has made significant progress. The most widely studied are single-image dehazing algorithms, such as image enhancement algorithms like histogram equalization[4] and wavelet transform[5], which enhance the image without considering the degradation nature. There are also dehazing algorithms based on prior knowledge that model the relevant parameters of foggy images, such as dark channel prior[6], context regularization constraints[7], color attenuation prior[8], and nonlocal prior[9]. However, these methods still have limitations in terms of scenarios.

[0004] In recent years, with the continuous development of polarization optical detection technology, the polarization-based fog reconstruction method has increased the dimension of information acquisition and obtained rich target features and environmental information by using higher-dimensional polarization information, which has significant technical advantages in the field of image dehazing. Some early algorithms were based on the polarization difference model [10, 11]. This model estimates atmospheric polarization degree by estimating atmospheric polarization difference information, thereby estimating atmospheric scattered light. However, when solving atmospheric polarization degree, it either relies on the sky region to solve atmospheric polarization degree [12-14], which makes it difficult to process fog scenes without sky region well. Or it introduces the bias coefficient [14-16] to stabilize atmospheric polarization degree and reduce atmospheric scattered light noise. The method that relies on sky region and bias coefficient has human interaction error, which leads to inaccurate estimation of atmospheric scattered light. In order to overcome the problem of sky region, LiangJetal.

[17] made full use of polarization angle information to estimate atmospheric polarization degree. Some algorithms have achieved certain results by combining this method and fusion strategy [18-20]. Dai et al.

[21] used the global features of the image to solve the atmospheric polarization degree by assuming that the atmospheric scattered light and the target reflected light are uncorrelated. However, the polarization characteristics of light in the medium are related to the medium concentration

[22] . As the fog concentration (transmission distance) increases to a certain extent, a depolarization effect will occur, that is, the atmospheric polarization degree will decrease. Therefore, since the fog concentration is different at different scene depths, its atmospheric polarization degree is also different.

[0005] While most existing methods have freed themselves from sky dependence, they often require bias coefficients to stabilize image restoration performance because atmospheric polarization is a globally invariant. Therefore, these methods have limitations in achieving target reconstruction tasks at different scene depths.

[0006] Some algorithms consider eliminating target information to estimate atmospheric scattered light based on median filtering and Gaussian filtering [23, 24]; some algorithms consider eliminating quantization noise to achieve accurate estimation of atmospheric scattered light [25-28]. Among them, Liang Jet et al.

[25] proposed to eliminate quantization noise based on frequency domain low-pass filtering to improve the accuracy of atmospheric scattered light estimation, and Liang Jet et al.

[28] again believed that the noise of the non-polarized part greatly affected the accurate estimation of atmospheric scattered light, so they used the global atmospheric polarization degree to solve the non-polarized part of atmospheric scattered light and performed Gaussian low-pass filtering to eliminate the noise influence. Liang Zet et al.

[26] optimized the polarization angle information based on the regularization constraint model to reduce the influence of noise. Shenet et al.

[27] proposed a transmission map optimization scheme with depth-chroma compensation regularization and an image dehazing optimization scheme with chroma-depth compensation regularization, respectively, and carried out denoising processing in the whole process.

[0007] Some algorithms avoid solving for atmospheric polarization and directly obtain atmospheric scattered light

[29] . Shao et al.

[29] used the prior information of atmospheric light gradient as a constraint to separate the target layer and atmospheric information, and used atmospheric information to realize polarization reconstruction. This type of method directly obtains atmospheric scattered light by filtering and denoising the original fog image.

[0008] In summary, existing methods rely on sky regions and manually selected bias coefficients to estimate and stabilize atmospheric polarization, which results in unsatisfactory performance in dense fog scenarios. Therefore, this embodiment proposes a polarization fog map reconstruction method and electronic device based on atmospheric light polarization orthogonal decomposition. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention discloses a fog image reconstruction method and electronic device based on orthogonal blind separation of atmospheric light polarization. The aim is to acquire polarization image data of different fog concentrations in real-world environments. In any scenario, it can estimate atmospheric polarization degree without relying on sky regions and bias coefficients, and estimate atmospheric scattered light and atmospheric light at infinity using a more reasonable and accurate method. This achieves excellent polarization reconstruction, solving the problems of existing defogging methods that rely on sky regions and manually selected bias coefficients to estimate and stabilize atmospheric polarization degree, and the unsatisfactory results in dense fog scenarios.

[0010] This invention is achieved through the following technical solution:

[0011] In a first aspect, the present invention provides a fog map reconstruction method based on atmospheric optical polarization orthogonal blind separation, comprising the following steps:

[0012] Multiple foggy polarization images with different polarization angles under the same exposure conditions were acquired using an imaging detection device.

[0013] The Stokes vector, total polarization degree, and total polarization angle are calculated from the polarization image of foggy weather, and the atmospheric polarization degree matrix related to depth is automatically estimated.

[0014] The polarization degree shift image and the gray-level histogram of the total light intensity image are obtained, and the atmospheric light at infinity is estimated using the OStu adaptive method.

[0015] Atmospheric scattered light is obtained by constructing a blind separation model of atmospheric scattered light with multiple regularization constraints from the perspective of polarization orthogonal decomposition.

[0016] The atmospheric scattering model is used to reconstruct the polarization fog map, and finally the fog-free map of the undegraded scene is obtained by inversion.

[0017] Furthermore, in the method, the Stokes vector S = (I, Q, U) T :

[0018]

[0019] In the formula, I represents the total intensity of the incident light, Q and U represent the polarization information of the incident light, and I0 and I... 60 and I 120 The image shows polarization images of foggy weather with polarizer angles of 0°, 60°, and 120°. The Stokes vector can describe all polarization states of light. The atmospheric polarization degree is automatically estimated by combining the atmospheric polarization characteristics of the Stokes vector, so as to get rid of the constraints of sky area and bias coefficient and reduce the estimation error of manual interaction.

[0020] Furthermore, in the method, obtaining the total polarization degree p and the total polarization angle θ image specifically involves:

[0021]

[0022]

[0023] Here, I contains all the information of the target scene, while the changes in Q and U are determined by atmospheric scattered light. The total polarization angle image θ can be approximated as the polarization angle distribution map of atmospheric scattered light, and the atmospheric polarization angle θ is used. A Estimated atmospheric polarization degree p A It is possible to obtain the atmospheric polarization degree as it varies with depth.

[0024] Furthermore, in the method, based on a frequency-first strategy, a grayscale histogram of the total polarization angle θ image is calculated, and the value with the highest frequency is taken as the atmospheric polarization angle θ. A Then, using the atmospheric polarization angle θ A The atmospheric polarization degree p varies with scene depth. A :

[0025]

[0026] Furthermore, in the method, polarization degree offset images are acquired respectively. The total light intensity image I grayscale histogram is obtained, and the segmentation thresholds for the sky and the target are obtained using the OStu adaptive method.

[0027] Furthermore, the method obtains the spatial coordinates of grayscale values ​​greater than a threshold in I, in Obtain the coordinates of grayscale values ​​less than the threshold in the coordinate space, and then take the intersection coordinates (x, y) of the two coordinate spaces. i y i That is, atmospheric light A at infinity. ∞ The positions with the highest probabilities are then averaged.

[0028]

[0029] Among them, the OTU histogram adaptive method is used to obtain the sky target segmentation threshold, which can effectively distinguish between the sky and the target area.

[0030] Furthermore, in the method, atmospheric scattered light is decomposed into maximum and minimum polarization orthogonal images A in polarization space. max and A min ;

[0031]

[0032] Obtain the maximum and minimum polarization orthogonal image A of atmospheric light at infinity. ∞max and A ∞min ;

[0033]

[0034] Where A represents atmospheric scattered light, and an intermediate quantity A is defined. ∞max With A ∞min Its meaning lies in A ∞ It has the same partial polarization characteristics as A. As the polarizer angle changes, its intensity value changes, and a maximum value A appears. ∞max and minimum value A ∞min The degree of atmospheric polarization changes with the depth of the scene target.

[0035] Furthermore, in A ∞ Given that T is unknown, it is necessary to construct an atmospheric scattering light-blind separation model from... A was separated from * This allows us to estimate atmospheric scattered light.

[0036] Furthermore, simplify the atmospheric scattering light model:

[0037]

[0038] In the formula, A * ∈{A max A min}, A ∞* ∈{A ∞max A ∞min}, T = 1 - t, which is related to t; The dot product symbol;

[0039] Furthermore, in the method, the maximum and minimum polarization orthogonal images of atmospheric light at infinity are synthesized according to Malus's law, and a blind separation objective function for atmospheric scattered light with multiple regularization constraints is constructed to obtain the atmospheric scattered light A as follows:

[0040]

[0041] in,

[0042]

[0043] In the formula, w represents different sets of weighted indices; W j D is the adaptive weight matrix; j It is a higher-order difference operator.

[0044] Furthermore, by employing a semi-quadratic variable separation method and FFT, the above problem is transformed into a series of simple subproblems by introducing auxiliary variables, and then iteratively optimized. Ultimately, the solutions to these subproblems will converge to the optimal solution of the original problem.

[0045] Furthermore, in the method, when reconstructing the polarization fog map using the atmospheric scattering model, the total light intensity I is expressed as: I = D + A;

[0046] Among them, D=R·t A=A ∞ ·(1-t)

[0047] In the formula, t is the transmittance, which is exponentially related to the scene depth z; D is the target reflected light directly reflected by the target scene; R is the undegraded haze-free scene image; A ∞ For atmospheric light at infinity, a polarization fog map is reconstructed to derive the expression for R:

[0048]

[0049] Furthermore, the orthogonal parameters of atmospheric light polarization at infinity are decomposed in polarization space, and a blind fog map reconstruction model with multiple regularization constraints is constructed to improve the estimation accuracy of atmospheric scattered light.

[0050] Furthermore, polarization difference information is incorporated to reduce the impact of bright white objects on atmospheric light estimation at infinity. This improves the visibility of target scenes at different depths and objectively enhances the quality of the reconstructed image, making it more adaptable to scenes with varying fog densities.

[0051] In a second aspect, the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, the memory being coupled to the processor, and the processor executing the computer program to implement the steps in the polarization fog map reconstruction method based on atmospheric optical polarization orthogonal decomposition described in the first aspect.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention obtains depth-dependent atmospheric polarization without the need for a bias coefficient, which is beneficial for handling targets at different scene depths, especially distant targets; the atmospheric light estimation method at infinity can reduce the influence of white, bright objects; and it achieves good reconstruction results under different fog concentrations, regardless of whether there is sky or no sky.

[0054] Subjectively, this invention greatly improves the visibility of target scenes at different depths, and objectively, it further enhances the quality of reconstructed images, making them more adaptable to scenes with different fog densities. Attached Figure Description

[0055] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of a fog map reconstruction method based on orthogonal blind separation of atmospheric optical polarization;

[0057] Figure 2 This is an overall framework diagram of the polarization fog reconstruction method according to an embodiment of the present invention;

[0058] Figure 3 This is a diagram showing the longitudinal and transverse selection areas according to an embodiment of the present invention;

[0059] Figure 4The longitudinal regions p and p in the embodiments of the present invention A Histogram comparison chart;

[0060] Figure 5 The horizontal regions p and p in the embodiments of the present invention A Histogram comparison chart;

[0061] Figure 6 This is a map of the sky target selection area according to an embodiment of the present invention;

[0062] Figure 7 The sky regions p and p in the embodiments of the present invention A Histogram comparison chart;

[0063] Figure 8 The target regions p and p in this embodiment of the invention A Histogram comparison chart;

[0064] Figure 9 A is higher than the threshold in the embodiments of the present invention. ∞ Distribution diagram of total light intensity I;

[0065] Figure 10 A is below the threshold in the embodiments of the present invention. ∞ In polarization difference Distribution map;

[0066] Figure 11 This is Embodiment A of the present invention. ∞ The spatial location map with the highest probability;

[0067] Figure 12 This is a flowchart of the atmospheric scattering light-blind separation iterative process according to an embodiment of the present invention;

[0068] Figure 13 These are polarization reconstruction results in regions with and without sky, based on an embodiment of the present invention.

[0069] Figure 14 This is a polarization reconstruction result image of a region with sky and a region without sky under dense fog weather conditions, according to an embodiment of the present invention;

[0070] Figure 15 This is a comparison of the results of different dehazing algorithms in the embodiments of the present invention, along with local comparison images;

[0071] Figure 16 This is a comparison chart of the results of different dehazing algorithms in an area with sky, according to an embodiment of the present invention;

[0072] Figure 17 This is a comparison chart of the results of different dehazing algorithms in the absence of sky in the embodiments of the present invention;

[0073] Figure 18This is a comparison chart of the results of different defogging algorithms under color polarized foggy data in an embodiment of the present invention. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example 1

[0076] Reference Figure 1 As shown, to address the problems of existing methods relying on sky regions and manually selected bias coefficients to estimate and stabilize atmospheric polarization, and the inaccurate estimation of atmospheric scattered light in dense fog scenarios, this embodiment provides a fog map reconstruction method based on orthogonal blind separation of atmospheric light polarization, including the following steps:

[0077] Multiple foggy polarization images with different polarization angles under the same exposure conditions were acquired using an imaging detection device.

[0078] The Stokes vector, total polarization degree, and total polarization angle are calculated from the polarization image of foggy weather, and the atmospheric polarization degree matrix related to depth is automatically estimated.

[0079] The polarization degree shift image and the gray-level histogram of the total light intensity image are obtained, and the atmospheric light at infinity is estimated using the OStu adaptive method.

[0080] Atmospheric scattered light is obtained by constructing a blind separation model of atmospheric scattered light with multiple regularization constraints from the perspective of polarization orthogonal decomposition.

[0081] The atmospheric scattering model is used to reconstruct the polarization fog map, and finally the fog-free map of the undegraded scene is obtained by inversion.

[0082] This embodiment automatically estimates the depth-related atmospheric polarization degree by combining the atmospheric polarization characteristics of the Stokes vector.

[0083] In this embodiment, as can be seen from the atmospheric scattering model

[30] , the total light intensity I of the image captured by the imaging detection system mainly consists of two parts; one part is the target reflected light D directly reflected by the target scene, which is related to the target's own material properties and the scene depth z. The other part is the atmospheric scattered light A caused by the scattering of sunlight by particles in the atmospheric environment, which is related to the scene depth z and the atmospheric light A at infinity. ∞ Regarding this, the total light intensity can be expressed as:

[0084] I = D + A (1)

[0085] in:

[0086] D=R·t (2)

[0087] A = A ∞ ·(1-t) (3)

[0088] In this embodiment, t represents transmittance, which is exponentially related to scene depth z; R represents the undegraded haze-free scene image; A ∞ Let R be atmospheric light at infinity. Combining equation (1-3), we can derive the expression for R by inverse derivation:

[0089]

[0090] In this embodiment, it can be seen from equation (4) that A is estimated. ∞ A can be used to recover the fog-free image.

[0091] This embodiment is based on the assumption that atmospheric light at infinity and atmospheric scattered light are both partially polarized. It decomposes the polarization orthogonal parameters of atmospheric light at infinity in polarization space. By constructing an objective function with gradient and L2 regularization constraints, atmospheric scattered light is accurately separated from the polarization orthogonal parameters of atmospheric light at infinity.

[0092] This embodiment combines polarization degree offset images to estimate the global atmospheric light value at infinity, effectively reducing the influence of bright white objects.

[0093] This embodiment effectively improves the visibility of target scenes at different depths, further enhancing the quality of reconstructed images, and is particularly applicable to foggy images taken in dense fog.

[0094] Example 2

[0095] Based on Example 1, the advantage of the polarization fog map reconstruction method is that it utilizes the partial polarization characteristics of atmospheric scattered light to achieve accurate estimation of atmospheric scattered light A. Therefore, exploring the atmospheric polarization degree estimation method is the key to accurately estimating atmospheric scattered light. Thus, this example provides an atmospheric polarization degree estimation method, since the polarization characteristics of light propagating in a medium are related to the medium concentration

[22] . As the fog concentration increases to a certain extent, a depolarization effect will occur, that is, the atmospheric polarization degree will decrease.

[0096] This embodiment is based on the definition of polarization degree, where the total polarization degree p and the atmospheric polarization degree p A The expression is:

[0097]

[0098]

[0099] In this embodiment, since the target reflected light exhibits a strong depolarization effect during propagation in the atmospheric scattering medium, the polarization characteristics of the target reflected light are often negligible when using polarization imaging technology for image dehazing, and only the polarization characteristics of the atmospheric scattered light are considered. Therefore, in a preferred implementation of this embodiment, the following can be obtained:

[0100]

[0101] In the formula, This is an atmospheric polarization difference image, representing the polarization portion of atmospheric scattered light, and is approximately a total polarization difference image. Therefore, from equations (1), (5), and (6), we can deduce that:

[0102]

[0103] This embodiment provides further analysis, where R reflects the inherent properties of the target scene, and A... ∞ The light is atmospheric light from infinity, and t decays exponentially with scene depth. Therefore, the degree of atmospheric polarization also varies with depth.

[0104] In this embodiment, when the scene approaches infinity, p A Approximately p. As the scene depth decreases, the fog concentration gradually decreases, p A It gradually becomes greater than p. Theoretical analysis can also well explain the above conclusion; therefore, p... A Treating it as a global invariant does not handle depth targets in different scenarios very well.

[0105] In this embodiment, the Stokes vector can describe all polarization states of light. By combining the atmospheric polarization characteristics of the Stokes vector, the atmospheric polarization degree can be automatically estimated, thereby getting rid of the constraints of the sky region and the bias coefficient and reducing the estimation error of manual interaction.

[0106] In this embodiment, when a Stokes vector is S = (I, Q, U) T When incident polarized light passes through a linear polarizer at an angle α, the intensity I of the polarized image is... α The expression for the polarizer angle α is:

[0107]

[0108] This embodiment acquires polarization images I0, I1, and I2 at polarizer angles of 0°, 60°, and 120°. 60 and I 120 Substituting this into the above formula, we can calculate the Stokes vector S = (I, Q, U). T :

[0109]

[0110] In the formula, I represents the total intensity of the incident light, and Q and U represent the polarization information of the incident light. Therefore, the total degree of polarization p and the total polarization angle θ of the incident light are expressed as:

[0111]

[0112]

[0113] In this embodiment, by analyzing equations (11) and (12), we can see that I contains all the information of the target scene, while the changes in Q and U are mainly caused by atmospheric scattered light. Therefore, the total polarization angle image θ can be approximated as the polarization angle distribution map of atmospheric scattered light.

[0114] This embodiment uses a frequency-first strategy to statistically analyze the grayscale histogram of the polarization angle image, and takes the value with the highest frequency as the atmospheric polarization angle θ. A Using the atmospheric polarization angle θ A Combining equations (11) and (12), we can obtain the atmospheric polarization degree p that varies with scene depth. A :

[0115]

[0116] In this embodiment, as Figure 3-8 The figure shows the total polarization p and atmospheric polarization p at different depths. A Comparison of statistical histograms. Among them, Figures 3-5 This indicates a histogram comparison of selected regions in both the vertical and horizontal directions. Figures 6-8 This indicates a histogram comparison of the selected target area in the sky.

[0117] In this embodiment, the probability distribution of grayscale values ​​is roughly similar in both the vertical and horizontal directions, as well as in the sky target area, and p A The red distribution is shifted to the right compared to the blue distribution, indicating that p A All are slightly greater than p.

[0118] In this embodiment, from the histogram Figure 7 and 8 It can be concluded that in the sky region, p A The offset between p and p is small, and their values ​​can be approximately equal. However, as the scene depth decreases and the target area is reached, the offset between p and p increases. A The value of will be greater than p, so the target area is more offset from the sky area, and its value is relatively larger.

[0119] In summary, this embodiment uses an atmospheric polarization angle θ. A Estimated atmospheric polarization degree p AIt is possible to conclude that the degree of atmospheric polarization changes with depth.

[0120] Example 3

[0121] Based on Example 1 and in accordance with the requirements of Example 2, this example provides a method for estimating atmospheric light at infinity, where atmospheric light A at infinity is... ∞ The estimation is often affected by bright white objects and strong light sources, leading to inaccurate estimates.

[0122] This embodiment of the study yielded that p A It is related to the concentration of smog, while in the sky (areas with high fog concentration) and the target area, p A The distribution distance between A and p is different, therefore this embodiment further proposes a method to estimate A using a polarization degree deviation image. ∞ .

[0123] In this embodiment, when the scene depth approaches infinity, I→A ∞ There is a noticeable difference between the sky (in areas with high fog density) and white objects. Figures 9-11 This is a diagram showing the estimation and analysis of atmospheric light at infinity.

[0124] In this embodiment, as Figure 10 Within the rectangular frame, the offset of the white object is greater than the offset of the sky. Therefore, this can be determined by combining I and To estimate atmospheric light at infinity.

[0125] This embodiment uses the OTU histogram adaptive method to obtain the sky target segmentation threshold, which can effectively distinguish between the sky and the target region, and respectively at I and Calculate the corresponding spatial coordinates, and then find the intersection point (x, y) of the two coordinate spaces. i y i The pixel value at that location is A. ∞ To maximize the probability, the average of the filtered pixel values ​​is calculated to obtain A. ∞ The expression is:

[0126]

[0127] Finally, based on equation (4), the fog-free map can be derived.

[0128] In this embodiment, Figure 9 It is the distribution of pixels with a light intensity greater than the threshold in the total light intensity map I; Figure 10 The pixels with values ​​less than a threshold are in the polarization difference map. Distribution; Figure 11 It is A ∞ The spatial location with the highest probability. As can be seen from the diagram, coordinate I includes the white building, while... White buildings were excluded, therefore, A was finally shown in the original image. ∞ The spatial coordinates with the highest probability are mostly in the sky area, and white highlighted objects are well excluded.

[0129] Example 4

[0130] In conjunction with Examples 2 and 3, to further achieve the requirements of Example 1, this example provides an orthogonal blind separation model for atmospheric scattered light polarization. Starting from the formation process of atmospheric scattered light in polarization space, because atmospheric scattered light has partial polarization characteristics, A is split into unpolarized atmospheric light and polarized atmospheric light. When the light passes through a polarizer with an angle of α, according to Malus's law, we can obtain:

[0131]

[0132] In this embodiment, when the polarizer angle is θ A When the atmospheric scattered light is at its maximum, it is considered that the influence is greatest at that moment, and vice versa. Thus, the maximum and minimum polarization orthogonal image A of the atmospheric scattered light can be synthesized. max and A min .

[0133]

[0134]

[0135] In a further implementation of this embodiment, the following occurs:

[0136] A = A max +A min (17)

[0137] In this embodiment, substituting equation (3) into equations (15) and (16) yields:

[0138] A max =A ∞max ·(1-t) (18)

[0139] A min =A ∞min ·(1-t) (19)

[0140] In the formula, A ∞max With A ∞min The orthogonal parameter for atmospheric light polarization at infinity means that A ∞ Having the same partial polarization characteristics as A, its intensity value changes with the angle of the polarizer, and a maximum polarization orthogonal value A appears. ∞max and minimum value A ∞min The atmospheric polarization degree varies with the depth of the scene target; therefore, the A value of each pixel in the image...∞max With A ∞min They are not the same; the specific expression is:

[0141]

[0142]

[0143] In this embodiment, equations (18) and (19) are simplified to the following equations:

[0144]

[0145] In the formula, A * ∈{A max A min}, T = 1 - t, which is related to t; · is the dot product symbol;

[0146] This embodiment is in A ∞ Given that T is unknown, it is necessary to construct an atmospheric scattering light-blind separation model from... A was separated from * This allows us to estimate atmospheric scattered light.

[0147] In this embodiment, considering the local smoothness of atmospheric scattered light, and the need to retain the abruptness of atmospheric scattered light at the target edge, the optimal solution A is obtained by minimizing equation (23). * This allows for the accurate separation of atmospheric scattered light and the suppression of atmospheric scattered light interference.

[0148]

[0149] in,

[0150]

[0151] In the formula, w represents different sets of weighted indices; W j D is the adaptive weight matrix; j It is a higher-order difference operator.

[0152] The first item in this embodiment is the data fidelity item, which ensures that the estimated A* is within the minimum error range of the data.

[0153] The second aspect of this embodiment introduces a gradient regularization term using a multi-difference operator to ensure the local smoothness and edge jumps of atmospheric scattered light.

[0154] This embodiment introduces Guided adaptive weights W j When the grayscale changes of adjacent pixels x and y are large, W jWhen (x, y) is very small, the regularization constraint at the edge will be canceled, thus achieving the effect of edge preservation.

[0155] This embodiment achieves a smoothing effect at non-edge areas. This embodiment uses different difference operators D. j To more effectively distinguish between edge regions and smooth regions, achieve edge sharpening, and reduce artifacts in smooth regions, the algorithm becomes more flexible. λ1 controls the degree of image smoothness; the larger λ1 is, the smoother the image.

[0156] The third term in this embodiment is an L2 regularization term for T, which can effectively suppress noise, and λ2 controls the noise level.

[0157] In this embodiment, the parameters λ1 and λ2 are set to 0.1, 0.02, and σ is set to 0.5. To obtain the optimal solution, equation (24) is transformed into the following two sub-problems for separate optimization:

[0158]

[0159]

[0160] This embodiment uses a semi-quadratic variable separation method and FFT. By introducing auxiliary variables, the above problem is transformed into a series of simple subproblems, and iterative optimization is performed. Finally, the solutions to these subproblems will converge to the optimal solution of the original problem.

[0161] In this embodiment, refer to Figure 12 This is a flowchart of the atmospheric scattering light-blind separation iterative process. and A represents the initially orthogonally polarized atmospheric scattered light, which is obtained after iterative optimization. max With A min It can be seen that atmospheric scattered light has depth similarity, that is, the estimated atmospheric scattered light values ​​at the same depth are similar.

[0162] Example 5

[0163] Referring to Examples 1-4 Figure 2 As shown, this embodiment provides a fog map reconstruction method based on orthogonal blind separation of atmospheric scattered light polarization. The main steps of this method include:

[0164] This embodiment utilizes a perspective polarization camera with a calibrated linear polarizer to simultaneously acquire foggy polarization images I0 and I1 at different polarization angles in a single exposure. 60 and I 120 And convert it to double format.

[0165] The calculated Stokes vector is S = (I, Q, U). T .

[0166]

[0167] Obtain images of the total polarization degree p and the total polarization angle θ.

[0168]

[0169]

[0170] This embodiment automatically estimates the depth-related atmospheric polarization matrix using the atmospheric polarization characteristics of the Stokes vector. Based on a frequency-first strategy, a grayscale histogram of the total polarization angle θ image is calculated, and the value with the highest frequency is taken as the atmospheric polarization angle θ. A .

[0171] This embodiment utilizes the atmospheric polarization angle θ A The atmospheric polarization degree p that varies with scene depth can be obtained. A This allows for more effective handling of targets at different depths in different scenarios, especially distant targets.

[0172]

[0173] This embodiment acquires polarization degree offset images respectively. The total light intensity image I grayscale histogram is used, and the OTU adaptive method is used to obtain the segmentation thresholds for the sky and the target. In I, the spatial coordinates of grayscale values ​​greater than the threshold are obtained. Obtain the coordinates of grayscale values ​​less than the threshold in the coordinate space, and then take the intersection coordinates (x, y) of the two coordinate spaces. * y * That is, atmospheric light A at infinity. ∞ The positions with the highest probability are then averaged.

[0174]

[0175] This embodiment decomposes atmospheric scattered light into orthogonal images A of maximum and minimum polarization in polarization space. max and A min .

[0176]

[0177] Obtain the maximum and minimum polarization orthogonal image A of atmospheric light at infinity. ∞max and A ∞min .

[0178]

[0179] This embodiment constructs a blind separation objective function for atmospheric scattered light with multiple regularization constraints, achieving accurate estimation of atmospheric scattered light from atmospheric light at infinity. Parameters λ1 are set to 0.1, λ2 to 0.02, and σ to 0.5.

[0180]

[0181] Finally, this embodiment utilizes an atmospheric scattering model to reconstruct the polarization fog map.

[0182] (9)

[0183] This embodiment can collect polarization image data of different fog concentrations in real environments. In any scenario, it can estimate the degree of atmospheric polarization without relying on the sky region and the bias coefficient, and estimate atmospheric scattered light and atmospheric light at infinity through a more reasonable and accurate method, thereby achieving a good polarization reconstruction task.

[0184] Example 6

[0185] At the implementation level, this embodiment reconstructs the polarization images of foggy days in areas with and without sky, as follows:

[0186] S1 acquires foggy polarization images I0 and I1 with and without sky areas, respectively. 60 and I 120 ;

[0187] S2 obtains the stokes vector S = (I, Q, U). T Images of total polarization degree p and total polarization angle θ;

[0188] S3 calculates the atmospheric polarization degree p as it varies with scene depth based on a frequency-priority strategy. A ;

[0189] S4 adaptively estimates atmospheric light A at infinity by combining polarization shift information. ∞ ;

[0190] S5 uses orthogonal blind separation of atmospheric scattered light polarization to accurately separate atmospheric scattered light;

[0191] S6 polarization without image reconstruction.

[0192] The implementation results of this embodiment are as follows: Figure 13 As shown, under normal conditions, this embodiment achieved good results in processing both areas with and without sky.

[0193] Example 7

[0194] At the implementation level, this embodiment reconstructs the polarization images of areas with and without sky in foggy weather, as follows:

[0195] S1 acquires polarized images I0 and I of areas with and without sky during dense fog conditions. 60 and I 120 ;

[0196] S2 obtains the stokes vector S = (I, Q, U). T Images of total polarization degree p and total polarization angle θ;

[0197] S3 calculates the atmospheric polarization degree p as it varies with scene depth based on a frequency-priority strategy. A ;

[0198] S4 adaptively estimates atmospheric light A at infinity by combining polarization shift information. ∞ ;

[0199] S5 uses orthogonal blind separation of atmospheric scattered light polarization to accurately separate atmospheric scattered light;

[0200] S6 polarization without image reconstruction.

[0201] The implementation results of this embodiment are as follows: Figure 14 As shown, this embodiment achieved good results in processing both areas with and without sky in a dense fog environment.

[0202] Example 8

[0203] This embodiment provides a qualitative analysis of the experimental results of Embodiments 6 and 7. The experiments in this embodiment were conducted on a PC with a Windows 10 system, an Intel Core i5-8250U CPU, 8GB RAM, and a Windows 10 64-bit operating system. The compilation software used in the experiments was Matlab R2018a.

[0204] To verify the effectiveness of the proposed algorithm, this embodiment uses polarized images of foggy weather taken in real-world scenarios for experimentation. Furthermore, this embodiment compares the algorithm with classic dehazing algorithms such as BCCR[7], Nonlocal[9], IPD

[10] , GPLPF

[25] , and PLF

[28] .

[0205] In this embodiment, the image acquisition device is a polarization camera with a calibrated linear polarizer. This device can simultaneously acquire three original polarized foggy images with different polarizer directions (0°, 60°, and 120°) in a single exposure under natural light, with an image resolution of 700×934. The total light intensity image obtained from the stokes vector is used as the input of BCCR[7] and Nonlocal[9]. IPD

[10] takes the synthesized maximum and minimum polarization images as its input, while GPLPF

[25] and PLF

[28] take the original polarized foggy images as their input.

[0206] In this embodiment, Figure 15 The figures show a comparison of the results of different dehazing algorithms and a comparison of the local magnified results. In the figure, (a) represents the original image and the local magnified image, and (b) to (g) represent the results of BCCR[7], Nonlocal[9], IPD

[10] , GPLPF

[25] , PLF

[28] and the algorithm processing results of this embodiment and the corresponding local magnified results, respectively.

[0207] Analysis of this embodiment shows that: BCCR[7] and Nonlocal[9] methods both leave some fog in the foreground and background, respectively, showing uneven and incomplete defogging; IPD

[10] , GPLPF

[25] and PLF

[28] methods are better than the first two methods in terms of overall defogging effect, but after local magnification, there is a lot of noise, which makes the gray value of the target details discontinuous.

[0208] In comparison, the algorithm proposed in this embodiment achieves good overall dehazing results and enhances the visibility of distant scenes. Simultaneously, noise is suppressed, more texture details are preserved, resulting in higher target fidelity. Therefore, the algorithm in this embodiment demonstrates excellent dehazing performance for targets at different depths.

[0209] In addition, to verify whether the algorithm is affected by the sky region, this embodiment divides the polarization test data into grayscale foggy scene images with sky regions and grayscale foggy scene images without sky regions for experimental comparison. Figure 16 and Figure 17 The images show a comparison of the results of different dehazing algorithms in areas with sky and in areas without sky, respectively.

[0210] In this embodiment, as Figure 16As shown, the polarization data of the sky area were experimentally compared under light fog (L1, L2), medium fog (M1, M2), dense fog (D1, D2), and heavy fog (H1, H2) conditions. Figure (a) shows the original polarized fog image, and (b) to (g) show the reconstruction results of BCCR[7], Nonlocal[9], IPD

[10] , GPLPF

[25] , PLF

[28] and the method of this embodiment, respectively. It can be seen that in light fog environment, several algorithms can effectively remove fog haze, but when the fog concentration is large, BCCR[7] and Nonlocal[9] methods lose the details of the target in some scenes, making some targets darker, such as the top of the tower in D1. At the same time, the scene after defogging is not natural. IPD

[10] and GPLPF

[25] methods still have fog characteristics, which means that the overall defogging is not thorough. PLF

[28] and the method of this embodiment have better overall processing effect than the previous ones, but for the processing of targets at different depths, the method of this embodiment is obviously better than PLF

[28] , especially for distant targets, the target edges are clearer.

[0211] Therefore, the method proposed in this embodiment has achieved good dehazing results for distant buildings, mid-ground lighthouses, and foreground shrubs. It not only preserves image details but also has a better dehazing effect, making the restored result more detailed. Therefore, it is superior to other methods in terms of detail restoration and fog removal.

[0212] In this embodiment, as Figure 17 As shown, polarized fog data in areas without sky were experimentally compared under light fog (L1, L2), moderate fog (M1, M2), and dense fog (D1, D2) conditions. The first and second rows compare light fog data, the third and fourth rows compare moderate fog data, and the fifth and sixth rows compare dense fog data.

[0213] As can be seen from the above: In light fog environment, BCCR[7] still has some haze remaining, and the processing of white objects is not good, so the processing result is relatively dark; other algorithms have a good defogging effect on the target; in medium fog and dense fog environment, Nonlocal[9] has exposure of near sky pixels and obvious artifacts; compared with the algorithm in this embodiment, the algorithm in this embodiment has a better overall target restoration effect, and the target processing effect of near and far scenes is better than the other methods.

[0214] In this embodiment, Figure 18The image shows a comparison of the results of different defogging algorithms under color polarized fog data. Since the atmospheric polarization degree of the method in this embodiment changes with the scene depth, it has a better defogging effect than IPD

[10] , GPLPF

[25] , and PLF

[28] when processing distant targets, making the details of distant targets richer and the noise relatively small; BCCR[7] still shows uneven and incomplete defogging; while in contrast entropy, the method in this embodiment is better than Nonlocal[9], and the processed result is more natural.

[0215] Example 9

[0216] This embodiment performs a quantitative analysis of the experimental results of Embodiments 6 and 7. The experiments in this embodiment were conducted on a PC with a Windows 10 system, an Intel Core i5-8250U CPU, 8GB RAM, and a Windows 10 64-bit operating system. The experimental compilation software was Matlab R2018a.

[0217] To verify the effectiveness of the proposed algorithm, this embodiment uses polarized images of foggy weather taken in real-world scenarios for experimentation. Furthermore, this embodiment compares the algorithm with classic dehazing algorithms such as BCCR[7], Nonlocal[9], IPD

[10] , GPLPF

[25] , and PLF

[28] .

[0218] In this embodiment, the image acquisition device is a polarization camera with a calibrated linear polarizer. This device can simultaneously acquire three original polarized foggy images with different polarizer directions (0°, 60°, and 120°) in a single exposure under natural light, with an image resolution of 700×934. The total light intensity image obtained from the stokes vector is used as the input of BCCR[7] and Nonlocal[9]. IPD

[10] takes the synthesized maximum and minimum polarization images as its input, while GPLPF

[25] and PLF

[28] take the original polarized foggy images as their input.

[0219] In order to objectively verify the effectiveness of the method proposed in the above embodiments, this embodiment uses non-reference quality evaluation indicators such as Natural Image Quality Evaluation Index (NIQE)

[31] , Fog Density Evaluation Index (FADE)

[32] , Image Entropy (IE) and Average Gradient (MG) to quantitatively analyze the experimental data.

[0220] In this embodiment, the Natural Image Quality Evaluation (NIQE) metric is calculated by statistically analyzing 36-dimensional features (NSS) that can measure the quality of natural, haze-free images. A multivariate Gaussian MVG model is used for modeling, and the differences between the test image and the MVG model parameters of the natural, haze-free image in terms of multivariate distribution are calculated. The smaller the NIQE, the closer the image quality is to its natural state, and the better the quality.

[0221]

[0222]

[0223] In the formula, f X (x1, x2, ... x k ) is the multivariate Gaussian distribution density function of the 36-dimensional NSS features, where υ1, υ2 and ∑1, ∑2 are the mean and covariance matrices of the fitted Gaussian distributions of the image to be tested and the natural haze-free image, respectively.

[0224] In this embodiment, the fog density assessment index FADE is a no-reference perceived fog density prediction index based on natural scene statistical features (36 dimensions) and fog perception statistical features (12 dimensions). The fog perception features are fitted into the MVG model, and then the differences in statistical patterns observed on the image under test, natural fog images, and fog-free images are calculated. Finally, the perceived fog density is expressed as the ratio of fog layer to fog-free layer.

[0225] In this embodiment, a reference index is used to describe the fog density of an image. The smaller the FADE value, the lower the fog density of the scene and the clearer the image.

[0226]

[0227]

[0228]

[0229] In the formula, f is the statistical characteristics of fog perception obtained from natural fog maps, and D... f The difference between the test image and the natural fog map in the MVG model parameters, D ff It is the difference in MVG model parameters between the image under test and the natural haze-free image.

[0230] In this embodiment, image entropy (IE) is a statistical form of feature that reflects the richness of image information based on the ordered distribution of grayscale values ​​among the pixels. The higher the information entropy in an image, the more information it contains.

[0231]

[0232] In the formula, P ij Let M be the probability of the grayscale value of the pixel at coordinate (i, j). M and N represent the height and width of the image, respectively.

[0233] In this embodiment, the average gradient (MG) is an important indicator for evaluating image sharpness, reflecting the details and texture variations in an image. The larger the average gradient, the sharper the image appears. The formula for calculating MG is:

[0234]

[0235] In the formula, Δ i G and Δ j G represents the gradient of the image in the horizontal and vertical directions, respectively.

[0236] Table 1 of this embodiment presents a comparison of the objective results of averaging the original image, BCCR[7], Nonlocal[9], IPD

[10] , GPLPF

[25] , PLF

[28] , and the data after dehazing by the method of this embodiment under four indicators: fog density evaluation index FADE

[32] , natural image quality evaluation index NIQE

[31] , information entropy IE, and average gradient MG. The bold text in the table represents the best result, and the underlined result represents the second best result.

[0237] As can be seen from the comprehensive results in Table 1, the algorithm in this embodiment is inferior to GPLPF

[25] and PLF

[28] in terms of IE and NIQE

[31] , but the difference in numerical values ​​is not significant. Obviously, FADE

[32] , NIQE

[31] and MG are superior to other algorithms in terms of average value, with numerical values ​​improving by an average of 28.8%, 0.8% and 18.0% respectively compared to the second-best results. This indicates that the reconstructed result has the lowest scene fog density, the image is closest to the natural image, and the quality is better. At the same time, the reconstructed image contains more edge detail information.

[0238] Table 1. Quantitative analysis results of different dehazing algorithms under single-channel polarization haze images.

[0239]

[0240]

[0241] Example 10

[0242] This embodiment provides an electronic device, characterized in that it includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. The memory is coupled to the processor, and when the processor executes the computer program, it implements the steps in the polarization fog map reconstruction method based on atmospheric optical polarization orthogonal decomposition of the above embodiment.

[0243] In summary, this invention addresses the problem of estimating atmospheric polarization degree and atmospheric scattered light under existing dense fog conditions by proposing a polarization fog map reconstruction method based on orthogonal decomposition of atmospheric light polarization at infinity.

[0244] The method of this invention makes full use of the polarization characteristics of light and combines the atmospheric polarization characteristics of the Stokes vector to solve the depth-related atmospheric polarization degree matrix, thereby reducing the error of manual interaction and improving the ability to restore targets at different scene depths.

[0245] This invention decomposes atmospheric light polarization orthogonal parameters at infinity in polarization space and constructs a blind fog map reconstruction model with multiple regularization constraints to improve the estimation accuracy of atmospheric scattered light.

[0246] This invention combines polarization difference information to reduce the impact of white, bright objects on the estimation of atmospheric light at infinity.

[0247] Experimental results show that the present invention significantly improves the visibility of target scenes at different depths subjectively, and also further enhances the quality of reconstructed images objectively, making it more adaptable to scenes with different fog densities.

[0248] The following references are used in this invention:

[0249] [1] HALLBERGT, ERIKSSONJ, et al. Optical polarization: background and camouflage [C]. Society of Photo-optical Instrumentation Engineers. Society of Photo-optical Instrumentation Engineers (SPIE) Conference Series, 2017: 1043204.

[0250] [2] NEVEND, DEBRABANDEREB, GEORGOULISS, et al. Towards end-to-endlane detection: an instance segmentation approach [C]. 2018 IEEE Intelligent Vehicles Symposium (IV), 2018: 286-291.

[0251] [3] PARNETF, FADEJ, ORTEGAQUIJANON, et al.Free-spaceactivepo1arimetricimagerperatingat1.55μmbyorthogonalitybreakingsensing[J].0pticsLetters, 2017, 42(4):723-726.

[0252] [4]JUNWL,RONGZ.Imagedefoggingalgorithmofsinglecolorimagebasedonwavelettransformandhistogramequalization[J].AppliedMathematicalSciences,2013,7(79):3913-3921.

[0253] [5]SHENHF,LIHF,QIANY,etal.Aneffectivethincloudremovalprocedureforvisibleremotesensingimages[J].ISPRSJournalofPhotogrammetryandRemoteSensing,2014,96:224-235.

[0254] [6]HEKM,SUNJ,TANGXO.Singleimagehazeremovalusingdarkchannelprior[J].IEEETransactionsonPatternAnalysisandMachineInte11igence,2011,33(12):2341-2353.

[0255] [7]MENGGF,WANGY,DUANJY,etal.Efficientimagedehazingwithboundaryconstraintandcontextualregularization[C].ProceedingsoftheIEEEInternationalConferenceonComputerVision,2013:617-624.

[0256] [8]ZHUQS,MAIJM,SHAOL.Afastsingleimagehazeremovalalgorithmusingcolorattenuationprior[J].IEEETransactionsonImageProcessing,2015,24(11):3522-3533.

[0257] [9]BERMAND,AVIDANS.Non-1ocalimagedehazing[C].ProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition,2016:1674-1682.

[0258]

[10] SCHECHNERYY,NARASIMHANSG,NAYARSK.Instantdehazingofimagesusingpo1arization[C].ProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition.CVPR2001,2001:1-8.

[0259]

[11] SCHECHNERYY,NARASIMHANSG,NAYARSK.Polarization-basedvisionthroughhaze[J].Appliedoptics,2003,42(3):511-525.

[0260]

[12] FANGS,XIAXS,HUOX,etal.Imagedehazingusingpolarizationeffectsofobjectsandairlight[J].OpticsExpress,2014,22(16):19523-19537.

[0261]

[13] ZHANGWF,LIANGJ,RENLY.Haze-removalpolarimetricimagingschemeswiththeconsiderationofairlight′scircularpolarizationeffect[J].Optik,2019,182:1099-1105.

[0262]

[14] HUANGF,KECZ,WUXY,etal.Polarizationdehazingmethodbasedonspatialfrequencydivisionandfusionforafar-fieldanddensehazyimage[J].AppliedOptics,2021,60(30):9319-9332.

[0263]

[15] LIANGJ,RENLY,JUHJ,etal.Polarimetricdehazingmethodfordensehazeremovalbasedondistributionanalysisofangleofpolarization[J].0pticsExpress,2015,23(20):26146-26157.

[0264]

[16] QUYF,ZOUZF.Non-skypo1arization-baseddehazingalgorithmfornon-specularobjectsusingpo1arizationdifferenceandgloba1scenefeature[J].0pticsExpress,2017,25(21):25004-25022.

[0265]

[17] LIANGJ,RENLY,JUHJ,etal.Visibilityenhancementofhazyimagesbasedonauniversalpolarimetricimagingmethod[J].JournalofAppliedPhysics,2014,116(17):820-827.

[0266]

[18] LIANGJ,ZHANGWF,RENLY,etal.Polarimetricdehazingmethodforvisibilityimprovementbasedonvisibleandinfraredimagefusion[J].AppliedOptics,2016,55(29):8221-8226.

[0267]

[19] SHENLH,REDAM,ZHAOYQ.Image-matchingenhancementusingapolarizedintensity-hue-saturationfusionmethod[J].AppliedOptics,2021,60(13):3699-3715.

[0268]

[20] LEIY, LEIB, CAIY, et al.PolarimetricDehazingMethodBasedonImageFusionandAdaptiveAdjustmentAlgorithm[J].AppliedSciences, 2021, 11(21):10040.

[0269]

[21] Dai Qingqing, Fan Zhiguo, Song Qiang, et al. A method for dehazing polarization of color images by automatic estimation of global parameters [J]. Applied Optics, 2018, 39(4): 511-517.

[0270]

[22] Xu Jingyi, Gao Jun, Zhang Yangyang, et al. Analysis of polarization characteristics of light scattering and transmission in media of different concentrations [J]. Acta Electronica Sinica, 2015, 43(08): 1604-1609.

[0271]

[23] Zhao Lujian, Gao Jun, Bi Ran, et al. Polarization dehazing method based on maximum and minimum light intensity images [J]. Applied Optics, 2017, 38(3): 415-420.

[0272]

[24] Zhou Wenzhou, Fan Chen, Hu Xiaoping, et al. Algorithm and experiment of polarization image fusion dehazing based on multi-scale singular value decomposition [J]. Chinese Optics, 2021, 14(2): 298-306.

[0273]

[25] LIANGJ, JUHJ, RENLY, et al. GeneralizedPolarimetricDehazingMethodBasedonLow-PassFilteringinFrequencyDomain[J]. Sensors, 2020, 20(6): 1-12.

[0274]

[26] LIANGZ, DINGXY, MIZ, et al. Effective polarization-based image dehazing with regularization constraint [J]. IEEE Geoscience and Remote Sensing Letters, 2020, 19: 1-5.

[0275]

[27] SHENLH, ZHAOYQ, PENGQN, et al. Aniterative image dehazing method with polarization [J]. IEEE Transaction son Multimedia, 2019, 21(5): 1093-1107.

[0276]

[28] LIANGJ, RENL, LIANGR. Low-pass filtering based polarimetric dehazing method fordense hazeremoval [J]. Optics Express, 2021, 29(18): 28178-28189.

[0277]

[29] Shao Ziqi, Jin Haihong, Qian Lijin, et al. Image reconstruction method for foggy weather based on atmospheric optical polarization tomography [J]. Progress in Laser & Optoelectronics, 2020, 57(22): 113-123.

[0278]

[30] COXLJ. Optics of the Atmosphere-Scattering by Molecules and Particles [J]. Optica Acta: International Journal of Optics, 1977, 24(7): 779-779.

[0279]

[31] MITTALA, SOUNDARARAJANR, BOVIKAC. Making a "completelyblind" image quality analyzer [J]. IEEE Signalprocessing letters, 2012, 20(3): 209-212.

[0280]

[32] CHOILK, YOUJ, BOVIKAC. Referencelesspredictionofperceptualfogdensityandperceptualimagedefogging[J]. IEEETransactionsonImageProcessing, 2015, 24(11): 3888-3901.

[0281] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fog map reconstruction method based on atmospheric light polarization orthogonal blind separation, characterized in that, The method comprises the following steps: S1, acquiring multiple fog-polarization images with different polarization angles under the same exposure condition by an imaging detection device; S2, calculating stokes vector, total polarization degree and total polarization angle information according to the fog-polarization images, and automatically estimating the depth-related atmospheric polarization degree matrix; S3, acquire the gray level histogram of the polarization degree offset image and the total light intensity image, and estimate the atmospheric light at infinity by using the Ostu adaptive method ; S4, atmospheric light at infinity A multi-regularization constrained blind separation model of atmospheric scattering light is constructed from the perspective of polarization orthogonal decomposition to obtain the atmospheric scattering light; the atmospheric scattering light is decomposed into maximum and minimum polarization orthogonal images in the polarization space and ; ； Acquiring maximum and minimum polarization orthogonal images of atmospheric light at infinity and ; ； Where, p A Atmospheric polarization degree Define intermediate quantities for atmospheric scattered light. and Its meaning is and Having the same partial polarization characteristics, their intensity values ​​change with the polarizer angle, and a maximum value is observed. and minimum value The degree of atmospheric polarization varies with the depth of the scene target; based on Malus's law, an orthogonal image of the maximum and minimum polarization of atmospheric light at infinity is synthesized, simplifying the atmospheric scattering light model: ； In the formula, , , related to , t is the transmittance; is the dot product symbol; and a multi-regularization constraint atmospheric scattering light blind source separation objective function is constructed to obtain the atmospheric scattering light as follows: ； wherein ; wherein are different sets of weight indices; is an adaptive weight matrix; is a high-order difference operator; S5, the polarization fog map is reconstructed by using the atmospheric scattering model, and the fog-free map of the undegraded scene is obtained by inversion; when the polarization fog map is reconstructed by using the atmospheric scattering model, the total light intensity is represented as: ; wherein , ; wherein is the transmittance, and the scene depth is exponentially related; is the target reflected light directly reflected by the target scene; is the undegraded scene haze-free map; is the atmospheric scattered light; is the atmospheric light at infinity, the polarization haze map reconstruction backpropagation evolves the expression: .

2. The method of claim 1, wherein the method is based on atmospheric light polarization orthogonal blind source separation. In the method, the stokes vector : ； wherein is the total light intensity information of the incident light, and is the polarization information of the incident light, , and are the fog polarization images with polarizer angles of 0°, 60° and 120°.

3. The method of claim 2, wherein the method is based on atmospheric light polarization orthogonal blind source separation. In the method, the total degree of polarization is obtained and the total angle of polarization The image is specifically: ； ； wherein, contains all the information of the target scene, and with is determined by the atmospheric scattered light.

4. The method of claim 3, wherein the method is based on atmospheric light polarization orthogonal blind source separation. In the method, and according to the frequency priority strategy, the total polarization angle is counted The gray histogram of the image, and the value with the highest frequency is taken as the atmospheric polarization angle Then the atmospheric polarization angle The atmospheric polarization degree varying with the scene depth is obtained : 。 5. The method of claim 1, wherein the method is based on atmospheric light polarization orthogonal blind source separation. In the method, the polarization degree offset image and the total light intensity image are acquired respectively and the gray scale histogram is acquired The sky and target segmentation threshold is acquired by using the Ostu adaptive method; wherein, ∇p represents the polarization degree difference, p A is the atmospheric polarization degree, and p is the total polarization degree.

6. The method of claim 5, wherein the method is based on atmospheric light polarization orthogonal blind source separation. The method obtains the gray value space coordinates greater than the threshold value in obtains the gray value space coordinates less than the threshold value in , and the intersection coordinates of the two coordinate spaces are that is, the atmospheric light at infinity the position with the highest probability, and finally taking the average, wherein .

7. An electronic device, comprising: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, the memory is coupled with the processor, and the processor executes the computer program to implement the steps in the polarization fog image reconstruction method based on atmospheric light polarization orthogonal decomposition according to any one of claims 1 to 6.

Images