A Single-Image Dehazing Algorithm Based on the Atmospheric Scattering Model

Abstract

The present invention discloses a single-image defogging algorithm based on an atmospheric scattering model, and the steps are as follows: S1. Obtain the RESIDE public dataset, and randomly divide the data in the ITS dataset and the OTS dataset respectively for training and verification; S2. Use the divided datasets to train the first stage of the model respectively, and obtain pre-trained models for indoor image defogging and outdoor image defogging respectively; S3. Load the pre-trained models obtained in step S2 respectively, and jointly train the entire model through the corresponding indoor and outdoor datasets to obtain an indoor image defogging model and an outdoor image defogging model with good defogging performance; S4. Use the indoor or outdoor image defogging model obtained in step S3 to test the indoor or outdoor images of the SOTS dataset respectively, and obtain the corresponding clear image output. The algorithm of the present invention uses the OSA module to replace the conventional dense residual block to build a transmission map estimation network, reducing the memory reading of the model and reducing the consumption of time and memory resources.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a single-image defogging algorithm based on an atmospheric scattering model. Background Art

[0002] In haze weather, atmospheric particles scatter some light, resulting in the degradation of the captured images, such as low contrast and color distortion. These degraded images will affect the performance of subsequent advanced vision processing tasks, such as image classification and image segmentation. Therefore, it is necessary to perform defogging preprocessing to restore clear images to improve the performance of these tasks.

[0003] Currently, image dehazing algorithms can be mainly divided into three categories: prior-based methods, learning methods based on the atmospheric scattering model, and end-to-end learning methods. Among the prior-based methods, the frequently used ones are the DCP (dark channel) prior and the CAP (color attenuation) prior. The DCP prior is obtained by observing external clear images, that is, at least one color channel value is close to 0 in the local area of the haze-free image. When using this method to process indoor images or images with pixel values in some areas close to the atmospheric light, the dehazed image tends to be dark. The CAP prior is obtained by analyzing the far, middle, and near views of multiple hazy images and finding that the concentration of haze is proportional to the differences in brightness and saturation. However, when the image contains thick fog, it is difficult to estimate the transmission map based on the prior, resulting in the dehazing process being unable to remove all fog. Due to the defects of the above types of methods, learning methods based on the atmospheric scattering model and end-to-end learning methods have emerged. These methods use the network to automatically learn the features related to haze. DCPDN (Densely connected pyramid dehazing network) and DRN (Single Image Dehazing with An Independent Detail-Recovery Network) are both methods based on the atmospheric scattering model. They use DenseNet (Densely connected convolutional networks) and UNet (U-net: Convolutional networks for biomedical image segmentation.) to predict the transmission map and the atmospheric light value respectively. These two methods only have one stage for dehazing, resulting in low dehazing accuracy of the model. Moreover, DenseNet is based on dense residual blocks, resulting in a large amount of memory access during operation. The learning methods based on the atmospheric scattering model no longer use manually set statistical features to estimate the transmission map and the global atmospheric light, but use the network to automatically learn the features related to haze. These methods rely on the atmospheric scattering model to restore clear images by learning intermediate variables, ignoring the error propagation caused by inaccurate variable estimation, resulting in low dehazing accuracy; and these methods generally use dense residual blocks to build the transmission map estimation network, increasing the consumption of memory and time resources during the training and testing processes. The end-to-end learning method is based on a neural network, improving the realism of the restored image. However, since this process directly restores the clear image from the input foggy image, ignoring the estimation of intermediate variables in the dehazing process, this process lacks physical interpretability. Summary of the Invention

[0004] In view of the above deficiencies in existing image defogging methods, the present invention provides a single-image defogging algorithm based on the atmospheric scattering model. This algorithm is a learning method based on the atmospheric scattering model, which improves to achieve a two-stage single-image defogging method to enhance the defogging accuracy; at the same time, it removes the commonly used dense residual blocks to build the transmission map estimation network, reducing the time and memory consumption during network training and testing.

[0005] The single-image defogging algorithm based on the atmospheric scattering model provided by the present invention is as follows:

[0006] S1. Obtain the RESIDE public dataset, randomly divide part of the data in the ITS dataset for training and the other part for validation, and randomly divide part of the data in the OTS dataset for training and the other part for validation.

[0007] S2. Use the datasets divided in step S1 to train the first stage of the model respectively to obtain two pre-trained models, one for indoor image defogging and one for outdoor image defogging. Specifically, it includes the following sub-steps:

[0008] S21. Build the transmission map estimation network in the model. Between the input layer and the output layer of this network are successively a basic convolution block, a first OSA downsampling module, a second OSA downsampling module, a third OSA downsampling module, a fourth OSA downsampling module, a residual attention module, a fourth OSA upsampling module, a third OSA upsampling module, a second OSA upsampling module, a first OSA upsampling module, and a basic convolution block.

[0009] Both the OSA downsampling module and the OSA upsampling module are constructed from basic convolution blocks. And a residual attention module is connected between the corresponding numbered OSA upsampling and downsampling modules, and the output of the OSA downsampling module is passed to the OSA upsampling as the input.

[0010] The input to the output of the OSA module passes through 5 basic convolution blocks in sequence, and the outputs of the 5 basic convolution blocks are stacked as the input of the sixth basic convolution block, and the input is processed by the sixth basic convolution block to obtain the output. The OSA module only stores and reads the output results of the first five basic convolution blocks as the input on the last basic convolution block, reducing the time and memory consumption.

[0011] The residual attention module is stacked by a basic convolution layer and a residual structure. The input to the output of the residual structure passes through a basic convolution layer, a batch normalization layer, an activation function, a basic convolution layer, and a batch normalization layer in sequence.

[0012] S22. Build the atmospheric light estimation network in the model. Between the input layer and the output layer of this neural network, there are successively a first basic convolutional block, a second basic convolutional block, a third basic convolutional block, a fourth basic convolutional block, a fifth basic convolutional block, a detail extraction module, a fifth basic deconvolutional block, a fourth basic deconvolutional block, a third basic deconvolutional block, a second basic deconvolutional block, and a sixth basic convolutional block. The basic convolutional block from input to output is successively a convolutional layer, batch normalization, and a linear activation function. The basic deconvolutional block from input to output is successively a deconvolutional layer, batch normalization, and a linear activation function. And the detail extraction module is connected between the corresponding convolutional block and deconvolutional block, and the output of the convolutional block is passed to the deconvolutional block as input. The detail extraction module is stacked by a convolutional layer with a large convolutional kernel and a residual structure.

[0013] S23. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated transmission map through the transmission map estimation network.

[0014] S24. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated global atmospheric light value through the atmospheric light estimation network.

[0015] S25. Based on the inverse operation of the atmospheric scattering model, calculate and obtain the defogged image output of the first stage of the model.

[0016] S3. Load the pre-trained models obtained in step S2 respectively, and jointly train the entire model through the corresponding indoor dataset and outdoor dataset to obtain an indoor image defogging model and an outdoor image defogging model with good defogging performance. Specifically, it includes the following sub-steps:

[0017] S31. Build the transmission map estimation network and the atmospheric light estimation network of the model according to steps S21 and S22. Then build a feature extraction and fusion module, which successively passes through a feature extraction module, a basic convolutional layer, a residual structure, a basic convolutional layer, and a residual structure from input to output. The feature extraction module contains three branches. The first branch includes two basic convolutional layers, the second branch does nothing, and the third branch contains a Fourier transform layer, two basic convolutional layers, and an inverse Fourier transform layer; thus forming the overall network structure of the model. The overall network structure includes two stages, and each stage includes a transmission map estimation network and an atmospheric light estimation network. There is a feature extraction and fusion module between the two stages for extracting and transmitting features.

[0018] S32. Load the parameters of the pre-trained model in step S2 for the first stage of the model.

[0019] S33. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model to obtain a relatively clear defogged image.

[0020] S34. Extract features from the input foggy image and the clear image output in the first stage of the joint model, and iteratively fuse them to obtain a high-quality feature map, which is fed into the second stage of the model.

[0021] S35. Use the high-quality feature map obtained in step S34 to input into the transmission map estimation network to obtain the predicted transmission map.

[0022] S36. Use the high-quality feature map obtained in step S34 to input into the global atmospheric light estimation network to obtain the predicted global atmospheric light value.

[0023] S37. Based on the inverse operation of the atmospheric scattering model, calculate to obtain the final defogged image of the model as the output.

[0024] S4. Use the indoor image defogging model or outdoor image defogging model obtained in step S3 to test the indoor or outdoor images of the SOTS dataset respectively, and obtain the corresponding clear image output.

[0025] Among them, a loss function for training is designed in step S2 and step S3. Through the limitation of the loss function, the network model automatically converts the input foggy image into a clear image.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] (1) The defogging method provided by the present invention performs image processing in two stages, improves the defogging accuracy of the model, obtains a good defogging effect, effectively improves the defogging accuracy of a single image, and is more robust for cross-dataset testing.

[0028] (2) In each stage, the OSA module is used to replace the dense residual block in the existing method to build the transmission map estimation network, reducing the memory reading of the model and the consumption of time and memory resources.

[0029] Other advantages, objectives and features of the present invention will be partially reflected by the following description, and partially will also be understood by those skilled in the art through the research and practice of the present invention. Brief Description of the Drawings

[0030] Figure 1 is the overall flowchart of the single-image defogging algorithm based on the atmospheric scattering model of the present invention.

[0031] Figure 2 is the overall network structure diagram of image defogging in the embodiment.

[0032] Figure 3 is the network structure diagram of the OSA module in the embodiment.

[0033] Figure 4It is the structure diagram of the transmission map estimation network in the embodiment.

[0034] Figure 5 It is the structure diagram of the atmospheric light estimation network in the embodiment.

[0035] Figure 6 It is the comparison chart of the effects of image dehazing in the embodiment. Specific implementation manners

[0036] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0037] The technical terms involved in the present invention are explained as follows:

[0038] Transmission map (t(x)): It represents the part of the light that the atmospheric light directly propagates to the camera without scattering, refraction, etc., and is closely related to the scene depth.

[0039] Global atmospheric light (A): It represents the ambient light intensity.

[0040] Atmospheric scattering model: I(x) = J(x)t(x) + A(1 - t(x)), where I(x) represents the pixel intensity at the x position of the hazy image, and J(x) represents the pixel intensity at the x position of the clear image.

[0041] Inverse operation of the atmospheric scattering model: The operation of taking the maximum value of the denominator is to prevent the predicted transmission map value from approaching 0 and causing the formula to go wrong.

[0042] PSNR: Peak signal-to-noise ratio, in dB, used to measure the similarity between two images.

[0043] SSIM: Structural similarity, used to compare the brightness, contrast, and structure of two images.

[0044] ITS dataset: The indoor dataset of the public dataset RESIDE, which contains 13,990 pairs of data. Each pair of data includes a clear image, a hazy image, a transmission map, and an atmospheric light value.

[0045] OTS dataset: The outdoor dataset of the public dataset RESIDE, which contains 72,135 pairs of data. The data includes a clear image, a hazy image, a depth map, an atmospheric light value, and a scattering coefficient.

[0046] SOTS-indoor dataset: The indoor dataset of the public dataset SOTS, used to test the performance of the indoor model, and contains paired clear images and hazy images.

[0047] SOTS-outdoor dataset: The outdoor dataset of the public dataset SOTS, used to test the performance of outdoor models, containing paired clear images and foggy images.

[0048] As Figure 1-6 shown, the single-image defogging algorithm based on the atmospheric scattering model provided by the present invention has the following detailed steps:

[0049] Step 1: Obtain the RESIDE public dataset, randomly divide the ITS dataset into 10,000 pairs of data for training and 3,990 pairs of data for validation, and randomly divide the OTS dataset into 54,110 pairs of data for training and 18,025 pairs of data for validation.

[0050] Step 2: Use the datasets divided in Step 1 to train the first stage of the model respectively to obtain two pre-trained models, one for indoor image defogging and one for outdoor image defogging. Specifically, it includes the following sub-steps:

[0051] (1) Build the transmission map estimation network in the model. As Figure 3 shown, the input to the output of the OSA (One-Shot Aggregation) module passes through 5 basic convolutional blocks in sequence, and the outputs of the 5 basic convolutional blocks are stacked as the input of the sixth basic convolutional block, and the input is processed by the sixth basic convolutional block to obtain the output. The OSA module only stores and reads the output results of the first five basic convolutional blocks as the input on the last basic convolutional block, reducing time and memory consumption.

[0052] As Figure 4 shown, between the input layer and the output layer of the transmission map estimation network are, in sequence, a basic convolutional block, a first OSA downsampling module, a second OSA downsampling module, a third OSA downsampling module, a fourth OSA downsampling module, a residual attention module, a fourth OSA upsampling module, a third OSA upsampling module, a second OSA upsampling module, a first OSA upsampling module, and a basic convolutional block. The OSA downsampling module and the OSA upsampling module are both constructed from basic convolutional blocks. And a residual attention module is connected between the corresponding numbered OSA upsampling and downsampling modules, and the output of the OSA downsampling module is passed to the OSA upsampling as the input.

[0053] The residual attention module is stacked by a basic convolutional layer and a residual structure. The input to the output of the residual structure passes through a basic convolutional layer, a batch normalization layer, an activation function, a basic convolutional layer, and a batch normalization layer in sequence.

[0054] (2) Build the atmospheric light estimation network in the model. As Figure 5As shown, between the input layer and the output layer of the neural network, there are successively a first basic convolutional block, a second basic convolutional block, a third basic convolutional block, a fourth basic convolutional block, a fifth basic convolutional block, a detail extraction module, a fifth basic transposed convolutional block, a fourth basic transposed convolutional block, a third basic transposed convolutional block, a second basic transposed convolutional block, and a sixth basic convolutional block. The basic convolutional block successively includes a convolutional layer, batch normalization, and a linear activation function from input to output. The basic transposed convolutional block successively includes a transposed convolutional layer, batch normalization, and a linear activation function from input to output. And the connection between the convolutional block and the transposed convolutional block with the same number is the detail extraction module, which passes the output of the convolutional block to the transposed convolutional block as input. The detail extraction module is stacked by a convolutional layer with a large convolutional kernel and a residual structure.

[0055] (3) Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated transmission map through the transmission map estimation network.

[0056] (4) Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated global atmospheric light value through the atmospheric light estimation network.

[0057] (5) Based on the inverse operation of the atmospheric scattering model, calculate and obtain the defogged image output of the first stage of the model.

[0058] Step 3: Load the pre-trained models obtained in Step 2 respectively, and jointly train the entire model through the corresponding indoor / outdoor dataset to obtain an indoor image defogging model and an outdoor image defogging model with good defogging performance. Specifically, it includes the following sub-steps:

[0059] (1) Build the transmission map estimation network and the atmospheric light estimation network of the model according to (1) and (2) in Step 2. Then build a feature extraction and fusion module, which successively passes through a feature extraction module, a basic convolutional layer, a residual structure, a basic convolutional layer, and a residual structure from input to output. The feature extraction module contains three branches. The first branch includes two convolutional layers, the second branch does nothing, and the third branch includes a Fourier transform layer, two convolutional layers, and an inverse Fourier transform layer; thus forming the overall network structure of the model, as shown in Figure 2 . The overall network structure includes two stages, and each stage includes a transmission map estimation network and an atmospheric light estimation network. There is a feature extraction and fusion module between the two stages for extracting and transmitting features.

[0060] (2) Load the parameters of the pre-trained model in Step 2 for the first stage of the model.

[0061] (3) Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model to obtain a relatively clear defogged image.

[0062] (4) Extract features from the clear image output by the first stage of the input fog image joint model and iteratively fuse them to obtain a high-quality feature map, which is passed into the second stage of the model.

[0063] (5) Use the high-quality feature map obtained in the previous step and input it into the transmission map estimation network to obtain the predicted transmission map.

[0064] (6) Use the high-quality feature map obtained in the previous step and input it into the atmospheric light estimation network to obtain the predicted global atmospheric light value.

[0065] (7) Based on the inverse operation of the atmospheric scattering model, calculate and obtain the final defogged image of the model as the output.

[0066] Step 4: Use the indoor image defogging model or outdoor image defogging model obtained in Step 3 to test the indoor or outdoor images of the SOTS dataset respectively, and obtain the corresponding clear image output.

[0067] Among them, the loss functions used for training in Step 2 and Step 3 are introduced as follows:

[0068] (1) Restoration loss: Compare the difference between the defogged image of the model and the real clear image, which can effectively capture the low-frequency information of the image and force the defogged image to be closer to the real clear image.

[0069]

[0070] Among them, J i (x) represents the pixel intensity at position x of the real clear image corresponding to the i-th fog image, represents the pixel intensity at position x of the defogged image generated in the first stage of the i-th fog image, represents the pixel intensity at position x of the defogged image obtained by the entire model for the i-th fog image, ||.|| represents the L1 norm, and N is the number of training samples.

[0071] (2) L1 loss: Used to limit the intermediate variable transmission map and atmospheric light value predicted by the model,

[0072]

[0073] Among them, t i (x) represents the pixel intensity at position x of the real transmission map corresponding to the i-th fog image; represents the pixel intensity at position x of the predicted transmission map generated in the first stage of the i-th fog image; represents the pixel intensity at position x of the predicted transmission map obtained by the entire model for the i-th fog image; A i represents the real atmospheric light value corresponding to the i-th fog image, denotes the predicted atmospheric light value generated by the i-th foggy image in the first stage, denotes the predicted atmospheric light value obtained by the i-th foggy image in the entire model. N is the number of training samples.

[0074] (3) Total variable loss: Ensure that the transmission map predicted by the model is smooth.

[0075]

[0076] where represent taking differences in the horizontal and vertical directions respectively.

[0077] (4) Gradient loss: Ensure that the clear image generated by the model retains more edge information:

[0078]

[0079] (5) Reconstruction loss: Ensure that the generated clear image is consistent with the content of the input image:

[0080]

[0081] where I i (x) represents the pixel intensity of the i-th foggy image at position x, represents the pixel intensity of the foggy image reconstructed based on the predicted value obtained by the i-th foggy image in the first stage of the model at position x, represents the pixel intensity of the foggy image reconstructed based on the predicted value obtained by the i-th foggy image in the second stage of the model at position x. The processes of the two reconstructed foggy images are shown in the following formulas:

[0082]

[0083]

[0084] (6) The loss function for the first-stage pre-training is:

[0085] L total = L rec + L1 + 0.05L TV + 0.1L edge + L recon

[0086] (7) The loss function for the joint training is:

[0087] L total = 0.1L rec + L1 + 0.05L TV + L edge + L recon

[0088] The present invention provides a single-image defogging method based on an atmospheric scattering model and detail supplementation. In one embodiment, as Figure 6 shown, the foggy image in the figure is input into the trained model, and the network model generates a defogged image. It can be seen that the two-stage defogging method based on the atmospheric scattering model of the present invention has achieved a good defogging effect; moreover, the present invention uses the OSA module to build a transmission map estimation network, which reduces the consumption of time and memory resources.

[0089] The defogging method of the present invention is compared with two existing model structures, DCPDN and DRN. For the DCPDN model, see the literature Zhang, H., Patel, V.M.: Densely connected pyramid dehazing network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. (2018) 3194–3203. For the DRN model, see the literature Li, Y., Cheng, D., Sun, J., Zhang, D., Wang, N., Gao, X.: Single image dehazing with an independent detail-recovery network. arXiv preprint arXiv:2109.10492 (2021). The comparison results are shown in Table 1. The two network structures of the model structures DCPDN and DRN also use a separate transmission map estimation network and an atmospheric light estimation network to obtain the estimated values of the intermediate variables, and then use the inverse transformation of the atmospheric scattering model to obtain the defogged image. However, the difference compared with the present invention is that the present invention is a two-stage defogging model, and compared with the resource-consuming dense residual blocks, the present invention uses the OSA module to construct a transmission map estimation network. The PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) are used to compare the defogging accuracy of the models. The higher the values of these two indicators, the better the defogging accuracy. As shown in Table 1, the experimental results of the present invention are better for different data sets, highlighting that the two-stage model structure effectively improves the accuracy of the model.

[0090] Table 1. Comparison results of experimental data of existing model methods and the method of the present invention

[0091]

[0092] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments of equivalent changes by using the above-disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A single-image defogging algorithm based on an atmospheric scattering model, characterized in that, The steps are as follows: S1. Obtain the RESIDE public dataset. Randomly divide part of the data in the ITS dataset for training and the other part for validation. Randomly divide part of the data in the OTS dataset for training and the other part for validation; S2. Use the datasets divided in step S1 to train the first stage of the model respectively, and obtain two pre-trained models, one for indoor image dehazing and one for outdoor image dehazing; Specifically, it includes the following sub-steps: S21. Build the transmission map estimation network in the model; from the input layer to the output layer of this network, there are successively a basic convolution block, a first OSA downsampling module, a second OSA downsampling module, a third OSA downsampling module, a fourth OSA downsampling module, a residual attention module, a fourth OSA upsampling module, a third OSA upsampling module, a second OSA upsampling module, a first OSA upsampling module, and a basic convolution block; a residual attention module is connected between the corresponding numbered OSA upsampling and downsampling modules, and the output of the OSA downsampling module is passed to the OSA upsampling module as input; S22. Build the atmospheric light estimation network in the model; S23. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated transmission map through the transmission map estimation network; S24. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model, and obtain the estimated global atmospheric light value through the atmospheric light estimation network; S25. Calculate and obtain the dehazed image output of the first stage of the model based on the inverse operation of the atmospheric scattering model; Both the OSA downsampling module and the OSA upsampling module are constructed by basic convolution blocks; from the input to the output of the OSA downsampling module and the OSA upsampling module, they successively pass through 5 basic convolution blocks, and the outputs of the 5 basic convolution blocks are stacked as the input of the sixth basic convolution block, and the input is processed by the sixth basic convolution block to obtain the output; S3. Load the pre-trained models obtained in step S2 respectively, and jointly train the entire model through the corresponding indoor dataset and outdoor dataset to obtain an indoor image dehazing model and an outdoor image dehazing model with good dehazing performance; Specifically, it includes the following sub-steps: S31. Build the transmission map estimation network and the atmospheric light estimation network of the model according to steps S21 and S22; then build the feature extraction and fusion module; form the overall network structure of the model; S32. Load the parameters of the pre-trained model in step S2 in the first stage of the model; S33. Randomly crop the paired training data into images with a length of 256 pixels and a width of 256 pixels and input them into the first stage of the model to obtain a relatively clear dehazed image; S34. Extract features from the input foggy image and the clear image output by the first stage of the model and iteratively fuse them to obtain a high-quality feature map, and input it into the second stage of the model; S35. Use the high-quality feature map obtained in step S34 and input it into the transmission map estimation network to obtain the predicted transmission map; S36. Using the high-quality feature map obtained in step S34, input it into the atmospheric light estimation network to obtain the predicted global atmospheric light value; S37. Based on the inverse operation of the atmospheric scattering model, calculate to obtain the final dehazed image of the model as the output; S4. Using the indoor image dehazing model or outdoor image dehazing model obtained in step S3, test the indoor or outdoor images of the SOTS dataset respectively, and obtain the corresponding clear image output.

2. The single-image defogging algorithm based on the atmospheric scattering model according to claim 1, characterized in that, The residual attention module is stacked by a basic convolutional layer and a residual structure. The input to the output of the residual structure sequentially passes through a basic convolutional layer, a batch normalization layer, an activation function, a basic convolutional layer, and a batch normalization layer.

Images