Deep learning aircraft skin damage detection method suitable for rain and fog conditions

Abstract

The invention discloses a deep learning aircraft skin damage detection method suitable for rain and fog conditions. The deep learning aircraft skin damage detection method comprises the following steps that an unmanned aerial vehicle obtains original image data; preprocessing the image data, and making a data set; inputting the image data of the training data set into the AOD-Net integrated defogging network; inputting the training data set into a CSPDarknet backbone network for feature extraction, and introducing an ECA channel attention mechanism; the method relates to the technical field of aircraft skin damage detection, the damage detection effect of the aircraft skin under the foggy condition can be effectively enhanced, meanwhile, local cross-channel interaction is captured in combination with an ECA attention mechanism, and the detection accuracy is improved. According to the method, the feature response of each channel can be adaptively adjusted, the channels related to the rain stripe features are enhanced, accurate damage detection under most rain and fog conditions can be ensured, and resources can be saved to the maximum extent and the detection cost can be reduced by relying on the lightweight features of AOD-Net and ECA.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft skin damage detection, and specifically to a deep learning aircraft skin damage detection method suitable for rain and fog conditions. Background Art

[0002] As a critical protective structure within the aircraft, the aircraft skin plays a crucial role in flight safety. It not only maintains the aircraft's aerodynamic shape and ensures aerodynamic performance during flight, but also protects the internal structure from environmental erosion, such as rain, fog, ultraviolet rays, and various pollutants. Therefore, ensuring the integrity and reliability of the aircraft skin is a core task in aviation safety.

[0003] However, in actual inspection scenarios, rain and fog frequently occur, posing a significant challenge to inspection work. Traditional aircraft skin inspection methods have exposed many drawbacks in rainy and foggy environments. Visual inspection relies on the visual observation of inspectors, which makes it difficult for manual work to continue. Although ultrasonic testing can detect internal damage, the equipment is complex to operate, requires high technical skills, and is slow, making it inefficient when inspecting large areas. Radiographic testing is costly, requires expensive equipment, and requires professional protective measures. It also poses a radiation hazard to the human body, making it unsuitable for frequent routine inspections.

[0004] The rise of deep learning technology in image detection has brought new hope for aircraft skin damage detection. However, existing deep learning algorithms still have significant shortcomings when applied to aircraft skin inspection in rain and fog. Specifically, rain and fog significantly reduce image contrast, reducing the grayscale difference between damaged and normal skin areas. This makes it difficult for deep learning algorithms that extract features based on grayscale differences to highlight damage features. At the same time, rain and fog are often accompanied by increased noise, which interferes with the algorithm's extraction of effective features, making it easy for the model to mistakenly identify noise as damage features or fail to accurately identify true damage features. In addition, aircraft skin damage types are diverse, including cracks, corrosion, pits, and debonding. Each type of damage manifests differently in rain and fog, increasing the difficulty of accurately classifying and locating damage. Moreover, the large area of aircraft skin requires detection algorithms to have efficient processing capabilities and be able to perform comprehensive inspections over large areas in a short period of time. This places extremely high demands on the algorithm's computational efficiency and resource utilization efficiency.

[0005] In summary, traditional aircraft skin inspection methods have many limitations under rain and fog conditions, and the existing deep learning algorithms still need to be further optimized and improved when applied to aircraft skin damage detection under rain and fog conditions to meet the strict requirements of the aviation field for aircraft skin safety inspection. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a deep learning aircraft skin damage detection method suitable for rain and fog conditions. The present invention aims to solve a series of difficult problems faced by aircraft skin damage detection under rain and fog conditions, namely, low detection accuracy, low efficiency and high cost. By introducing the AOD-Net integrated defogging network to preprocess the rain and fog images, and combining the ECA attention mechanism to improve the YOLOv8 network, an efficient and accurate aircraft skin damage detection method is proposed. The K-estimating module of AOD-Net is used to estimate the fog depth and relative level, generate a restored image, improve the image quality, and provide a clear data basis for damage detection. With the help of the ECA attention mechanism, channels related to rain streak features are enhanced, irrelevant channels are suppressed, and the detection capability of rain streaks and damage features is improved. At the same time, the algorithm is deployed on a drone, and the maneuverability of the drone is used for detection, which reduces manpower and material resources, reduces detection costs, and realizes efficient and accurate detection of aircraft skin under rain and fog conditions, thereby ensuring flight safety.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A deep learning aircraft skin damage detection method suitable for rain and fog conditions, comprising the following steps:

[0008] S1. The UAV flies around the aircraft to obtain original image data of aircraft skin damage;

[0009] S2. Preprocess the collected image data, complete screening and labeling, and create the aircraft skin damage dataset required for training the network;

[0010] S3. Input the training dataset image data into the AOD-Net integrated defogging network. Through the multi-scale feature fusion, fog depth and relative level estimation in the K-estimating module of the network, combined with the atmospheric scattering model and residual connection optimization of the defogging image generation module, a clear restored image is generated.

[0011] S4. Input the training data set processed by the AOD-Net integrated defogging network into the CSPDarknet backbone network of yolov8 for feature extraction;

[0012] S5. The target features extracted by the CSPDarknet backbone network are introduced into the ECA channel attention mechanism. Through global average pooling, one-dimensional convolution, and sigmoid weight generation, cross-channel interaction information is captured to enhance the channel features related to rain streaks and damage.

[0013] S6. Using the aircraft skin damage features obtained after the above processing, a target detection network is trained to obtain a target detector;

[0014] S7. Deploy the trained target detector onto the UAV, and use the UAV to inspect the aircraft skin to obtain target detection results and the target's appearance features.

[0015] Preferably, in step S1, data augmentation is used to expand the data, and a trimming operation is performed on the collected image or another part of data is obtained from an existing database, wherein the trimming operation includes rotation, flipping, and cropping.

[0016] Preferably, the pre-processing in step S2 includes screening out clear and usable data and eliminating fuzzy, damaged or non-compliant data;

[0017] Use professional annotation tools to annotate the filtered images. The annotation content includes the location information and identification number of the damaged target. During annotation, it is necessary to ensure the balance of data in each category, so as to produce the aircraft skin damage dataset required for training the network.

[0018] Preferably, the specific operation steps of step S3 include:

[0019] S3.1. In the K-estimating module, a set of initial convolutional layers is used to process the input image. Feature information, including texture, edges, and low-level features related to fog, is initially extracted from the image to obtain a preliminary feature map. Convolution operations are performed on the preliminary feature map at different scales using convolution kernels of different sizes or dilated convolution. Dilated convolution is also used to increase the receptive field and obtain more global fog feature information. These feature maps of different scales are then concatenated and fused into a multi-scale feature map containing multi-scale information.

[0020] S3.2. Based on the fused multi-scale feature map, a fully connected layer is used to perform regression calculations. The fully connected layer integrates the information in the multi-scale feature map and outputs an estimate of the fog depth and relative level based on pre-trained parameters.

[0021] S3.3. The fog depth and relative level estimation values obtained by the K-estimating module are input into the defogging image generation module as adaptive parameters. Based on the preliminary processing based on the physical model, the defogging image generation module further uses a convolutional neural network to optimize the image. Through multi-layer convolution operations, the feature extraction and reconstruction of the image after preliminary defogging are performed. At the same time, residual connections are used to ensure the information integrity of the image during the processing process to avoid image distortion caused by excessive defogging.

[0022] Preferably, the CSPDarknet backbone network in step S4 adopts a combination structure of multiple convolutional layers and residual blocks, wherein the convolutional layer adopts a 3x3 or 5x5 convolution kernel with a step size of 1 or 2, and the residual block consists of two convolutional layers and a jump connection. The input features are added to the convolutional layer output features through the jump connection, so that the network can learn deeper features.

[0023] Preferably, the specific operation steps of step S5 include:

[0024] S5.1. Input the feature map extracted from the CSPDarknet backbone network into the ECA attention mechanism module and perform a global average pooling operation. The global average pooling operation compresses the two-dimensional feature map of each channel in the spatial dimension to obtain a scalar value. The scalar value corresponding to each channel forms a one-dimensional vector. The one-dimensional vector contains the global information of the entire feature map in each channel.

[0025] S5.2. The one-dimensional vector enters a one-dimensional convolutional layer, and the convolution kernel size of the one-dimensional convolutional layer is automatically adjusted according to the number of channels of the feature map to capture local cross-channel interaction information and mine the correlation between different channels;

[0026] S5.3. Process the output of the one-dimensional convolutional layer through a Sigmoid activation function, mapping the output value to a range between 0 and 1. The output value is the attention weight corresponding to each channel, and the attention weight is used to indicate the importance of each channel in the feature representation;

[0027] The closer the attention weight is to 1, the more important the channel is to the subsequent detection task;

[0028] The closer the attention weight is to 0, the less important the channel is;

[0029] S5.4. Multiply the generated attention weights by the original feature map element by element, so that the channels related to the rain streak features are enhanced and their feature information is highlighted, while the features of irrelevant channels are suppressed, thereby improving the detection capability of rain streak features and aircraft skin damage features.

[0030] Preferably, in step S6, the loss function of the network model is continuously adjusted during the training process so that the trained network detection is more consistent with the data set.

[0031] Beneficial effects

[0032] The present invention provides a deep learning aircraft skin damage detection method suitable for rain and fog conditions, which has the following beneficial effects:

[0033] The present invention introduces the AOD-Net integrated defogging network into the YOLOv8 network, estimates the fog depth and relative level through the K-estimating module, and inputs the estimated parameters as adaptive parameters into the defogging image generation module to generate a restored image, effectively enhancing the damage detection effect of aircraft skin in foggy conditions. At the same time, the Efficient Channel Attention (ECA) attention mechanism is combined to capture local cross-channel interactions, enabling it to adaptively adjust the feature response of each channel, enhance channels related to rain streak features, and suppress irrelevant channels, thereby improving the detection capability of rain streak features. Deploying the algorithm on a drone can not only ensure accurate damage detection in most rain and fog conditions, but also achieve maximum resource conservation by relying on the lightweight features of AOD-Net and ECA, reducing the consumption of manpower and material resources and lowering the detection cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the process of the present invention.

[0035] Figure 2 This is a schematic diagram of the process of inputting training dataset image data into the AOD-Net integrated defogging network in the present invention.

[0036] Figure 3 Schematic diagram of the detection process of the target detector of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] Example: The core idea of the method of the present invention is to use a drone to collect aircraft skin damage data under rain and fog conditions, train a target detector, and use a combination of the target detector and the drone to detect aircraft skin damage. First, use a drone to fly around the aircraft to collect source data on aircraft skin damage under rain and fog conditions, and use annotation tools to process the collected data to create a dataset of aircraft skin damage under rain and fog conditions. By collecting aircraft skin image data under rain and fog conditions and training a target detector, the aircraft skin damage is detected by combining the target detector with the drone. In this process, in order to cope with the interference of rain and fog on aircraft skin damage detection, the AOD-Net integrated defogging network and the ECA attention mechanism are introduced to improve the YOLOv8 network.

[0039] First, the AOD-Net integrated defogging network plays a crucial role. Its K-estimating module estimates the depth and relative level of fog in an image. These estimated parameters are then fed as adaptive inputs into the defogging image generation module. Based on these parameters, the module processes the captured foggy aircraft skin image to generate a clear restored image, effectively improving image quality and highlighting aircraft skin details. This provides a high-quality data foundation for subsequent damage detection, making damage features more clearly discernible in the image.

[0040] Next, the ECA attention mechanism is introduced. This mechanism can keenly capture local cross-channel interaction information. When processing aircraft skin images, it can intelligently and adaptively adjust the feature response of each channel. For channels closely related to rain streak characteristics, the ECA mechanism enhances their feature expression, thereby strengthening the information of these key channels; while for channels unrelated to rain streak characteristics, it suppresses their contribution, reducing interference information. This significantly improves the network's ability to detect rain streak characteristics, enabling the entire detection system to more accurately identify aircraft skin damage in rain and fog environments.

[0041] Finally, the improved algorithm is deployed on the UAV. With its flexible maneuverability, the UAV flies around the aircraft in rain and fog conditions to collect aircraft skin image data. Thanks to the lightweight design of AOD-Net and ECA, the improved algorithm can not only ensure the accuracy of aircraft skin damage detection under most rain and fog conditions, but also save computing resources to the greatest extent, significantly reduce the investment of manpower and material resources, thereby effectively reducing the detection cost and meeting the strict requirements of the aviation field for efficient, accurate and low-cost aircraft skin damage detection. A detection method for aircraft skin damage under rain and fog conditions in this embodiment, such as Figure 1 As shown, the following steps are included:

[0042] Step 1: The UAV flies around the aircraft to obtain original image data of aircraft skin damage;

[0043] In this implementation, sufficient data is collected to avoid a monotonous background image of aircraft skin damage, imbalanced classifications, and to prevent model overfitting. Data augmentation is used to expand the data, performing operations such as rotation, flipping, and cropping on the collected images, or obtaining additional data from an existing database.

[0044] Step 2: Preprocess the collected image data, complete screening and labeling, and create the aircraft skin damage dataset required for training the network.

[0045] In this implementation, the collected image data is preprocessed, including filtering out clear, usable data and removing blurry, damaged, or non-compliant data. Professional annotation tools are used to annotate the filtered images, including the location information (bounding box coordinates) and identification number of the damaged targets. At the same time, the data is balanced across categories to avoid poor training results caused by imbalanced data. This creates the aircraft skin damage dataset required for network training.

[0046] Step 3: Input the training dataset image data into the AOD-Net integrated dehazing network.

[0047] In this embodiment, the training set data is input into the AOD-Net integrated defogging network, and a clear restored image is generated through the K-estimating module and the defogging image generation module in the network. The lightweight and image quality are improved through parameter estimation and adaptive parameter re-input, thereby improving the performance of subsequent detectors. Figure 2 shown.

[0048] Step 3.1: Generate and concatenate feature maps using a multi-scale feature fusion strategy. In the K-estimating module, a set of initial convolutional layers is used to process the input image, initially extracting feature information from the image, including texture, edges, and low-level features related to fog, to generate a preliminary feature map. By using convolution kernels of different sizes or dilated convolutions, the preliminary feature map is convolved at different scales. Dilated convolutions are then used to increase the receptive field and obtain more global fog feature information. These feature maps of different scales are then concatenated and fused into a feature map containing multi-scale information, enabling more accurate estimation of fog-related parameters.

[0049] Step 3.2: Estimating fog depth and relative level. Based on the fused multi-scale feature map, a fully connected layer performs regression. The fully connected layer integrates the information from the fused feature map and, based on pre-trained parameters, outputs estimates of fog depth and relative level. During training, the weights of the fully connected layer are continuously adjusted by comparing with a large amount of image data accurately annotated with fog parameters, resulting in increasingly accurate estimates. These estimates reflect information such as the density of fog in the image and its relative positional distribution within the image, providing key information for subsequent dehazing operations.

[0050] Step 3.3: Defogging image generation is performed based on the atmospheric scattering model. The fog depth and relative level estimation parameters obtained by the K-estimating module are input into the defogging image generation module as adaptive parameters to adjust the various processing intensities and methods in the defogging process to adapt to the specific conditions of fog in different images. Based on the preliminary processing based on the physical model, the defogging image generation module further uses a convolutional neural network to optimize the image. Through multi-layer convolution operations, the image after preliminary defogging is feature extracted and reconstructed. The convolution layer will adjust according to the local features of the image, enhance the detailed information of the image, and repair the texture and edges that may be lost in the defogging process. At the same time, the residual connection is used to ensure the information integrity of the image during the processing process to avoid image distortion caused by excessive defogging.

[0051] Step 4: Input the training data set processed by the AOD-Net integrated defogging network into the CSPDarknet backbone network of yolov8 for feature extraction.

[0052] In this implementation, data is input into the CSPDarknet network for feature extraction. The network employs a combination of multiple convolutional layers and residual blocks. The convolutional layers use 3x3 or 5x5 kernels with a step size of 1 or 2. The residual block consists of two convolutional layers and a skip connection. This skip connection adds the input features to the convolutional layer output features, effectively resolving the vanishing gradient problem caused by increased network depth and enabling the network to learn deeper features. During this process, the image resolution gradually decreases, the number of channels in the feature map gradually increases, and the network continuously extracts local and global features of the image, providing rich feature information for subsequent object detection.

[0053] Step 5: The target features extracted by the CSPDarknet backbone network are introduced into the ECA channel attention mechanism to enhance the rain streak features.

[0054] In this implementation, the target features extracted by CSPDarknet are introduced into the ECA channel attention mechanism to capture cross-channel information interactions. A global average pooling operation compresses the two-dimensional feature map of each channel into a scalar, yielding global channel feature information. The globally average pooled channel features are then processed using a one-dimensional convolution to capture local cross-channel interaction information. The output of the one-dimensional convolution is then passed through a sigmoid activation function to generate attention weights for each channel. Finally, the generated attention weights are element-wise multiplied with the original feature map, enhancing channels related to rain streak characteristics and suppressing irrelevant channels. This improves the detection of rain streak characteristics while also enhancing the recognition of aircraft skin damage features.

[0055] Step 5.1: Global Average Pooling to Obtain Global Channel Information. The feature maps extracted from the CSPDarknet backbone are fed into the ECA attention mechanism. First, a global average pooling operation is performed. This operation compresses the two-dimensional feature maps of each channel in the spatial dimension, averaging the feature maps along the height and width to obtain a scalar value. Each channel corresponds to a scalar representing its global feature information. These scalars form a one-dimensional vector that contains the global information of the entire feature map across each channel.

[0056] Step 5.2: One-dimensional convolution captures cross-channel interaction information. The one-dimensional vector obtained after global average pooling enters the one-dimensional convolution layer. The convolution kernel size of the one-dimensional convolution layer is automatically adjusted according to the number of channels in the feature map to capture local cross-channel interaction information. Through the one-dimensional convolution operation, the correlation between different channels is explored, and the channel combination information that is important for aircraft skin damage detection and rain streak feature recognition is identified.

[0057] Step 5.3: Sigmoid activation generates attention weights. The output of the one-dimensional convolution is processed through a sigmoid activation function, mapping the output value to a range between 0 and 1. This output value is the attention weight corresponding to each channel. The attention weight indicates the importance of each channel in the feature representation. The closer the weight is to 1, the more important the channel is to the subsequent detection task; the closer the weight is to 0, the less important the channel is.

[0058] Step 5.4: Feature Weighting to Enhance Important Channels. The generated attention weights are element-wise multiplied with the original feature map. This enhances channels related to rain streak features, highlighting their characteristic information, while suppressing features from irrelevant channels, reducing interference with the detection task. This improves the detection capabilities of rain streak features and aircraft skin damage characteristics, providing a more discriminative feature representation for subsequent object detection tasks.

[0059] Step 6: Use the aircraft skin damage features obtained after the above processing to train the target detection network. During the training process, continuously adjust the loss function of the network model to make the trained network detection more consistent with the data set, and finally obtain a trained target detector. The detection steps are shown in the attached figure. Figure 3 shown.

[0060] Step 7: Deploy the trained target detector onto the drone and use it to fly around the aircraft skin to obtain target detection results and the target's appearance features.

[0061] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprise," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations. The phrase "includes an element defined by..." does not exclude the presence of other identical elements in the process, method, article, or device that includes the element.

[0062] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning aircraft skin damage detection method suitable for rain and fog conditions, characterized by: The following steps are involved: S1. The UAV flies around the aircraft to obtain original image data of aircraft skin damage; S2. Preprocess the collected image data, complete screening and labeling, and create the aircraft skin damage dataset required for training the network; S3. Input the training dataset image data into the AOD-Net integrated defogging network. Through the multi-scale feature fusion, fog depth and relative level estimation in the K-estimating module of the network, combined with the atmospheric scattering model and residual connection optimization of the defogging image generation module, a clear restored image is generated. S4. Input the training data set processed by the AOD-Net integrated defogging network into the CSPDarknet backbone network of yolov8 for feature extraction; S5. The target features extracted by the CSPDarknet backbone network are introduced into the ECA channel attention mechanism. Through global average pooling, one-dimensional convolution, and sigmoid weight generation, cross-channel interaction information is captured to enhance the channel features related to rain streaks and damage. S6. Using the aircraft skin damage features obtained after the above processing, a target detection network is trained to obtain a target detector; S7. Deploy the trained target detector onto the UAV, and use the UAV to inspect the aircraft skin to obtain target detection results and the target's appearance features.

2. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: The specific operation steps of step S3 include: S3.

1. In the K-estimating module, a set of initial convolutional layers is used to process the input image. Feature information, including texture, edges, and low-level features related to fog, is initially extracted from the image to obtain a preliminary feature map. Convolution operations are performed on the preliminary feature map at different scales using convolution kernels of different sizes or dilated convolution. Dilated convolution is also used to increase the receptive field and obtain more global fog feature information. These feature maps of different scales are then concatenated and fused into a multi-scale feature map containing multi-scale information. S3.

2. Based on the fused multi-scale feature map, a fully connected layer is used to perform regression calculations. The fully connected layer integrates the information in the multi-scale feature map and outputs an estimate of the fog depth and relative level based on pre-trained parameters. S3.

3. The fog depth and relative level estimation values obtained by the K-estimating module are input into the defogging image generation module as adaptive parameters. Based on the preliminary processing based on the physical model, the defogging image generation module further uses a convolutional neural network to optimize the image. Through multi-layer convolution operations, the feature extraction and reconstruction of the image after preliminary defogging are performed. At the same time, residual connections are used to ensure the information integrity of the image during the processing process to avoid image distortion caused by excessive defogging.

3. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: The specific operation steps of step S5 include: S5.

1. Input the feature map extracted from the CSPDarknet backbone network into the ECA attention mechanism module and perform a global average pooling operation. The global average pooling operation compresses the two-dimensional feature map of each channel in the spatial dimension to obtain a scalar value. The scalar value corresponding to each channel forms a one-dimensional vector. The one-dimensional vector contains the global information of the entire feature map in each channel. S5.

2. The one-dimensional vector enters a one-dimensional convolutional layer, and the convolution kernel size of the one-dimensional convolutional layer is automatically adjusted according to the number of channels of the feature map to capture local cross-channel interaction information and mine the correlation between different channels; S5.

3. Process the output of the one-dimensional convolutional layer through a Sigmoid activation function, mapping the output value to a range between 0 and 1. The output value is the attention weight corresponding to each channel, and the attention weight is used to indicate the importance of each channel in the feature representation; The closer the attention weight is to 1, the more important the channel is to the subsequent detection task; The closer the attention weight is to 0, the less important the channel is; S5.

4. Multiply the generated attention weights by the original feature map element by element, so that the channels related to the rain streak features are enhanced and their feature information is highlighted, while the features of irrelevant channels are suppressed, thereby improving the detection capability of rain streak features and aircraft skin damage features.

4. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: In step S1, data augmentation is used to expand the data, and a trimming operation is performed on the collected image or another part of the data is obtained from the existing database. The trimming operation includes rotation, flipping, and cropping.

5. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: The CSPDarknet backbone network in step S4 adopts a combination of multiple convolutional layers and residual blocks, wherein the convolutional layer adopts a 3x3 or 5x5 convolution kernel with a step size of 1 or 2, and the residual block consists of two convolutional layers and a skip connection. The input features are added to the convolutional layer output features through the skip connection, so that the network can learn deeper features.

6. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: The pre-processing in step S2 includes screening out clear and usable data and eliminating fuzzy, damaged or non-compliant data; Use professional annotation tools to annotate the filtered images. The annotation content includes the location information and identification number of the damaged target. During annotation, it is necessary to ensure the balance of data in each category, so as to produce the aircraft skin damage dataset required for training the network.

7. The deep learning aircraft skin damage detection method applicable to rain and fog conditions according to claim 1 is characterized in that: In step S6, during the training process, the loss function of the network model is continuously adjusted to make the trained network detection more consistent with the data set.

Images