A method for removing non-uniform fog from remote sensing images

Abstract

The present invention discloses a method for removing non-uniform fog from remote sensing images. The method steps are: constructing a fog removal network for remote sensing images and performing supervised learning: obtaining a synthetic foggy remote sensing image data set and a fog-free label image; inputting the synthetic foggy remote sensing image data set into the remote sensing image fog removal network for processing, and outputting a synthetic defogging remote sensing image data set; calculating the loss value between the synthetic defogging remote sensing image data set and the fog-free label image; reversely transmitting the loss value to the remote sensing image fog removal network for optimization, and obtaining an optimized remote sensing image fog removal network; obtaining a real foggy remote sensing image and inputting it into the optimized remote sensing image fog removal network for processing, and obtaining a real defogging remote sensing image. The present invention uses a dual attention defogging module for feature extraction, so that the network has stronger non-uniform fog perception and removal capabilities; the fog perception features between different levels are directly transmitted, which improves the efficiency of extracting fog features and obtains a better defogging effect.

Description

Technical Field

[0001] The invention relates to a method for removing fog from an image, and in particular to a method for removing non-uniform fog from a remote sensing image. Background Art

[0002] Remote sensing technology has been used in various fields including military, agriculture, and forestry. However, the quality of optical remote sensing images is often degraded by fog. The imaging light will be scattered, resulting in low brightness and low contrast. In addition, due to the wide field of view covered by remote sensing imaging, the fog on remote sensing images is often non-uniform. This will increase the difficulty of performing various key tasks, such as target detection and surface classification. Removing non-uniform fog from remote sensing images can not only improve image quality, but also enhance the accuracy of various tasks.

[0003] Remote sensing defogging methods can generally be divided into traditional methods and deep learning methods. Traditional methods use strong priors or assumptions to constrain optimization results, such as methods based on inferred fog thickness maps, methods based on dark channel priors, methods based on haze optimized transformation (HOT), and methods based on elliptical boundary priors. However, the priors based on these methods may not be valid in all cases, so they sometimes produce artifacts and lack the universality of handling all cases. Methods based on deep learning have better reconstruction effects, but similar methods in the past were mainly based on convolutional neural networks, and there was no effective removal strategy for non-uniform fog. Convolution is a method of extracting object edges, so previous convolutional neural network methods tend to enhance ground edge details to achieve defogging effects. However, this strategy does not perform well in removing non-uniform fog, because simply stacking convolutional layers will make the neural network lack the ability to perceive fog and not know which details need to be enhanced and which details do not need to be enhanced. Summary of the invention

[0004] In order to solve the problems existing in the background technology, the present invention provides a method for removing non-uniform fog from remote sensing images, the purpose of which is to enhance the perception ability of neural networks to fog, thereby achieving better results in removing non-uniform fog from remote sensing images.

[0005] The technical solution adopted by the present invention is:

[0006] The method for removing non-uniform fog from remote sensing images of the present invention comprises the following steps:

[0007] 1) Construct a fog removal network for remote sensing images;

[0008] 2) Supervised learning of the fog removal network:

[0009] 2.1) Obtain synthetic foggy remote sensing images, establish synthetic foggy remote sensing image datasets, and obtain supervised learning labels as fog-free label images;

[0010] 2.2) Inputting the synthetic foggy remote sensing image dataset into the remote sensing image fog removal network for processing, and outputting the synthetic defogging remote sensing image dataset;

[0011] 2.3) Calculate the loss value between the synthetic dehazed remote sensing image dataset and the haze-free labeled image. The loss value is the weighted sum of the pixel-level loss and the content loss.

[0012] 2.4) The loss value is transmitted back to the remote sensing image defogging network, and the remote sensing image defogging network is optimized to obtain an optimized remote sensing image defogging network;

[0013] 3) Obtain a real foggy remote sensing image, input the real foggy remote sensing image into the optimized remote sensing image fog removal network for processing, and obtain a real defogging remote sensing image.

[0014] In the step 1), the remote sensing image fog removal network includes a first 3×3 convolutional layer, K encoders, (K+1)(K+2) / 2 dual attention defogging modules DAD, K decoders and a second 3×3 convolutional layer; each encoder corresponds to a decoder, K+2-n dual attention defogging modules DAD are connected in sequence between the nth encoder and the nth decoder, and another dual attention defogging module DAD is connected between the Kth encoder and the Kth decoder, wherein K is the total number of encoders, n is an integer parameter, K and n are both positive integers, n=1, 2, 3...K.

[0015] In the step 2.2), the synthetic foggy remote sensing image data set is input into the remote sensing image fog removal network for processing, and the following operations are performed for each synthetic foggy remote sensing image in the synthetic foggy remote sensing image data set:

[0016] 2.2.1) Multi-scale mapping

[0017] The synthetic foggy remote sensing image is input into the first 3×3 convolutional layer for processing, and the image features with increased dimension are output; the image features with increased dimension are mapped to K+1 scales after being mapped through K encoders in sequence, and are divided into K+1 scale layers. The nth encoder and the nth decoder and the K+2-n dual attention defogging modules DAD connected in sequence between the two constitute the nth scale layer; another dual attention defogging module DAD connected between the Kth encoder and the Kth decoder constitutes the K+1th scale layer; the input scale and output scale of each scale layer are the same;

[0018] Each encoder is mapped into an output of the same scale and an output of half scale, and the half scale output of each encoder is used as the input of the next encoder;

[0019] When n is not equal to K, the output of the nth encoder with the same scale is downsampled and spliced ​​with the output of the n+1th to Kth encoders with the same scale as the nth splicing result, and the nth splicing result is input into the first dual attention defogging module DAD of the n+1th scale layer; the scale of each downsampled result during splicing is the same as the scale of the spliced ​​output with the same scale;

[0020] When n is equal to K, the outputs of the 1st to Kth encoders with the same scale are all downsampled and spliced ​​with the half-scale output of the Kth encoder and then input into the dual attention dehazing module DAD of the K+1th scale layer; the scale of each downsampled result during splicing is the same as the scale of the half-scale output of the Kth encoder being spliced;

[0021] If the input of an encoder is an image feature E0 (with a scale of height H × width W), then its output is E′0 (with a scale of height H × width W) and E1 (with a scale of ).

[0022] K+1 scales are H×W,

[0023] 2.2.2) Feature extraction and feature transfer

[0024] The Kth splicing result is directly input into the dual attention defogging module DAD in the K+1th scale layer for processing; the processing result of the dual attention defogging module DAD in the K+1th scale layer is upsampled and scaled twice before feature transfer. The feature transfer is specifically performed by attention mechanism processing with the processing result of the first dual attention defogging module DAD in the Kth scale layer;

[0025] When n is not equal to 1, the nth splicing result is directly input into the first dual attention defogging module DAD of the n+1th scale layer for processing; the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is upsampled and scaled twice before feature transfer, and the feature transfer is specifically performed with the processing result of the mth dual attention defogging module DAD in the n-1th scale layer. At the same time, the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is spliced ​​with the attention mechanism results of the mth dual attention defogging module DAD in the second scale layer to the n-1th scale layer, and then input into the corresponding m+1th dual attention defogging module DAD in the second scale layer to the n-1th scale layer;

[0026] When n is equal to 1, the output of the first encoder with the same scale is directly input into the first dual attention defogging module DAD of the first scale layer for processing; the attention mechanism result of the mth dual attention defogging module DAD in the first scale layer is directly input into the m+1th dual attention defogging module DAD in the first scale layer;

[0027] The output of the last dual attention defogging module DAD in the nth scale layer is input to the nth decoder for processing; the output of the dual attention defogging module DAD in the K+1th scale layer is input to the Kth decoder for processing; where m is an integer parameter, m=1, 2, 3...K+1-n;

[0028] Upsampling refers to bilinear interpolation of image features, and then the output after processing by a 1×1 convolution layer is processed by an activation function.

[0029] 2.2.3) Feature decoding

[0030] The output of the last dual attention defogging module DAD in the K+1th scale layer and the Kth scale layer is used as the input of the Kth decoder, and is input into the K-1th decoder after feature decoding processing by the Kth decoder; except for the Kth decoder, the input of other nth decoders is the output of the last dual attention defogging module DAD in the nth scale layer and the output of the n+1th decoder; finally, the 1st decoder performs feature decoding on the input and outputs the high-order image feature F;

[0031] The high-order image feature F is mapped through the second 3×3 convolutional layer and output as a synthetic dehazed remote sensing image. Each synthetic dehazed remote sensing image constitutes a synthetic dehazed remote sensing image dataset.

[0032] The dual attention defogging module DAD includes a channel attention branch, a spatial attention branch, and a first 1×1 convolutional layer; the input of the dual attention defogging module DAD is respectively input into the attention branch and the spatial attention branch for processing, the output of the attention branch and the output of the spatial attention branch are spliced, and the splicing result is then processed by the attention mechanism with the input of the dual attention defogging module DAD after passing through the first 1×1 convolutional layer, and the attention mechanism result is used as the output of the dual attention defogging module DAD;

[0033] The attention branch extracts feature information in the input channel dimension and performs a "compression-excitation" operation; the spatial attention branch extracts feature information in the input spatial dimension and performs a "compression-excitation" operation. By integrating channel information and spatial information, the perception ability of the remote sensing image fog removal network is enhanced, and fog perception features are extracted.

[0034] The channel attention branch includes a third 3×3 convolutional layer, a fourth 3×3 convolutional layer, a maximum pooling layer, a second 1×1 convolutional layer, and a third 1×1 convolutional layer; the input of the channel attention branch is sequentially input into the third 3×3 convolutional layer, the fourth 3×3 convolutional layer, the maximum pooling layer, the second 1×1 convolutional layer, and the third 1×1 convolutional layer for processing, the output of the third 1×1 convolutional layer is multiplied by the output of the fourth 3×3 convolutional layer, the multiplication result is then added to the input of the channel attention branch, and the addition result is used as the output of the channel attention branch;

[0035] The spatial attention branch includes a fifth 3×3 convolutional layer, a sixth 3×3 convolutional layer, a seventh 3×3 convolutional layer, an eighth 3×3 convolutional layer, and a ninth 3×3 convolutional layer; the input of the spatial attention branch is sequentially input into the fifth 3×3 convolutional layer, the sixth 3×3 convolutional layer, the seventh 3×3 convolutional layer, the eighth 3×3 convolutional layer, and the ninth 3×3 convolutional layer for processing, the output of the ninth 3×3 convolutional layer is multiplied by the output of the sixth 3×3 convolutional layer, the multiplication result is then added to the input of the spatial attention branch, and the addition result is used as the output of the spatial attention branch.

[0036] In the step 2.3), for each synthetic defogging remote sensing image in the synthetic defogging remote sensing image dataset, the following operations are performed:

[0037] The loss value L between the synthetic dehazed remote sensing image and the haze-free label image is calculated as shown below:

[0038] L=L pixel +0.04×L context

[0039] Among them, L pixel Represents pixel-level loss, which is a smooth L1 loss function; L context Represents content loss, which is a perceptual loss function.

[0040] In the step 2.4), all loss values ​​L are simultaneously transmitted back to the remote sensing image defogging network using the ADAM optimizer for optimization, thereby obtaining an optimized remote sensing image defogging network. The initial learning rate of the ADAM optimizer is 0.0003.

[0041] Aiming at the non-uniformity of remote sensing fog, a remote sensing image defogging network with stronger perception ability of non-uniform fog is designed; a large-scale dataset is constructed to train and optimize the remote sensing image defogging network; the trained optimized remote sensing image defogging network is used to defog real foggy remote sensing images.

[0042] The beneficial effects of the present invention are:

[0043] When processing foggy remote sensing images, considering that the fog on remote sensing images is often non-uniform, in order to enhance the perception of non-uniform fog, the present invention first performs multi-scale processing on foggy remote sensing images, and uses a dual attention defogging module to extract fog perception features at each scale, so that the network has a stronger fog perception ability; in the process of transmitting fog features, the upsampled small-scale fog perception feature map is used as the attention image of the large-scale fog perception feature map, so that the perception of fog is directly transferred from the small scale to the large scale, that is, the fog perception features between different levels can be directly transferred, which improves the efficiency of extracting fog features. The method proposed by the present invention has a stronger ability to perceive and remove non-uniform fog, and can obtain better defogging processing effect than other similar methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the method of the present invention;

[0045] Figure 2 This is a schematic diagram of the dual attention dehazing module structure;

[0046] Figure 3 Schematic diagram of the channel attention branch of the dual attention dehazing module;

[0047] Figure 4 Schematic diagram of the spatial attention branch of the dual-attention dehazing module;

[0048] Figure 5 Schematic diagram of extracting fog perception features for the dual-attention dehazing module;

[0049] Figure 5 (a) is an example input remote sensing image;

[0050] Figure 5 (b) is a schematic diagram of image features before entering the dual attention dehazing module;

[0051] Figure 5 (c) is a schematic diagram of fog perception features extracted by the dual attention defogging module;

[0052] Figure 5 (d) is a schematic diagram of fog attention features after small-scale feature upsampling;

[0053] Figure 6 This is a schematic diagram of an example of defogging a synthetic foggy remote sensing image;

[0054] Figure 6 (a) is a schematic diagram of an example of synthetic foggy remote sensing image;

[0055] Figure 6 (b) is a schematic diagram of an example of synthesizing a dehazed remote sensing image;

[0056] Figure 6 (c) is a schematic diagram of an example of a haze-free label image;

[0057] Figure 7 This is a schematic diagram of an example of defogging a real foggy remote sensing image;

[0058] Figure 7 (a) is a schematic diagram of an example of a real foggy remote sensing image;

[0059] Figure 7 (b) is a schematic diagram of an example of a real dehazed remote sensing image. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] The specific embodiments are as follows:

[0062] 1) Construct a fog removal network for remote sensing images;

[0063] like Figure 1 As shown in the figure, the remote sensing image fog removal network includes a first 3×3 convolutional layer, 3 encoders, 10 dual attention defogging modules DAD, 3 decoders and a second 3×3 convolutional layer; each encoder corresponds to a decoder, 5-n dual attention defogging modules DAD are connected between the nth encoder and the nth decoder in sequence, and another dual attention defogging module DAD is connected between the 3rd encoder and the 3rd decoder, where the total number of encoders is 3, n is an integer parameter, n is a positive integer, n=1, 2, 3.

[0064] 2) Supervised learning of the fog removal network:

[0065] 2.1) Obtain synthetic foggy remote sensing images, establish a synthetic foggy remote sensing image dataset, and obtain supervised learning labels as fog-free label images.

[0066] 2.2) If Figure 6 As shown, the synthetic foggy remote sensing image dataset is input into the remote sensing image fog removal network for processing, and the synthetic defogging remote sensing image dataset is output.

[0067] The synthetic foggy remote sensing image dataset is input into the remote sensing image fog removal network for processing. For each synthetic foggy remote sensing image in the synthetic foggy remote sensing image dataset, the following operations are performed:

[0068] 2.2.1) Multi-scale mapping

[0069] The synthetic foggy remote sensing image is input into the first 3×3 convolutional layer for processing, and the image features with increased dimension are output; the image features with increased dimension are mapped to 4 scales after being mapped through 3 encoders in sequence, and are divided into 4 scale layers. The nth encoder and the nth decoder and the 5-n dual attention defogging modules DAD connected in sequence between the two constitute the nth scale layer; another dual attention defogging module DAD connected between the 3rd encoder and the 3rd decoder constitutes the 4th scale layer; the input scale and output scale of each scale layer are the same. Each encoder is mapped to the same scale output and the scale-halved output, and the scale-halved output of each encoder is used as the input of the next encoder.

[0070] When n is not equal to 3, the output of the nth encoder with the same scale is downsampled and spliced ​​with the output of the n+1th to 3rd encoders with the same scale as the nth splicing result, and the nth splicing result is input into the first dual attention dehazing module DAD of the n+1th scale layer; the scale of each downsampled result during splicing is the same as the scale of the spliced ​​output with the same scale.

[0071] When n is equal to 3, the outputs of the first to third encoders with the same scale are all downsampled and spliced ​​with the halved scale output of the third encoder and then input into the dual attention dehazing module DAD of the fourth scale layer; the scale of each downsampling result during splicing is the same as the scale of the halved scale output of the spliced ​​third encoder.

[0072] If the input of an encoder is an image feature E0 (with a scale of height H × width W), then its output is E′0 (with a scale of height H × width W) and E1 (with a scale of ). The four scales are H×W,

[0073] 2.2.2) Feature extraction and feature transfer

[0074] The third splicing result is directly input into the dual attention dehazing module DAD in the fourth scale layer for processing; the processing result of the dual attention dehazing module DAD in the fourth scale layer is upsampled and scaled twice before feature transfer. The feature transfer is specifically performed by attention mechanism processing with the processing result of the first dual attention dehazing module DAD in the third scale layer.

[0075] When n is not equal to 1, the nth splicing result is directly input into the first dual attention defogging module DAD of the n+1th scale layer for processing; the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is upsampled and scaled twice before feature transfer, and the feature transfer is specifically performed with the processing result of the mth dual attention defogging module DAD in the n-1th scale layer by the attention mechanism; at the same time, the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is spliced ​​with the attention mechanism results of the mth dual attention defogging module DAD in the second scale layer to the n-1th scale layer, and then input into the corresponding m+1th dual attention defogging module DAD in the second scale layer to the n-1th scale layer.

[0076] When n is equal to 1, the scale-identical output of the first encoder is directly input into the first dual-attention dehazing module DAD of the first scale layer for processing; the attention mechanism result of the mth dual-attention dehazing module DAD in the first scale layer is directly input into the m+1th dual-attention dehazing module DAD in the first scale layer.

[0077] The output of the last dual attention dehazing module DAD in the nth scale layer is input into the nth decoder for processing; the output of the dual attention dehazing module DAD in the fourth scale layer is input into the third decoder for processing; wherein m is an integer parameter, m=1, 2, 3.

[0078] The output of the dual attention defogging module DAD and the attention mechanism in each scale layer is the fog perception feature of the foggy remote sensing image. The fog perception feature in this scale layer is upsampled to a larger scale, which is used as the attention image of the larger scale fog perception feature. Upsampling refers to the bilinear interpolation of the image features, and then the output after a 1×1 convolution layer is processed and then processed by the activation function.

[0079] 2.2.3) Feature decoding

[0080] The output of the last dual attention defogging module DAD in the 4th scale layer and the 3rd scale layer is used as the input of the 3rd decoder, and is input into the 2nd decoder after feature decoding by the 3rd decoder; except for the 3rd decoder, the input of other nth decoders is the output of the last dual attention defogging module DAD in the nth scale layer and the output of the n+1th decoder; finally, the 1st decoder performs feature decoding on the input and outputs the high-order image feature F.

[0081] The high-order image feature F is mapped through the second 3×3 convolutional layer and output as a synthetic dehazed remote sensing image. Each synthetic dehazed remote sensing image constitutes a synthetic dehazed remote sensing image dataset.

[0082] like Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, the dual attention dehazing module DAD includes a channel attention branch, a spatial attention branch, and a first 1×1 convolutional layer; the input of the dual attention dehazing module DAD is respectively input into the attention branch and the spatial attention branch for separate processing, and the output of the attention branch and the output of the spatial attention branch are spliced, and the splicing result is then processed by the attention mechanism with the input of the dual attention dehazing module DAD after passing through the first 1×1 convolutional layer, and the attention mechanism result is used as the output of the dual attention dehazing module DAD.

[0083] The attention branch extracts feature information in the input channel dimension and performs a "compression-excitation" operation; the spatial attention branch extracts feature information in the input spatial dimension and performs a "compression-excitation" operation. By integrating channel information and spatial information, the perception ability of the remote sensing image fog removal network is enhanced, and fog perception features are extracted.

[0084] The channel attention branch includes the third 3×3 convolution layer, the fourth 3×3 convolution layer, the maximum pooling layer, the second 1×1 convolution layer and the third 1×1 convolution layer; the input of the channel attention branch is sequentially input into the third 3×3 convolution layer, the fourth 3×3 convolution layer, the maximum pooling layer, the second 1×1 convolution layer and the third 1×1 convolution layer for processing, the output of the third 1×1 convolution layer is multiplied by the output of the fourth 3×3 convolution layer, and the multiplication result is added to the input of the channel attention branch, and the addition result is used as the output of the channel attention branch, as shown in Figure 3 shown.

[0085] The spatial attention branch includes the fifth 3×3 convolution layer, the sixth 3×3 convolution layer, the seventh 3×3 convolution layer, the eighth 3×3 convolution layer, and the ninth 3×3 convolution layer; the input of the spatial attention branch is sequentially input into the fifth 3×3 convolution layer, the sixth 3×3 convolution layer, the seventh 3×3 convolution layer, the eighth 3×3 convolution layer, and the ninth 3×3 convolution layer for processing, and the output of the ninth 3×3 convolution layer is multiplied by the output of the sixth 3×3 convolution layer, and the multiplication result is added to the input of the spatial attention branch, and the addition result is used as the output of the spatial attention branch. Figure 4 shown.

[0086] 2.3) Calculate the loss value between the synthetic dehazed remote sensing image dataset and the haze-free label image. The loss value is the weighted sum of the pixel-level loss and the content loss. For each synthetic dehazed remote sensing image in the synthetic dehazed remote sensing image dataset, perform the following operations:

[0087] The loss value L between the synthetic dehazed remote sensing image and the haze-free label image is calculated as shown in the following formula:

[0088] L=L pixel +0.04×L context

[0089] Among them, L pixel Represents pixel-level loss, which is a smooth L1 loss function L1 smooth ; L context Represents content loss, which is the perceptual loss function L per .

[0090] Loss value L, smooth L1 loss function L1 smooth And the perceptual loss function L per The specific definition of is:

[0091] L(I d , I gt )=L pixel (I d , I gt )+0.04×L context (I d , I gt )

[0092]

[0093]

[0094] Among them, I d represents the dehazed remote sensing image, I gt represents a haze-free label image, Represents the backbone neural network The jth layer is from the dehazed remote sensing image I d The extracted features, Represents the backbone neural network The jth layer is from the haze-free label image I gt The extracted features, represents the L2 norm, C j , H j and W j Respectively or The number of channels, height and width of the backbone neural network Select VGG19.

[0095] 2.4) The loss value is transmitted back to the remote sensing image defogging network, the remote sensing image defogging network is optimized, and the optimized remote sensing image defogging network is obtained.

[0096] The ADAM optimizer is used to simultaneously transmit all loss values ​​L back to the remote sensing image defogging network for optimization, and the optimized remote sensing image defogging network is obtained. The initial learning rate of the ADAM optimizer is 0.0003.

[0097] 3) If Figure 7 As shown, a real foggy remote sensing image is obtained, and the real foggy remote sensing image is input into the optimized remote sensing image fog removal network for processing to obtain a real defogging remote sensing image.

Claims

1. A method for removing non-uniform fog from remote sensing images, characterized in that: The following steps are involved: 1) Construct a fog removal network for remote sensing images; 2) Supervised learning of the fog removal network: 2.1) Obtain synthetic foggy remote sensing images, establish synthetic foggy remote sensing image datasets, and obtain supervised learning labels as fog-free label images; 2.2) Inputting the synthetic foggy remote sensing image dataset into the remote sensing image fog removal network for processing, and outputting the synthetic defogging remote sensing image dataset; 2.3) Calculate the loss value between the synthetic dehazed remote sensing image dataset and the haze-free label image; 2.4) The loss value is transmitted back to the remote sensing image defogging network, and the remote sensing image defogging network is optimized to obtain an optimized remote sensing image defogging network; 3) Obtain a real foggy remote sensing image, input the real foggy remote sensing image into the optimized remote sensing image fog removal network for processing, and obtain a real defogging remote sensing image; In the step 1), the remote sensing image fog removal network includes a first 3×3 convolutional layer, K encoders, (K+1)(K+2) / 2 dual attention defogging modules DAD, K decoders and a second 3×3 convolutional layer; each encoder corresponds to a decoder, K+2-n dual attention defogging modules DAD are connected in sequence between the nth encoder and the nth decoder, and another dual attention defogging module DAD is connected between the Kth encoder and the Kth decoder, wherein K is the total number of encoders, n is an integer parameter, K and n are both positive integers, n=1, 2, 3...K.

2. The method for removing non-uniform fog from a remote sensing image according to claim 1, characterized in that: In the step 2.2), the synthetic foggy remote sensing image data set is input into the remote sensing image fog removal network for processing, and the following operations are performed for each synthetic foggy remote sensing image in the synthetic foggy remote sensing image data set: 2.2.1) Multi-scale mapping The synthetic foggy remote sensing image is input into the first 3×3 convolutional layer for processing, and the image features with increased dimension are output; The dimensionally enhanced image features are mapped to K+1 scales after being mapped through K encoders in sequence, and are divided into K+1 scale layers. The nth encoder and the nth decoder and the K+2-n dual attention defogging modules DAD connected in sequence between the two constitute the nth scale layer; another dual attention defogging module DAD connected between the Kth encoder and the Kth decoder constitutes the K+1th scale layer; the input scale and output scale of each scale layer are the same; Each encoder is mapped into an output of the same scale and an output of half scale, and the half scale output of each encoder is used as the input of the next encoder; When n is not equal to K, the output of the nth encoder with the same scale is downsampled and then spliced ​​with the output of the n+1th to Kth encoders with the same scale as the nth splicing result; the scale of each downsampled result during splicing is the same as the scale of the spliced ​​output with the same scale; When n is equal to K, the outputs of the first to Kth encoders with the same scale are all downsampled and then spliced ​​with the half-scale output of the Kth encoder; the scale of each downsampled result during splicing is the same as the scale of the half-scale output of the Kth encoder being spliced; 2.2.2) Feature extraction and feature transfer The Kth splicing result is directly input into the dual attention defogging module DAD in the K+1th scale layer for processing; the processing result of the dual attention defogging module DAD in the K+1th scale layer is upsampled and scaled twice before feature transfer. The feature transfer is specifically performed by attention mechanism processing with the processing result of the first dual attention defogging module DAD in the Kth scale layer; When n is not equal to 1, the nth splicing result is directly input into the first dual attention defogging module DAD of the n+1th scale layer for processing; the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is upsampled and scaled twice before feature transfer, and the feature transfer is specifically performed with the processing result of the mth dual attention defogging module DAD in the n-1th scale layer. At the same time, the attention mechanism result of the mth dual attention defogging module DAD in the nth scale layer is spliced ​​with the attention mechanism results of the mth dual attention defogging module DAD in the second scale layer to the n-1th scale layer, and then input into the corresponding m+1th dual attention defogging module DAD in the second scale layer to the n-1th scale layer; When n is equal to 1, the output of the first encoder with the same scale is directly input into the first dual attention defogging module DAD of the first scale layer for processing; the attention mechanism result of the mth dual attention defogging module DAD in the first scale layer is directly input into the m+1th dual attention defogging module DAD in the first scale layer; The output of the last dual attention defogging module DAD in the nth scale layer is input to the nth decoder for processing; the output of the dual attention defogging module DAD in the K+1th scale layer is input to the Kth decoder for processing; where m is an integer parameter, m=1, 2, 3...K+1-n; 2.2.3) Feature decoding The output of the last dual attention defogging module DAD in the K+1th scale layer and the Kth scale layer is used as the input of the Kth decoder, and is input into the K-1th decoder after feature decoding processing by the Kth decoder; except for the Kth decoder, the input of other nth decoders is the output of the last dual attention defogging module DAD in the nth scale layer and the output of the n+1th decoder; finally, the 1st decoder performs feature decoding on the input and outputs the high-order image feature F; The high-order image feature F is mapped through the second 3×3 convolutional layer and output as a synthetic dehazed remote sensing image. Each synthetic dehazed remote sensing image constitutes a synthetic dehazed remote sensing image dataset.

3. The method for removing non-uniform fog from remote sensing images according to claim 2, characterized in that: The dual attention defogging module DAD includes a channel attention branch, a spatial attention branch, and a first 1×1 convolutional layer; the input of the dual attention defogging module DAD is respectively input into the attention branch and the spatial attention branch for processing, the output of the attention branch and the output of the spatial attention branch are spliced, and the splicing result is then processed by the attention mechanism with the input of the dual attention defogging module DAD after passing through the first 1×1 convolutional layer, and the attention mechanism result is used as the output of the dual attention defogging module DAD; The channel attention branch includes a third 3×3 convolutional layer, a fourth 3×3 convolutional layer, a maximum pooling layer, a second 1×1 convolutional layer, and a third 1×1 convolutional layer; the input of the channel attention branch is sequentially input into the third 3×3 convolutional layer, the fourth 3×3 convolutional layer, the maximum pooling layer, the second 1×1 convolutional layer, and the third 1×1 convolutional layer for processing, the output of the third 1×1 convolutional layer is multiplied by the output of the fourth 3×3 convolutional layer, the multiplication result is then added to the input of the channel attention branch, and the addition result is used as the output of the channel attention branch; The spatial attention branch includes a fifth 3×3 convolutional layer, a sixth 3×3 convolutional layer, a seventh 3×3 convolutional layer, an eighth 3×3 convolutional layer, and a ninth 3×3 convolutional layer; the input of the spatial attention branch is sequentially input into the fifth 3×3 convolutional layer, the sixth 3×3 convolutional layer, the seventh 3×3 convolutional layer, the eighth 3×3 convolutional layer, and the ninth 3×3 convolutional layer for processing, the output of the ninth 3×3 convolutional layer is multiplied by the output of the sixth 3×3 convolutional layer, the multiplication result is then added to the input of the spatial attention branch, and the addition result is used as the output of the spatial attention branch.

4. The method for removing non-uniform fog from remote sensing images according to claim 1, characterized in that: In the step 2.3), for each synthetic defogging remote sensing image in the synthetic defogging remote sensing image dataset, the following operations are performed: The loss value L between the synthetic dehazed remote sensing image and the haze-free label image is calculated as shown below: L=L pixel +0.04×L context Among them, L pixel Represents pixel-level loss, which is a smooth L1 loss function; L context Represents content loss, which is a perceptual loss function.

5. The method for removing non-uniform fog from remote sensing images according to claim 1, characterized in that: In the step 2.4), all loss values ​​L are simultaneously transmitted back to the remote sensing image defogging network for optimization using the ADAM optimizer to obtain an optimized remote sensing image defogging network.

Images