An aircraft skin damage detection method and system based on an FC-YOLO network model

Abstract

A method and system for aircraft skin damage detection based on an FC-YOLO network model is disclosed. The method includes steps such as obtaining the original dataset; dividing the dataset into training and validation sets; constructing an initial FC-YOLO network model; obtaining the final FC-YOLO network model; and outputting the aircraft skin damage identification results from the image to be detected. This invention utilizes the ELAN-C module to focus attention on the channel and location information of the target using global pooling and unidirectional pooling, thereby assigning higher weights to key channels and locations of features and improving the target detection capability. The FC-YOLO neck feature fusion network avoids the reduction or loss of feature information of small targets during multi-scale feature fusion, redistributing weights to the inputs of different layers and strengthening the FC-YOLO network model's feature extraction capability for multi-scale targets. The SI oU loss function can reconstruct the penalty term of the loss function through the angle, distance, and shape losses between the target bounding box and the ground truth bounding box, accelerating the convergence of the loss function and optimizing the training process.

Description

Technical Field

[0001] This invention belongs to the field of target detection and image classification technology in computer vision, and specifically relates to a method and system for detecting aircraft skin damage based on the FC-YOLO network model. Background Technology

[0002] Aircraft skin is a crucial component affecting the structural safety of an aircraft. Fixed to the aircraft frame, it shapes the aircraft's aerodynamic shape and bears and transmits aerodynamic loads. During flight, due to weather conditions and complex loads, damage to the aircraft skin is inevitable, posing significant safety hazards to normal aircraft operation. Therefore, regular damage inspection of aircraft skin is of paramount importance for safe aircraft operation. Traditional damage inspection methods are complex, costly, require highly skilled personnel, and are difficult to implement on a large scale. Deep learning-based inspection technology has emerged as a significant trend in the automated inspection of aircraft skin damage.

[0003] Object detection is an important branch of computer vision. In recent years, deep learning-based object detection algorithms have developed rapidly and achieved good detection results. Deep learning-based object detection algorithms are divided into two categories: single-stage detection and two-stage detection. Two-stage object detection algorithms first generate candidate boxes and then classify these candidate boxes; typical algorithms include the Region Convolutional Neural Network (R-CNN) series. Single-stage object detection algorithms do not need to generate candidate boxes; they directly generate the object's location and confidence score, and can obtain the prediction result in a single detection; typical algorithms include the SSD and YOLO series.

[0004] The YOLOv7 network model, as a state-of-the-art single-stage object detection algorithm, boasts high accuracy and high detection speed, thus being widely used in industrial fields. However, due to the varying sizes of damage on aircraft skin, for small damage, during feature extraction, in order to obtain higher-level semantic information, the positional information of lower-level small targets is gradually lost as the number of convolutions and downsampling iterations increases. This feature inconsistency can cause small targets to be mistaken for background when the algorithm detects images containing targets of different sizes. Secondly, because aircraft skin has a large surface area, the target localization performance is also crucial during detection. To avoid false positives and false negatives, attention must be paid to the target's positional information. However, no effective method has yet been found to address this. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a method and system for detecting aircraft skin damage based on the FC-YOLO network model.

[0006] To achieve the above objectives, the aircraft skin damage detection method based on the FC-YOLO network model provided by this invention includes the following steps performed in sequence:

[0007] 1) Collect raw images of aircraft skin damage, and then filter, classify, and manually annotate them to obtain the raw dataset;

[0008] 2) Use data augmentation techniques to augment the images in the original dataset, and divide the augmented images into training and validation sets proportionally;

[0009] 3) The original YOLOv7 model was improved to construct the initial FC-YOLO network model;

[0010] 4) Input the training set obtained in step 2) into the initial FC-YOLO network model above for training, and then use the validation set obtained in step 2) to evaluate the training results, and finally obtain the FC-YOLO network model.

[0011] 5) The camera is used to acquire images of the aircraft skin damage to be detected in real time and input them into the FC-YOLO network model mentioned above for damage identification. Finally, the FC-YOLO network model outputs the aircraft skin damage identification results in the images of the aircraft skin damage to be detected.

[0012] In step 1), the method for acquiring raw images of aircraft skin damage and then filtering, classifying, and manually annotating them to obtain the raw dataset is as follows:

[0013] The original images of aircraft skin damage were acquired using a camera and then filtered to retain the clear images. The damage in the clear images was then classified into four categories: paint peeling, cracks, deformation, and rivet damage. The LabelImg tool was then used manually to select the damage area with the minimum bounding rectangle and set the classification label. This generated a txt image containing information on the damage category, location, and label size. All the labeled txt images constituted the original dataset.

[0014] In step 2), the method of using data augmentation techniques to perform data augmentation processing on the images in the original dataset and then dividing the augmented images into training and validation sets proportionally is as follows:

[0015] The images in the original dataset above are subjected to data augmentation processing including cropping, translation, brightness adjustment, noise addition, rotation, mirroring, random cropping, and mosaic enhancement. Mosaic enhancement involves stitching together four images and then dividing the enhanced images into training and validation sets in a 4:1 ratio.

[0016] In step 3), the method for improving the original YOLOv7 model to construct the initial FC-YOLO network model is as follows:

[0017] 3.1) The CCA attention module is integrated into the ELAN module in the original YOLOv7 model to form the ELAN-C module, thereby obtaining the FC-YOLO backbone network;

[0018] The calculation formula for the CCA attention module is as follows:

[0019]

[0020] Where, x c Indicates the input, y c Indicates output, g c Indicates channel characteristics; These represent spatial features along the X and Y directions, respectively;

[0021] The channel feature g c The specific calculation process is as follows:

[0022] The input image features are simultaneously subjected to max pooling and mean pooling, followed by a convolution operation with a kernel size of 1 to adjust the number of channels. Then, the features are activated using the Swish activation function. The two features are summed, and then a convolution operation with a kernel size of 1 is performed to restore the number of channels. Finally, the features are activated using the sigmoid activation function, resulting in the channel feature g. c ;

[0023] Spatial features along the X and Y directions The specific calculation process is as follows:

[0024] The input image features are subjected to unidirectional pooling along the X and Y directions. The resulting features are concatenated and then the number of channels is adjusted using a convolution operation with a kernel size of 1. The Swish activation function is then applied. The resulting features are then divided into two branches, and the number of channels is restored using a convolution operation with a kernel size of 1 on each branch. Finally, the Sigmoid activation function is applied to activate the features, resulting in the spatial features along the X and Y directions.

[0025] 3.2) Add an adaptive feature fusion module after the ELAN-C module and replace the PAFPN structure in the original YOLOv7 model to form the FC-YOLO neck feature fusion network;

[0026] The first and second features output by the FC-YOLO backbone network are adaptively fused using the adaptive feature fusion module, and then the third feature output by the FC-YOLO backbone network is fused again. After two adaptive feature fusions, three feature maps of different sizes are output for detecting targets of different sizes.

[0027] After each adaptive feature fusion, feature enhancement is performed using an ELAN module. For the l-th layer, the calculation process of the adaptive feature fusion module is as follows:

[0028] First, the three features x of the input image 1→l x 2→l x 3→l Each feature is convolved with a kernel size of 1 and then concatenated. The resulting features are then subjected to dimensionality reduction through a convolution operation with a kernel size of 1, and the variables are calculated using the softmax function. in, Let each represent the weight of the input feature at layer l, and satisfy the following conditions: Weight The expression is as follows:

[0029]

[0030]

[0031]

[0032] Let x be the feature input from layer n to layer l. n→l The eigenvector at (i,j) is The feature vector obtained after adaptive weight allocation is but and The relationship can be expressed as follows:

[0033]

[0034] 3.3) Replace the CIoU loss function of the original YOLOv7 model with the SIoU loss function as the regression loss function;

[0035] The specific calculation process of the SIoU loss function is as follows:

[0036]

[0037] Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. S B , , respectively, represent the areas of the predicted bounding box and the ground truth bounding box; Δ is the distance loss, used to measure the distance difference between the center points of the predicted bounding box and the ground truth bounding box. γ = 2 - sin2α, where α is the angle between the line connecting the centers of the predicted and ground truth boxes and the horizontal line; Ω is the shape loss, used to measure the shape difference between the predicted and ground truth boxes. In the formula, w and h are the width and height of the maximum bounding rectangles of the predicted bounding box and the ground truth bounding box, respectively; w gt h gt These are the width and height of the actual bounding box, w. p h p These represent the width and height of the prediction box, respectively.

[0038] In step 4), the method of inputting the training set obtained in step 2) into the initial FC-YOLO network model for training, and then using the validation set obtained in step 2) to evaluate the training results, and finally obtaining the FC-YOLO network model, is as follows:

[0039] The initial FC-YOLO model is deployed to a computer environment. The initial FC-YOLO model is trained using the training set obtained in step 2). During each training round, the network parameters are adjusted using the training results of the previous round. After each training round, the training results of the round are evaluated using the validation set obtained in step 2), and the training parameters for the next round are optimized. After training, the generated weights are output to obtain the FC-YOLO network model, which is used to detect other relevant aircraft skin damage images.

[0040] In step 5), the aircraft skin damage identification result includes the following three items:

[0041] 1. Bounding box: For each detected aircraft skin damage, output a bounding box to indicate the location of the damage in the image of the aircraft skin damage to be detected;

[0042] 2. Category: Identify which of the four categories of damage it belongs to: paint peeling, cracks, deformation, and rivet damage;

[0043] 3. Confidence score: Output a confidence score for each detected aircraft skin damage, which represents the degree of confidence of the FC-YOLO network model in the detection result of the aircraft skin damage.

[0044] The aircraft skin damage detection system based on the FC-YOLO network model provided by this invention includes:

[0045] Data acquisition module: used to acquire raw images of aircraft skin damage and perform filtering, classification, and manual annotation to obtain the raw dataset;

[0046] Data augmentation module: Uses data augmentation techniques to augment the images in the original dataset and divides the augmented images into training and validation sets proportionally;

[0047] The initial FC-YOLO network model is as follows: The CCA attention module is integrated into the ELAN module of the original YOLOv7 model to form the ELAN-C module, thus obtaining the FC-YOLO backbone network; an adaptive feature fusion module is added after the ELAN-C module and the PAFPN structure in the original YOLOv7 model is replaced to form the FC-YOLO neck feature fusion network; the SIoU loss function is used to replace the CIoU loss function of the original YOLOv7 model as the regression loss function.

[0048] Network training module: Trains the initial FC-YOLO network model using the training set and evaluates the training results using the validation set to obtain the FC-YOLO network model;

[0049] The results output module uses the FC-YOLO network model to perform damage identification on the real-time acquired images of aircraft skin damage to be detected, and finally outputs the aircraft skin damage identification results.

[0050] The aircraft skin damage detection method and system based on the FC-YOLO network model provided by this invention has the following beneficial effects: The ELAN-C module utilizes global pooling and unidirectional pooling to focus attention on the target's channel and position information, thereby assigning higher weights to key channels and positions of features and improving target detection capabilities; the FC-YOLO neck feature fusion network can avoid the reduction or loss of feature information of small targets during multi-scale feature fusion, redistributing weights to the inputs of different layers and strengthening the FC-YOLO network model's feature extraction capability for multi-scale targets; the SIoU loss function can reconstruct the penalty term of the loss function through the angle, distance, and shape losses between the target box and the ground truth box, accelerating the convergence of the loss function and optimizing the training process. Attached Figure Description

[0051] Figure 1 This is a structural diagram of the FC-YOLO network model in this invention.

[0052] Figure 2 This is a schematic diagram of the CCA attention module processing in this invention.

[0053] Figure 3 This is a schematic diagram of the ELAN-C module structure in this invention.

[0054] Figure 4 This is a simplified diagram of the FC-YOLO neck feature fusion network in this invention.

[0055] Figure 5 This is a schematic diagram of the adaptive feature fusion module in this invention.

[0056] Figure 6 (a), (b), and (c) are visual comparison charts of aircraft skin damage detection results using the YOLOv5 model, the YOLOv7 model, and the FC-YOLO network model in this invention, respectively. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0058] The aircraft skin damage detection method based on the FC-YOLO network model provided by this invention includes the following steps performed in sequence:

[0059] 1) Collect raw images of aircraft skin damage, and then filter, classify, and manually annotate them to obtain the raw dataset;

[0060] The original images of aircraft skin damage were acquired using a camera and then filtered to retain the clear images. The damage in the clear images was then classified into four categories: paint peeling, cracks, deformation, and rivet damage. The LabelImg tool was then used manually to select the damage area with the minimum bounding rectangle and set the classification label. This generated a txt image containing information on the damage category, location, and label size. All the labeled txt images constituted the original dataset.

[0061] In this invention, a total of 668 original images of aircraft skin damage were collected, with a resolution of 3840×2160, which included 1369 aircraft skin damages of various types.

[0062] 2) Use data augmentation techniques to augment the images in the original dataset, and divide the augmented images into training and validation sets according to the proportions.

[0063] To improve the generalization performance and robustness of the model, data augmentation processing was performed on the images in the original dataset, including cropping, translation, brightness adjustment, noise addition, rotation, mirroring, random cropping, and mosaic enhancement. Mosaic enhancement involves stitching together four images to increase the number of images, and then dividing the enhanced images into training and validation sets in a 4:1 ratio.

[0064] In this invention, a total of 2672 enhanced images were prepared.

[0065] 3) such as Figure 1 As shown in the dashed box, the original YOLOv7 model was improved to construct the FC-YOLO network model;

[0066] The specific method is as follows:

[0067] 3.1) As Figure 2 As shown, integrating the CCA attention module into the ELAN module of the original YOLOv7 model results in the following structure: Figure 3 The ELAN-C module shown is used to obtain the FC-YOLO backbone network;

[0068] The CCA attention module encodes features of the input image along different spatial directions, obtaining features in two directions. These features can capture the precise positional information of the target in both directions. Simultaneously, it encodes the target's channel information using mean pooling and max pooling operations, collecting key channel information and reallocating weights to the feature channels. Therefore, the CCA attention module can extract both the target's positional information and retain its channel information. The calculation formula for the CCA attention module is as follows:

[0069]

[0070] Where, x c Indicates the input, y c Indicates output, g c Indicates channel characteristics; These represent spatial features along the X and Y directions, respectively;

[0071] The channel feature g c The specific calculation process is as follows:

[0072] The input image features are simultaneously subjected to max pooling and mean pooling, followed by a convolution operation with a kernel size of 1 to adjust the number of channels. Then, the features are activated using the Swish activation function. The two features are summed, and then a convolution operation with a kernel size of 1 is performed to restore the number of channels. Finally, the features are activated using the sigmoid activation function, resulting in the channel feature g. c ;

[0073] Spatial features along the X and Y directions The specific calculation process is as follows:

[0074] The input image features are subjected to unidirectional pooling along the X and Y directions. The resulting features are concatenated and then the number of channels is adjusted using a convolution operation with a kernel size of 1. The Swish activation function is then applied. The resulting features are then divided into two branches, and the number of channels is restored using a convolution operation with a kernel size of 1 on each branch. Finally, the Sigmoid activation function is applied to activate the features, resulting in the spatial features along the X and Y directions.

[0075] 3.2) Add an adaptive feature fusion module after the ELAN-C module and replace the PAFPN structure in the original YOLOv7 model, forming the following structure: Figure 4 The FC-YOLO neck feature fusion network shown;

[0076] To mitigate the channel characteristics g output from the FC-YOLO backbone network c Spatial features along the X and Y directions The inconsistency among these three features is addressed by using an adaptive feature fusion module. The first and second features output from the FC-YOLO backbone network are adaptively fused, and then a second adaptive feature fusion is performed with the third feature output from the FC-YOLO backbone network. This step-by-step fusion method adaptively assigns different weights to each channel to enhance the feature fusion effect. After two adaptive feature fusions, three feature maps of different sizes are output for detecting targets of different sizes. It is worth noting that the adaptive feature fusion module can adaptively allocate weights between different input features, suppressing contradictory information from other input features and preserving key information.

[0077] After each adaptive feature fusion, feature enhancement is performed using an ELAN module; taking the l-th layer as an example, such as... Figure 5 As shown, the calculation process of the adaptive feature fusion module is as follows:

[0078] First, the three features x of the input image 1→l x 2→l x 3→l Each feature is convolved with a kernel size of 1 and then concatenated. The resulting features are then subjected to dimensionality reduction through a convolution operation with a kernel size of 1, and the variables are calculated using the softmax function. in, Let each represent the weight of the input feature at layer l, and satisfy the following conditions: Weight The expression is as follows:

[0079]

[0080]

[0081]

[0082] Let x be the feature input from layer n to layer l. n→l The eigenvector at (i,j) is The feature vector obtained after adaptive weight allocation is but and The relationship can be expressed as follows:

[0083]

[0084] 3.3) Considering the orientation issue between the predicted bounding box and the ground truth bounding box, the CIoU loss function of the original YOLOv7 model is replaced with the SIoU loss function as the regression loss function; the SIoU loss function introduces an angle loss to allocate the weight of the distance loss. As the angle increases, the distance loss will be assigned a higher weight to improve the regression accuracy as much as possible.

[0085] The specific calculation process of the SIoU loss function is as follows:

[0086]

[0087] Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. S B , , respectively, represent the areas of the predicted bounding box and the ground truth bounding box; Δ is the distance loss, used to measure the distance difference between the center points of the predicted bounding box and the ground truth bounding box. γ = 2 - sin2α, where α is the angle between the line connecting the centers of the predicted and ground truth boxes and the horizontal line; Ω is the shape loss, used to measure the shape difference between the predicted and ground truth boxes. In the formula, w and h are the width and height of the maximum bounding rectangles of the predicted bounding box and the ground truth bounding box, respectively; w gt h gt These are the width and height of the actual bounding box, w. p h p These represent the width and height of the prediction box, respectively.

[0088] 4) Input the training set obtained in step 2) into the initial FC-YOLO network model above for training, and then use the validation set obtained in step 2) to evaluate the training results, and finally obtain the FC-YOLO network model.

[0089] The initial FC-YOLO model is deployed to a computer environment. The initial FC-YOLO model is trained using the training set obtained in step 2). During each training round, the network parameters are adjusted using the training results of the previous round. After each training round, the training results of the round are evaluated using the validation set obtained in step 2), and the training parameters for the next round are optimized. After training, the generated weights are output to obtain the FC-YOLO network model, which is used to detect other relevant aircraft skin damage images.

[0090] The initial parameters for training are as follows: input image size is 640×640, number of iterations is 200, initial learning rate is 0.01, weight decay coefficient is 0.0005, optimizer is SGD, momentum parameter is 0.937, and batch size is 16.

[0091] 5) Use a camera to acquire images of the aircraft skin damage to be detected in real time and input them into the above FC-YOLO network model for damage identification. Finally, the trained FC-YOLO network model outputs the aircraft skin damage identification results in the images of the aircraft skin damage to be detected.

[0092] The aircraft skin damage identification results include the following three items:

[0093] 1. Bounding box: For each detected aircraft skin damage, output a bounding box to indicate the location of the damage in the image of the aircraft skin damage to be detected;

[0094] 2. Category: Identify which of the four categories of damage it belongs to: paint peeling, cracks, deformation, and rivet damage;

[0095] 3. Confidence score: Output a confidence score for each detected aircraft skin damage, which represents the degree of confidence of the FC-YOLO network model in the detection result of the aircraft skin damage.

[0096] To verify the beneficial effects of this invention, the inventors conducted the following ablation experiments and evaluated the performance of the FC-YOLO network model based on mean accuracy (mAP), number of parameters, computational cost (Flops), model size (MB), and detection time (ms). The formula for calculating mAP is as follows:

[0097]

[0098]

[0099]

[0100]

[0101] In the formula, P represents precision; R represents recall; TP represents a positive sample predicted as a positive sample; FP represents a negative sample predicted as a positive sample; FN represents a positive sample predicted as a negative sample; P(r) represents the PR curve; and n represents the number of detected categories.

[0102] The results of the ablation experiment training are shown in Table 1.

[0103] Table 1

[0104]

[0105] As can be seen, after introducing the SIoU loss function (Network 1), the mAP value improved by 0.9% and the detection time was extended by 0.2ms compared to the original YOLOv7 model; after introducing the feature fusion network proposed in this invention (Network 2), the mAP value improved by 1.9% and the detection time was extended by 0.8ms compared to the original YOLOv7 model; after introducing the ELAN-C module (Network 3) in the backbone, the mAP value improved by 1.1% and the detection time was extended by 0.6ms compared to the original YOLOv7 model; after replacing the SloU loss function and introducing the feature fusion network proposed in this invention (Network 4), the mAP value improved by 2.4% and the detection time was extended by 0.9ms compared to the original YOLOv7 model; finally, after incorporating all improvements, compared to Network 4, the mAP value improved by 0.7% and the detection time was extended by 0.7ms, and compared to the original YOLOv7 model, the mAP value improved by 3.1% and the detection time was extended by 1.6ms.

[0106] In summary, replacing the loss function of the original YOL0v7 model with the SloU loss function improved the mAP value and accelerated the detection speed. Because the adaptive feature fusion module increases the network's floating-point computation, the detection speed decreased slightly after introducing the designed feature fusion network, but the mAP value improved, and memory usage decreased by 14.9MB. Introducing the ELAN-C module into the backbone network further improved the mAP value. Experiments demonstrate that the improvements made in this invention have good detection performance on aircraft skin damage datasets.

[0107] The results of the comparative experiment training are shown in Table 2;

[0108] Table 2

[0109] YOLOv5 80.4 13.6 8.8 YOLOv7 83.9 71.3 7.3 YOLOv8 85.6 49.6 7.4 RT-DETR 85.4 63.1 11.4 FC-YOLO 87.0 65.6 8.9

[0110] As can be seen, the mAP values ​​in this embodiment are all higher than those of other comparative network models. Comparative experiments demonstrate that the improvements made in this invention have better detection performance than other network models on aircraft skin damage datasets.

[0111] To more intuitively demonstrate the effectiveness of the FC-YOLO network model compared to the original YOLOv7 model, this embodiment visualizes the experimental results. For example... Figure 6 As shown, the FC-YOL0 network model has a higher target confidence level than the YOLOv5 and YOLOv7 models during detection, and it is more capable of detecting small targets.

Claims

1. A method for detecting aircraft skin damage based on an FC-YOLO network model, characterized in that: The aircraft skin damage detection method includes the following steps performed in sequence: 1) Collect raw images of aircraft skin damage, and then filter, classify, and manually annotate them to obtain the raw dataset; 2) Use data augmentation techniques to augment the images in the original dataset, and then divide the augmented images into training and validation sets according to the proportions. 3) The original YOLOv7 model was improved to construct the initial FC-YOLO network model; 4) Input the training set obtained in step 2) into the initial FC-YOLO network model above for training, and then use the validation set obtained in step 2) to evaluate the training results, and finally obtain the FC-YOLO network model; 5) The camera is used to acquire images of the aircraft skin damage to be detected in real time and input them into the above FC-YOLO network model for damage identification. Finally, the FC-YOLO network model outputs the aircraft skin damage identification results in the images of the aircraft skin damage to be detected. In step 3), the method for improving the original YOLOv7 model to construct the initial FC-YOLO network model is as follows: 3.1) The CCA attention module is integrated into the ELAN module in the original YOLOv7 model to form the ELAN-C module, thereby obtaining the FC-YOLO backbone network; The calculation formula for the CCA attention module is as follows: ； in, Indicates input, Indicates the output. Indicates channel characteristics; , These represent spatial features along the X and Y directions, respectively; The channel features The specific calculation process is as follows: The input image features are simultaneously subjected to max pooling and mean pooling. Then, a convolution operation with a kernel size of 1 is performed to adjust the number of channels. After activation using the Swish activation function, the two features are summed, and then a convolution operation with a kernel size of 1 is performed to restore the number of channels. Finally, the sigmoid activation function is applied to obtain the channel features. ; Spatial features along the X and Y directions , The specific calculation process is as follows: The input image features are subjected to unidirectional pooling along the X and Y directions. The resulting features are concatenated and then the number of channels is adjusted using a convolution operation with a kernel size of 1. The Swish activation function is then applied. The resulting features are then divided into two branches, and the number of channels is restored using a convolution operation with a kernel size of 1 on each branch. Finally, the Sigmoid activation function is applied to activate the features, resulting in the spatial features along the X and Y directions. , ; 3.2) Add an adaptive feature fusion module after the ELAN-C module and replace the PAFPN structure in the original YOLOv7 model to form the FC-YOLO neck feature fusion network; The first and second features output by the FC-YOLO backbone network are adaptively fused using the adaptive feature fusion module, and then the third feature output by the FC-YOLO backbone network is fused again. After two adaptive feature fusions, three feature maps of different sizes are output for detecting targets of different sizes. After each adaptive feature fusion, feature enhancement needs to be performed using an ELAN module; for the th The calculation process of the adaptive feature fusion module in the layer is as follows: First, the three features of the input image , , Each feature is convolved with a kernel size of 1 and then concatenated. The resulting features are then subjected to dimensionality reduction through a convolution operation with a kernel size of 1, and the variables are calculated using the softmax function. , , ;in, They represent the first The weights of the input features in the layer and satisfying Weight The expression is as follows: ； ； ； Order from the first layer to the first Features of layer input exist The eigenvector at that location is The feature vector obtained after adaptive weight allocation is ,but and The relationship is expressed as follows: ； 3.3) Replace the CIoU loss function of the original YOLOv7 model with the SIoU loss function as the regression loss function; The specific calculation process of the SIoU loss function is as follows: ； in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. ; , These are the areas of the predicted bounding box and the ground truth bounding box, respectively; Distance loss measures the distance difference between the center point of the predicted bounding box and the center point of the ground truth bounding box. , , The angle between the line connecting the center points of the predicted bounding box and the ground truth bounding box and the horizontal line; Shape loss is used to measure the shape difference between the predicted bounding box and the ground truth bounding box. In the formula, , , , These are the width and height of the maximum bounding rectangles of the predicted bounding box and the ground truth bounding box, respectively; , These are the width and height of the actual bounding box, respectively. , These represent the width and height of the prediction box, respectively.

2. The aircraft skin damage detection method based on the FC-YOLO network model according to claim 1, characterized in that: In step 1), the method for acquiring raw images of aircraft skin damage and then filtering, classifying, and manually annotating them to obtain the raw dataset is as follows: The original images of aircraft skin damage were acquired using a camera and then filtered to retain the clear images. The damage in the clear images was then classified into four categories: paint peeling, cracks, deformation, and rivet damage. The LabelImg tool was then used manually to select the damage area with the minimum bounding rectangle and set the classification label. This generated a txt image containing information on the damage category, location, and label size. All the labeled txt images constituted the original dataset.

3. The aircraft skin damage detection method based on the FC-YOLO network model according to claim 1, characterized in that: In step 2), the method of using data augmentation techniques to perform data augmentation processing on the images in the original dataset and then dividing the augmented images into training and validation sets proportionally is as follows: The images in the original dataset above are subjected to data augmentation processing including cropping, translation, brightness adjustment, noise addition, rotation, mirroring, random cropping, and mosaic enhancement. Mosaic enhancement involves stitching together four images and then dividing the enhanced images into training and validation sets in a 4:1 ratio.

4. The aircraft skin damage detection method based on the FC-YOLO network model according to claim 1, characterized in that: In step 4), the method of inputting the training set obtained in step 2) into the initial FC-YOLO network model for training, and then using the validation set obtained in step 2) to evaluate the training results, and finally obtaining the FC-YOLO network model, is as follows: The initial FC-YOLO model is deployed to a computer environment. The initial FC-YOLO model is trained using the training set obtained in step 2). During each training round, the network parameters are adjusted using the training results of the previous round. After each training round, the training results of the round are evaluated using the validation set obtained in step 2), and the training parameters for the next round are optimized. After training, the generated weights are output to obtain the FC-YOLO network model, which is used to detect other relevant aircraft skin damage images.

5. The aircraft skin damage detection method based on the FC-YOLO network model according to claim 1, characterized in that: In step 5), the aircraft skin damage identification result includes the following three items: 1) Bounding box: For each detected aircraft skin damage, output a bounding box to indicate the location of the damage in the image of the aircraft skin damage to be detected; 2) Category: Identify which of the four categories of damage it belongs to: paint peeling, cracks, deformation, and rivet damage; 3) Confidence score: Output a confidence score for each detected aircraft skin damage, which represents the degree of confidence of the FC-YOLO network model in the detection result of the aircraft skin damage.

6. An aircraft skin damage detection system based on an FC-YOLO network model, employing the method described in any one of claims 1 to 5, characterized in that: The aircraft skin damage detection system based on the FC-YOLO network model includes: Data acquisition module: used to acquire raw images of aircraft skin damage and perform filtering, classification, and manual annotation to obtain the raw dataset; Data augmentation module: Uses data augmentation techniques to augment the images in the original dataset and divides the augmented images into training and validation sets proportionally; The initial FC-YOLO network model is as follows: The CCA attention module is integrated into the ELAN module of the original YOLOv7 model to form the ELAN-C module, thus obtaining the FC-YOLO backbone network; an adaptive feature fusion module is added after the ELAN-C module and the PAFPN structure in the original YOLOv7 model is replaced to form the FC-YOLO neck feature fusion network; the SIoU loss function is used to replace the CIoU loss function of the original YOLOv7 model as the regression loss function. Network training module: Trains the initial FC-YOLO network model using the training set and evaluates the training results using the validation set to obtain the FC-YOLO network model; The results output module uses the FC-YOLO network model to perform damage identification on the real-time acquired images of aircraft skin damage to be detected, and finally outputs the aircraft skin damage identification results.

Images