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

Abstract

The application discloses a kind of deep learning aircraft skin damage detection methods suitable for rain and fog conditions, comprising the following steps: unmanned aerial vehicle obtains original image data;Image data is preprocessed, and data set is made;Training data set image data is input into AOD-Net integrated fog removal network;Training data set is input into CSPDarknet main network to carry out feature extraction and introduce ECA channel attention mechanism;The target detector obtained is deployed to unmanned aerial vehicle, and target detection result is obtained, the present application relates to aircraft skin damage detection technical field, the present application can effectively enhance the damage detection effect of aircraft skin under foggy condition, simultaneously, combined with ECA attention mechanism, capture local cross-channel interaction, so that it can adaptively adjust the feature response of each channel, enhance the channel related to rain feature, not only can guarantee the damage accurate detection of most rain and fog conditions, but also rely on the light weight characteristics of AOD-Net and ECA can realize the maximum degree of resource saving, reduce detection cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aircraft skin damage detection, in particular to a deep learning aircraft skin damage detection method suitable for rain and fog conditions. BACKGROUND

[0002] The aircraft skin, as a key protective structure of the aircraft body, plays a decisive role in flight safety. It not only maintains the aerodynamic shape of the aircraft and ensures the aerodynamic performance during flight, but also protects the internal structure from the erosion of complex external environments such as rain, fog, ultraviolet light, and various pollutants. Therefore, ensuring the integrity and reliability of the aircraft skin is one of the core tasks in the field of aviation safety.

[0003] However, in actual detection scenarios, rain and fog conditions frequently occur, posing a great challenge to detection work. Traditional aircraft skin detection methods have many drawbacks in rain and fog environments. Visual detection relies on the visual observation of detection personnel, but it is difficult for humans to work continuously in rain and fog. Ultrasonic detection can detect internal damage, but the equipment is complex to operate, requires high technical skills of detection personnel, and has slow detection speed, low efficiency in large-area detection, high cost, expensive equipment, and the need for professional protection measures. In addition, it poses radiation hazards to the human body and is not suitable for frequent routine detection.

[0004] With the rise of deep learning technology in the field of image detection, it brings new hope for aircraft skin damage detection. However, existing deep learning algorithms still have significant shortcomings when applied to aircraft skin detection in rain and fog conditions. Specifically, rain and fog can significantly reduce the contrast of images, making the gray difference between damaged areas and normal skin areas smaller, making it difficult for deep learning algorithms based on gray difference feature extraction to highlight damage features. At the same time, rain and fog environments often come with increased noise, which can interfere with the extraction of effective features by the algorithm, causing 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, dents, and delamination, each with different manifestations in rain and fog conditions, increasing the difficulty of accurately classifying and locating damage by the model. Moreover, the aircraft skin area is huge, requiring detection algorithms to have high processing power to comprehensively detect large areas in a short time, which poses high requirements on the computational efficiency and resource utilization efficiency of the algorithm.

[0005] In summary, traditional aircraft skin detection methods have many limitations in rain and fog conditions, and existing deep learning algorithms still need further optimization and improvement when applied to aircraft skin damage detection in rain and fog conditions to meet the strict requirements of the aviation field for aircraft skin safety detection. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a deep learning aircraft skin damage detection method suitable for rain and fog conditions, which aims to solve a series of problems faced by aircraft skin damage detection under rain and fog conditions, i.e., low detection accuracy, low efficiency and high cost. The method introduces an AOD-Net integrated defogging network for preprocessing rain and fog images, improves the YOLOv8 network by combining the ECA attention mechanism, and proposes an efficient and accurate aircraft skin damage detection method. The K-estimating module of the AOD-Net is used to estimate the fog depth and relative level, generate a restored image, and improve the image quality to provide a clear data basis for damage detection. The ECA attention mechanism is used to enhance the channels related to rain and damage features and suppress irrelevant channels to improve the detection capability of rain and damage features. At the same time, the algorithm is deployed on a UAV, and the maneuverability of the UAV is used for detection to reduce the investment of manpower and resources and reduce the detection cost, thereby realizing efficient and accurate detection of aircraft skin under rain and fog conditions and ensuring flight safety.

[0007] To achieve the above object, the present application is implemented by the following technical scheme: a deep learning aircraft skin damage detection method suitable for rain and fog conditions, comprising the following steps:

[0008] S1, the UAV obtains original image data of aircraft skin damage by flying around the aircraft;

[0009] S2, preprocessing the collected image data, completing the labeling of screening and labels, and making an aircraft skin damage data set required for training network;

[0010] S3, inputting the training data set image data into the AOD-Net integrated defogging network, estimating the multi-scale feature fusion, fog depth and relative level in the K-estimating module in the network, combining the atmospheric scattering model and residual connection optimization of the defogging image generation module to generate a clear restored image;

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

[0012] S5, introducing the target features extracted by the CSPDarknet main network into the ECA channel attention mechanism, capturing cross-channel interaction information through global average pooling and one-dimensional convolution, and generating Sigmoid weight to enhance the rain and damage related channel features;

[0013] S6, training the target detection network using the aircraft skin damage features obtained after the above processing to obtain a target detector;

[0014] S7, deploy the trained target detector to the UAV, use the UAV to detect the aircraft skin, and obtain the target detection result and the appearance feature of the target.

[0015] Preferably, the data is expanded in step S1 by using data enhancement, and the collected image is trimmed or another part of data is obtained from an existing database, and the trimming operation includes rotation, flipping and cropping.

[0016] Preferably, the preprocessing in step S2 includes screening clear and usable data, and removing blurred, damaged or non-compliant data.

[0017] The screened image is labeled using a professional labeling tool, and the labeled content includes position information and identity number of the damaged target. When labeling, the balance of various types of data is ensured, so as to make an aircraft skin damage data set required for training network.

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

[0019] S3.1, in the K-estimating module, an initial convolutional layer is used to process the input image, and the feature information including texture, edge and low-level features related to fog in the image is preliminarily extracted, a preliminary feature map is obtained, different size convolution kernels or hole convolution are used to perform convolution operation on the preliminary feature map at different scales, and the receptive field is increased by combining hole convolution, so as to obtain more global fog feature information, and then the feature maps at different scales are spliced and fused into a multi-scale feature map containing multi-scale information;

[0020] S3.2, based on the fused multi-scale feature map, a regression calculation is performed by using a fully connected layer, the information in the multi-scale feature map is integrated by the fully connected layer, and according to the parameters learned by pre-training, an estimated value of fog depth and relative level is output;

[0021] S3.3, the estimated values of fog depth and relative level obtained by the K-estimating module are input as adaptive parameters into a dehazing image generation module, on the basis of preliminary processing based on a physical model, the dehazing image generation module further optimizes the image by using a convolutional neural network, extracts and reconstructs the image after preliminary dehazing by multi-layer convolution operation, and uses residual connection to ensure the information integrity of the image in the processing process, and avoids image distortion caused by excessive dehazing.

[0022] Preferably, the CSPDarknet backbone network in step S4 adopts a combined structure of multiple convolutional layers and residual blocks, wherein the convolutional layers adopt a 3x3 or 5x5 convolution kernel with a stride of 1 or 2, and the residual blocks are composed of two convolutional layers and a skip connection, and the input features are added to the output features of the convolutional layers through the skip 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 global average pooling operation, which will compress the two-dimensional feature map of each channel in the spatial dimension to obtain a scalar value, and each channel corresponds to a one-dimensional vector composed of scalar values, and the one-dimensional vector contains global information of the entire feature map in each channel;

[0025] S5.2, the one-dimensional vector enters a one-dimensional convolutional layer, which automatically adjusts the size of the one-dimensional convolutional layer kernel 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, the output of the one-dimensional convolutional layer is processed by a Sigmoid activation function to map the output value to between 0 and 1, and the output value is the attention weight corresponding to each channel, which is used to represent the importance of each channel in 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 lower the importance of the channel;

[0029] S5.4, multiply the generated attention weight with the original feature map element by element, so that the channel related to the rain streak feature is enhanced, and its feature information is highlighted, while the feature of the irrelevant channel is suppressed, thereby improving the detection ability of the rain streak feature and the aircraft skin damage feature.

[0030] Preferably, step S6 adjusts the loss function of the network model during training to make the trained network detection more suitable for the data set.

[0031] Advantages

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

[0033] The application effectively enhances the damage detection effect of the aircraft skin under foggy conditions by introducing an AOD-Net integrated defogging network into the YOLOv8 network, estimating the fog depth and relative level through the K-estimating module, inputting the estimated parameters as adaptive parameters into the defogging image generation module to generate a restored image, and at the same time, combining the ECA (Efficient Channel Attention) attention mechanism to capture local cross-channel interaction, so that it can adaptively adjust the feature response of each channel, enhance the channel related to the rain streak feature, and suppress the irrelevant channel, thereby improving the detection capability of the rain streak feature. Deploying the algorithm on the unmanned aerial vehicle can not only ensure accurate detection of damage under most rain and fog conditions, but also realize maximum resource saving, reduce manpower and material consumption, and reduce detection cost relying on the lightweight features of AOD-Net and ECA. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The flowchart of the application is shown.

[0035] Figure 2 The flowchart of the application is shown.

[0036] Figure 3 The detection flowchart of the target detector is shown. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0038] Embodiment: The core idea of the method of the application is to use an unmanned aerial vehicle to collect aircraft skin damage data under rain and fog conditions and train a target detector, and to use the combination of the target detector and the unmanned aerial vehicle to detect aircraft skin damage. First, the unmanned aerial vehicle is used to fly around the aircraft, collect rain and fog condition aircraft skin damage source data, and use a labeling tool to process the collected data to make a rain and fog condition aircraft skin damage data set. The collected rain and fog condition aircraft skin image data is used to train a target detector, and the combination of the target detector and the unmanned aerial vehicle is used to detect aircraft skin damage. In this process, to deal with the interference of rain and fog on aircraft skin damage detection, an AOD-Net integrated defogging network and an ECA attention mechanism are introduced to improve the YOLOv8 network.

[0039] Firstly, the AOD-Net integrated defogging network plays an important role. The K-estimating module estimates the fog depth and relative level in the image, and these estimated parameters are input as adaptive parameters into the defogging image generation module. This module processes the collected foggy aircraft skin image according to these parameters to generate a clear restored image, thereby effectively improving the image quality and highlighting the aircraft skin details, providing a high-quality data basis for subsequent damage detection, and making the damage features more clear and identifiable in the image.

[0040] Next, the ECA attention mechanism is introduced. This mechanism can sensitively capture local cross-channel interaction information and intelligently adaptively adjust the feature response of each channel when processing aircraft skin images. For channels closely related to rain streak features, the ECA mechanism enhances their feature expression, strengthening the information of these key channels. For channels unrelated to rain streak features, the ECA mechanism suppresses their role, reducing interference information. In this way, the detection ability of the network for rain streak features is significantly improved, enabling the entire detection system to more accurately identify aircraft skin damage in rain and fog conditions.

[0041] Finally, the improved algorithm is deployed on the UAV. The UAV, with its flexible maneuverability, 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 ensure the accuracy of aircraft skin damage detection under most rain and fog conditions while maximizing the saving of computing resources, significantly reducing 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. The aircraft skin damage detection method in rain and fog conditions according to the present embodiment, as shown in Figure 1 includes the following steps:

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

[0043] In this embodiment, to avoid single background and class imbalance of aircraft skin damage and prevent model overfitting, a sufficient amount of data needs to be collected. Data augmentation is used to expand the data, and the collected images are rotated, flipped, cropped, or another part of data is obtained from an existing database.

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

[0045] In the embodiment, the collected image data is preprocessed, including screening out clear and usable data, and rejecting blurred, damaged or non-compliant data. The screened images are labeled using professional labeling tools, and the labeling content includes position information (bounding box coordinates) of the damage target and identity number, etc. At the same time, the balance of various types of data is ensured to avoid poor training results due to unbalanced class data, so as to produce the aircraft skin damage data set required for training network.

[0046] Step 3: input the training data set image data into the AOD-Net integrated defogging network.

[0047] In the embodiment, the training set data is input into the AOD-Net integrated defogging network, and a clear recovered image is generated through the K-estimating module and the defogging image generation module in the network. Through parameter estimation and adaptive parameter re-input, light weight and image quality improvement are achieved, and the performance of the subsequent detector is improved. As shown in the accompanying Figure 2

[0048] Step 3.1: multi-scale feature fusion strategy is adopted for feature map generation and splicing. In the K-estimating module, a group of initial convolution layers is used to process the input image, and the feature information in the image is preliminarily extracted, including texture, edge and low-level features related to fog, to obtain a preliminary feature map. By using convolution kernels or atrous convolution of different sizes, the preliminary feature map is convolved on different scales, and the receptive field is increased by combining atrous convolution to obtain more global fog feature information. Then the feature maps of different scales are spliced and fused into a feature map containing multi-scale information, so as to more accurately estimate the related parameters of the fog in the subsequent process.

[0049] Step 3.2: estimate the fog depth and relative level. Based on the fused multi-scale feature map, a regression calculation is performed using a fully connected layer. The fully connected layer will integrate the information in the fused feature map, and output the estimated values of the fog depth and relative level according to the parameters learned by pre-training. In the training process, by comparing with a large number of image data with accurate fog parameter labels, the weights of the fully connected layer are constantly adjusted, so that the estimation result becomes more and more accurate. These estimated values reflect the density of the fog in the image and the relative position distribution of the fog in the image, etc., which provides a key basis for the subsequent defogging operation.

[0050] ​Step 3.3: Fog image generation based on atmospheric scattering model. The fog depth and relative level estimation parameters obtained by the K-estimating module are input as adaptive parameters into the fog image generation module to adjust the processing intensity and method during the fog removal process to adapt to the specific situation of fog in different images. Based on the preliminary processing based on the physical model, the fog image generation module further optimizes the image using a convolutional neural network. Through multi-layer convolution operation, the image after preliminary fog removal is feature extracted and reconstructed. The convolutional layer will adjust according to the local features of the image, enhance the detail information of the image, repair the texture and edges that may be lost in the fog removal process, and at the same time use residual connection to ensure the integrity of the image information during processing, avoid over-fogging leading to image distortion.

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

[0052] In this embodiment, the data is input into the CSPDarknet network for feature extraction, which adopts a combination structure of multiple convolutional layers and residual blocks. The convolutional layer uses a 3x3 or 5x5 convolution kernel with a step size of 1 or 2. The residual block is composed of two convolutional layers and a jump connection, which adds the input features and the convolutional layer output features through the jump connection, effectively solving the gradient vanishing problem caused by the increase in network depth, enabling the network to learn deeper features. In this process, the image resolution gradually decreases, the number of channels of the feature map gradually increases, and the network continuously extracts local and global features of the image, providing rich feature information for subsequent target detection.

[0053] Step 5: Introduce the target features extracted by the CSPDarknet backbone network into the ECA channel attention mechanism for rain streak feature enhancement.

[0054] In this embodiment, the target features extracted by the CSPDarknet are introduced into the ECA channel attention mechanism for cross-channel information interaction capture. Through global average pooling operation, the two-dimensional feature map of each channel is compressed into a scalar to obtain the global feature information of the channel. Then, a one-dimensional convolution is used to process the channel features after global average pooling to capture local cross-channel interaction information. Then the output of the one-dimensional convolution is passed through the Sigmoid activation function to generate attention weights for each channel. Finally, the generated attention weights are multiplied element-wise with the original feature map to enhance the channels related to rain streak features and suppress irrelevant channels, improving the detection ability of rain streak features and also enhancing the recognition ability of aircraft skin damage features.

[0055] Step 5.1: Global average pooling obtains channel global information. The feature map extracted from the CSPDarknet backbone network is input into the ECA attention mechanism module. First, the global average pooling operation is performed, which compresses the two-dimensional feature map in the spatial dimension for each channel, i.e., the feature map is averaged along the height and width directions to obtain a scalar value. Each channel corresponds to a scalar representing its global feature information, and these scalars form a one-dimensional vector containing the global information of the entire feature map in each channel.

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

[0057] Step 5.3: Sigmoid activation generates attention weights. The output of one-dimensional convolution is processed by the Sigmoid activation function to map the output value to between 0 and 1, and this output value is the attention weight corresponding to each channel. Attention weight represents the importance of each channel in feature representation, and the weight closer to 1 indicates that the channel is more important for subsequent detection tasks; the weight closer to 0 indicates that the channel is less important.

[0058] Step 5.4: Feature weighting enhances important channels. Multiply the generated attention weights with the original feature map element by element, so that the channels related to rain streak features are enhanced, highlighting their feature information, while irrelevant channels are suppressed, reducing interference with the detection task. In this way, the detection ability of rain streak features and aircraft skin damage features is improved, providing more discriminative feature representation for subsequent target detection tasks.

[0059] Step 6: Use the aircraft skin damage features obtained after the above processing to train the target detection network. During training, the loss function of the network model is constantly adjusted to make the trained network detection more consistent with the data set, and finally a trained target detector is obtained. The detection step is shown in FIG. 8. Figure 3

[0060] Step 7: Deploy the trained target detector to the unmanned aerial vehicle, use the unmanned aerial vehicle to inspect the aircraft skin, and obtain the target detection result and the appearance feature of the target.

[0061] ​It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify different entities or actions and do not necessarily require or imply any actual relationship or order between these entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. In other words, without further restriction, reference to elements will not, without more limitations, exclude additional, unrecited elements of a process, method, article, or apparatus.

[0062] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to these embodiments. Rather, it is the intention to cover any variations and modifications, provided they fall within the scope of the application as 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 in that, The method comprises the following steps: S1, the unmanned aerial vehicle flies around the aircraft to obtain original image data of the aircraft skin damage; S2, preprocessing the collected image data, completing the screening and label annotation, and making an aircraft skin damage data set required for training network; S3, inputting the training data set image data into the AOD-Net integrated defogging network, estimating the fog depth and relative level through the multi-scale feature fusion in the K-estimating module in the network, combining the atmospheric scattering model and residual connection optimization of the defogging image generation module to generate a clear restored image; S4, inputting the training data set processed by the AOD-Net integrated defogging network into the CSPDarknet backbone network of yolov8 for feature extraction; S5, introducing the target features extracted by the CSPDarknet backbone network into the ECA channel attention mechanism, capturing cross-channel interaction information through global average pooling and one-dimensional convolution, and generating Sigmoid weight to enhance the features of rain streaks and damage-related channels; S6, training the target detection network using the aircraft skin damage features obtained after the above processing to obtain a target detector; S7, deploying the trained target detector to the unmanned aerial vehicle, using the unmanned aerial vehicle to fly around the aircraft skin to obtain the target detection result and the appearance feature of the target.

2. The deep learning based aircraft skin damage detection method for rain and fog conditions according to claim 1, wherein, The specific operation steps of step S3 include: S3.1, in the K-estimating module, a set of initial convolution layers is used to process the input image, and the feature information including texture, edge and low-level features related to fog in the image is preliminarily extracted to obtain a preliminary feature map. Different size convolution kernels or atrous convolution are used to perform convolution operation on the preliminary feature map at different scales, and the receptive field is increased by combining atrous convolution to obtain more global fog feature information. Then, the feature maps at different scales are spliced and fused into a multi-scale feature map containing multi-scale information; S3.2, based on the fused multi-scale feature map, a regression calculation is performed using a fully connected layer, which integrates the information in the multi-scale feature map and outputs the estimated values of the fog depth and relative level according to the parameters learned by pre-training; S3.3, the estimated values of the fog depth and relative level obtained by the K-estimating module are input as adaptive parameters into the defogging image generation module, which further optimizes the image using convolutional neural network based on the preliminary processing based on the physical model. Through multi-layer convolution operation, the preliminary defogging image is extracted and reconstructed, and residual connection is used to ensure the integrity of the image information in the processing process to avoid image distortion caused by excessive defogging.

3. The deep learning based aircraft skin damage detection method for rain and fog conditions according to claim 1, wherein, 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 global average pooling operation, which will compress the two-dimensional feature map of each channel in the spatial dimension to obtain a scalar value, and the scalar value corresponding to each channel forms a one-dimensional vector, which contains the global information of the entire feature map in each channel; S5.2, the one-dimensional vector enters the one-dimensional convolution layer, which automatically adjusts the convolution kernel size of the one-dimensional convolution layer according to the channel number of the feature map, to capture local cross-channel interaction information and mine the correlation between different channels; S5.3, the output of the one-dimensional convolution layer is processed by the Sigmoid activation function to map the output value to between 0 and 1, which is the attention weight corresponding to each channel, which is used to represent the importance of each channel in 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 lower the importance of the channel; S5.4, multiply the generated attention weight with the original feature map element by element, so that the channel related to the rain feature is enhanced, highlighting its feature information, while the feature of irrelevant channel is suppressed, improving the detection ability of rain feature and aircraft skin damage feature.

4. The deep learning based aircraft skin damage detection method for rain and fog conditions of claim 1, wherein, In step S1, data augmentation is used to expand the data, and the collected images are trimmed or another part of data is obtained from the existing database, the trimming operation includes rotation, flipping and cropping.

5. The deep learning based aircraft skin damage detection method for rain and fog conditions of claim 1, wherein, In step S4, the CSPDarknet backbone network uses a combination structure of multiple convolution layers and residual blocks, wherein the convolution layer uses a 3x3 or 5x5 convolution kernel with a stride of 1 or 2, and the residual block is composed of two convolution layers and a jump connection, which adds the input features and the convolution layer output features through the jump connection, so that the network can learn deeper features.

6. The deep learning based aircraft skin damage detection method for rain and fog conditions of claim 1, wherein In step S2, the preprocessing includes screening out clear and usable data, and rejecting fuzzy, damaged or non-compliant data; Use professional annotation tools to label the screened images, the labeled content includes position information and identity number of damage target, and ensure the balance of each class during labeling, so as to make the aircraft skin damage data set required for training network.

7. The deep learning based aircraft skin damage detection method for rain and fog conditions of claim 1, wherein In step S6, during the training process, the loss function of the network model is constantly adjusted to make the trained network detection more suitable for the data set.

Images