Deep learning aircraft skin damage detection method suitable for dark light condition

Abstract

The invention discloses a deep learning aircraft skin damage detection method suitable for a dark light condition, and the method comprises the following steps: enabling an unmanned plane to fly around an aircraft to obtain the original image data of the aircraft skin damage, carrying out the preprocessing of the collected image data, completing the screening and label marking, and carrying out the detection of the aircraft skin damage. The invention relates to the technical field of aircraft skin detection, in particular to an aircraft skin damage detection method based on a PENet network, an aircraft skin damage data set needed by a training network is made, image data of the training data set is input into the PENet network, the dark light information capture rate is increased, and the training data set processed by the PENet is input into an improved yolov8 network and a CSPDarknet backbone network for feature extraction. The SCConv module is used for carrying out CNN compression by using space and channel redundancy between features, redundancy calculation is reduced, learning of representative features is promoted, the training calculation amount is reduced, the model is lightened, the model training time is shortened, an algorithm is deployed on the unmanned aerial vehicle, consumption of manpower and material resources can be greatly reduced, and the detection cost is reduced.

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 low-light conditions. Background Art

[0002] Aircraft skin is a key component of the aircraft body, responsible for maintaining the aircraft's aerodynamic shape and protecting the internal structure from environmental corrosion. Therefore, ensuring the integrity and reliability of the aircraft skin is crucial to flight safety.

[0003] In the field of traditional aircraft skin inspection, visual inspection is one of the most common methods. Inspectors observe the skin surface directly with the naked eye, or use simple optical tools (such as magnifying glasses) to assist in inspection. This method can play a certain role when there is sufficient light and the damage is obvious. However, when faced with a dim light environment, the visual perception ability of the human eye is greatly reduced, and it is difficult to detect subtle signs of damage, such as tiny cracks on the surface, early corrosion areas, and imperceptible dents. These seemingly minor damages may gradually expand during the flight of the aircraft due to the influence of complex factors such as airflow impact, air pressure changes, and temperature fluctuations, and eventually pose a serious threat to the integrity of the aircraft structure.

[0004] Ultrasonic testing technology utilizes the propagation characteristics of ultrasound waves through materials to detect internal defects. When ultrasound waves encounter defects, they reflect, refract, and scatter. By analyzing these signal changes, the defect's location, size, and nature can be determined. However, this technology requires high operator expertise, requiring extensive experience and precise techniques to accurately interpret the detection signals. Furthermore, ultrasonic testing is relatively slow, requiring significant time to inspect large areas of aircraft skin, which undoubtedly increases testing costs and reduces efficiency.

[0005] Radiographic testing uses the attenuation characteristics of radiation as it penetrates an object to create images and detect internal structural defects. Common radiographic testing methods include X-ray testing and gamma-ray testing. Radiographic testing can provide relatively clear images of internal structures, offering advantages for detecting defects hidden deep within the skin. However, radiographic testing equipment is expensive, with extremely high purchase, maintenance, and operating costs. Furthermore, radiation poses a risk to the human body, necessitating strict protective measures. This not only complicates testing but also poses a potential threat to the health of testers. Furthermore, the relatively complex process of radiographic testing requires specialized technicians to operate and interpret the test results, further limiting its widespread use in routine inspections.

[0006] With the rapid development of artificial intelligence, deep learning has achieved remarkable success in image recognition and detection. Deep learning algorithms can automatically learn characteristic patterns from large amounts of data, demonstrating powerful feature extraction and classification capabilities. In the field of object detection, the YOLO (You Only Look Once) family of algorithms stands out for its speed and efficiency, becoming one of the preferred algorithms for numerous detection tasks. Since its inception, the YOLO family has undergone numerous iterations, each one building upon previous versions to address limitations and improve performance.

[0007] The YOLO family of detection networks includes multiple versions. Some of these models, while requiring fewer parameters and requiring less computation, still meet detection accuracy requirements. Subsequent YOLO versions have introduced improvements and enhancements to address these limitations while maintaining the framework's real-time performance. For example, YOLOv2 (YOLO9000) introduced anchor boxes and penetration layers to improve object localization, thereby increasing accuracy. Furthermore, YOLOv3 enhanced the model's performance by adopting a multi-scale feature extraction architecture, enabling better object detection at various scales. Models such as YOLOv4 and YOLOv5 introduced innovations such as a new network backbone, improved data augmentation techniques, and optimized training strategies. YOLOv8 utilizes an advanced network architecture, including an efficient backbone network for feature extraction, a multi-scale feature fusion strategy to enhance detection capabilities for objects of varying sizes, and a concise and effective detection head for predicting object location, category, and confidence. Trained on large amounts of image data, it can quickly and accurately identify a wide range of objects in images.

[0008] Images captured under dim conditions present unique characteristics and challenges. First, the image contrast is significantly reduced, and the grayscale difference between the damaged area and the normal skin area becomes smaller, making it difficult to highlight the damage features. This poses difficulties for deep learning algorithms that extract features based on grayscale differences. Secondly, dim environments are often accompanied by an increase in noise, which interferes with the algorithm's extraction of effective features, causing the model to mistakenly identify noise as damage features, or fail to accurately identify true damage features. In addition, aircraft skin damage types vary, including cracks, corrosion, pits, debonding, etc. Each type of damage manifests differently under dim conditions, increasing the difficulty of the model accurately classifying and locating damage. Moreover, the aircraft skin area is huge, requiring the detection algorithm to have efficient processing capabilities and be able to perform comprehensive inspections of large areas in a short period of time. This places extremely high demands on the algorithm's computational efficiency and resource utilization efficiency.

[0009] In summary, traditional aircraft skin inspection methods have many limitations under dark light conditions, and the existing deep learning algorithms still need to be further optimized and improved when applied to aircraft skin damage detection under dark light conditions to meet the strict requirements of the aviation field for aircraft skin safety inspection. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this paper proposes a low-light aircraft skin damage detection method based on an improved YOLO network. This method specifically targets low-light images of aircraft skin. By integrating multiple image processing techniques, it effectively reduces noise interference, allowing damage features that would otherwise be obscured in low light to be clearly visualized, improving detection accuracy. Furthermore, the paper proposes the use of drones for patrol inspections to collect images, saving significant manpower and resources and reducing the cost of aircraft skin damage detection.

[0011] 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 low-light conditions, comprising the following steps:

[0012] S1. The UAV flies around the aircraft to obtain original image data of aircraft skin damage;

[0013] S2. Preprocess the collected raw image data, complete screening and labeling, and create the aircraft skin damage dataset required for training the network;

[0014] S3. Input the training dataset image data into the PENet network to improve the dark light information capture rate. The image components are optimized jointly by the detail processing module (DPM) and the low-frequency enhancement filter (LEF). The detail processing module (DPM) adopts a dual-branch structure to perform edge enhancement and upper and lower feature extraction respectively. The low-frequency enhancement filter (LEF) captures and filters low-frequency information through dynamic low-pass filtering and average pooling.

[0015] S4. Input the training data set processed by PENet into the improved yolov8 network, perform feature extraction through the CSPDarknet backbone network, and introduce the extracted target features into the PSA channel attention mechanism for feature enhancement. The PSA channel attention mechanism includes the SPC module and the SEWeight module;

[0016] S5. In the neck and head networks of YOLOv8, SCConv modules are used to replace 30% of the c2f convolution blocks to optimize the feature extraction capabilities of specific layers and reduce spatial and channel redundancy.

[0017] S6. Using the obtained aircraft skin damage features, the target detection network is trained to obtain a trained target detector;

[0018] 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.

[0019] Preferably, the processing process of the PENet network in step S3 includes:

[0020] S3.1. Use Gaussian extraction on the input image. After each Gaussian pyramid operation, the width and height of the image are halved, and the resolution is reduced to 1 / 4 of the original. The lost information constitutes the components of the Laplacian pyramid. When reconstructing the image, the corresponding reverse operation is performed to restore the high-resolution image.

[0021] S3.2. Different components in the pyramid enter the detail processing module (DPM) and low-frequency enhancement filter (LEF) in parallel. The detail processing module (DPM) includes an edge enhancement branch and a context branch. The context branch uses a residual block to process features before and after obtaining long-range dependencies. The first residual block changes the number of feature channels from 3 to 32, and the second residual block changes the number of feature channels from 32 to 3, thereby globally enhancing information. The edge enhancement branch uses the Sobel operator to calculate image gradients in both horizontal and vertical directions to obtain edges and enhance the texture of components.

[0022] S3.3. The low-frequency enhancement filter (LEF) is used to capture low-frequency information in image components. First, the input components are converted into different scales through the convolution layer, and then the low-frequency information is captured and filtered using a dynamic low-pass filter and average pooling. The specific pooling size is combined with bilinear interpolation sampling to form low-pass filters of different scales. The low-pass filters are finally integrated through tensor splicing to restore the original image size.

[0023] Preferably, the specific operations of step S4 are:

[0024] S4.1. The CSPDarknet backbone network adopts a combination of multiple convolutional layers and residual blocks. The residual block consists of two convolutional layers and a skip connection. The convolutional layer adds the input features to the convolutional layer output features through the skip connection, enabling the network to learn deeper features. After the input image is processed by the initial convolutional layer, its size is halved and the number of channels is increased. It then enters the residual block structure. As the network depth increases, the image resolution gradually decreases and the number of channels in the feature map gradually increases. In this process, the network continuously extracts local and global features of the image, providing rich feature information for subsequent target detection.

[0025] S4.2. The learned features are input into the PSA channel. The SPC module first performs channel segmentation on the input feature map, evenly dividing the number of channels into several groups. For each group of channels, convolution kernels of different sizes are used to perform convolution operations to obtain spatial feature information of different scales. The feature maps output by the convolution kernels of different scales are spliced together to obtain a feature map that integrates multi-scale spatial information, so that the network can simultaneously focus on skin damage features of different sizes.

[0026] S4.3, SEWeight module performs global average pooling operation on the feature map output by SPC module to obtain global feature information of each channel. Then the global feature information is processed by two fully connected layers. The first fully connected layer maps the global feature information to a lower dimension, and the activation function is ReLU; the second fully connected layer maps it back to the original number of channels, and the activation function is Sigmoid. Through the two fully connected layers, attention weights are generated for each channel.

[0027] S4.4. Multiply the attention weights generated by the SEWeight module by the feature map output by the SPC module element by element to obtain the weighted feature map. Perform a Softmax operation on the weighted feature map in the channel dimension to obtain the final attention weights after multi-scale channel interaction. Finally, multiply the attention weights by the feature map element by element again to obtain the feature map after multi-scale feature information attention weighting, which effectively enhances the detection capability of skin damage at different scales.

[0028] Preferably, the convolution layer in step S4.1 uses a 3x3 or 5x5 convolution kernel with a step size of 1 or 2.

[0029] Preferably, the size of the Gaussian kernel in step S3.1 is 5×5.

[0030] Preferably, in step S2, position information is marked on the target, and the position information includes bounding box coordinates and an identity number.

[0031] Beneficial effects

[0032] The present invention discloses a deep learning aircraft skin damage detection method suitable for low-light conditions, which has the following beneficial effects:

[0033] This paper introduces PENet into the YOLOv8 network, decomposing images into components of different resolutions using a Laplacian pyramid. This method captures the potential information of objects in low-light images while suppressing noise in high-frequency components, effectively enhancing detection performance under low-light conditions. Furthermore, the proposed method combines the PSA channel attention mechanism with the SPC module, SEWeight module, and weight recalculation to effectively process the spatial information of multi-scale input feature maps and establish long-term dependencies between multi-scale channel attentions. Furthermore, the proposed method modifies the C2F convolution block and uses the SCConv (spatial and channel reconstruction convolution) module as one of the initial components of the backbone network. The SCConv module exploits spatial and channel redundancy between features for CNN compression, reducing redundant computation and facilitating the learning of representative features. This reduces the amount of training computation, making the model lightweight and shortening model training time. Deploying the algorithm on drones can significantly reduce labor and resource consumption and lower detection costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Flowchart of the present invention.

[0035] Figure 2 This is a flowchart of the PENet network processing of the present invention.

[0036] Figure 3 2 is a diagram of the detection steps 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 concept of this method is to collect dark-light data, train an object detector, and then combine the object detector with a drone to detect aircraft skin damage. First, a drone is used to fly around the aircraft to collect dark-light aircraft skin damage source data. This data is then processed using annotation tools to create a dark-light aircraft skin damage dataset. After preprocessing, the image is fed into a PENet for component decomposition. Detail Processing Modules (DPMs) and Low-Frequency Enhancement Filters (LEFs) are used to enhance the components, improving the potential information capture of objects in low-light images. The image is then fed into a CSPDarknet backbone for feature extraction. The extracted target features are then fed into the PSA channel attention mechanism. The SPC module splits the channels, extracting multi-scale features based on the spatial information in each channel's feature map. The SEWeight module extracts channel attention from feature maps at different scales to obtain enhanced aircraft skin damage features. Softmax is then used to obtain new attention weights after multi-scale channel interaction, resulting in a feature map weighted by multi-scale feature information. This effectively detects skin damage at different scales. Furthermore, the SCConv module replaces the C2F convolutional block to achieve lightweight performance. Finally, an improved YOLOv8 network is trained to obtain an object detector. During training, the network model's loss function is adjusted to ensure that the trained network's detection performance is more consistent with the dataset. The trained object detector is then deployed on a drone, which performs a flyby inspection of aircraft skin damage and obtains detection results.

[0039] This embodiment provides a method for detecting aircraft skin damage under dark light conditions, such as Figure 1 As shown, the following steps are included:

[0040] Step 1: The UAV flies around the aircraft to obtain original image data of aircraft skin damage;

[0041] In this implementation, the aircraft skin damage data collected by the drone during its flight is used as part of the data. To avoid a single background of aircraft skin damage, imbalanced categories, and prevent model overfitting, sufficient data must be collected. Therefore, data augmentation or additional data from existing databases is employed.

[0042] Step 2: Preprocess the collected image data, complete screening and labeling, and create the aircraft skin damage dataset required for training the network.

[0043] In this embodiment, the image is annotated to ensure that the data of each category is balanced to avoid imbalance of category data that may lead to poor training results, and invalid data is eliminated. The location information of the annotated object, namely the bounding box coordinates and identity serial number, is annotated.

[0044] Step 3: Input the training dataset image data into the PENet network to improve the dark light information capture rate.

[0045] In this embodiment, the training set data is input into the PENet network, and the components are enhanced by the detail processing module (DPM) and low frequency enhancement filter (LEF) in the network. By decomposing and reconstructing the Laplacian pyramid, lightweight and effective capture are achieved, thereby improving the performance of subsequent detectors. Figure 2 shown.

[0046] Step 3.1: Apply a Gaussian filter (with a 5×5 kernel) to the input image. Each time a Gaussian pyramid operation is performed, the image width and height are halved, reducing the resolution to 1 / 4 of the original. The lost information forms the components of the Laplacian pyramid. During image reconstruction, the corresponding reverse operation is performed to restore the high-resolution image.

[0047] Step 3.2: Different components in the pyramid enter the DPM and LEF modules in parallel. The DPM is designed into two branches: an edge enhancement branch and a context branch. The context branch uses a residual block to process features before and after acquiring long-range dependencies. The first residual block changes the number of feature channels from 3 to 32, and the second residual block changes the number of feature channels from 32 to 3, effectively enhancing information globally. The edge branch uses the Sobel operator (a discrete operator) to calculate image gradients in both the horizontal and vertical directions to obtain edges and enhance the texture of components.

[0048] Step 3.3: Different components in the pyramid enter the DPM and LEF modules in parallel. The main purpose of the low-frequency enhancement filter (LEF) is to capture low-frequency information in image components. The input components are first converted to different scales through the convolution layer. Then, a dynamic low-pass filter and average pooling are used to capture and filter low-frequency information. Specific pooling sizes (such as 1×1, 2×2, 3×3, and 6×6) are used, combined with bilinear interpolation sampling to form low-pass filters of different scales. These filters are finally combined through tensor splicing to restore the original image size.

[0049] Step 4: Input the training dataset processed by PENet into the improved yolov8 network and CSPDarknet backbone network for feature extraction, and introduce the extracted target features into the PSA channel attention mechanism for feature enhancement.

[0050] In this embodiment, the data is input into the CSPDarknet network for feature extraction, and then the extracted target features are introduced into the PSA channel attention mechanism to enhance the model's ability to model spatial context.

[0051] Step 4.1: The CSPDarknet backbone network uses a combination of multiple convolutional layers and residual blocks. The convolutional layers use 3x3 or 5x5 convolution kernels with a stride of 1 or 2. The residual block consists of two convolutional layers and a skip connection. The convolutional layer adds the input features to the output features of the convolutional layer through the skip connection, effectively solving the gradient vanishing problem caused by increased network depth and enabling the network to learn deeper features. After the input image is processed by the initial convolutional layer, its size is halved and the number of channels is increased. It then enters the residual block structure. As the network depth increases, the image resolution gradually decreases and the number of channels in the feature map gradually increases. In this process, the network continuously extracts local and global features of the image, providing rich feature information for subsequent object detection.

[0052] Step 4.2: Input the learned features into the PSA channel. The PSA channel attention mechanism consists of two parts: the SPC and SEWeight modules. The SPC module first splits the input feature map into several channels, evenly dividing the number of channels into several groups. For each channel group, convolution operations are performed using convolution kernels of different sizes (such as 3x3, 5x5, 7x7, etc.) to obtain spatial feature information at different scales. The feature maps output by the convolution kernels of different scales are spliced together to obtain a feature map that integrates multi-scale spatial information, allowing the network to simultaneously focus on skin damage features of different sizes.

[0053] Step 4.3: The SEWeight module performs a global average pooling operation on the feature map output by the SPC module to obtain global feature information for each channel (where and are the height and width of the feature map, respectively). This global feature information is then processed through two fully connected layers. The first fully connected layer maps the global feature information to a lower dimension using the ReLU activation function; the second fully connected layer maps it back to the original number of channels using the Sigmoid activation function. Through these two fully connected layers, attention weights are generated for each channel.

[0054] Step 4.4: Multiply the attention weights generated by the SEWeight module by the feature map output by the SPC module element-by-element to obtain a weighted feature map. A Softmax operation is performed on the weighted feature map along the channel dimension to obtain the final attention weights after multi-scale channel interaction. Finally, the attention weights are again element-wise multiplied by the feature map to obtain a feature map after multi-scale feature information attention weighting, effectively enhancing the detection capability of skin damage at different scales.

[0055] Step 5: Use the SCConv module to replace 30% of the c2f convolution blocks in the neck and head networks, which are responsible for abstracting semantic information and are more sensitive to multi-scale feature fusion. This module optimizes the feature extraction capabilities of specific layers and reduces spatial and channel redundancy.

[0056] In this implementation, the improved YOLOv8 network uses SCConv modules to replace 30% of the C2F convolutional blocks in the NECK and HEAD networks, which are responsible for abstracting semantic information and are more sensitive to multi-scale feature fusion. However, since the shallow layers are responsible for extracting underlying features such as edges and textures and are insensitive to changes in convolution kernel scale, the C2F residual connections and CSP structure are retained to maintain the stability of underlying features. This enables the network to learn richer features at different scales, improves its ability to express skin damage characteristics, and reduces spatial and channel redundancy.

[0057] Step 6: Use the obtained aircraft skin damage features to train the target detection network to obtain a trained target detector. The detection steps are shown in the attached figure. Figure 3 shown.

[0058] 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.

[0059] 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.

[0060] 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 dark light 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 raw 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 PENet network to improve the dark light information capture rate. The image components are optimized jointly by the detail processing module (DPM) and the low-frequency enhancement filter (LEF). The detail processing module (DPM) adopts a dual-branch structure to perform edge enhancement and upper and lower feature extraction respectively. The low-frequency enhancement filter (LEF) captures and filters low-frequency information through dynamic low-pass filtering and average pooling. S4. Input the training data set processed by PENet into the improved yolov8 network, perform feature extraction through the CSPDarknet backbone network, and introduce the extracted target features into the PSA channel attention mechanism for feature enhancement. The PSA channel attention mechanism includes the SPC module and the SEWeight module; S5. In the neck and head networks of YOLOv8, SCConv modules are used to replace 30% of the c2f convolution blocks to optimize the feature extraction capabilities of specific layers and reduce spatial and channel redundancy. S6. Using the obtained aircraft skin damage features, the target detection network is trained to obtain a trained 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 suitable for dark light conditions according to claim 1 is characterized in that: The processing of the PENet network in step S3 includes: S3.

1. Use Gaussian extraction on the input image. After each Gaussian pyramid operation, the width and height of the image are halved, and the resolution is reduced to 1 / 4 of the original. The lost information constitutes the components of the Laplacian pyramid. When reconstructing the image, the corresponding reverse operation is performed to restore the high-resolution image. S3.

2. Different components in the pyramid enter the detail processing module (DPM) and low-frequency enhancement filter (LEF) in parallel. The detail processing module (DPM) includes an edge enhancement branch and a context branch. The context branch uses a residual block to process features before and after obtaining long-range dependencies. The first residual block changes the number of feature channels from 3 to 32, and the second residual block changes the number of feature channels from 32 to 3, thereby globally enhancing information. The edge enhancement branch uses the Sobel operator to calculate image gradients in both horizontal and vertical directions to obtain edges and enhance the texture of components. S3.

3. The low-frequency enhancement filter (LEF) is used to capture low-frequency information in image components. First, the input components are converted into different scales through the convolution layer, and then the low-frequency information is captured and filtered using a dynamic low-pass filter and average pooling. The specific pooling size is combined with bilinear interpolation sampling to form low-pass filters of different scales. The low-pass filters are finally integrated through tensor splicing to restore the original image size.

3. The deep learning aircraft skin damage detection method applicable to dark light conditions according to claim 2 is characterized in that: The specific operations of step S4 are: S4.

1. The CSPDarknet backbone network adopts a combination of multiple convolutional layers and residual blocks. The residual block consists of two convolutional layers and a skip connection. The convolutional layer adds the input features to the convolutional layer output features through the skip connection, enabling the network to learn deeper features. After the input image is processed by the initial convolutional layer, its size is halved and the number of channels is increased. It then enters the residual block structure. As the network depth increases, the image resolution gradually decreases and the number of channels in the feature map gradually increases. In this process, the network continuously extracts local and global features of the image, providing rich feature information for subsequent target detection. S4.

2. The learned features are input into the PSA channel. The SPC module first performs channel segmentation on the input feature map, evenly dividing the number of channels into several groups. For each group of channels, convolution kernels of different sizes are used to perform convolution operations to obtain spatial feature information of different scales. The feature maps output by the convolution kernels of different scales are spliced together to obtain a feature map that integrates multi-scale spatial information, so that the network can simultaneously focus on skin damage features of different sizes. S4.3, SEWeight module performs global average pooling operation on the feature map output by SPC module to obtain global feature information of each channel. Then the global feature information is processed by two fully connected layers. The first fully connected layer maps the global feature information to a lower dimension, and the activation function is ReLU; the second fully connected layer maps it back to the original number of channels, and the activation function is Sigmoid. Through the two fully connected layers, attention weights are generated for each channel. S4.

4. Multiply the attention weights generated by the SEWeight module by the feature map output by the SPC module element by element to obtain the weighted feature map. Perform a Softmax operation on the weighted feature map in the channel dimension to obtain the final attention weights after multi-scale channel interaction. Finally, multiply the attention weights by the feature map element by element again to obtain the feature map after multi-scale feature information attention weighting, which effectively enhances the detection capability of skin damage at different scales.

4. The deep learning aircraft skin damage detection method suitable for dark light conditions according to claim 3 is characterized in that: The convolution layer described in step S4.1 uses a 3x3 or 5x5 convolution kernel with a step size of 1 or 2.

5. The deep learning aircraft skin damage detection method applicable to dark light conditions according to claim 2 is characterized in that: The size of the Gaussian kernel in step S3.1 is 5×5.

6. The deep learning aircraft skin damage detection method applicable to dark light conditions according to claim 1 is characterized in that: In step S2, the target is marked with position information, which includes bounding box coordinates and an identity number.

Images