A defogging method and system based on atmospheric scattering model parameter estimation

Abstract

The present application relates to the technical field of image defogging, in particular to a defogging method and system based on atmospheric scattering model parameter estimation. The present application calculates the sky feature confidence function value of each pixel based on the gradient, brightness and color saturation of each pixel of the foggy image to be processed, to screen the initial sky candidate pixels; removes the pixels with brightness exceeding the brightness overexposure threshold in the initial sky candidate pixels to obtain the diffuse reflection sky area mask, calculates the local spatial variance of each pixel in the diffuse reflection sky area mask, sorts each pixel in the diffuse reflection sky area mask in ascending order according to the size of the local spatial variance, and takes the pixel mean value of the front set number of pixels on the three color channels of RGB as the global atmospheric light vector; based on the global atmospheric light vector and the transmittance of each pixel position in the foggy image to be processed, the defogging image is obtained, which effectively reduces the phenomenon of large-area blackening or abnormal color drift of the defogging image.

Description

Technical Field

[0001] This invention relates to the field of image dehazing technology, and in particular to a dehazing method and system based on atmospheric scattering model parameter estimation. Background Technology

[0002] Image dehazing is an important research direction in image processing and computer vision, mainly used to improve image degradation problems such as reduced contrast, color shift, and blurred details at long distances under atmospheric conditions like fog. Early dehazing research was primarily based on atmospheric scattering models, restoring clear images by estimating atmospheric light and transmittance. Representative methods included dark channel priors, color attenuation priors, and fog line priors. With the development of deep learning, convolutional neural networks, generative adversarial networks, and attention mechanisms have been gradually introduced into image dehazing tasks, driving the development of dehazing methods from manual prior modeling to data-driven mapping. In recent years, researchers have begun to focus on the constraints of depth information, scene structure, and physical imaging models on the dehazing process to improve the stability and interpretability of dehazing results in complex outdoor scenes.

[0003] Current mainstream image dehazing techniques fall into three categories, each relying on its core principles to restore the clarity of foggy images. Image enhancement-based methods do not require explicit construction of a foggy imaging degradation model. They primarily utilize conventional image processing techniques such as histogram equalization, Retinex theory, local contrast enhancement, and image sharpening to directly optimize the grayscale, color, and contrast distribution of the image, thereby improving the overall visual effect of foggy images. This type of method boasts advantages such as simplicity and high computational speed. Physical model-based image restoration methods use atmospheric scattering models as their core foundation, accurately depicting the imaging degradation process of foggy images. By estimating the two core parameters of atmospheric light and transmittance, and combining model inversion calculations, they restore fog-free, clear images. Dark channel priors, color attenuation priors, and fog line priors are representative algorithms in this category, and they generally possess a certain degree of physical interpretability. Deep learning-based dehazing methods rely on deep learning models such as convolutional neural networks, generative adversarial networks, attention mechanisms, and Transformers to autonomously learn the mapping relationship between foggy and clear images, or directly learn intermediate imaging parameters such as transmittance and atmospheric light. They achieve image dehazing using massive training data, achieving good dehazing recovery results in synthetic datasets. In addition, some recent studies have introduced binocular images, depth estimation, and stereo matching results to assist in single-image dehazing, supplementing scene depth information to compensate for the lack of information in a single image. However, existing methods mostly focus on jointly estimating disparity maps and dehazed images, or directly predicting transmittance and atmospheric light, without fully decoupling and modeling the multiple unknown parameters in the atmospheric scattering model.

[0004] Existing physical model-based dehazing solutions generally suffer from poor anti-interference performance in atmospheric light estimation, thus affecting the overall restoration effect. Traditional atmospheric light estimation methods mostly select the brightest pixel in the entire image as the atmospheric light value. In real outdoor scenes, there are generally bright areas such as sky areas, solar halos, white buildings, road surfaces, water reflections, and overexposed strong light. Algorithms are very prone to misselecting these bright interference areas as atmospheric light sampling locations. Sky texture fluctuations, object reflections, and sunlight highlights can all seriously interfere with atmospheric light estimation, causing significant deviations in atmospheric light parameters. Incorrect atmospheric light values ​​interfere with subsequent calculations, ultimately inducing various image distortion defects, causing large areas of the image to turn black and color drift after dehazing, such as abnormal color casts like fluorescent green appearing in tree areas. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the defects in the prior art that interference factors such as bright objects, road surface reflections, and sunlight halos can easily cause errors in atmospheric light estimation, which in turn lead to blacking of the image and abnormal color drift.

[0006] To address the aforementioned technical problems, this invention provides a dehazing method based on atmospheric scattering model parameter estimation, comprising: Based on the gradient, brightness, and color saturation of each pixel in the foggy image to be processed, calculate the sky feature confidence function value for each pixel; Based on the confidence function value of the sky features of each pixel, the initial sky candidate pixels in the foggy image to be processed are selected. Remove pixels from the initial sky candidate pixels whose brightness exceeds the brightness overexposure threshold to obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask; Based on the magnitude of the local spatial variance, each pixel in the diffuse sky region mask is sorted in ascending order, and the average pixel value of the first set number of pixels in the three RGB color channels is taken as the global atmospheric light vector. Based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed, a dehazed image is obtained.

[0007] Preferably, the sky feature confidence function value for each pixel is calculated. The formula is: , in, Let q be the confidence function value of the sky features. It is an exponential function with the natural constant e as its base. This indicates taking the maximum value. and The threshold constant is Let be the gradient of pixel q. The brightness of pixel q The average brightness of all pixels in the hazy image to be processed. Let q represent the color saturation of pixel q.

[0008] Preferably, the method for selecting initial sky candidate pixels in a hazy image to be processed based on the sky feature confidence function value of each pixel includes: The target selection threshold is obtained by multiplying the maximum sky feature confidence function value in the foggy image to be processed by a preset percentage; Pixels whose sky feature confidence function value is greater than the target screening threshold are used as initial sky candidate pixels.

[0009] Preferably, the formula for calculating the local spatial variance of each pixel in the diffuse sky region mask is: , in, Let be the local spatial variance of pixel q. A local window centered at pixel q. For pixel p in the color channel The strength value on, for The total number of pixels in the array, p is Any pixel inside, These are the image color channels, where R, G, and B represent the red, green, and blue channels, respectively.

[0010] Preferably, the method for obtaining the transmittance at each pixel location in the hazy image to be processed includes: Obtain the foggy image to be processed and its depth map; Extract the scattering coefficient response map corresponding to the foggy image to be processed, and perform a global average to obtain the global scattering coefficient; Based on the depth map and global scattering coefficient of the foggy image to be processed, the transmittance at each pixel location in the foggy image to be processed is calculated.

[0011] Preferably, the method for obtaining the foggy image to be processed and its depth map includes: The foggy images in the left and right eyes were acquired using a binocular camera. The foggy image in the left eye or the foggy image in the right eye is used as the foggy image to be processed, and the other image is used as the matching reference image. The disparity map between the foggy image to be processed and the matching reference image is calculated by using a binocular stereo matching algorithm. The disparity map is calculated based on the focal length and baseline distance of the binocular camera to obtain the depth map corresponding to the foggy image to be processed.

[0012] Preferably, the method for extracting the scattering coefficient response map corresponding to the foggy image to be processed includes: The foggy image to be processed is passed through multiple convolutional layers connected in sequence to obtain fog features at different scales; Fog features at different scales are fused using the RefineNet module to obtain fused features; The fused features are passed through a projection head to obtain the scattering coefficient response map corresponding to the foggy image to be processed.

[0013] Preferably, after calculating the transmittance at each pixel location in the hazy image to be processed, the method further includes optimizing the transmittance at each pixel location in the hazy image to be processed, specifically including: Set a lower limit for transmittance. For pixels in the foggy image to be processed whose transmittance is lower than the lower limit, use the lower limit as their optimized transmittance.

[0014] Preferably, the dehazed image is obtained based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed, using the following formula: , in, This represents the pixel brightness value at pixel location x in the dehazed image. Let x be the transmittance of the hazy image to be processed at pixel location x. Let x be the pixel intensity at pixel location x in the hazy image to be processed. This is the global atmospheric light vector.

[0015] This invention also provides a defogging system based on atmospheric scattering model parameter estimation, comprising: The Sky Feature Confidence Function Calculation Module is used to calculate the sky feature confidence function value for each pixel based on the gradient, brightness, and color saturation of each pixel in the foggy image to be processed. The filtering module is used to filter initial sky candidate pixels in the hazy image to be processed based on the sky feature confidence function value of each pixel. The local spatial variance calculation module is used to remove pixels whose brightness exceeds the brightness overexposure threshold from the initial sky candidate pixels, obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask. The global atmospheric light vector acquisition module is used to sort each pixel in the diffuse sky region mask in ascending order according to the size of the local spatial variance, and take the average value of the first set number of pixels in the three color channels of RGB as the global atmospheric light vector. The dehazing module is used to obtain a dehazed image based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed.

[0016] This application provides a defogging method and system based on atmospheric scattering model parameter estimation, which has the following beneficial effects: 1. To address the problem that traditional methods easily misidentify bright areas as atmospheric light, this invention constructs a multi-dimensional feature joint purification strategy that integrates image gradient, brightness, color saturation, and local spatial variance, and sets a strict brightness overexposure threshold to forcibly remove abnormal pixels. This technique accurately locks in the atmospheric light benchmark representing true diffuse reflection, successfully avoiding interference from sunlight halos, white building reflections, or road surface reflections in outdoor scenes, fundamentally eliminating the situation of large areas of blackening or abnormality in the image after dehazing.

[0017] 2. To address the problem of ill-posed parameter estimation in traditional single-image dehazing due to the lack of reliable depth constraints, this invention extracts an absolutely dense depth map through binocular stereo matching and directly injects it as a geometric hard constraint into the transmittance calculation equation. This technique effectively removes the information entanglement between transmittance and medium concentration (scattering coefficient), transforming the highly ill-conditioned blind mapping into a physically well-posed solution process. This fundamentally avoids the distant haze and local over-enhancement phenomena caused by transmittance being affected by object color and texture.

[0018] 3. To address the problem that existing methods, by setting the scattering coefficient to a fixed constant, cannot adapt to changing weather conditions, this invention designs a scattering coefficient prediction module. This module fully utilizes the multi-scale features of foggy images, keenly captures the intensity of spatial degradation, and adaptively outputs the scattering coefficient response map of the current environment. This breaks through the limitations of existing methods that rely on a single prior or fixed parameters, enabling the algorithm's defogging intensity to dynamically adapt to different outdoor scenes with varying concentrations of fog, such as light fog, heavy fog, and dense fog. This effectively avoids excessive defogging in light fog scenes and fog residue in dense fog scenes. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the steps of a defogging method based on atmospheric scattering model parameter estimation according to the present invention. Figure 2 This is a structural diagram of a defogging method based on atmospheric scattering model parameter estimation according to the present invention; Figure 3 Here is a structural diagram of the scattering coefficient prediction module; Figure 4 This is a schematic diagram illustrating the segmentation effect of a mask for a diffuse sky region.

[0020] Figure 5 This is a schematic diagram comparing the atmospheric light estimation method of the present invention with the traditional dark channel method.

[0021] Figure 6 This is a schematic diagram of the training process for the scattering coefficient prediction module.

[0022] Figure 7 These are comparison images of dehazing effects on the KITTI binocular synthetic fog dataset. Figure 7 In the image (a), the input hazy image to be processed is shown. Figure 7 (b) in the figure is a schematic diagram of the dehazing effect of DCP on the KITTI stereo synthetic fog dataset. Figure 7 (c) in the figure is a schematic diagram of the dehazing effect of D4 on the KITTI stereo synthetic fog dataset. Figure 7 (d) in the diagram is a schematic of the dehazing effect of Cycle-SNSPGAN on the KITTI stereo synthetic fog dataset. Figure 7 (e) in the figure is a schematic diagram of the dehazing effect of UCL-Dehaze on the KITTI stereo synthetic fog dataset. Figure 7 (f) in the figure is a schematic diagram of the dehazing effect of combining the transmittance acquisition method of the present invention with the DCP atmospheric light method on the KITTI binocular synthetic fog dataset. Figure 7 (g) in the figure is a schematic diagram of the defogging effect of the method of the present invention on the KITTI binocular synthetic fog dataset.

[0023] Figure 8 These are comparison images of defogging effects on the DrivingStereo real traffic dataset. Figure 8 In the image (a), the input hazy image to be processed is shown. Figure 8 (b) in the diagram illustrates the defogging effect of DCP on the DrivingStereo real traffic dataset. Figure 8 (c) in the figure is a schematic diagram of the defogging effect of D4 on the DrivingStereo real traffic dataset. Figure 8 (d) in the diagram illustrates the dehazing effect of Cycle-SNSPGAN on the DrivingStereo real traffic dataset. Figure 8 (e) in the diagram is a schematic representation of the dehazing effect of UCL-Dehaze on the DrivingStereo real traffic dataset. Figure 8 (f) in the figure is a schematic diagram of the defogging effect of combining the transmittance acquisition method of the present invention with the DCP atmospheric light method on the DrivingStereo real traffic dataset. Figure 8 (g) in the figure is a schematic diagram of the defogging effect of the method of the present invention on the DrivingStereo real traffic dataset. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0025] refer to Figure 1 , Figure 1 The dehazing method based on atmospheric scattering model parameter estimation provided in this embodiment includes: This invention's method is based on an atmospheric scattering model. It models and estimates scene depth, atmospheric light, scattering coefficient, and transmittance during the restoration of foggy images, and then retrieves a clear image based on the estimated physical parameters. For example... Figure 2 As shown, Figure 2 This is a structural diagram of a defogging method based on atmospheric scattering model parameter estimation according to the present invention.

[0026] The atmospheric scattering model is expressed as: , in, Let x be the pixel intensity at pixel location x in the hazy image to be processed. This represents the pixel brightness value at pixel location x in the dehazed image. Let x be the transmittance of the hazy image to be processed at pixel location x. is the global atmospheric light vector, and x represents the pixel position.

[0027] Transmittance at pixel location x of the hazy image to be processed for: , in, The scattering coefficient is... Let x be the depth of the foggy image to be processed at pixel location x, and e be the natural constant. Depth represents the proportion of scene radiation that remains after it propagates from the object's surface to the imaging device. The larger the value, the longer the light propagation path and the lower the transmittance; the scattering coefficient... The larger the value, the higher the fog concentration, and the more significant the attenuation of transmittance.

[0028] This invention will , and Instead of directly estimating all unknowns by mixing them into a single black-box network, different technical approaches are used to obtain them. The specific solution is as follows: Step S1: Based on the gradient, brightness, and color saturation of each pixel in the hazy image to be processed, calculate the sky feature confidence function value for each pixel; The formula for calculating the sky feature confidence function value for each pixel is: , in, Let q be the confidence function value of the sky features. It is an exponential function with the natural constant e as its base. This indicates taking the maximum value. and These are threshold constants, used to control the influence of the gradient term and color saturation term on the selection of sky candidate regions, respectively. In this embodiment, Take 0.005, The value is set to 0.04. This value is based on empirical settings used in sky region segmentation methods, where sky regions typically exhibit low gradient, high brightness, and low saturation characteristics. Used to filter out high-gradient areas such as building edges and tree textures. Used to filter out high-saturation areas such as vehicles and traffic signs. Let be the gradient of pixel q. The brightness of pixel q The average brightness of all pixels in the hazy image to be processed. Let q represent the color saturation of pixel q.

[0029] In outdoor scenes that include sky areas, atmospheric light is typically associated with distant skies or low-texture, high-brightness background areas. To accurately separate the sky area from the ground foreground, this invention directly extracts light from the original foggy image. Features are extracted based on pixel attributes.

[0030] The sky feature confidence function utilizes exponential decay to amplify the weights of pixels that meet the conditions of "low gradient, high brightness, and low saturation".

[0031] Step S2: Based on the sky feature confidence function value of each pixel, filter the initial sky candidate pixels in the hazy image to be processed; In this embodiment, preferably, the method for selecting initial sky candidate pixels in the hazy image to be processed based on the sky feature confidence function value of each pixel includes: The maximum sky feature confidence function value in the foggy image to be processed Multiply by the preset percentage to obtain the target screening threshold; Pixels with a sky feature confidence function value greater than the target screening threshold are used as initial sky candidate pixels. In this embodiment, the preset percentage ranges from 90% to 98%. A higher preset percentage results in stricter screening and reduces interference from bright objects and reflective areas on the ground; a lower preset percentage results in a more complete candidate region but may introduce non-sky areas. Therefore, this embodiment selects a preset percentage of 95% to balance the integrity of the candidate region and its anti-interference capability.

[0032] Step S3: Remove pixels whose brightness exceeds the brightness overexposure threshold from the initial sky candidate pixels to obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask; Furthermore, in order to eliminate pure white overexposed areas caused by direct sunlight or limited camera dynamic range (these areas are not the result of atmospheric scattering), a strict brightness overexposure threshold is set. Pixels in the initial sky candidate pixels whose brightness exceeds the brightness overexposure threshold are marked as abnormal highlight areas and removed. The remaining area is the mask for the diffuse sky area unaffected by direct light. In this embodiment, the brightness overexposure threshold... The value range is 0.90 to 0.99. Brightness overexposure threshold. When the value is too low, it is easy to accidentally delete normal sky areas and cause overexposure of the brightness threshold. If the value is too high, it may retain overexposed pixels and cause deviations in atmospheric light estimation. Therefore, the brightness overexposure threshold... The preferred value is 0.95.

[0033] like Figure 4 As shown, Figure 4 This is a schematic diagram illustrating the segmentation effect of the diffuse sky region mask. The first column shows the original images, containing two original images. The second column shows the sky feature maps corresponding to the original images, i.e., the confidence functions of the sky features for each pixel in the original images. A visualized heatmap, where colors represent the confidence function of sky features for pixels. The numerical values ​​are indicated by color bars, and the third column is the diffuse sky region mask corresponding to the original image. In the mask, the white area is the diffuse sky region obtained by judgment, and the black area is the non-diffuse sky region.

[0034] In physics, the purest atmospheric background scattered light should possess extremely high spatial homogeneity. Therefore, in the diffuse sky region mask... Within this framework, the local spatial variance of each pixel is calculated using a fixed-size sliding window (15×15): The formula for calculating the local spatial variance of each pixel in the diffuse sky region mask is: , in, Let be the local spatial variance of pixel q. A local window centered at pixel q. For pixel p in the color channel The strength value on, for The total number of pixels in the array, p is Any pixel inside, These are the image color channels, where R, G, and B represent the red, green, and blue channels, respectively.

[0035] The smaller the local variance, the smoother the color change of the point and its surrounding neighborhood, and the more it conforms to the ideal diffuse reflection characteristics in physics.

[0036] Step S4: Sort each pixel in the diffuse sky region mask in ascending order according to the magnitude of the local spatial variance, and take the first set number of pixels. The average pixel value across the RGB color channels is used as the global atmospheric light vector. ; Preset quantity The average pixel value of the pixel in the red channel. Preset quantity The average pixel value of the pixels in the green channel. Preset quantity The average pixel value of the pixels in the blue channel.

[0037] In this embodiment, The value is determined by 0.1% of the total number of pixels in the foggy image to be processed, which is a reference to the common setting in the dark channel prior for estimating atmospheric light by selecting the first 0.1% of pixels from the candidate region. Values ​​that are too small are easily affected by a single abnormal pixel. If the value is too large, it may introduce building edges, reflective areas, or uneven sky areas.

[0038] like Figure 5 As shown, Figure 5 This is a schematic diagram comparing the atmospheric light estimation method of the present invention with the traditional dark channel method.

[0039] Step S5: Based on the global atmospheric light vector and the transmittance of each pixel in the foggy image to be processed, obtain the dehazed image.

[0040] In this embodiment, preferably, the method for obtaining the transmittance of each pixel location in the foggy image to be processed includes: Obtain the foggy image to be processed and its depth map; Extract the scattering coefficient response map corresponding to the foggy image to be processed, and perform a global average to obtain the global scattering coefficient. The formula is as follows: , in, This represents the total number of pixels in the foggy image to be processed. This represents the value at pixel position x in the scattering coefficient response map corresponding to the hazy image to be processed; this global scattering coefficient... Used to describe the overall fog concentration in the current outdoor scene.

[0041] Based on the depth map and global scattering coefficient of the foggy image to be processed, the transmittance at each pixel location in the foggy image to be processed is calculated.

[0042] In this embodiment, preferably, the method for obtaining the foggy image to be processed and its depth map includes: Image of foggy conditions in the left eye was captured using a binocular camera. Image with fog on the right eye ; The foggy image in the left eye or the foggy image in the right eye is used as the foggy image to be processed, and the other image is used as the matching reference image. The disparity map between the foggy image to be processed and the matching reference image is calculated by using a binocular stereo matching algorithm. Through stereo matching, based on the focal length of the stereo cameras... and baseline distance disparity map The depth map corresponding to the foggy image to be processed is obtained by performing calculations, and the formula is: , in, Let x be the depth of the foggy image to be processed at pixel location x. Let x be the disparity of the foggy image to be processed at pixel location x.

[0043] For invalid parallax, occluded areas, or areas with abrupt changes in depth, median filtering, guided filtering, or confidence screening methods can be used for correction.

[0044] In other implementations, the depth map corresponding to the foggy image to be processed can also be obtained from an existing depth estimation model, a depth sensor, or offline depth data.

[0045] The lack of adaptive estimation for the scattering coefficient makes it difficult to adapt to different fog concentrations. Some existing methods set the scattering coefficient to a fixed empirical value or directly predict the transmittance without modeling the scattering coefficient separately. Since fog concentration, imaging distance, and lighting conditions vary in different outdoor scenes, fixing or implicitly processing the scattering coefficient reduces the physical consistency of transmittance calculations, leading to over-defogging in light fog scenes and residual fog in dense fog scenes.

[0046] To address the above problems, in this embodiment, preferably, the method for extracting the scattering coefficient response map corresponding to the foggy image to be processed includes: like Figure 3 As shown, Figure 3 This is a structural diagram of the scattering coefficient prediction module.

[0047] The hazy image to be processed is passed through the scattering coefficient prediction module to obtain the corresponding scattering coefficient response map of the hazy image to be processed, including: The foggy image to be processed is passed through multiple convolutional layers connected in sequence to obtain fog features at different scales; Fog features at different scales are fused using the RefineNet module to obtain fused features; The fused features are passed through a projection head to obtain the scattering coefficient response map corresponding to the foggy image to be processed.

[0048] It should be noted that the input to the scattering coefficient prediction module is a foggy image. Image features extracted from foggy images; scene depth map and atmospheric light It is not used as input to this module. Depth map For subsequent transmittance calculations, atmospheric light Used for subsequent image inversion.

[0049] The foggy image undergoes a feature encoding process to obtain multi-scale features, followed by a feature refinement process to enhance the fog concentration-related expression. Finally, a scattering coefficient response map is generated from the scattering coefficient output. .

[0050] In atmospheric scattering models, multiple unknown parameters are heavily coupled, leading to insufficient inversion stability. Physically-based image dehazing methods typically require simultaneous estimation of parameters such as the clear image, atmospheric light, transmittance, scene depth, and scattering coefficient from a hazy image. Due to the limited observable information in a single image, these parameters exhibit strong coupling, resulting in significant ill-posedness in the dehazing inversion process. When the scene depth is unknown or inaccurately estimated, transmittance is easily affected by object color, brightness, and texture distribution, leading to residual fog in distant areas, local over-enhancement, or loss of detail.

[0051] To address the problems in existing methods where scene depth, atmospheric light, scattering coefficient, and transmittance are interdependent, leading to ill-posed inversion processes and issues such as residual fog, color cast, and over-enhancement in defogging results, this invention introduces a binocular stereo vision-based absolutely dense depth map as a geometric prior to improve the stability of transmittance calculation under scene depth constraints. Furthermore, by fusing multi-dimensional features such as image gradient, brightness, and spatial variance for joint optimization, interference from the sky, bright objects, and reflective areas on atmospheric light estimation is reduced. In addition, this invention implements adaptive network prediction of the scattering coefficient, enabling the defogging intensity to adapt to outdoor scenes with varying fog concentrations and imaging distances.

[0052] like Figure 6 As shown, Figure 6 This is a schematic diagram of the training process for the scattering coefficient prediction module.

[0053] In this embodiment, optionally, the scattering coefficient prediction module is trained using an unsupervised closed-loop framework. The training process is as follows: During the training phase, fog-free stereo image pairs are input and the true depth map is calculated. Randomly generate true scattering coefficients within a given range. A physically consistent hazy image is synthesized using a forward rendering equation. This synthesized hazy image is then fed into a scattering coefficient prediction module to predict the corresponding scattering coefficients. Obtain the reconstructed dehazed image The joint loss function during training includes: (1) Parameter monitoring loss : To force the network to learn the nonlinear mapping between image blurriness and physical scattering intensity, the L1 number is used as the optimization criterion, and the formula is: , in, The scattering coefficients predicted by the scattering coefficient prediction module. These are the true scattering coefficients from random sampling.

[0054] (2) Defogging monitoring loss : To constrain the network to reconstruct a clear image of the initial input without loss after undergoing a complete forward and reverse physical cycle, a formula combining pixel-level absolute error and structural similarity (SSIM) is used, which is: , in, The input is a haze-free, original, clear image. The image is a clear image reconstructed after fogging and defogging. To balance the weights, The value range is 0.1 to 1.0, preferably 0.5. It is used to balance pixel-level error and structural similarity constraints. Its specific value can be adjusted and determined according to the verification results during the model training process, so as to ensure that the reconstructed image has both small pixel error and good edge and texture structure.

[0055] (3) Geometric supervision loss : It not only focuses on brightness adjustment but also requires maintaining the scene's geometric topology and edge sharpness. The images before and after dehazing are then re-entered into the binocular stereo matching module to calculate the restored disparity and depth maps, using the following formula:

[0056] in, To estimate the depth map for a synthetic foggy image, For the depth map estimated from the reconstructed image after dehazing, This is a reference depth map of the original, clear scene. These are the weighting coefficients. The value ranges from 0.3 to 0.7, and it is used to balance the constraint strength between the depth estimation results of the foggy image and the depth estimation results of the defogging image relative to the reference depth map. Its specific value can be determined by adjusting the parameters during the model training process. In this embodiment... We set it to 0.5. This constraint forces the geometric depth information before and after the dehazing process to remain highly consistent with the real scene.

[0057] In this embodiment, preferably, the method for optimizing the transmittance of each pixel location in the hazy image to be processed after calculating the transmittance includes: A lower limit value for transmittance is set. For pixel locations in the foggy image to be processed whose transmittance is lower than the lower limit value, the lower limit value is used as their optimized transmittance. In this embodiment, optionally, the lower limit value for transmittance is 0.10. If the lower limit value is too low, it will reduce numerical stability; if it is too high, it may cause residual fog in the distance.

[0058] To avoid excessive noise amplification during the inversion process due to low transmittance in distant regions, a lower limit for transmittance is set. The optimized transmittance is obtained as follows:

[0059] in, The optimized transmittance at pixel location x for the hazy image to be processed. This is used to ensure the numerical stability of subsequent image inversion.

[0060] In this embodiment, specifically, the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed are substituted into the inverse solution equation of the atmospheric scattering model to perform the final dehazing reconstruction, resulting in a dehazed image. The formula is as follows: , in, This represents the pixel brightness value at pixel location x in the dehazed image. Let x be the transmittance of the hazy image to be processed at pixel location x. Let x be the pixel intensity at pixel location x in the hazy image to be processed. This is the global atmospheric light vector.

[0061] Through the above steps, the present invention will increase scene depth. Global atmospheric light and scattering coefficient Complete decoupling was achieved, avoiding mutual interference of parameters caused by direct mapping, thus realizing high-fidelity image reconstruction.

[0062] Taking autonomous driving or intelligent transportation visual perception scenarios as an example, in specific applications, firstly, binocular cameras deployed on the vehicle or roadside acquire left and right views, and obtain the absolute density depth map of the physical scene through a stereo matching algorithm (or directly reading radar depth); subsequently, the edge computing platform inputs the fog image into a pre-trained scattering coefficient prediction module, which adaptively outputs the global scattering coefficient of the current foggy environment. Simultaneously, features such as gradient and brightness are extracted from the image to generate a sky mask and forcibly remove overexposed pixels such as sun halos. The atmospheric light vector without color cast is then accurately calculated through local variance optimization. Finally, by combining depth map, scattering coefficient and atmospheric light, a clear traffic scene image is reconstructed through atmospheric scattering model and input into the downstream lane detection or target recognition module.

[0063] To verify the effectiveness of the method described in this invention, experiments were conducted on real autonomous driving datasets (KITTI and DrivingStereo datasets). The specific experimental results are as follows: (1) Qualitative visual effect verification like Figure 7 As shown, Figure 7 These are comparison images of dehazing effects on the KITTI binocular synthetic fog dataset. Figure 7 In the image (a), the input hazy image to be processed is shown. Figure 7 (b) in the figure is a schematic diagram of the dehazing effect of DCP on the KITTI stereo synthetic fog dataset. Figure 7 (c) in the figure is a schematic diagram of the dehazing effect of D4 on the KITTI stereo synthetic fog dataset. Figure 7 (d) in the diagram is a schematic of the dehazing effect of Cycle-SNSPGAN on the KITTI stereo synthetic fog dataset. Figure 7 (e) in the figure is a schematic diagram of the dehazing effect of UCL-Dehaze on the KITTI stereo synthetic fog dataset. Figure 7 (f) in the figure is a schematic diagram of the dehazing effect of combining the transmittance acquisition method of the present invention with the DCP atmospheric light method on the KITTI binocular synthetic fog dataset. Figure 7 (g) in the figure is a schematic diagram of the defogging effect of the method of the present invention on the KITTI binocular synthetic fog dataset.

[0064] like Figure 8 As shown, Figure 8 These are comparison images of defogging effects on the DrivingStereo real traffic dataset. Figure 8 In the image (a), the input hazy image to be processed is shown. Figure 8 (b) in the diagram illustrates the defogging effect of DCP on the DrivingStereo real traffic dataset. Figure 8 (c) in the figure is a schematic diagram of the defogging effect of D4 on the DrivingStereo real traffic dataset. Figure 8 (d) in the diagram illustrates the dehazing effect of Cycle-SNSPGAN on the DrivingStereo real traffic dataset. Figure 8 (e) in the diagram is a schematic representation of the dehazing effect of UCL-Dehaze on the DrivingStereo real traffic dataset. Figure 8 (f) in the figure is a schematic diagram of the defogging effect of combining the transmittance acquisition method of the present invention with the DCP atmospheric light method on the DrivingStereo real traffic dataset. Figure 8 (g) in the figure is a schematic diagram of the defogging effect of the method of the present invention on the DrivingStereo real traffic dataset.

[0065] The method of this invention is visually compared with traditional Dark Channel Prior (DCP) and existing contrastive learning unsupervised dehazing networks. Experimental results show that in traffic scenes containing large areas of sky, white vehicles, and other high-brightness interference, traditional methods exhibit severe overall underexposure and color distortion, while existing unsupervised networks produce structural artifacts in the distant scene. In contrast, this invention, by employing a strategy of "multi-dimensional features + variance optimization" to lock in real atmospheric light and supplemented by absolute depth constraints, demonstrates its absolute advantage in image quality restoration.

[0066] (2) Verification of quantitative indicator improvement: As shown in Table 1, Table 1 presents the quantitative comparison results of different methods on the KITTI binocular synthetic fog dataset.

[0067] Table 1

[0068] In quantitative evaluation on the KITTI binocular synthetic fog dataset, the method of this invention achieved peak signal-to-noise ratios (PSNR) of 25.2689 dB and 26.1811 dB in the left and right views, respectively, and structural similarity scores (SSIM) of 0.8889 and 0.8947, respectively. Compared with existing advanced contrastive learning networks (such as UCL-Dehaze), the PSNR is improved by nearly 3–4 dB, and the mean squared error shows an order-of-magnitude reduction, demonstrating the absolute advantage of this invention in image quality restoration.

[0069] This second embodiment provides a defogging system based on atmospheric scattering model parameter estimation, including: The Sky Feature Confidence Function Calculation Module is used to calculate the sky feature confidence function value for each pixel based on the gradient, brightness, and color saturation of each pixel in the foggy image to be processed. The filtering module is used to filter initial sky candidate pixels in the hazy image to be processed based on the sky feature confidence function value of each pixel. The local spatial variance calculation module is used to remove pixels whose brightness exceeds the brightness overexposure threshold from the initial sky candidate pixels, obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask. The global atmospheric light vector acquisition module is used to sort each pixel in the diffuse sky region mask in ascending order according to the size of the local spatial variance, and take the average value of the first set number of pixels in the three color channels of RGB as the global atmospheric light vector. The dehazing module is used to obtain a dehazed image based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed.

[0070] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A dehazing method based on atmospheric scattering model parameter estimation, characterized in that, include: Based on the gradient, brightness, and color saturation of each pixel in the foggy image to be processed, calculate the sky feature confidence function value for each pixel; Based on the confidence function value of the sky features of each pixel, the initial sky candidate pixels in the foggy image to be processed are selected. Remove pixels from the initial sky candidate pixels whose brightness exceeds the brightness overexposure threshold to obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask; Based on the magnitude of the local spatial variance, each pixel in the diffuse sky region mask is sorted in ascending order, and the average pixel value of the first set number of pixels in the three RGB color channels is taken as the global atmospheric light vector. Based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed, a dehazed image is obtained.

2. The dehazing method based on atmospheric scattering model parameter estimation according to claim 1, characterized in that, Calculate the sky feature confidence function value for each pixel. The formula is: ， in, Let q be the confidence function value of the sky features. It is an exponential function with the natural constant e as its base. This indicates taking the maximum value. and The threshold constant is Let be the gradient of pixel q. The brightness of pixel q The average brightness of all pixels in the hazy image to be processed. Let q represent the color saturation of pixel q.

3. The dehazing method based on atmospheric scattering model parameter estimation according to claim 1, characterized in that, Methods for selecting initial sky candidate pixels in a hazy image to be processed based on the sky feature confidence function values ​​of each pixel include: The target selection threshold is obtained by multiplying the maximum sky feature confidence function value in the foggy image to be processed by a preset percentage; Pixels whose sky feature confidence function value is greater than the target screening threshold are used as initial sky candidate pixels.

4. The dehazing method based on atmospheric scattering model parameter estimation according to claim 1, characterized in that, Calculate the local spatial variance of each pixel in the diffuse sky region mask. The formula is: ， in, Let be the local spatial variance of pixel q. A local window centered at pixel q. For pixel p in the color channel The strength value on, for The total number of pixels in the array, p is Any pixel inside, These are the image color channels, where R, G, and B represent the red, green, and blue channels, respectively.

5. The dehazing method based on atmospheric scattering model parameter estimation according to claim 1, characterized in that, Methods for obtaining the transmittance at each pixel location in the hazy image to be processed include: Obtain the foggy image to be processed and its depth map; Extract the scattering coefficient response map corresponding to the foggy image to be processed, and perform a global average to obtain the global scattering coefficient; Based on the depth map and global scattering coefficient of the foggy image to be processed, the transmittance at each pixel location in the foggy image to be processed is calculated.

6. The dehazing method based on atmospheric scattering model parameter estimation according to claim 5, characterized in that, Methods for obtaining the foggy image to be processed and its depth map include: The foggy images in the left and right eyes were acquired using a binocular camera. The foggy image in the left eye or the foggy image in the right eye is used as the foggy image to be processed, and the other image is used as the matching reference image. The disparity map between the foggy image to be processed and the matching reference image is calculated by using a binocular stereo matching algorithm. The disparity map is calculated based on the focal length and baseline distance of the binocular camera to obtain the depth map corresponding to the foggy image to be processed.

7. The dehazing method based on atmospheric scattering model parameter estimation according to claim 5, characterized in that, Methods for extracting the scattering coefficient response map corresponding to the foggy image to be processed include: The foggy image to be processed is passed through multiple convolutional layers connected in sequence to obtain fog features at different scales; Fog features at different scales are fused using the RefineNet module to obtain fused features; The fused features are passed through a projection head to obtain the scattering coefficient response map corresponding to the foggy image to be processed.

8. The dehazing method based on atmospheric scattering model parameter estimation according to claim 5, characterized in that, After calculating the transmittance at each pixel location in the hazy image to be processed, the method also includes optimizing the transmittance at each pixel location in the hazy image to be processed, specifically including: Set a lower limit for transmittance. For pixel locations in the foggy image to be processed where the transmittance is lower than the lower limit, use the lower limit as their optimized transmittance.

9. The dehazing method based on atmospheric scattering model parameter estimation according to claim 1, characterized in that, Based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed, the formula for obtaining the dehazed image is as follows: ， in, This represents the pixel brightness value at pixel location x in the dehazed image. Let x be the transmittance of the hazy image to be processed at pixel location x. Let x be the pixel intensity at pixel location x in the hazy image to be processed. This is the global atmospheric light vector.

10. A defogging system based on atmospheric scattering model parameter estimation, characterized in that, include: The Sky Feature Confidence Function Calculation Module is used to calculate the sky feature confidence function value for each pixel based on the gradient, brightness, and color saturation of each pixel in the foggy image to be processed. The filtering module is used to filter initial sky candidate pixels in the hazy image to be processed based on the sky feature confidence function value of each pixel. The local spatial variance calculation module is used to remove pixels whose brightness exceeds the brightness overexposure threshold from the initial sky candidate pixels, obtain the diffuse sky region mask, and calculate the local spatial variance of each pixel in the diffuse sky region mask. The global atmospheric light vector acquisition module is used to sort each pixel in the diffuse sky region mask in ascending order according to the size of the local spatial variance, and take the average value of the first set number of pixels in the three color channels of RGB as the global atmospheric light vector. The dehazing module is used to obtain a dehazed image based on the global atmospheric light vector and the transmittance of each pixel in the hazy image to be processed.

Images