Urban foggy day scene visibility estimation method based on physical perception multi-task regression network

Abstract

The invention discloses an urban foggy day scene visibility estimation method based on a physical perception multi-task regression network. The method comprises the following steps: acquiring an urban scene image collected in sunny weather; preprocessing and data enhancement are carried out on an urban scene image collected in sunny weather; constructing a data set based on the preprocessed and data enhanced images; constructing a physical perception multi-task regression network model for estimating visibility parameters of the urban foggy day scene; training the physical perception multi-task regression network based on a training set in the data set to obtain a trained physical perception multi-task regression network model; inputting a test set image into the trained physical perception multi-task regression network model to realize estimation of visibility parameters of the urban foggy day scene; and calculating the visibility of the urban foggy day scene by adopting a visibility estimation algorithm on the basis of estimating the visibility parameters of the urban foggy day scene.

Description

Technical Field

[0001] This invention belongs to the field of urban fog visibility prediction technology, and relates to a method for estimating visibility in urban fog scenes based on a physical perception multi-task regression network. Background Technology

[0002] Scene perception is crucial for driving decisions and traffic safety. Fog, defined as a type of cloud on the ground formed by tiny water droplets suspended in the air, is one of the most common causes of low visibility. Low visibility caused by fog has multiple consequences, such as making it difficult for drivers to perceive their driving environment and make driving decisions, thus threatening traffic safety. Studies have confirmed that traffic accident rates increase significantly when visibility is below 200 meters. In addition to traffic safety, some equipment (such as cameras and lidar) may also be affected by low visibility weather, hindering the development of intelligent transportation systems. However, fog is a common weather phenomenon, especially frequently occurring in urban environments. Besides being prone to systemic, large-scale fog events, it also often presents as localized patches of fog on some main roads, making environmental observation difficult. Therefore, accurate estimation of visibility in foggy weather can significantly improve traffic management and safety levels.

[0003] Accurately estimating visibility in foggy weather is essential but extremely challenging: (1) Visibility is a complex parameter affected by multiple atmospheric conditions such as light scattering, atmospheric light, and light absorption; (2) Unless the fog is evenly distributed throughout the image, it is difficult to characterize visibility with a single numerical value. Currently, there are two main types of commonly used detection methods: one is equipment detection, where the commonly used visibility detection instrument is mainly a laser visibility meter. Although the detection results are accurate, it has the following drawbacks: (1) poor real-time performance; (2) small detection range; and (3) high maintenance costs. The other is image-based visibility detection methods. Although they are low-cost, simple to operate, and have low space requirements, they are difficult to accurately estimate visibility because they are based on an indirect calculation method. Moreover, most algorithms are limited to fixed scenes and have limited applicability.

[0004] Artificial intelligence technology is developing rapidly, especially deep learning, which has made significant progress in image processing and has been widely applied in many fields. Performance improvements in various computer vision tasks largely depend on the integration of physical laws and deep learning methods. A visibility estimation algorithm can be constructed by integrating physical laws with the features of foggy images to build a deep learning network framework that focuses on the characteristics of foggy images. By learning the features of foggy images and integrating physical laws, it can achieve defogging and visibility estimation. Our designed physical perception multi-task regression network algorithm has the advantage of better handling non-uniform fog conditions and outputting a more robust visibility distribution map. Summary of the Invention

[0005] To address the aforementioned problems, the technical solution adopted in this invention is: a visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network, comprising the following steps: Acquire urban scene images under clear weather conditions; Preprocessing and data augmentation of urban scene images acquired under clear weather conditions; A dataset was constructed based on preprocessed and data-enhanced images; A physical perception multi-task regression network model was constructed to estimate visibility parameters in urban foggy scenes; The physical perception multi-task regression network is trained based on the training set in the dataset to obtain a trained physical perception multi-task regression network model. The test set images are input into a trained physical perception multi-task regression network model to estimate the visibility parameters of urban fog scenes. Based on the estimated visibility parameters of urban foggy scenes, a visibility estimation algorithm is used to calculate the visibility of urban foggy scenes.

[0006] Furthermore: the data augmentation employs a data augmentation algorithm, the specific process of which is as follows: Acquire urban scene images and depth maps of urban scene images captured under clear weather conditions; Samples are taken from a uniform distribution U as a unified value for the visibility map V; The visibility map V is calculated using the following formula:

[0007] in: It is the meteorological visibility distance threshold defined by the International Commission on Illumination; A transmission map T is generated using depth D and visibility V; The red (R), green (G), and blue (B) values ​​of generated atmospheric light A; A fogged image is generated using the following formula. The specific formula for generating is as follows: ,

[0008] in: is the atmospheric scattering coefficient.

[0009] 3. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: red (R), green (G), and blue (B) values ​​of atmospheric light A are generated; The value of channel B is sampled from U(180,255). The value of channel G is sampled from min(U(B-5,B+2),255). The value of channel R is min(U( -5, +2),255).

[0010] Furthermore: the physical perception multi-task regression network model includes: Encoder: Used with ResNet 18. Network extraction of multi-scale features F; Decoder A: Used for atmospheric light estimation based on the multi-scale features output by the encoder. Decoder T: Used for transmittance estimation based on the multi-scale features output by the encoder. Decoder D: Used for depth estimation based on the multi-scale features output by the encoder.

[0011] Furthermore: the encoder employs ResNet. The process of extracting multi-scale features F from the network is as follows: For an input image of size H×W×3, the processing order is as follows. A convolutional layer with a kernel size of 7×7, an output dimension of 64, and a stride of 2 is used, along with a 3×3 max pooling layer stacked with a stride of 2, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The first convolutional module uses four 3×3 convolutional layers, stacked with an output dimension of 64 and a stride of 1, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The second set of convolutional modules, based on the feature map transmitted by the first set of convolutional modules, uses four 3×3 convolutional modules stacked together. The output dimension of the convolutional layer is 128 and the stride is 1, which realizes one downsampling and feature extraction. The extracted output feature map is H / 8×W / 8×128. The third convolutional module, based on the feature map transmitted by the second convolutional module, uses four 3×3 convolutional layers stacked together. The output dimension of the convolutional layer is 256 and the stride is 1, which realizes downsampling and feature extraction in one step, and obtains a feature map of H / 16×W / 16×256. The fourth convolutional module, based on the feature map transmitted by the fourth convolutional module, uses four 3×3 convolutional layers stacked together. The output dimension of the convolutional layer is 512 and the stride is 1, to achieve downsampling and feature extraction in one step, and obtain the final multi-scale feature map of H / 32×W / 32×512.

[0012] Furthermore, the process by which decoder A estimates atmospheric light based on the multi-scale features output by the encoder is as follows: First convolutional layer: Based on a convolutional kernel of size 1×1, output dimension 256, and stride 1, a feature map of H / 32×W / 32×512 is obtained; The second convolutional layer: Based on the feature map transmitted from the first convolutional layer, it uses two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, and stacks them with a convolutional kernel with a size of 1×1, an output dimension of 3, and a stride of 1 to obtain an output of H / 32×W / 32×3. Based on the stacking of a Sigmoid activation layer and a Mean pooling layer, the output result H / 288×W / 512×3 is obtained, which is the atmospheric light vector.

[0013] Furthermore: the decoder D and decoder T have the same structure, and the process of decoder D and decoder T processing multi-scale feature maps is as follows; The first set of convolutional layers: Based on the stacking of two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, one upsampling and feature fusion are achieved to obtain a feature map of H / 16×W / 16×256. The second set of convolutional layers: Based on the feature map transmitted by the first set of convolutional layers, two convolutional kernels with a size of 3×3, an output dimension of 128, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 8×W / 8×128. The third convolutional layer: Based on the feature map transmitted by the second convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 64, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 4×W / 4×64. The fourth convolutional layer: Based on the feature map transmitted by the third convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 32, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 2×W / 2×32. The fifth convolutional layer: Based on the feature map transmitted by the third convolutional layer, it uses two convolutional kernels of size 3×3 with an output dimension of 16 and a stride of 1, as well as a convolutional kernel of size 3×3 with an output dimension of 1 and a stride of 1, and a stacked Sigmoid activation layer to achieve one upsampling and obtain an H×W×1 result map.

[0014] Furthermore: the loss function of the physical perception multi-task regression network model for different tasks is based on coupling the outputs of the three decoders with Koschmieder's law, and the expression of the loss function is as follows:

[0015] in: Represents the final loss function. This indicates the estimated atmospheric light loss. This indicates that the loss is estimated using a transmittance map. This indicates that the disparity map is used to estimate the loss. This represents the loss for estimating the dehazed image. This indicates the estimated loss of visibility. , , , and This is a hyperparameter.

[0016] This invention provides a visibility estimation method for urban fog scenes based on a physical perception multi-task regression network. It proposes an end-to-end fog visibility estimation method based on a physical perception multi-task regression network algorithm, utilizing Koschmieder's law to achieve visibility estimation using only image data. This method incorporates a physical model into the framework for visibility extrapolation, rather than directly predicting visibility values ​​through convolutional neural networks. Furthermore, it innovatively proposes a new pixel-level visibility estimation paradigm, generating pixel-by-pixel visibility maps. This overcomes the limitation of traditional methods that only output a single value for the entire image. The results are richer in information under non-uniform fog conditions, contributing to the construction of more accurate fog warning systems, thereby better protecting intelligent transportation infrastructure and promoting its development. This method not only provides more accurate visibility estimates but also exhibits good real-time performance and adaptability, making it suitable for various scenarios.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Improve performance by integrating deep neural networks with physical laws; (2) A pixel-level visibility estimation paradigm is proposed for the first time to deal with uneven fog conditions. It is more informative and practical than single numerical estimation. (3) A new way of constructing datasets is proposed, which uses data augmentation algorithms to obtain datasets containing multi-scale information. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is the overall process of the present invention; Figure 2 This is a schematic diagram of the network structure; Figure 3The diagram shows the visibility estimation results, where (a) is the original foggy image, (b) is the image-level visibility estimation result, (c) is the transmittance map, and (d) is the pixel-level visibility distribution histogram. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Figure 1 This is the overall process of the present invention; A visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network includes the following steps: S1: Acquire urban scene images under clear weather conditions; S2: Preprocessing and data augmentation of urban scene images acquired under clear weather conditions; S3: Constructing a dataset based on preprocessed and data-enhanced images; S4: Construct a physical perception multi-task regression network model for estimating visibility parameters in urban foggy scenes; S5: Train the physical perception multi-task regression network based on the training set in the dataset to obtain a trained physical perception multi-task regression network model. S6: Input the test set images into the trained physical perception multi-task regression network model to estimate the visibility parameters of urban fog scenes; S7: Based on the estimated visibility parameters of urban foggy scenes, a visibility estimation algorithm is used to calculate the visibility of urban foggy scenes.

[0023] Steps S1 / S2 / S3 / S4 / S5 / S6 / S7 are executed sequentially; The preprocessing process for urban scene images acquired under clear weather conditions is as follows: For urban scene images acquired under clear weather conditions, depth maps and semantic segmentation maps are obtained through depth estimation algorithms and semantic segmentation algorithms, respectively.

[0024] The data augmentation employs a data augmentation algorithm. Inputs: fog-free image J, depth map D of the fog-free image; Output: Visibility map V, atmospheric light A, transmission map T, and fogging image I under current atmospheric light and visibility; The specific process is as follows: Acquire urban scene images and depth maps of urban scene images captured under clear weather conditions; Samples are taken from a uniform distribution U as a unified value for the visibility map V; The visibility map V is calculated using the following formula:

[0025] in: It is the meteorological visibility distance threshold defined by the International Commission on Illumination (CIE); A transmission map T is generated using depth D and visibility V; The red (R), green (G), and blue (B) values ​​of the generated atmospheric light A; A = [R, G, B] A fogged image is generated using the following formula. The specific formula for generating is as follows: ,

[0026] in: The atmospheric extinction coefficient (or scattering coefficient) is measured in m. -1 It comprehensively reflects the scattering and absorption of light by the atmosphere. The extinction coefficient β is a key parameter connecting image features and visibility; its physical meaning is the relative proportion of light intensity attenuation per unit distance. I is generated by combining J, A, and T; the values ​​of V, A, T, and I are returned.

[0027] Furthermore, the red (R), green (G), and blue (B) values ​​of atmospheric light A are generated; The value of channel B is sampled from U(180,255). The value of channel G is sampled from min(U(B-5,B+2),255). The value of channel R is min(U( -5, +2),255), to make the fog effect more realistic.

[0028] Figure 2 This is a schematic diagram of the network structure; Furthermore: the physical perception multi-task regression network model includes: Encoder: Used with ResNet 18. Network extraction of multi-scale features F; Decoder A: Used for atmospheric light estimation based on the multi-scale features output by the encoder. Decoder T: Used for transmittance estimation based on multi-scale features of the encoder output. Decoder D: Used for depth estimation based on the multi-scale features output by the encoder.

[0029] Furthermore: the encoder employs ResNet. The process of extracting multi-scale features F from the network is as follows: For an input image of size H×W×3, the processing order is as follows. A convolutional layer with a kernel size of 7×7, an output dimension of 64, and a stride of 2 is used, along with a 3×3 max pooling layer stacked with a stride of 2, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The first convolutional module uses four 3×3 convolutional layers, stacked with an output dimension of 64 and a stride of 1, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The second convolutional module, based on the feature map transmitted from the first convolutional layer, uses four 3×3 convolutional layers with an output dimension of 128 and a stride of 1. These layers are stacked to achieve downsampling and feature extraction in one operation, and the extracted output feature map is H / 8×W / 8×128. The third convolutional module, based on the feature map transmitted by the second convolutional module, uses four 3×3 convolutional layers with an output dimension of 256 and a stride of 1. These layers are stacked to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 16×W / 16×256. The fourth convolutional module, based on the feature map transmitted by the third convolutional module, uses four 3×3 convolutional layers with an output dimension of 512 and a stride of 1, stacked together to achieve downsampling and feature extraction in one operation, resulting in a final multi-scale feature map of H / 32×W / 32×512.

[0030] Furthermore, the process by which decoder A estimates atmospheric light based on the multi-scale features output by the encoder is as follows: First convolutional layer: Based on a convolutional kernel of size 1×1, output dimension 256, stride 1, a feature map of H / 32×W / 32×512 is obtained; The second convolutional layer: Based on the feature map transmitted from the first convolutional layer, it uses two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, and stacks them with a convolutional kernel with a size of 1×1, an output dimension of 3, and a stride of 1 to obtain an output of H / 32×W / 32×3. Based on the stacking of a Sigmoid activation layer and a Mean pooling layer, the output result H / 288×W / 512×3 is obtained, which is the atmospheric light vector.

[0031] Furthermore, the decoder D and decoder T have the same structure, but have independent class definitions and weight matrices, and have different loss functions to update the weight matrix of each convolutional kernel.

[0032] The process of decoder D and decoder T processing multi-scale feature maps is as follows; The first set of convolutional layers: Based on the stacking of two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, one upsampling and feature fusion are achieved to obtain a feature map of H / 16×W / 16×256. The second set of convolutional layers: Based on the feature map transmitted by the first set of convolutional layers, two convolutional kernels with a size of 3×3, an output dimension of 128, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 8×W / 8×128. The third convolutional layer: Based on the feature map transmitted by the second convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 64, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 4×W / 4×64. The fourth convolutional layer: Based on the feature map transmitted by the third convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 32, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 2×W / 2×32. The fifth convolutional layer: Based on the feature map transmitted by the third convolutional layer, it uses two convolutional kernels of size 3×3 with an output dimension of 16 and a stride of 1, as well as a convolutional kernel of size 3×3 with an output dimension of 1 and a stride of 1, and a stacked Sigmoid activation layer to achieve one upsampling and obtain an H×W×1 result map.

[0033] The weight matrices and loss functions of each layer of decoder D and decoder T are different, resulting in different transmittance maps and depth maps. Furthermore, the framework proposed in this invention first performs multi-task learning, and then couples multiple results with Koschmieder's law. Therefore, the loss functions of different tasks are used to train the network together. The expression of the loss function is as follows:

[0034] in: Represents the final loss function. This indicates the estimated atmospheric light loss. This indicates that the loss is estimated using a transmittance map. This indicates that the disparity map is used to estimate the loss. This represents the loss for estimating the dehazed image. This indicates the estimated loss of visibility. , , , and This is a hyperparameter.

[0035] Example 1 A visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network includes the following steps: S1: Acquire urban scene images under clear weather conditions; S2: Preprocessing and data augmentation of urban scene images acquired under clear weather conditions; S3: Constructing a dataset based on preprocessed and data-enhanced images; The training set, validation set, and test set are distributed as follows:

[0036] S4: Construct a physical perception multi-task regression network model for estimating visibility parameters in urban foggy scenes; S5: The physical perception multi-task regression network is trained based on the training set in the dataset to obtain a trained physical perception multi-task regression network model; during the training phase, the image and its corresponding pixel-level label are adjusted to 288×512, and the following data augmentation operations are performed: random cropping (relative size between 0.75 and 1), horizontal flipping with a 50% probability, and random brightness, contrast, saturation, and hue adjustments, with ranges of ±0.2, ±0.2, ±0.2, and ±0.1, respectively; S6: Input the test set images into the trained physical perception multi-task regression network model to estimate the visibility parameters of urban fog scenes; S7: Based on the estimated visibility parameters of urban foggy scenes, a visibility estimation algorithm is used to calculate the visibility of urban foggy scenes.

[0037] The physical perception multi-task regression network model can obtain three results by simply integrating the encoder and decoder: atmospheric light for defogging, and transmittance and depth as inputs to the visibility estimation algorithm; Encoder: , Decoder: , , , Furthermore, the physical perception multi-task regression network model was trained and validated: 4-1 According to Koschmieder's law, the formula for calculating pixel-level visibility can be obtained:

[0038] The pixel-level visibility is converted to image-level visibility based on the following principles: (1) Not all pixels in the visibility map are valid. For example, when the transmittance T of the sky region pixels is 0 and the depth D is infinite, the visibility V cannot be calculated using the formula; (2) Extreme outliers in the visibility map caused by human visual limitations should be ignored. Therefore, a minimum threshold for the transmittance map is set. Pixel-level visibility highest threshold Minimum threshold for pixel-level visibility Calculate the average pixel-level visibility map, where transmittance is above a threshold. And the estimated visibility is lower than the maximum value. If the number of valid pixels in the visibility map is 0, then the visibility will be set to the minimum value. .

[0039] 4-2 Use the dataset obtained in step 2 to train and test the physical perception multi-task regression network to obtain the optimal weights; 4-3 Using the trained physical perception multi-task regression network weights, the input foggy image to be estimated is dehazed and parameters are estimated to obtain the input parameters required for the pixel-level visibility calculation formula. Substituting these parameters into the formula, a pixel-level visibility map is obtained, and after conversion, an image-level visibility estimate is obtained. Furthermore: Regarding evaluation indicators: 5-1: For pixel-level visibility estimation, we use absolute relative error (AbsRel), squared relative error (SqRel), root mean square error (RMSE), and their logarithmic form (RMSElog) as indicators, defined as follows: , ,and These represent the predicted value, the actual value, and the number of valid pixels, respectively.

[0040]

[0041]

[0042]

[0043]

[0044] Image-level visibility estimation uses accuracy as the metric. To convert regression results into classification results, we divide visibility into 5 levels, defined as follows:

[0045] It is an indicator function: it returns 1 if the index condition is true, otherwise it returns 0; It is an equivalence relation: determining whether the predicted category is equal to the true category;

[0046] Finally, based on the obtained pixel-level visibility map and image-level visibility estimate, and through the mentioned evaluation metrics, the accuracy of the results is evaluated.

[0047] Figure 3 The diagram shows the visibility estimation results, where (a) is the original foggy image, (b) is the image-level visibility estimation result, (c) is the transmittance map, and (d) is the pixel-level visibility distribution histogram.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network, characterized in that: Includes the following steps: Acquire urban scene images under clear weather conditions; Preprocessing and data augmentation of urban scene images acquired under clear weather conditions; A dataset was constructed based on preprocessed and data-enhanced images; A physical perception multi-task regression network model was constructed to estimate visibility parameters in urban foggy scenes; The physical perception multi-task regression network is trained based on the training set in the dataset to obtain a trained physical perception multi-task regression network model. The test set images are input into a trained physical perception multi-task regression network model to estimate the visibility parameters of urban fog scenes. Based on the estimated visibility parameters of urban foggy scenes, a visibility estimation algorithm is used to calculate the visibility of urban foggy scenes.

2. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The data augmentation employs a data augmentation algorithm, and the specific process is as follows: Acquire urban scene images and depth maps of urban scene images captured under clear weather conditions; Samples are taken from a uniform distribution U as a unified value for the visibility map V; The visibility map V is calculated using the following formula: in: It is the meteorological visibility distance threshold defined by the International Commission on Illumination; A transmission map T is generated using depth D and visibility V; The red (R), green (G), and blue (B) values ​​of generated atmospheric light A; A fogged image is generated using the following formula. The specific formula for generating is as follows: ， in: is the atmospheric scattering coefficient.

3. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The red (R), green (G), and blue (B) values ​​of generated atmospheric light A; The value of channel B is sampled from U(180,255). The value of channel G is sampled from min(U(B-5,B+2),255). The value of channel R is min(U( -5, +2),255).

4. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The physical perception multi-task regression network model includes: Encoder: Used with ResNet 18. Network extraction of multi-scale features F; Decoder A: Used for atmospheric light estimation based on the multi-scale features output by the encoder. Decoder T: Used for transmittance estimation based on the multi-scale features output by the encoder. Decoder D: Used for depth estimation based on the multi-scale features output by the encoder.

5. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The encoder uses ResNet. The process of extracting multi-scale features F from the network is as follows: For an input image of size H×W×3, the processing order is as follows. A convolutional layer with a kernel size of 7×7, an output dimension of 64, and a stride of 2 is used, along with a 3×3 max pooling layer stacked with a stride of 2, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The first convolutional module uses four 3×3 convolutional layers, stacked with an output dimension of 64 and a stride of 1, to achieve downsampling and feature extraction in one operation, resulting in a feature map of H / 4×W / 4×64. The second set of convolutional modules, based on the feature map transmitted by the first set of convolutional modules, uses four 3×3 convolutional modules stacked together. The output dimension of the convolutional layer is 128 and the stride is 1, which realizes one downsampling and feature extraction. The extracted output feature map is H / 8×W / 8×128. The third convolutional module, based on the feature map transmitted by the second convolutional module, uses four 3×3 convolutional layers stacked together. The output dimension of the convolutional layer is 256 and the stride is 1, which realizes downsampling and feature extraction in one step, and obtains a feature map of H / 16×W / 16×256. The fourth convolutional module, based on the feature map transmitted by the fourth convolutional module, uses four 3×3 convolutional layers stacked together. The output dimension of the convolutional layer is 512 and the stride is 1, to achieve downsampling and feature extraction in one step, and obtain the final multi-scale feature map of H / 32×W / 32×512.

6. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The process by which decoder A performs atmospheric light estimation based on the multi-scale features output by the encoder is as follows: First convolutional layer: Based on a convolutional kernel of size 1×1, output dimension 256, and stride 1, a feature map of H / 32×W / 32×512 is obtained; The second convolutional layer: Based on the feature map transmitted from the first convolutional layer, it uses two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, and stacks them with a convolutional kernel with a size of 1×1, an output dimension of 3, and a stride of 1 to obtain an output of H / 32×W / 32×3. Based on the stacking of a Sigmoid activation layer and a Mean pooling layer, the output result H / 288×W / 512×3 is obtained, which is the atmospheric light vector.

7. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The decoder D and decoder T have the same structure, and the process of decoder D and decoder T processing multi-scale feature maps is as follows; The first set of convolutional layers: Based on the stacking of two convolutional kernels with a size of 3×3, an output dimension of 256, and a stride of 1, one upsampling and feature fusion are achieved to obtain a feature map of H / 16×W / 16×256. The second set of convolutional layers: Based on the feature map transmitted by the first set of convolutional layers, two convolutional kernels with a size of 3×3, an output dimension of 128, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 8×W / 8×128. The third convolutional layer: Based on the feature map transmitted by the second convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 64, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 4×W / 4×64. The fourth convolutional layer: Based on the feature map transmitted by the third convolutional layer, two convolutional kernels with a size of 3×3, an output dimension of 32, and a stride of 1 are stacked and used to achieve one upsampling and feature fusion, resulting in a feature map of H / 2×W / 2×32. The fifth convolutional layer: Based on the feature map transmitted by the third convolutional layer, it uses two convolutional kernels of size 3×3 with an output dimension of 16 and a stride of 1, as well as a convolutional kernel of size 3×3 with an output dimension of 1 and a stride of 1, and a stacked Sigmoid activation layer to achieve one upsampling and obtain an H×W×1 result map.

8. The visibility estimation method for urban foggy scenes based on a physical perception multi-task regression network according to claim 1, characterized in that: The loss function of the physical perception multi-task regression network model for different tasks is based on coupling the outputs of the three decoders with Koschmieder's law. The expression of the loss function is as follows: in: Represents the final loss function. This indicates the estimated atmospheric light loss. This indicates that the loss is estimated using a transmittance map. This indicates that the disparity map is used to estimate the loss. This represents the loss for estimating the dehazed image. This indicates the estimated loss of visibility. , , , and This is a hyperparameter.

Images