Unmanned aerial vehicle aerial image defogging method based on improved AOD-Net

Abstract

The invention provides an unmanned aerial vehicle aerial image defogging method based on improved AOD-Net, and aims to improve the defogging effect of an unmanned aerial vehicle aerial image in a haze environment. According to the method, a Pseudo-Huber loss function and total variation regularization are combined to construct a PH-TV composite loss function, and a traditional L2 loss function is improved; the AOD-Net is inserted into a space channel interaction cooperation SCSIA module, multi-semantic-level cooperation of space and channels is achieved, and extraction of the model on the feature dependency relation and space structure of unmanned aerial vehicle aerial images is optimized; a multi-scale expansion fusion module is added, different expansion rates are set, normalization processing is carried out, and semantic information of aerial images of the unmanned aerial vehicle under different scales is captured. The method can effectively recover the color and structure details of the aerial image of the unmanned aerial vehicle under the foggy weather condition, and has a wide application prospect in the field of aerial photography of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and particularly to a method for removing haze from UAV aerial images based on an improved AOD-Net. Background Art

[0002] With the continuous progress of UAV technology, its application fields have expanded rapidly, covering multiple key areas such as intelligent agriculture, large-scale scene monitoring, disaster rescue, and intelligent traffic management. Especially when performing complex environment monitoring and large-scale dynamic tasks, UAVs, with their high mobility, flexibility, and wide observation range, can capture key scene data in real time, providing indispensable technical support for tasks such as target detection and scene analysis. However, during actual flight, UAVs are often interfered by complex weather conditions, such as haze, rain, and snow. The water vapor or particulate matter in these weather phenomena will seriously affect the image quality. Specifically, it is manifested as a decrease in image contrast, blurred details, color distortion, etc. This not only masks key visual features but also may introduce useless information such as uneven brightness and background artifacts, further affecting subsequent visual processing tasks. Especially when detecting small targets and complex background scenes, the degradation of image quality is likely to lead to missed detections or false detections, severely weakening the reliability and accuracy of the detection algorithm.

[0003] When a UAV is performing a task, the dynamic changes in its flight altitude and viewing angle further exacerbate the impact of atmospheric scattering on image quality. To solve the above problems, image dehazing technology provides an effective way to improve the quality of UAV aerial images. This technology can restore clear haze-free images by combining physical models and data-driven methods, effectively restoring the true texture and color of the scene, providing high-quality input data for target detection tasks, and thereby enhancing the robustness and stability of the detection algorithm. Summary of the Invention

[0004] The present invention proposes a method for removing haze from UAV aerial images based on an improved AOD-Net, aiming to solve problems such as poor quality of UAV aerial images in foggy conditions affecting subsequent visual processing tasks. The present invention provides a method for removing haze from UAV aerial images based on an improved AOD-Net, and the specific steps are as follows:

[0005] Step 1: Based on the atmospheric scattering model, use a physical simulation method to perform image fogging processing on the VisDrone2019 public dataset to generate a dataset containing foggy images and their corresponding clear labels;

[0006] Step 2: Replace the L2 loss function of AOD-Net with a PH-TV composite loss function that combines the Pseudo-Huber loss function and total variation regularization, and adjust the corresponding weight parameters to better restore the clarity of foggy aerial images and reduce model errors;

[0007] Step 3: To address the problems of insufficient edge restoration ability, limited fog structure perception ability, and lack of channel feature compression in the original AOD-Net model, a Spatial Channel Interaction and Synergy SCSIA module is inserted after Concat2 to fuse the semantic information of the UAV aerial images extracted by the front-end network layer of the AOD-Net model;

[0008] Step 4: To solve the redundant calculations generated during the extraction of shallow features, different dilation rates are set for Conv1, Concat1, and Concat2 of the AOD-Net model and normalized, and a multi-scale dilated fusion module is formed to jump connect to Concat3 to capture the semantic information of UAV aerial images at different scales;

[0009] Step 5: Use the synthesized foggy images as training data, input them into the improved AOD-Net model for training, and optimize the AOD-Net parameters according to the PH-TV function; after training, use the model to perform defogging tests on the foggy image test set to evaluate the image restoration quality.

[0010] In the above step 1, an atmospheric scattering model is used to synthesize the UAV aerial image dataset required for model training, effectively simulating the images obtained by UAVs under foggy conditions. Based on a simplified physical model of atmospheric scattering, it describes the phenomenon of energy attenuation in the propagation of light in the air in nature. The medium in the atmosphere absorbs the energy of light and converts it into other energy, and at the same time, due to the uneven distribution of the transmission path, the light changes direction during propagation, resulting in the attenuation of light energy. The model can be expressed by the following formula:

[0011] I(x) = J(x)t(x) + A(1 - t(x))

[0012] where x represents the image pixel coordinates, I(x) represents the foggy image obtained by the camera, J(x) represents the clear fog-free image to be restored, A represents the atmospheric light value, and t(x) is the transmittance, indicating the proportion of direct energy retained after the light passes through the atmospheric medium. In a uniform atmosphere, t(x) can be expressed as:

[0013] t(x) = e -βd(x)

[0014] where β is the atmospheric scattering coefficient, and d(x) is the scene depth, that is, the distance from the camera to the object being photographed. According to the model, the parameters such as the atmospheric light intensity, fog concentration coefficient, and azimuth are randomly adjusted to simulate the foggy images obtained by UAVs in the real environment.

[0015] In step 2, the L2 loss function is sensitive to the texture details and brightness changes of the image. Especially in the smooth areas of the image, it is prone to over-smoothing or detail loss. Therefore, combining the Pseudo-Huber loss and the TV loss, the composite loss function is beneficial for training and can also provide better visual image quality. The definition of the Pseudo-Huber loss function:

[0016]

[0017] where x is the predicted value, y is the true value, and δ is an adjustable parameter used to control the curvature of the curve. The total variation regularization TV loss function is defined as:

[0018]

[0019] where x i,j represents the pixel value of the pixel at position (i, j). The expression of the composite function is as follows:

[0020] L PHTV (x, y) = αL Pseudo-Huber (δ) + βL TV (x)

[0021] where α and β are weight coefficients used to balance the contributions of the two losses.

[0022] In step 3, the AOD-Net model is improved to enhance the defogging performance of UAV aerial images. Most deep learning-based defogging networks are based on the atmospheric scattering model and estimate the light intensity and transmittance of the atmosphere to remove fog. According to the atmospheric scattering physical model, the restored clear image J(x) can be easily obtained:

[0023]

[0024] During the process of the defogging network model estimating the illumination value and the refractive index respectively, the problem of mutual amplification of estimation errors occurs. The core idea of AOD-Net (All-in-One Dehazing Network) is to integrate the two key parameters of the transmittance t(x) and the atmospheric illumination value A into a unified K(x) expression, and achieve defogging by directly minimizing the reconstruction error in the image pixel domain. AOD-Net is as Figure 1 shown. The K(x) expression and the expression of the fog-free image output are as follows:

[0025] J(x) = K(x)I(x) - K(x) + b

[0026]

[0027] where b is a constant bias with a default value of 1.

[0028] In step 3, the spatial channel interaction and collaborative SCSIA module is inserted after the Concat2 layer of the AOD-Net model to integrate the semantic information of the drone aerial image from the first few network layers of AOD-Net. SCSIA consists of two parts: Shared Multi-Semantic Spatial Attention (SMSA) and Progressive Channel-by-Channel Self-Attention (PCSA). First, Concat2 is mean pooled to obtain the horizontal X h and vertical direction X w The global features of UAV aerial images; secondly, by h and X w Apply convolution operations separately and then concatenate to generate the spatial attention weight map M s , M s Weighted with the input feature map cat2 to obtain Concat2 spatial prior information M s Can guide PCSA to fine-tune the channel, using Downsampling is performed, and then query Q, key K, value V operations are performed to calculate Dependencies between channels, calculating the channel-level weight A c and Weighted to finally get the optimized feature map It can retain the key information in the aerial foggy image. Feature map after spatial and channel attention optimization It can retain the key information in the input image and further process it in the subsequent network.

[0029] In step 4, to address the redundant computations generated during shallow feature extraction, the AOD-Net model is improved by adding a multi-scale dilation fusion module. Initially, the input image is subjected to preliminary feature extraction through the basic convolutional layer, and features are gradually extracted from the underlying texture to the mid-level semantics through multi-layer convolution operations. A multi-scale dilation fusion module is constructed by performing dilated convolution operations on the three feature maps x1, cat1, and cat2, using convolution kernels with dilation rates of 1, 2, and 3, respectively. The three process feature maps are then compressed through 1x1 convolution. By applying dilated convolution to the process feature maps at different dilation rates, basic features in drone aerial images can be gradually extracted, mid-scale spatial information and boundary regions can be further mined, and semantic associations in larger regions can be effectively captured. Through this fusion process, the network can effectively combine feature information at different levels, thereby improving image restoration and resolving issues caused by changes in target scale.

[0030] Based on the above technical solution, the detail restoration ability and color reproduction of the image can be significantly improved. Compared with the traditional defogging method, the beneficial effects of the present invention are as follows:

[0031] By introducing a composite loss function and multi-scale feature fusion, the present invention can more effectively restore the detailed parts of the image and avoid the problem of excessive image smoothing in traditional methods. The spatial-channel interaction and collaboration module enables the network to pay more attention to the key regions in the image, especially in areas with severe haze, further optimizing the defogging performance. The multi-scale dilation fusion module expands the receptive field of the network through dilated convolution operations, improves the accuracy of image restoration, and makes the visual effect of the defogged image closer to the real image. Brief Description of the Drawings

[0032] Figure 1 is the image defogging process;

[0033] Figure 2 is the AOD-Net network diagram;

[0034] Figure 3 is the structure diagram of the K(x) module in the improved AOD-Net of the present invention;

[0035] Figure 4 is the schematic diagram of the spatial-channel interaction and collaboration SCSIA module in the present invention. Detailed Embodiment

[0036] To better illustrate the present invention, the following elaborates on the implementation steps of the present invention in combination with examples. This implementation scheme is only one application mode of the present invention and does not limit the protection scope of the present invention. The present invention proposes a method for defogging UAV aerial images based on an improved AOD-Net, which improves the defogging effect of UAV aerial images in a haze environment by improving the loss function, introducing an attention mechanism, and a multi-scale dilation fusion module. The brief defogging process is as Figure 1 shown, and the specific steps are as follows:

[0037] Step 1: Based on the atmospheric scattering model, use a physical simulation method to perform image fogging processing on the VisDrone2019 public dataset to generate a dataset containing foggy images and their corresponding clear labels;

[0038] Step 2: Replace the L2 loss function of AOD-Net with a PH-TV composite loss function combining the Pseudo-Huber loss function and total variation regularization, and adjust the corresponding weight parameters to better restore the clarity of aerial foggy images and reduce model errors;

[0039] Step 3: To address the problems of insufficient edge restoration ability, limited fog structure perception ability, and lack of channel feature compression in the original AOD-Net model, a Spatial Channel Interaction and Synergy SCSIA module is inserted after Concat2 to fuse the semantic information of the UAV aerial images extracted by the front-end network layer of the AOD-Net model;

[0040] Step 4: To solve the redundant calculations generated during the extraction of shallow features, different dilation rates are set for Conv1, Concat1, and Concat2 of the AOD-Net model and normalized, and a multi-scale dilated fusion module is connected to Concat3 through skip connections to capture the semantic information of UAV aerial images at different scales;

[0041] Step 5: Use the synthesized foggy images as training data, input them into the improved AOD-Net model for training, and optimize the AOD-Net parameters according to the PH-TV function; after training, use the model to perform defogging tests on the foggy image test set to evaluate the image restoration quality.

[0042] In the said Step 1, first, the VisDrone2019 UAV aerial public dataset is collected. This dataset contains UAV aerial images from different cities, environments, and lighting conditions, covering various object categories (such as pedestrians, vehicles, bicycles, etc.) and scene densities (including sparse and crowded scenes). This dataset has a total of 10,209 images, covering various climate, lighting, and weather change conditions. To generate the foggy environment images required for training, the atmospheric scattering model is used to perform image synthesis processing on this dataset. This model simulates the propagation process of light in the atmosphere in nature, considering the attenuation phenomenon of light when passing through the atmosphere. By introducing a simplified physical model of atmospheric scattering, images simulating UAVs taken in foggy weather can be effectively generated. This model can consider the scattering and absorption effects of fog particles on light, thereby generating a highly realistic foggy image dataset for training the defogging algorithm. This synthesis process includes processing images under different weather conditions, such as foggy, cloudy, and other environmental conditions, to ensure that the model can adapt to diverse foggy scenes during the training process. The atmospheric scattering model can be expressed by the following formula:

[0043] I(x) = J(x)t(x) + A(1 - t(x))

[0044] where x represents the image pixel coordinates, I(x) is the foggy image captured by the camera, J(x) represents the clear fog-free image to be restored, A represents the atmospheric light value, and t(x) is the transmittance, representing the proportion of direct energy retained after light passes through the atmospheric medium. In a uniform atmosphere, t(x) can be expressed as:

[0045] t(x) = e-βd(x)

[0046] Among them, β is the atmospheric scattering coefficient, and d(x) is the scene depth, that is, the distance from the camera to the object being photographed. Based on the atmospheric scattering model, the atmospheric light intensity A is randomly set within the range of [0.5, 0.6], and the fog concentration coefficient (representing the density of haze) is randomly taken within the range of [0.001, 0.150]. Finally, the above model is used to process the VisDrone2019 dataset to generate UAV aerial images with different fog effects for training the defogging algorithm.

[0047] In step 2, the traditional L2 loss function is sensitive to the texture details and brightness changes of the image. Especially in the smooth areas of the image, it is easy to cause over-smoothing or loss of details. To improve this problem, the present invention combines the Pseudo-Huber loss function and the total variation (TV) loss function to construct a composite loss function. This composite loss function can not only effectively improve the training effect but also bring a significant improvement in visual quality. The definition of the Pseudo-Huber loss function is:

[0048]

[0049] Among them, x is the predicted value, y is the true value, and δ is an adjustable parameter used to control the curvature of the loss function curve. The TV loss function is defined as:

[0050]

[0051] where x i,j represents the pixel value of the pixel at position (i, j). The expression of the composite function is as follows:

[0052] L(x, y) = αL Pseudo-Huber (δ) + βL TV (x)

[0053] where α and β are weight coefficients used to balance the contributions of the two losses. By adjusting these two parameters, the performance can be optimized according to specific application requirements. Adjusting the α parameter is beneficial to minimizing the overall error between the prediction and the actual image, especially in the edge regions. Carefully adjust the β parameter to ensure that the TV loss does not overly suppress the important textures and structures in the image. By adjusting these two parameters, the performance can be optimized according to specific application requirements. Initialize the weight parameters of the loss function α = 0.2 and β = 0.8. During the model training process, the loss function is enabled in stages: in the first stage, only the Pseudo-Huber loss function is enabled to quickly restore the overall contrast and brightness of the image; in the second stage, joint optimization is performed, and the composite function L PHTV(x, y), add gradient clipping (threshold 1e-3), adjust the weights according to the SSIM index of the validation set every 50 epochs. If the SSIM improvement is <0.5%, increase α by 0.05.

[0054] In step 3, the AOD-Net model is improved to enhance the defogging performance of UAV aerial images. Most deep learning-based defogging networks are based on the atmospheric scattering model and estimate the light intensity and transmittance of the atmosphere to remove fog. According to the atmospheric scattering physical model, the restored clear image J(x) can be easily obtained:

[0055]

[0056] During the process of the defogging network model estimating the illumination value and refractive index respectively, the problem of mutual amplification of estimation errors occurs. The core idea of AOD-Net (All-in-One Dehazing Network) is to integrate the two key parameters, the transmittance t(x) and the atmospheric light value A, into a unified K(x) expression. The model structure is as Figure 1 shown. It realizes defogging by directly minimizing the reconstruction error in the image pixel domain. Through this parameterization method, the network can directly learn the non-linear mapping relationship between K(x) and the clear image J(x), avoiding the error propagation path of multi-parameter cascade estimation in the traditional model. The K(x) estimation module design includes five convolutional layers. The first three layers gradually fuse feature maps at different levels through cascade operations (concat1-concat3), and finally output K(x) through a non-linear regression layer. This design not only reduces the model complexity but also significantly improves the robustness in non-uniform haze scenes by directly minimizing the difference between the defogged image and the real clear image. The K(x) expression and the expression for outputting the fog-free image are as follows:

[0057] J(x) = K(x)I(x) - K(x) + b

[0058]

[0059] Among them, b is a constant bias with a default value of 1.

[0060] In step 3, insert the Spatial Channel Interaction and Synergy SCSIA module after the Concat2 layer of the AOD-Net model, as Figure 2 shown, to fuse the semantic information of the UAV aerial image from the previous several network layers of the model. As Figure 3 shown, SCSIA consists of two parts: shareable multi-semantic spatial attention (SMSA) and progressive per-channel self-attention (PCSA). SMSA performs average pooling (AvgPool) on Concat2 to obtain X in the horizontal direction h and X in the vertical directionw Features:

[0061]

[0062] Enhance the extracted spatial information through convolutions at multiple scales, thereby generating the spatial feature M s Can effectively guide the channel attention (PCSA) to focus on the key regions in the image. Fog usually makes the images of distant regions blurred, while the regions closer to the camera are relatively clearer. SMSA (Spatial Multi-Semantic Attention) can enhance the defogging effect and guide PCSA to perform fine calibration of channels by focusing on these regions with significant features. M s Weight with the input feature map cat2 to obtain spatial prior information:

[0063] M s = σ(Conv 1×1 (f local (X h ) + f global (X w ))) ∈ R 1×H×W

[0064]

[0065] PCSA suppresses redundant information by enhancing the semantic dependencies between channels. After downsampling the input feature map , the network can focus on the global information of the image. The network calculates the dependencies between channels through operations of query, key, and value, and uses the attention formula to calculate the weighted relationships between channels, thereby achieving the enhancement of important features and the suppression of redundant features:

[0066]

[0067] where d is the scaling factor that controls the scale of matrix calculation. Calculate the average of the attention weights for each channel, and obtain the final channel-level attention weights by summing over the spatial dimensions (height and width). Use the calculated channel attention weights A c and to perform weighting:

[0068]

[0069] The feature map after spatial and channel attention optimization Can retain the key information in the input image and be further processed in the subsequent network.

[0070] In step 4, since the distance between the drone and the target changes during high-altitude shooting, it will directly affect the scale representation of the target in the image, thus affecting the overall coverage of the aerial image. To solve this problem, this paper introduces a multi-scale dilated convolution feature fusion module to enhance the expressive power and visual quality of the image. During the stage of extracting initial features in the basic convolutional layer, the input image x ∈ R B×3×H×W , Conv1, Concat1, and Concat2 gradually extract features from the underlying texture to the middle-level semantics:

[0071] x1 = ReLU(Conv1(x)) ∈ R B×3×H×W

[0072] x2 = ReLU(Conv2(x1))

[0073] cat1 = Concat(x1, x2) ∈ R B×6×H×W

[0074] x3 = ReLU(Conv3(cat1))

[0075] cat2 = Concat(x2, x3) ∈ R B×6×H×W

[0076] To expand the receptive field and enhance the context information, dilated convolution operations are performed on the three feature maps x1, cat1, and cat2. Convolution kernels with dilation rates of 1, 2, and 3 are used to obtain richer context information, and these features are concatenated and then compressed through a 1x1 convolution to obtain the fused features:

[0077] x1_DW = ReLU(DWConv(x1, r = 1))

[0078] cat1_DW = ReLU(DWConv(x1,, r = 2))

[0079] cat2_DW = ReLU(DWConv(x1, r = 3))

[0080] f dilated = Concat(x1_DW, cat1_DW, cat2_DW) ∈ R B×9×H×W

[0081] f fused = Conv 1×1 (f dilated ) ∈ R B×3×H×W

[0082] After introducing the feature fusion sub-module, Conv1, Conv2, Conv3, Conv4 are concatenated with the feature layer after feature fusion to obtain the final feature set cat3:

[0083] x4 = ReLU(Conv4(X ECPA )

[0084] cat3 = Concat(x1,x2,x3,x4,f fused ) ∈ R B×15×H×W

[0085] Finally, through convolutional operations on the fused feature set, an intermediate estimation function K(x) is obtained for image restoration and to enhance the network's ability to model image degradation information.

[0086] In step 5, synthetic hazy images are used as training data and input into the improved AOD-Net model for training, thereby optimizing the model parameters and improving the defogging effect. After training is completed, the model is used to perform inference processing on the test set images to obtain the improved defogging results. The specific operation steps and technical solutions are as follows:

[0087] (1) First, the training data loads the original images and hazy images from local through a data loader. The images will be adjusted to 480×640 pixels, and the image channel order is adjusted to C×H×W to adapt to the input format of the deep learning model. The image pixel values are normalized by dividing by 255 to ensure that the pixel values are in the range of [0,1], thereby avoiding excessive numerical fluctuations during gradient calculation. The dataset is divided into a training set and a validation set in a ratio of 8:1:1, and random allocation is used to ensure data diversity and avoid negative impacts of data bias on the training effect.

[0088] (2) In each training epoch, the image batches are fed into the model for forward propagation, the loss is calculated, and the model parameters are optimized through backpropagation. The optimizer used during training is the Adam optimizer, and the initial learning rate is set to 1×10 -4 , which is used to optimize all parameters of the network. To avoid over-adjusting the learning rate, the gradient clipping threshold is set to 1×10 -3 , and the batch size is set to 64. In the initial stage, only the Pseudo-Huber loss function is used for training to quickly restore the brightness and contrast of the image; a composite loss function is introduced for joint optimization, and the details of the image are restored by adjusting the SSIM metric. If the SSIM improvement on the validation set is less than 0.5%, the α parameter is gradually increased.

[0089] (3) After each training epoch, the model will be evaluated on the validation set, and the mean squared error (MSE) is used to calculate the difference between the dehazed image and the original image. After every 50 training rounds, the performance of the model is evaluated according to the structural similarity index (SSIM) of the validation set. If the increase in the SSIM value is less than 0.5%, the learning rate parameter is adjusted.

[0090]

[0091] Among them, X(i,j) and Y(i,j) represent the pixel values of the dehazed image and the original image at the position (i,j) respectively, and H and W are the height and width of the image. By calculating the MSE, the performance of the model in the dehazing task is evaluated.

[0092] In addition, after every 50 training rounds, the structural similarity index (SSIM) is used to further evaluate the dehazing effect of the image to ensure the consistency of the brightness, contrast, and structural information between the restored image and the original image. The formula for SSIM is as follows:

[0093]

[0094] SSIM(x,y) = I(x,y)·c(x,y)·s(x,y)

[0095] Among them, μ represents the mean value of the pixel group, σ represents the variance of the pixel group, and c1, c2, and c3 are all constants. SSIM evaluates the image similarity through brightness, contrast, and structure, and the larger the value, the closer the quality is to the reference image. During the training process, if the increase in the SSIM value is less than 0.5%, the learning rate is adjusted. Through appropriate learning rate adjustment, the performance of the model on the validation set is further improved to ensure that the model can be continuously optimized within multiple training epochs.

[0096] After the training is completed, the improved AOD-Net model can effectively improve the clarity of the image and restore more details and realism. Through the inference process, the model can effectively remove haze and restore the natural color and structural details of the image.

Claims

1. A method for removing haze from UAV aerial images based on improved AOD-Net, characterized in that, The method includes the following steps: Step 1: Based on the atmospheric scattering model, use the physical simulation method to perform image fogging processing on the VisDrone2019 public dataset to generate a dataset containing foggy images and their corresponding clear labels; Step 2: Replace the L2 loss function of AOD-Net with the PH-TV composite loss function combining the Pseudo-Huber loss function and the total variation regularization, and adjust the corresponding weight parameters to restore the clarity of aerial foggy images; Step 3: Insert the Spatial Channel Interaction and Synergy SCSIA module after the Concat2 layer of the AOD-Net model; Step 4: Set different dilation rates for Conv1, Concat1, and Concat2 of the AOD-Net model and perform normalization to form a multi-scale dilated fusion module and jump connect to Concat3; Step 5: Use the synthesized foggy images as training data, input them into the improved AOD-Net model for training, and optimize the AOD-Net parameters according to the PH-TV function; After training is completed, use the model to perform defogging tests on the foggy image test set to evaluate the image restoration quality.

2. The method for removing haze from UAV aerial images based on the improved AOD-Net according to claim 1, characterized in that: In the said Step 1, perform image fogging processing and label preprocessing on the VisDrone2019 public dataset based on the atmospheric scattering physical model; the simplified atmospheric scattering physical model can be expressed by the following formula: I(x) = J(x)t(x) + A(1 - t(x)) where x represents the image pixel coordinates, I(x) represents the foggy image obtained by camera shooting, J(x) represents the clear fog-free image to be restored, A represents the atmospheric light value, and t(x) is the transmittance, indicating the proportion of direct energy retained after light passes through the atmospheric medium; in a uniform atmosphere, t(x) can be expressed as: t(x) = e -βd(x) 0where β is the atmospheric scattering coefficient and d(x) is the scene depth; adjust the random values of the atmospheric light intensity and fog concentration coefficient parameters according to the model to simulate the foggy images obtained by drones in the real environment.

3. The method for dehazing UAV aerial images based on the improved AOD-Net according to claim 1, characterized in that: In the said Step 2, combine the Pseudo-Huber loss and the TV loss to construct a composite loss function. This composite loss function improves the training efficiency and also brings a significant improvement in visual quality; the definition of the Pseudo-Huber loss function is: where x is the predicted value, y is the true value, and δ is an adjustable parameter used to control the curvature of the loss function curve; the TV loss function is defined as: where x i,j represents the pixel value of the pixel at position (i, j), and the composite function expression is as follows: L(x,y) = αL Pseudo-Huber (δ) + βL TV (x) where α and β are weight coefficients used to balance the contributions of the two losses.

4. The method for removing haze from UAV aerial images based on the improved AOD-Net according to claim 1, wherein: In step 3, a Spatial Channel Interaction and Synergy Module (SCSIA) is inserted after Concat2. The module consists of two parts: a Multi-Semantic Module (SMSA) and a Progressive Channel-by-Channel Module (PCSA) for processing. First, average pooling is performed on Concat2 to obtain the global features of the UAV aerial images in the horizontal direction \(X\) h and the vertical direction \(X\) w . Secondly, after applying convolution operations to \(X\) h and \(X\) w respectively and then splicing them, a spatial attention weight map \(M\) s is generated. \(M\) s is weighted with the input feature map cat2 to obtain the spatial prior information of Concat2 \(M\) s can guide PCSA to perform fine calibration of channels. Downsampling is performed on , and then query \(Q\), key \(K\), and value \(V\) operations are performed to calculate the dependencies between channels, and the channel-level weight \(A\) c is calculated. Weighting is performed with to finally obtain the optimized feature map which can retain the key information in the aerial foggy images.

5. The method for removing haze from UAV aerial images based on the improved AOD-Net according to claim 1, wherein: In the basic convolutional layer stage of step 4, the input image x ∈ R B×3×H×W , Conv1, Concat1, and Concat2 gradually extract features from the underlying texture to the middle-level semantics: x1 = ReLU(Conv1(x)) ∈ R B×3×H×W x2 = ReLU(Conv2(x1)) cat1 = Concat(x1, x2) ∈ R B×6×H×W x3 = ReLU(Conv3(cat1)) cat2 = Concat(x2, x3) ∈ R B×6×H×W Perform dilated convolution operations on the three feature maps of x1, cat1, and cat2 using convolution kernels with dilation rates of 1, 2, and 3 respectively; compress the three process feature maps through 1x1 convolution to form a multi-scale dilated fusion module: x1_DW = ReLU(DWConv(x1, r = 1)) cat1_DW = ReLU(DWConv(x1, r = 2)) It should be noted that there seems to be a missing comma in the formula in line . It should be "cat1_DW = ReLU(DWConv(x1, r = 2))". cat2_DW = ReLU(DWConv(x1, r = 3)) f dilated = Concat(x1_DW, cat1_DW, cat2_DW) ∈ R B×9×H×W f fused = Conv 1×1 (f dilated ) ∈ R B×3×H×W 。

Images