Aircraft skin damage detection method and system based on improved YOLO11

Abstract

The invention belongs to the technical field of aircraft damage detection, and discloses an improved YOLO11-based aircraft skin damage detection method and system, and the method comprises the steps: obtaining an aircraft skin image; constructing an aircraft skin damage detection model; and based on the aircraft skin damage detection model, analyzing the aircraft skin image, and identifying different types of damages such as paint layer peeling, dents, cracks, corrosion and scratches. The aircraft skin damage detection efficiency can be effectively improved, manual intervention is reduced, and the flight safety is improved.

Description

Technical Field

[0001] This invention belongs to the field of aircraft damage detection technology, specifically relating to an aircraft skin damage detection method and system based on the improved YOLO11. Background Technology

[0002] As aircraft age, damage to aircraft skin has become a significant factor affecting flight safety. Previous research has addressed the occurrence and treatment of aircraft skin damage. Zhou Yuxuan proposed that aircraft skin, as a major external component, bears loads and maintains aerodynamic performance. Compared to traditional visual inspection, automated ultrasonic testing of internal damage and machine vision testing of external damage offer higher efficiency and consistency. Yang Zeqing analyzed the metal corrosion characteristics and typical corrosion sites on wing surfaces, reviewed corrosion identification methods based on visual inspection, and focused on the research progress in traditional image processing, machine learning, and deep learning. Shen proposed that the damage process of multiple parts of an aircraft under the same operating environment is dependent on the load mechanism. Chen Daixin addressed the inadequacy and low efficiency of current methods for detecting common aircraft surface damage by designing two types of rapid detection tools and conducting field applications. Bu Xiaoyan addressed the shortcomings of existing algorithms in detecting small targets on aircraft skin by proposing an improved EfficientDet model. Liu Qingsong proposed an aircraft surface damage detection method based on spectral matching technology. Zhang Dongping points out that traditional aircraft skin damage detection relies on human visual observation, which is prone to efficiency reduction due to fatigue and cognitive limitations. To address this issue, the paper proposes an aircraft skin damage detection algorithm based on an improved YOLOv8. Li Bo addresses the problems of insufficient accuracy and large model parameter count in existing target detection methods by proposing a method called GSV-YOLO for aircraft landing gear damage detection. This method is based on YOLOv7-tiny and has been improved.

[0003] The research on an automated aircraft skin damage detection method based on the improved YOLOv11 is of great significance. This method, by leveraging the powerful feature extraction and pattern recognition capabilities of deep learning models, can automatically identify various types of damage on aircraft skin, greatly improving detection efficiency and accuracy. Compared with traditional manual inspection or inefficient detection methods, automated inspection can significantly reduce human error, improve flight safety, and reduce maintenance costs. The application of this technology will facilitate rapid detection in aircraft manufacturing and maintenance processes, providing strong support for regular aircraft inspections and safety assurance. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an aircraft skin damage detection method and system based on the improved YOLO11. Based on deep learning technology, it analyzes aircraft skin images to accurately identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches. This effectively improves the detection efficiency of aircraft skin damage, reduces manual intervention, and enhances flight safety.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for detecting aircraft skin damage based on the improved YOLO11, the method comprising:

[0007] Obtain aircraft skin images;

[0008] Construct an aircraft skin damage detection model;

[0009] Based on the aircraft skin damage detection model, aircraft skin images are analyzed to identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches.

[0010] Preferably, the aircraft skin damage detection model is constructed using an improved YOLO11s network; the improved YOLO11s network uses ADNet as the initialization layer, replaces some convolutions with OREPA modules, and combines SPPF with C2PSA_LSKA.

[0011] Preferably, the ADNet architecture includes sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB).

[0012] Sparse blocks (SB) are used to combine dilated convolution and regular convolution to improve denoising performance while maintaining computational efficiency.

[0013] Feature Enhancement Blocks (FEBs) are used to integrate global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs.

[0014] Attention blocks (AB) are used to process real-world noisy images and blind denoising tasks by extracting noise information from complex backgrounds, focusing on relevant features and suppressing irrelevant noise.

[0015] The Reconstructed Block (RB) is used to combine noise mapping with the input image to ensure that the structure of the output image remains unchanged while noise is minimized.

[0016] Preferably, the core of the OREPA module is a 3x3 convolution operation used to extract local features from the input data;

[0017] The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers.

[0018] Among them, convolutional layers are used to increase the non-linear expressive power of the network; pooling layers are used to reduce the computational cost and enhance the invariance of features by reducing the size of feature maps; frequency prior layers add prior information about the frequency of data to help the network respond to features of different frequencies; scaling layers are used to adjust the scale of feature maps, and through the squeezing operation, they compress and integrate the feature maps generated by multiple convolutional layers and pooling layers to finally output a feature representation.

[0019] Preferably, the LSKA attention mechanism decomposes the traditional 2D convolution kernel into two 1D convolution kernels, which perform convolution operations in the horizontal and vertical directions respectively.

[0020] The present invention also provides an aircraft skin damage detection system based on an improved YOLO11, the system being used to implement the aforementioned method, the system comprising: an acquisition module, a construction module, and a detection module;

[0021] The acquisition module is used to acquire images of the aircraft skin;

[0022] The building module is used to build an aircraft skin damage detection model;

[0023] The detection module is used to analyze aircraft skin images based on an aircraft skin damage detection model to identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches.

[0024] Preferably, the aircraft skin damage detection model is constructed using an improved YOLO11s network; the improved YOLO11s network uses ADNet as the initialization layer, replaces some convolutions with OREPA modules, and combines SPPF with C2PSA_LSKA.

[0025] Preferably, the ADNet architecture includes sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB).

[0026] Sparse blocks (SB) are used to combine dilated convolution and regular convolution to improve denoising performance while maintaining computational efficiency.

[0027] Feature Enhancement Blocks (FEBs) are used to integrate global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs.

[0028] Attention blocks (AB) are used to process real-world noisy images and blind denoising tasks by extracting noise information from complex backgrounds, focusing on relevant features and suppressing irrelevant noise.

[0029] The Reconstructed Block (RB) is used to combine noise mapping with the input image to ensure that the structure of the output image remains unchanged while noise is minimized.

[0030] Preferably, the core of the OREPA module is a 3x3 convolution operation used to extract local features from the input data;

[0031] The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers.

[0032] Among them, convolutional layers are used to increase the non-linear expressive power of the network; pooling layers are used to reduce the computational cost and enhance the invariance of features by reducing the size of feature maps; frequency prior layers add prior information about the frequency of data to help the network respond to features of different frequencies; scaling layers are used to adjust the scale of feature maps, and through the squeezing operation, they compress and integrate the feature maps generated by multiple convolutional layers and pooling layers to finally output a feature representation.

[0033] Preferably, the LSKA attention mechanism decomposes the traditional 2D convolution kernel into two 1D convolution kernels, which perform convolution operations in the horizontal and vertical directions respectively.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention proposes an improved YOLO11s model. By using ADNet as the initialization layer, introducing the OREPA module to replace some convolutions, and combining SPPF and C2PSA_LSKA, the accuracy and efficiency of aircraft skin damage detection are significantly improved. Experimental results show that compared to the basic YOLO11s model, the improved model achieves better performance across various metrics, especially in aircraft skin damage detection, where it exhibits stronger robustness and accuracy. This improved approach optimizes feature extraction and information fusion capabilities, providing solid technical support for the practical application of intelligent aircraft skin damage detection technology. Attached Figure Description

[0036] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the basic framework of YOLO11 according to an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the improved YOLO11 model according to an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of ADNet according to an embodiment of the present invention;

[0040] Figure 4This is a schematic diagram of an embodiment of the OREPA invention;

[0041] Figure 5 This is a schematic diagram of an LSKA embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of the performance result analysis curves of an embodiment of the present invention;

[0043] Figure 7 This is a visual diagram illustrating the recognition results of an embodiment of the present invention;

[0044] Figure 8 This is a visualization diagram of Grad_CAM according to an embodiment of the present invention, wherein (a) is the original image, (b) is CAM, (c) is the gradient map, (d) is CAM + gradient, and (e) is CAM color space conversion. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Example 1

[0048] This invention provides a method for detecting aircraft skin damage based on an improved YOLO11, the method comprising:

[0049] Obtain aircraft skin images;

[0050] Construct an aircraft skin damage detection model;

[0051] Based on the aircraft skin damage detection model, aircraft skin images are analyzed to identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches.

[0052] In this embodiment, the aircraft skin damage dataset includes three parts: a training set, a validation set, and a test set. The training set contains 10,842 images for model training and learning; the validation set contains 1,250 images for evaluating model performance and adjusting hyperparameters during training; and the test set contains 553 images for final model evaluation and performance verification.

[0053] As shown in Table 1, the dataset is categorized into several different types of damage, including: "Paint Peeling": Coating peeling off the aircraft surface, which may be caused by external impact, environmental factors, or coating quality issues. It affects appearance and can lead to corrosion problems. "Dents or Deformations": Physical deformation of the aircraft's outer surface caused by impact or pressure. This type of damage typically affects structural integrity, potentially leading to uneven load distribution on the aircraft structure and thus affecting flight safety. More severe dents or deformations may affect the aircraft's aerodynamic performance or cause fatigue damage. "Cracks": Tiny cracks or fractures that appear in materials under external force or during long-term use. If cracks are not repaired in time, they can lead to material breakage or destruction, potentially threatening the structural safety of the aircraft. "Corrosion": Reaction of metal surfaces with moisture, salt, or chemicals, leading to material degradation, impairing structural strength, and increasing maintenance costs. "Scratches": Shallow surface damage that, while not directly affecting the structure, weakens coating protection and increases the risk of corrosion. Each type of damage is used as a key type to train the model, helping the model identify and classify various types of damage to the aircraft surface.

[0054] Table 1 Dataset Classification

[0055]

[0056] In this embodiment, the aircraft skin damage detection model is built using an improved YOLO11s network. The YOLO11s architecture integrates an improved backbone network and a neck structure. The backbone network uses depthwise separable convolutions and cross-stage partial connections, with the C3K2 module enhancing feature extraction capabilities. As an evolution of the CSP residual module, the C3K2 module optimizes information flow and reduces computational cost by using smaller convolutional kernels, while simultaneously improving feature representation capabilities. The neck structure introduces the SPFF (Spatial Pyramid Fast Pooling) module, enhancing the detection capability for objects of different sizes, especially excelling in small object recognition. Furthermore, the addition of the C2PSA (Cross-Stage Partial Spatial Attention) module further optimizes the fusion of channel and spatial information in the feature maps, improving the model's perception capabilities in complex scenes. Through a refined architecture design and optimized training process, YOLO11s achieves faster processing speed and higher average accuracy while maintaining real-time performance. The basic framework of YOLO11s is as follows: Figure 1 As shown.

[0057] The improved YOLO11s model design includes: The improved model enhances dynamic adaptability to input images through an ADNet initialization layer; the OREPA module replaces some convolutional layers to strengthen feature extraction capabilities of downsampling paths, improving small object detection performance; and it combines SPPF and C2PSA_LSKA modules, utilizing spatial pyramid pooling for multi-scale feature aggregation and long-distance dependency modeling through attention mechanisms to significantly optimize multi-scale feature fusion, thereby improving the model's ability to represent complex scenes while maintaining a lightweight design. The improved YOLO11s model is shown below. Figure 2 As shown.

[0058] The ADNet architecture comprises four main modules: Sparse Block (SB), Feature Augmentation Block (FEB), Attention Block (AB), and Reconstruction Block (RB). The Sparse Block combines dilated convolutions and regular convolutions to effectively maintain computational efficiency while improving denoising performance. The Feature Augmentation Block integrates global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs. The Attention Block accurately extracts noise information from complex backgrounds, handling both real-world noisy images and blind denoising tasks, focusing on relevant features and suppressing irrelevant noise. The Reconstruction Block combines noise mapping with the input image, ensuring the output image structure remains unchanged while minimizing noise. Compared to other CNN-based models, ADNet maintains high image quality while removing noise. The integration of global and local features through the FEB, along with its robust attention mechanism, makes the model's denoising performance in complex scenes even more outstanding. The structure of ADNet is as follows: Figure 3 As shown.

[0059] OREPA (Online Convolutional Reparameterization) is a method that optimizes the cost and complexity of deep learning model training, primarily through two stages. The first stage employs an online optimization phase, utilizing special linear scaling layers to enhance the performance of online modules, thereby improving training efficiency. The second stage significantly reduces memory consumption and computational overhead during training by compressing complex modules into single convolutional operations, thus accelerating the training process. The structure of OREPA is as follows: Figure 4 As shown.

[0060] The core of the OREPA module is a 3x3 convolutional operation used to extract local features from the input data. The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers. Convolutional layers increase the network's non-linear expressive power; pooling layers reduce computation and enhance feature invariance by decreasing the size of the feature maps; and frequency prior layers incorporate prior information about data frequencies, helping the network better respond to features of different frequencies. Scaling layers help adjust the scale of the feature maps, thereby improving sensitivity to features at different scales and enhancing the model's robustness. A squeezing operation compresses and integrates the feature maps generated by multiple convolutional and pooling layers, ultimately outputting a simplified feature representation.

[0061] The structure of the LSKA attention mechanism is as follows: Figure 5 As shown.

[0062] The LSKA attention mechanism aims to address the high computational and memory requirements faced by traditional Large Kernel Attention (LKA) modules when applied to Visual Attention Networks (VANs). The core innovation of the LSKA attention mechanism lies in decomposing the traditional 2D convolutional kernel into two 1D convolutional kernels, performing convolution operations in the horizontal and vertical directions respectively. This significantly reduces the number of parameters and computational complexity, while improving computational efficiency.

[0063] In this embodiment, to ensure the repeatability and environmental consistency of the experiment, Table 2 details the hardware and software configuration for this experiment. Through a unified hardware and software configuration, the repeatability of the experimental results is effectively guaranteed, which helps subsequent researchers to verify and reproduce the experiment.

[0064] Table 2 Experimental Environment

[0065]

[0066] To comprehensively evaluate the effectiveness of the proposed method, this invention employs some common evaluation criteria.

[0067]

[0068] Table 3 presents the comparative experimental results of different base models on multiple evaluation metrics. By comparing the performance metrics of each model, it can be seen that YOLO11s performs excellently across all performance metrics, especially in recall (0.942) and F1 score (0.975), far surpassing other models and demonstrating its powerful capabilities in object detection tasks. Compared with other models, YOLO11s has a more reasonable configuration in terms of parameter count (9,414,735) and FLOPs (21.3G), and a moderate model size (19.2MB), which can optimize computational and memory overhead while maintaining high performance.

[0069] Table 3 Comparison Experiments of Basic Models

[0070]

[0071] As shown in the results of the optimizer comparison experiment in Table 4, the SGD optimizer performed best in all performance indicators, demonstrating its advantage in improving model accuracy.

[0072] Table 4. Comparison of Optimizers

[0073]

[0074] As shown in Table 5, the training results of 640×640 images are significantly better than those of 320×320 images. A reasonable input image size can improve the detection performance of the model.

[0075] Table 5 Image Size Comparison Test

[0076]

[0077] As shown in Table 6, the ablation experiment results demonstrate that the AOL-YOLO algorithm performs excellently across various performance metrics, particularly in accuracy, where P (0.987), R (0.935), and F1 (0.976) all achieve high levels. It also exhibits superior processing speed in FPS (144.43 frames / second). Compared to other models, AOL-YOLO achieves a good balance between accuracy and speed. Although its parameter count and FLOPs / G value are similar to other models, it performs better in terms of memory usage (28.4MB) and frame rate. By optimizing its architecture and algorithm, the AOL-YOLO algorithm improves the overall performance of the model, reducing computational resource consumption while ensuring efficient detection, making it suitable for scenarios with high real-time requirements.

[0078] Table 6 Ablation Experiment

[0079]

[0080] In the results visualization and analysis section, this invention visually displays the performance of each model on different evaluation metrics by plotting performance result analysis curves. These include precision (P), recall (R), and mean precision (mAP@50). By comparing the trends of different models on these metrics, the differences in performance of each model during training and testing are reflected. Figure 6 As shown.

[0081] Figure 7 A visualization of the recognition results is shown. Figure 7The results of class identification for aircraft skin images and their corresponding confidence scores are shown in the visualization. These visualizations clearly show the model's prediction accuracy and confidence score distribution for different classes.

[0082] Figure 8 It is a heatmap visualization. Figure 8 (a) is an image generated using Class Activation Mapping (CAM) technology, used to visualize which regions the model focuses on; (b) is a gradient map, calculated using gradients or other methods, showing the model's response to the input image; (c) is an image combining Class Activation Mapping and Gradient Mapping, providing a more comprehensive view of the model's focus areas and gradient information; and (d) is an image generated after color space conversion of the CAM image, ensuring that the image can be correctly displayed in OpenCV.

[0083] This invention proposes an improved YOLO11s model. By using ADNet as the initialization layer, introducing the OREPA module to replace some convolutions, and combining SPPF and C2PSA_LSKA, the accuracy and efficiency of aircraft skin damage detection are significantly improved. Experimental results show that compared to the basic YOLO11s model, the improved model achieves better performance across various metrics, especially in aircraft skin damage detection, where it exhibits stronger robustness and accuracy. This improved approach optimizes feature extraction and information fusion capabilities, providing solid technical support for the practical application of intelligent aircraft skin damage detection technology.

[0084] This invention has achieved significant results, with the improved model showing marked improvements in accuracy and efficiency. However, there is still room for optimization in terms of computational resource consumption and real-time performance. Future research will focus on enhancing the model's real-time detection capabilities, improving its adaptability to diverse damage types, and further enhancing its stability, laying a more solid foundation for intelligent detection systems in practical applications.

[0085] Example 2

[0086] The present invention also provides an aircraft skin damage detection system based on an improved YOLO11, the system being used to implement the method of Embodiment 1, the system comprising: an acquisition module, a construction module, and a detection module;

[0087] The acquisition module is used to acquire images of the aircraft skin.

[0088] Modules for building aircraft skin damage detection models;

[0089] The detection module is used to analyze aircraft skin images based on the aircraft skin damage detection model and identify different types of damage such as paint peeling, dents, cracks, corrosion, and scratches.

[0090] In this embodiment, the aircraft skin damage detection model is constructed using an improved YOLO11s network; the improved YOLO11s network uses ADNet as the initialization layer, replaces some convolutions with OREPA modules, and combines SPPF with C2PSA_LSKA.

[0091] In this embodiment, the ADNet architecture includes sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB).

[0092] Sparse blocks (SB) are used to combine dilated convolution and regular convolution to improve denoising performance while maintaining computational efficiency.

[0093] Feature Enhancement Blocks (FEBs) are used to integrate global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs.

[0094] Attention blocks (AB) are used to process real-world noisy images and blind denoising tasks by extracting noise information from complex backgrounds, focusing on relevant features and suppressing irrelevant noise.

[0095] The Reconstructed Block (RB) is used to combine noise mapping with the input image to ensure that the structure of the output image remains unchanged while noise is minimized.

[0096] In this embodiment, the core of the OREPA module is a 3x3 convolution operation used to extract local features from the input data;

[0097] The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers.

[0098] Among them, convolutional layers are used to increase the non-linear expressive power of the network; pooling layers are used to reduce the computational cost and enhance the invariance of features by reducing the size of feature maps; frequency prior layers add prior information about the frequency of data to help the network respond to features of different frequencies; scaling layers are used to adjust the scale of feature maps, and through the squeezing operation, they compress and integrate the feature maps generated by multiple convolutional layers and pooling layers to finally output a feature representation.

[0099] In this embodiment, the LSKA attention mechanism decomposes the traditional 2D convolution kernel into two 1D convolution kernels, which perform convolution operations in the horizontal and vertical directions, respectively.

[0100] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for detecting aircraft skin damage based on an improved YOLO11, characterized in that, The method includes: Obtain aircraft skin images; Construct an aircraft skin damage detection model; Based on the aircraft skin damage detection model, aircraft skin images are analyzed to identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches.

2. The method according to claim 1, characterized in that, The aircraft skin damage detection model is constructed using an improved YOLO11s network. The improved YOLO11s network uses ADNet as the initialization layer, replaces some convolutions with OREPA modules, and combines SPPF with C2PSA_LSKA.

3. The method according to claim 2, characterized in that, The architecture of ADNet includes sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB). Sparse blocks (SB) are used to combine dilated convolution and regular convolution to improve denoising performance while maintaining computational efficiency. Feature Enhancement Blocks (FEBs) are used to integrate global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs. Attention blocks (AB) are used to process real-world noisy images and blind denoising tasks by extracting noise information from complex backgrounds, focusing on relevant features and suppressing irrelevant noise. The Reconstructed Block (RB) is used to combine noise mapping with the input image to ensure that the structure of the output image remains unchanged while noise is minimized.

4. The method according to claim 2, characterized in that, The core of the OREPA module is a 3x3 convolution operation used to extract local features from the input data; The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers. Among them, convolutional layers are used to increase the non-linear expressive power of the network; pooling layers are used to reduce the computational cost and enhance the invariance of features by reducing the size of feature maps; frequency prior layers add prior information about the frequency of data to help the network respond to features of different frequencies; scaling layers are used to adjust the scale of feature maps, and through the squeezing operation, they compress and integrate the feature maps generated by multiple convolutional layers and pooling layers to finally output a feature representation.

5. The method according to claim 2, characterized in that, The LSKA attention mechanism decomposes the traditional 2D convolution kernel into two 1D convolution kernels, which perform convolution operations in the horizontal and vertical directions, respectively.

6. An aircraft skin damage detection system based on an improved YOLO11, said system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: an acquisition module, a construction module, and a detection module; The acquisition module is used to acquire images of the aircraft skin; The building module is used to build an aircraft skin damage detection model; The detection module is used to analyze aircraft skin images based on an aircraft skin damage detection model to identify different types of damage, such as paint peeling, dents, cracks, corrosion, and scratches.

7. The system according to claim 6, characterized in that, The aircraft skin damage detection model is constructed using an improved YOLO11s network. The improved YOLO11s network uses ADNet as the initialization layer, replaces some convolutions with OREPA modules, and combines SPPF with C2PSA_LSKA.

8. The system according to claim 7, characterized in that, The architecture of ADNet includes sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB). Sparse blocks (SB) are used to combine dilated convolution and regular convolution to improve denoising performance while maintaining computational efficiency. Feature Enhancement Blocks (FEBs) are used to integrate global and local features through long paths, enhancing the model's ability to reconstruct noisy inputs. Attention blocks (AB) are used to process real-world noisy images and blind denoising tasks by extracting noise information from complex backgrounds, focusing on relevant features and suppressing irrelevant noise. The Reconstructed Block (RB) is used to combine noise mapping with the input image to ensure that the structure of the output image remains unchanged while noise is minimized.

9. The system according to claim 7, characterized in that, The core of the OREPA module is a 3x3 convolution operation used to extract local features from the input data; The OREPA module consists of a combination of convolutional layers, pooling layers, and frequency prior layers. Among them, convolutional layers are used to increase the non-linear expressive power of the network; pooling layers are used to reduce the computational cost and enhance the invariance of features by reducing the size of feature maps; frequency prior layers add prior information about the frequency of data to help the network respond to features of different frequencies; scaling layers are used to adjust the scale of feature maps, and through the squeezing operation, they compress and integrate the feature maps generated by multiple convolutional layers and pooling layers to finally output a feature representation.

10. The system according to claim 7, characterized in that, The LSKA attention mechanism decomposes the traditional 2D convolution kernel into two 1D convolution kernels, which perform convolution operations in the horizontal and vertical directions, respectively.

Images