A haze image synthesis method based on deep guidance and domain alignment

Abstract

The application discloses a haze image synthesis method based on depth guidance and domain alignment, which comprises the following steps: 1, depth information acquisition and processing, obtaining depth normalization value; 2, through domain alignment, the real haze image and the clear image to be synthesized are obtained to update the transmittance; 3, based on the real haze image, the atmospheric light value A' is obtained; 4, based on the depth normalization value and the atmospheric scattering coefficient multilevel strategy, the transmittance under different atmospheric scattering coefficients is obtained, and the optimized transmittance under different atmospheric scattering coefficients is obtained by combining the updated transmittance weighting; 5, based on the atmospheric light value A' and the optimized transmittance under different atmospheric scattering coefficients, the haze synthesis image under different atmospheric scattering coefficients is obtained. The haze synthesis image with real depth effect and natural haze concentration distribution is generated through the depth information guidance and the domain alignment physical model constraint, and the diversity and authenticity of the synthesized image are improved.

Description

Technical Field

[0001] This invention belongs to the field of haze image synthesis technology, specifically relating to a haze image synthesis method based on depth guidance and domain alignment. Background Technology

[0002] With the widespread application of computer vision technology in fields such as autonomous driving, video surveillance, and remote sensing mapping, image quality has become a key factor affecting the performance of vision systems. However, under adverse weather conditions such as fog and haze, suspended particles in the atmosphere absorb and scatter light, leading to problems such as decreased contrast, color distortion, and loss of detail in the acquired images, severely restricting the reliability and robustness of vision systems. Currently, deep learning-based image dehazing methods have demonstrated superior performance on several benchmark datasets, but the effectiveness of these methods largely depends on large-scale, high-quality training data. Obtaining pairs of clear images and hazy images in real-world scenarios is extremely challenging, and hazy conditions vary significantly across different regions and seasons, making it exceptionally difficult to construct sufficiently diverse real-world hazy datasets. This data bottleneck severely restricts the further development and practical application of dehazing networks.

[0003] Existing methods for synthesizing hazy images mainly fall into two categories: physics-based methods and data-driven methods. Physics-based methods primarily rely on atmospheric scattering models to synthesize hazy images by estimating transmittance maps and atmospheric light values. However, transmittance estimation often employs simplified assumptions, such as uniform fog concentration distribution, ignoring the geometric structure of the scene, resulting in synthesized hazy images lacking realism. Data-driven methods, such as generative adversarial networks (GANs), can learn the distribution characteristics of hazy images from data, generating visually more realistic hazy images. However, these methods lack explicit physical constraints, and the generated images may be insufficient in terms of physical plausibility. Furthermore, existing datasets of synthesized hazy images still lag behind real hazy images in terms of visual realism and physical plausibility. This leads to poor performance and limited generalization ability of dehazing networks trained on these datasets in real-world scenarios.

[0004] Therefore, there is currently a lack of a simple and well-designed haze image synthesis method based on depth guidance and domain alignment. By using depth information guidance and domain alignment physical model constraints, a haze composite image with realistic depth effect and natural fog concentration distribution can be generated, which improves the diversity and realism of the composite image, thereby enhancing the training effect and generalization ability of the subsequent dehazing network. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a haze image synthesis method based on depth guidance and domain alignment. The method has simple steps and reasonable design. By guiding with depth information and constraining with a domain alignment physical model, it generates haze synthesized images with realistic depth effect and natural fog concentration distribution, thereby improving the diversity and realism of the synthesized images and thus enhancing the training effect and generalization ability of the subsequent dehazing network.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a haze image synthesis method based on depth guidance and domain alignment, the method comprising the following steps:

[0007] Step 1: Depth information acquisition and processing to obtain the normalized depth value:

[0008] The DPT-Hybrid model is used to perform depth processing on the sharp image to be synthesized to obtain a depth map; and the depth map is then normalized to obtain a depth normalization value.

[0009] Step 2: Align the real haze image and the clear image to be synthesized through domain alignment to obtain the updated transmittance; where the updated transmittance of the j-th pixel is... j is a positive integer, and the value of j ranges from 1 to J, where J is the total number of pixels;

[0010] Step 3: Obtain the atmospheric light value A′ based on real haze images;

[0011] Step 4: Based on the depth normalization value and the multi-level strategy of atmospheric scattering coefficient, the transmittance under different atmospheric scattering coefficients is obtained, and combined with the updated transmittance weighting, the optimized transmittance under different atmospheric scattering coefficients is obtained.

[0012] Step 5: Based on the atmospheric light value A′ and the optimized transmittance under different atmospheric scattering coefficients, obtain the haze composite image under different atmospheric scattering coefficients.

[0013] The aforementioned method for synthesizing haze images based on depth guidance and domain alignment further includes the following step one:

[0014] Step 101: Extract the R component, B component and G component from the clear image to be synthesized to obtain the R component image, B component image and G component image respectively.

[0015] Step 102: Perform standard normalization on each pixel value in the R component image, B component image and G component image respectively to obtain the normalized R component image, normalized B component image and normalized G component image; and merge the normalized R component image, normalized B component image and normalized G component image to obtain the normalized image.

[0016] Step 103: Process the normalized image using the DPT-Hybrid model to obtain a depth map;

[0017] Step 104: Record the pixel value of the j-th pixel in the depth map as the j-th depth value dj;

[0018] Step 105: Obtain the j-th depth normalization value dj′ according to dj′=(dj-dmin) / (dmax-dmin); where dmax represents the maximum depth value on the depth map and dmin represents the minimum depth value on the depth map.

[0019] The above-mentioned method for synthesizing haze images based on depth guidance and domain alignment further includes the following step two:

[0020] Step 201: Process the real haze image into grayscale to obtain a real haze grayscale image;

[0021] Step 202: For the real haze grayscale image, construct the first Gaussian mixture model using the expectation-maximization algorithm. ;in, This represents data points composed of pixel values ​​in a real grayscale image of haze. This represents the first Gaussian distribution. This represents the mixture weights of the first Gaussian distribution. This represents the mean of the first Gaussian distribution. This represents the variance of the first Gaussian distribution; This represents the second Gaussian distribution. This represents the mixture weights of the second Gaussian distribution. This represents the mean of the second Gaussian distribution. This represents the variance of the second Gaussian distribution; and The value of is between 0 and 1, and ;

[0022] Step 203: Apply the first Gaussian mixture model As a probability density function, The cumulative distribution probability P1 of the real haze grayscale image is obtained within the range of 0 to 255.

[0023] Step 204: Process the clear image to be synthesized according to the methods in steps 201 to 203 to obtain the second Gaussian mixture model. The cumulative probability distribution P2 of the clear image to be synthesized;

[0024] Step 205: Process the real haze image using the dark channel prior method to obtain the true transmittance map; where the initial transmittance of the j-th pixel on the true transmittance map is denoted as... ;

[0025] Step 206, according to The updated transmittance of the j-th pixel is obtained. .

[0026] The above-mentioned method for synthesizing haze images based on depth guidance and domain alignment further includes step three, which is as follows:

[0027] Step 301: Use the quadtree segmentation method to recursively segment the grayscale image corresponding to the real haze image to obtain each sub-region;

[0028] Step 302: Record the average value of all pixel values ​​in each sub-region as the average brightness, and record the variance of all pixel values ​​in each sub-region as the brightness variance;

[0029] Step 303: During the recursive segmentation of the grayscale image corresponding to the real haze image, if there is a sub-region with the smallest brightness variance and the highest average brightness, then stop the segmentation.

[0030] Step 304: Select the sub-region with the smallest brightness variance and the highest average brightness as candidate atmospheric light regions.

[0031] Step 305: Sort the pixel values ​​of the candidate atmospheric light regions from largest to smallest. The average value of the R, G, and B components of the first [0.1% × J] pixels in the real haze image is denoted as the atmospheric light value A′; where, [ [] indicates rounding; J is the total number of pixels.

[0032] The above-mentioned method for synthesizing haze images based on depth guidance and domain alignment further includes step four, which is as follows:

[0033] Step 401: Calculate the atmospheric scattering coefficient. Set within the range of 0.8 to 1.5, and increasing in steps of 0.1, the nth atmospheric scattering coefficient is... n is a positive integer, ranging from 1 to 8, representing the first atmospheric scattering coefficient. The value is 0.8; where n ranges from 1 to 7. , This is the (n+1)th atmospheric scattering coefficient;

[0034] Step 402, according to The transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. Where e is the natural constant;

[0035] Step 403, according to The optimized transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. ;in, The random weighting coefficient ranges from 0.2 to 0.8.

[0036] The above-mentioned method for synthesizing haze images based on depth guidance and domain alignment further includes step five, which is as follows:

[0037] Step 501: Based on the atmospheric scattering physics model Perform haze synthesis to obtain the updated R component of the j-th pixel under the n-th atmospheric scattering coefficient. ;in, Represents the R component of the j-th pixel in the sharp image to be synthesized;

[0038] Step 502: Following the method in step 501, process the G component and B component of the j-th pixel in the clear image to be synthesized, and merge the updated R component, G component and B component of the j-th pixel under the n-th atmospheric scattering coefficient to obtain the haze composite image under the n-th atmospheric scattering coefficient.

[0039] Step 503: Repeat steps 501 to 502 multiple times to adjust the atmospheric scattering coefficient and obtain haze composite images under different atmospheric scattering coefficients.

[0040] The aforementioned method for synthesizing haze images based on depth guidance and domain alignment further includes, after step five, color correction and detail enhancement processing on the synthesized haze image, as detailed below:

[0041] Step A1: Convert the haze composite image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the first lightness component, the first A component and the first B component.

[0042] Step A2: Convert the real haze image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the second A component and the second B component.

[0043] Step A3: Based on the histogram of the second A component, perform histogram matching on the histogram of the first A component to obtain the remapped first A component;

[0044] Based on the histogram of the second B component, histogram matching is performed on the histogram of the first B component to obtain the remapped first B component.

[0045] Step A4: Convert the first brightness component, the remapped first A component, and the remapped first B component back to the RGB color space to obtain the remapped haze composite image;

[0046] Step A5: The edge perception enhancement algorithm based on guided filtering is used to enhance the remapped haze composite image to obtain the enhanced haze composite image.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] 1. The method of this invention has simple steps and a reasonable design. First, depth information is acquired and processed to obtain a depth normalization value. Second, the real haze image and the clear image to be synthesized are aligned by domain to obtain the updated transmittance. Next, the atmospheric light value A′ is obtained based on the real haze image. Then, based on the depth normalization value and a multi-level strategy of atmospheric scattering coefficient, the transmittance under different atmospheric scattering coefficients is obtained. Combined with the updated transmittance weighting, the optimized transmittance under different atmospheric scattering coefficients is obtained. Finally, based on the atmospheric light value A′ and the optimized transmittance under different atmospheric scattering coefficients, haze composite images under different atmospheric scattering coefficients are obtained, synthesizing a more realistic and higher quality haze image. This provides a higher quality dataset for subsequent image dehazing networks and improves the network's dehazing effect.

[0049] 2. This invention obtains the depth information of the clear image to be synthesized through the DPT-Hybrid model, generates an accurate depth map, and performs normalization processing on the depth map so that when the transmittance is obtained based on the normalized depth value and the atmospheric scattering coefficient, the correlation between the geometric structure information of the scene and the fog concentration distribution is considered, thereby improving the realism of the subsequent fog and haze synthesis.

[0050] 3. When generating depth-guided transmittance, this invention calculates transmittance based on normalized depth values, extracts initial transmittance from real haze images, and obtains updated transmittance based on domain alignment. Then, the calculated transmittance and the updated transmittance are weighted to obtain optimized transmittance under different atmospheric scattering coefficients, ensuring that the transmittance distribution meets physical constraints and statistical characteristics, thus improving the realism of subsequent haze synthesis.

[0051] 4. This invention introduces domain alignment technology. Based on the cumulative distribution probability P1 of the real haze grayscale image and the cumulative distribution probability P2 of the clear image to be synthesized, the transmittance distribution of the clear image to be synthesized and the real haze image are aligned to obtain the updated transmittance, which effectively reduces the domain difference between the synthesized image and the real image.

[0052] 5. This invention employs a multi-level atmospheric scattering coefficient strategy to generate haze composite images with different transmittances (i.e., different fog concentrations) by adjusting the atmospheric scattering coefficient, significantly enhancing the diversity of the dataset.

[0053] 6. After obtaining the haze composite image, the present invention performs color correction and detail enhancement to further ensure that the haze composite image has high visual quality and realism.

[0054] In summary, the method of this invention is simple in steps and reasonable in design. By guiding with depth information and constraining with a domain-aligned physical model, it generates a haze composite image with realistic depth effect and natural fog concentration distribution, which improves the diversity and realism of the composite image, thereby enhancing the training effect and generalization ability of the subsequent dehazing network.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention.

[0057] Figure 2 This is a comparison diagram showing the effects of the present invention. Detailed Implementation

[0058] like Figure 1 As shown, the present invention provides a method for synthesizing haze images based on depth guidance and domain alignment, comprising the following steps:

[0059] Step 1: Depth information acquisition and processing to obtain the normalized depth value:

[0060] The DPT-Hybrid model is used to perform depth processing on the sharp image to be synthesized to obtain a depth map; and the depth map is then normalized to obtain a depth normalization value.

[0061] Step 2: Align the real haze image and the clear image to be synthesized through domain alignment to obtain the updated transmittance; where the updated transmittance of the j-th pixel is... j is a positive integer, and the value of j ranges from 1 to J, where J is the total number of pixels;

[0062] Step 3: Obtain the atmospheric light value A′ based on real haze images;

[0063] Step 4: Based on the depth normalization value and the multi-level strategy of atmospheric scattering coefficient, the transmittance under different atmospheric scattering coefficients is obtained, and combined with the updated transmittance weighting, the optimized transmittance under different atmospheric scattering coefficients is obtained.

[0064] Step 5: Based on the atmospheric light value A′ and the optimized transmittance under different atmospheric scattering coefficients, obtain the haze composite image under different atmospheric scattering coefficients.

[0065] In this embodiment, step one, the specific process is as follows:

[0066] Step 101: Extract the R component, B component and G component from the clear image to be synthesized to obtain the R component image, B component image and G component image respectively.

[0067] Step 102: Perform standard normalization on each pixel value in the R component image, B component image and G component image respectively to obtain the normalized R component image, normalized B component image and normalized G component image; and merge the normalized R component image, normalized B component image and normalized G component image to obtain the normalized image.

[0068] Step 103: Process the normalized image using the DPT-Hybrid model to obtain a depth map;

[0069] Step 104: Record the pixel value of the j-th pixel in the depth map as the j-th depth value dj;

[0070] Step 105: Obtain the j-th depth normalization value dj′ according to dj′=(dj-dmin) / (dmax-dmin); where dmax represents the maximum depth value on the depth map and dmin represents the minimum depth value on the depth map.

[0071] In this embodiment, step two is as follows:

[0072] Step 201: Process the real haze image into grayscale to obtain a real haze grayscale image;

[0073] Step 202: For the real haze grayscale image, construct the first Gaussian mixture model using the expectation-maximization algorithm. ;in, This represents data points composed of pixel values ​​in a real grayscale image of haze. This represents the first Gaussian distribution. This represents the mixture weights of the first Gaussian distribution. This represents the mean of the first Gaussian distribution. This represents the variance of the first Gaussian distribution; This represents the second Gaussian distribution. This represents the mixture weights of the second Gaussian distribution. This represents the mean of the second Gaussian distribution. This represents the variance of the second Gaussian distribution; and The value of is between 0 and 1, and ;

[0074] Step 203: Apply the first Gaussian mixture model As a probability density function, The cumulative distribution probability P1 of the real haze grayscale image is obtained within the range of 0 to 255.

[0075] Step 204: Process the clear image to be synthesized according to the methods in steps 201 to 203 to obtain the second Gaussian mixture model. The cumulative probability distribution P2 of the clear image to be synthesized;

[0076] Step 205: Process the real haze image using the dark channel prior method to obtain the true transmittance map; where the initial transmittance of the j-th pixel on the true transmittance map is denoted as... ;

[0077] Step 206, according to The updated transmittance of the j-th pixel is obtained. .

[0078] In this embodiment, step three is as follows:

[0079] Step 301: Use the quadtree segmentation method to recursively segment the grayscale image corresponding to the real haze image to obtain each sub-region;

[0080] Step 302: Record the average value of all pixel values ​​in each sub-region as the average brightness, and record the variance of all pixel values ​​in each sub-region as the brightness variance;

[0081] Step 303: During the recursive segmentation of the grayscale image corresponding to the real haze image, if there is a sub-region with the smallest brightness variance and the highest average brightness, then stop the segmentation.

[0082] Step 304: Select the sub-region with the smallest brightness variance and the highest average brightness as candidate atmospheric light regions.

[0083] Step 305: Sort the pixel values ​​of the candidate atmospheric light regions from largest to smallest. The average value of the R, G, and B components of the first [0.1% × J] pixels in the real haze image is denoted as the atmospheric light value A′; where, [ [] indicates rounding; J is the total number of pixels.

[0084] In this embodiment, step four is as follows:

[0085] Step 401: Calculate the atmospheric scattering coefficient. Set within the range of 0.8 to 1.5, and increasing in increments of 0.1, the nth atmospheric scattering coefficient is... n is a positive integer, ranging from 1 to 8, representing the first atmospheric scattering coefficient. The value is 0.8; where n ranges from 1 to 7. , This is the (n+1)th atmospheric scattering coefficient;

[0086] Step 402, according to The transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. Where e is the natural constant;

[0087] Step 403, according to The optimized transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. ;in, The random weighting coefficient ranges from 0.2 to 0.8.

[0088] In this embodiment, step five is as follows:

[0089] Step 501: Based on the atmospheric scattering physics model Perform haze synthesis to obtain the updated R component of the j-th pixel under the n-th atmospheric scattering coefficient. ;in, Represents the R component of the j-th pixel in the sharp image to be synthesized;

[0090] Step 502: Following the method in step 501, process the G component and B component of the j-th pixel in the clear image to be synthesized, and merge the updated R component, G component and B component of the j-th pixel under the n-th atmospheric scattering coefficient to obtain the haze composite image under the n-th atmospheric scattering coefficient.

[0091] Step 503: Repeat steps 501 to 502 multiple times to adjust the atmospheric scattering coefficient and obtain haze composite images under different atmospheric scattering coefficients.

[0092] In this embodiment, after step five, the composite haze image undergoes color correction and detail enhancement processing, as detailed below:

[0093] Step A1: Convert the haze composite image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the first lightness component, the first A component and the first B component.

[0094] Step A2: Convert the real haze image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the second A component and the second B component.

[0095] Step A3: Based on the histogram of the second A component, perform histogram matching on the histogram of the first A component to obtain the remapped first A component;

[0096] Based on the histogram of the second B component, histogram matching is performed on the histogram of the first B component to obtain the remapped first B component.

[0097] Step A4: Convert the first brightness component, the remapped first A component, and the remapped first B component back to the RGB color space to obtain the remapped haze composite image;

[0098] Step A5: The edge perception enhancement algorithm based on guided filtering is used to enhance the remapped haze composite image to obtain the enhanced haze composite image.

[0099] In this embodiment, DPT-Hybrid (Dense Prediction Transformer-Hybrid) is a model used for depth estimation, which combines the features of Vision Transformer (ViT) and Convolutional Neural Network (CNN).

[0100] In this embodiment, the size of the clear image to be synthesized is 512×512.

[0101] In this embodiment, it should be noted that the total number of pixels in the clear image to be synthesized, the real haze image, the real transmittance map, and the depth map is the same.

[0102] In this embodiment, it should be noted that the average value of the R, G, and B components corresponding to the first [0.1%×J] pixels in the real haze image is the sum of the R, G, and B components corresponding to the first [0.1%×J] pixels divided by 3×[0.1%×J].

[0103] In this embodiment, during the standard normalization step 102, the mean and standard deviation of the R component map, B component map and G component map of the dataset where the clear image to be synthesized is located are obtained in advance.

[0104] In this embodiment, eight different concentration levels are set at intervals of 0.1, namely: 1 = 0.8 2 = 0.9 3 = 1.0 4 = 1.1 5 = 1.2 6 = 1.3 7 = 1.4 8 = 1.5.

[0105] In this embodiment, during the synthesis process in step 501, parallel computing technology is employed to improve synthesis efficiency. The image is divided into multiple blocks, and these blocks are processed simultaneously on the GPU. For large images, a block-based processing strategy is used to avoid memory overflow issues.

[0106] In this embodiment, the haze composite image is converted from RGB color space to LAB color space while keeping the luminance channel unchanged. Histogram matching is performed on the color channels to make the color distribution of the composite image closer to that of a real haze image. An edge perception enhancement algorithm based on guided filtering is used to enhance the remapped haze composite image, enhancing important edges and texture details while maintaining the overall consistency of the haze effect.

[0107] like Figure 2 As shown, (a) is the clear image to be synthesized, (b) is the depth map, and (c) is the enhanced haze composite image, which can intuitively display the synthesis effect.

[0108] In this embodiment, the OTS dataset, HazeGEN dataset, and the dataset obtained by the method of this invention (i.e., the dataset of this invention) are used as the training set, and clear images and haze images provided in the BeDDE, MRFID, O-Haze, and NH-Haze datasets are used as the validation set. The AOD-Net network (integrated dehazing network) and the DEA-Net network (a single-image dehazing network based on detail enhancement convolution and content-guided attention) are trained on the training set and then quantitatively evaluated using the validation set. The quantitative evaluation results are shown in Table 1.

[0109] Table 1 Quantitative Assessment Results

[0110]

[0111] As indicated by 1, under the training of the dataset of this invention, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) obtained subsequently through the validation set are relatively high. Therefore, the dataset generated by this method is used for training and shows excellent performance compared with other datasets, which fully demonstrates the effectiveness and practicality of this method and meets the actual requirements.

[0112] In summary, the method of this invention is simple in steps and reasonable in design. By guiding with depth information and constraining with a domain-aligned physical model, it generates a haze composite image with realistic depth effect and natural fog concentration distribution, which improves the diversity and realism of the composite image, thereby enhancing the training effect and generalization ability of the subsequent dehazing network.

[0113] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for synthesizing haze images based on depth guidance and domain alignment, characterized in that, The method includes the following steps: Step 1: Depth information acquisition and processing to obtain the normalized depth value: The DPT-Hybrid model is used to perform depth processing on the sharp image to be synthesized to obtain a depth map; and the depth map is then normalized to obtain a depth normalization value. Step 2: Align the real haze image and the clear image to be synthesized through domain alignment to obtain the updated transmittance; where the updated transmittance of the j-th pixel is... j is a positive integer, and the value of j ranges from 1 to J, where J is the total number of pixels; Step 3: Obtain the atmospheric light value A′ based on real haze images; Step 4: Based on the depth normalization value and the multi-level strategy of atmospheric scattering coefficient, the transmittance under different atmospheric scattering coefficients is obtained, and combined with the updated transmittance weighting, the optimized transmittance under different atmospheric scattering coefficients is obtained. Step 5: Based on the atmospheric light value A′ and the optimized transmittance under different atmospheric scattering coefficients, obtain the haze composite image under different atmospheric scattering coefficients; Step two, the specific process is as follows: Step 201: Process the real haze image into grayscale to obtain a real haze grayscale image; Step 202: For the real haze grayscale image, construct the first Gaussian mixture model using the expectation-maximization algorithm. ;in, This represents data points composed of pixel values ​​in a real grayscale image of haze. This represents the first Gaussian distribution. This represents the mixture weights of the first Gaussian distribution. This represents the mean of the first Gaussian distribution. This represents the variance of the first Gaussian distribution; This represents the second Gaussian distribution. This represents the mixture weights of the second Gaussian distribution. This represents the mean of the second Gaussian distribution. This represents the variance of the second Gaussian distribution; and The value of is between 0 and 1, and ; Step 203: Apply the first Gaussian mixture model As a probability density function, The cumulative distribution probability P1 of the real haze grayscale image is obtained within the range of 0 to 255. Step 204: Process the clear image to be synthesized according to the methods in steps 201 to 203 to obtain the second Gaussian mixture model. The cumulative probability distribution P2 of the clear image to be synthesized; Step 205: Process the real haze image using the dark channel prior method to obtain the true transmittance map; where the initial transmittance of the j-th pixel on the true transmittance map is denoted as... ; Step 206, according to The updated transmittance of the j-th pixel is obtained. ; Step four, the specific process is as follows: Step 401: Calculate the atmospheric scattering coefficient. Set within the range of 0.8 to 1.5, and increasing in increments of 0.1, the nth atmospheric scattering coefficient is... n is a positive integer, ranging from 1 to 8, representing the first atmospheric scattering coefficient. The value is 0.8; where n ranges from 1 to 7. , This is the (n+1)th atmospheric scattering coefficient; Step 402, according to The transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. Where e is the natural constant; Step 403, according to The optimized transmittance of the j-th pixel under the n-th atmospheric scattering coefficient is obtained. ;in, The random weighting coefficient ranges from 0.2 to 0.

8.

2. The haze image synthesis method based on depth guidance and domain alignment according to claim 1, characterized in that: Step one, the specific process is as follows: Step 101: Extract the R component, B component and G component from the clear image to be synthesized to obtain the R component image, B component image and G component image respectively. Step 102: Perform standard normalization on each pixel value in the R component map, B component map and G component map respectively to obtain the normalized R component map, normalized B component map and normalized G component map. Then, the normalized R component image, the normalized B component image, and the normalized G component image are merged to obtain the normalized image; Step 103: Process the normalized image using the DPT-Hybrid model to obtain a depth map; Step 104: Record the pixel value of the j-th pixel in the depth map as the j-th depth value dj; Step 105: Obtain the j-th depth normalization value dj′ according to dj′=(dj-dmin) / (dmax-dmin); where dmax represents the maximum depth value on the depth map and dmin represents the minimum depth value on the depth map.

3. A haze image synthesis method based on depth guidance and domain alignment according to claim 1, characterized in that: Step three, the specific process is as follows: Step 301: Use the quadtree segmentation method to recursively segment the grayscale image corresponding to the real haze image to obtain each sub-region; Step 302: Record the average value of all pixel values ​​in each sub-region as the average brightness, and record the variance of all pixel values ​​in each sub-region as the brightness variance; Step 303: During the recursive segmentation of the grayscale image corresponding to the real haze image, if there is a sub-region with the smallest brightness variance and the highest average brightness, then stop the segmentation. Step 304: Select the sub-region with the smallest brightness variance and the highest average brightness as candidate atmospheric light regions. Step 305: Sort the pixel values ​​of the candidate atmospheric light regions from largest to smallest. The average value of the R, G, and B components of the first [0.1% × J] pixels in the real haze image is denoted as the atmospheric light value A′; where, [ [] indicates rounding; J is the total number of pixels.

4. A haze image synthesis method based on depth guidance and domain alignment according to claim 1, characterized in that: Step five, the specific process is as follows: Step 501: Based on the atmospheric scattering physics model Perform haze synthesis to obtain the updated R component of the j-th pixel under the n-th atmospheric scattering coefficient. ;in, Represents the R component of the j-th pixel in the sharp image to be synthesized; Step 502: Following the method in step 501, process the G component and B component of the j-th pixel in the clear image to be synthesized, and merge the updated R component, G component and B component of the j-th pixel under the n-th atmospheric scattering coefficient to obtain the haze composite image under the n-th atmospheric scattering coefficient. Step 503: Repeat steps 501 to 502 multiple times to adjust the atmospheric scattering coefficient and obtain haze composite images under different atmospheric scattering coefficients.

5. A haze image synthesis method based on depth guidance and domain alignment according to claim 1, characterized in that: After step five, the composite image of the haze is subjected to color correction and detail enhancement processing, as follows: Step A1: Convert the haze composite image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the first lightness component, the first A component and the first B component. Step A2: Convert the real haze image from RGB space to LAB color space to obtain the lightness component, A component and B component, and denot them as the second A component and the second B component. Step A3: Based on the histogram of the second A component, perform histogram matching on the histogram of the first A component to obtain the remapped first A component; Based on the histogram of the second B component, histogram matching is performed on the histogram of the first B component to obtain the remapped first B component. Step A4: Convert the first brightness component, the remapped first A component, and the remapped first B component back to the RGB color space to obtain the remapped haze composite image; Step A5: The edge perception enhancement algorithm based on guided filtering is used to enhance the remapped haze composite image to obtain the enhanced haze composite image.

Images