An unmanned aerial vehicle detection method based on wing-body fusion features

Abstract

The application provides a kind of UAV detection method based on wing-body fusion characteristics, it is related to UAV detection technical field.The method first collects and constructs the UAV sample data set conforming to YOLO label format, pre-processes data set, converts the UAV sample data set format into the format that YOLO can train and identify, and respectively constructs fuselage data set and wing data set;YOLOv5 model is then used to train the fuselage data set of UAV, and the structure and weight parameters of YOLOv5 model are optimized and adjusted;YOLOv7 model is used to train the wing data set of UAV, and the structure and weight parameters of YOLOv7 model are optimized and adjusted;Finally, the trained YOLOv5 model and YOLOv7 model are used to detect the fuselage and wing of UAV respectively;The final UAV detection result is obtained by distance and confidence level determination level using the fuselage and wing detection result of UAV.The method improves the detection accuracy of UAV, has the advantages of high detection efficiency, high recognition accuracy, strong robustness and the like.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) detection technology, and in particular to a UAV detection method based on blended wing-body characteristics. Background Technology

[0002] With the rapid development of information, communication, and control technologies, drones, originally excelling in the military field, have opened up the consumer market. Currently, drones have a wide range of applications, including but not limited to large-scale drone light shows, drone forest fire prevention and power line inspection, drone filming and short video creation, drone delivery and logistics, and drone aerial surveying. Drones are widely loved for their high maneuverability, small size, insensitivity to complex terrain, and low cost, and their adoption rate continues to increase.

[0003] However, the emergence of new technologies has also brought a series of safety issues. Crimes committed using drones and incidents endangering public safety due to users' lack of safety awareness are commonplace. Although some regulations have been introduced to regulate the drone market, such as setting up restricted and no-fly zones to manage drone flight activities, numerous accidents still occur. These problems indicate, on the one hand, that the new management methods are incomplete, and on the other hand, that most local authorities, due to considerations of labor or equipment costs, have not closely supervised the process. The reasonable use of drones still relies on users' awareness of "compliance with the law," resulting in frequent drone accidents, causing significant economic losses, and posing a threat to national security. Therefore, using technology to counter drones is essential. Drone detection and identification is a crucial part of drone countermeasures.

[0004] With the rapid development of image processing and computer vision, using images to detect targets has become a simple and effective method, widely used in various fields. Compared with general target detection tasks, UAV detection technology has the following challenges:

[0005] (1) The drone images are blurry and have low resolution. The number of pixels in the whole image is very small, resulting in a very limited amount of information. Due to the lack of detailed information about the drone, it is much more difficult to distinguish between objects with similar appearances.

[0006] (2) Target detection by UAVs is affected by complex background noise. Complex environmental noise has a large number of high-frequency components, which can easily overwhelm the negligible amount of information from the UAV, resulting in unsatisfactory detection results. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a drone detection method based on blended wing-body features to overcome the shortcomings of the prior art and achieve drone detection.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a UAV detection method based on blended wing-body features, comprising the following steps:

[0009] Step 1: Collect and construct a UAV sample dataset that conforms to the YOLO annotation format, preprocess the dataset, convert the UAV sample dataset format to the YOLO-trainable and recognizable .txt format, and construct the fuselage dataset and wing dataset respectively.

[0010] The collection of datasets mainly involves searching for drone images or videos on domestic and international websites, or taking drone videos and images with a mobile phone. These images are then filtered, integrated, and processed to create drone sample images, forming a drone sample dataset. First, the overall fuselage data of the drone is labeled to form a fuselage dataset. Then, images of each rotor of the drone are cropped out and labeled to form a wing dataset.

[0011] The annotation of the drone sample dataset uses XML files. For each image, the detected object is labeled with category information and detection region BoundingBox data. The data format is VOC format. The YOLO dataset recognition format is Label, X_center / Width, Y_center / High, Bbox / Width, and Bbox / High.

[0012] The labeled drone sample dataset is preprocessed by reshaping the images and performing data augmentation operations, namely rotation, translation, and cropping, to increase the diversity of the dataset. The transformed data information is stored in .txt format files, with one .txt file for each image sample. The number of annotations in the .txt file is the number of target samples in the image. Each target sample occupies one line and includes category information, the ratio of the x-coordinate of the target sample center point to the overall width of the image, the ratio of the y-coordinate of the target sample center point to the overall width of the image, the ratio of the width of the recognition region to the overall width of the image, and the ratio of the height of the recognition region to the overall width of the image.

[0013] Step 2: Train the UAV's fuselage dataset using the YOLOv5 model, and optimize and adjust the YOLOv5 model structure and weight parameters;

[0014] The fuselage dataset is divided into training, validation, and test sets. The images in the fuselage dataset are preprocessed to meet the input requirements of the YOLOv5 model. The fuselage is then trained using the YOLOv5 model.

[0015] After preprocessing, the fuselage images in the fuselage dataset are used for feature extraction through the backbone network structure of the YOLOv5 model. Then, the neck network enhancement model is used to detect drones of different sizes, thus enabling the identification of drones of different sizes. Finally, the bounding box prediction results are obtained in the detection head part.

[0016] After multiple iterations, the YOLOv5 model network parameters are continuously updated until the predetermined requirements are met;

[0017] Step 3: Train the YOLOv7 model on the drone wing dataset, and optimize and adjust the YOLOv7 model structure and weight parameters to make the YOLOv7 model network structure better adapt to learning small sample datasets.

[0018] The wing dataset is divided into a training set, a validation set, and a test set;

[0019] Since the wing is a small target to be detected, the YOLOv7 model was selected for training. First, the image was preprocessed through a series of operations such as input data augmentation, and then fed into the backbone network of the YOLOv7 model to extract features from the processed image. Subsequently, the extracted features were fused by the neck module to obtain features of three sizes: large, medium, and small. Finally, the fused features were fed into the detection head, and the results were output after detection.

[0020] After multiple iterations, the YOLOv7 model network parameters are continuously updated until the predetermined requirements are met.

[0021] Step 4: Use the trained YOLOv5 and YOLOv7 models to detect the UAV fuselage and wings respectively; use the UAV fuselage and wing detection results to obtain the final UAV detection result through distance and confidence level.

[0022] Step 4.1: Perform distance determination on the drone fuselage and wings identified by the YOLOv5 and YOLOv7 models;

[0023] The Euclidean distance is used to determine the distance between the fuselage and wings identified by the YOLOv5 and YOLOv7 models. The center positions of the UAV fuselage and wings are represented as two-dimensional coordinates (x1, y1) and (x2, y2), respectively. The Euclidean distance between the UAV fuselage and wings is calculated as follows:

[0024] d = sqrt((x2-x1)) 2 +(y2+y1) 2 )

[0025] Where d represents the Euclidean distance between the fuselage and the wing;

[0026] If the Euclidean distance between the fuselage and wings is within the set range, the fuselage and wings are determined to be the same drone; otherwise, proceed to step 4.2 and use the confidence level formula to determine whether the identified fuselage and wings belong to the same drone.

[0027] Step 4.2: Use the confidence level formula to associate the fuselage and wings with the target and determine whether the identified fuselage and wings belong to the same UAV.

[0028] The training weights obtained in steps 2 and 3 are used to detect the drone image to obtain the drone's position, as well as the confidence scores for the fuselage and wings. The following confidence score determination formula is then used to perform target association between the fuselage and wings:

[0029] p = w1*X1 + w2*X2

[0030] Where w1 is the weight of the wing, X1 is the confidence level of the wing; w2 is the weight of the fuselage, X2 is the confidence level of the fuselage, and p is the confidence level of the UAV.

[0031] Wing weight Where n is the number of drone wings, α is a constant set according to the actual situation of the drone, 0 < α < 1; the weight of the fuselage w2 = 1 - w1; and the confidence level of the wings. That is, the average confidence level of each wing of the drone; therefore, the final confidence level determination formula for the drone is:

[0032]

[0033] Finally, the calculated final confidence score of the drone is compared with the confidence score threshold. If the final confidence score of the drone is higher than the confidence score threshold, the identified fuselage and wings belong to the same drone, and the detection result is considered to be a drone with high credibility.

[0034] The beneficial effects of adopting the above technical solution are as follows: The UAV detection method based on wing-body blending features provided by the present invention selects a suitable network for training according to the characteristics of the fuselage and wings, which can improve the detection accuracy of UAVs. The method has the advantages of high detection efficiency, high recognition accuracy and strong robustness.

[0035] This invention incorporates distance and confidence level judgments for joint fuselage and wing discrimination. It can detect wings even in complex environments such as those with obstructions. If a wing is detected, a judgment mechanism can determine the presence of a drone, improving the accuracy of detection in complex environments. Simultaneously, the increased processing speed enables the detection of illegal "black flight" drones. Attached Figure Description

[0036] Figure 1A flowchart illustrating a UAV detection method based on blended wing-body features, provided as an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the YOLOv5 model structure provided in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the YOLOv7 model structure provided in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the drone confidence determination mechanism provided in an embodiment of the present invention. Detailed Implementation

[0040] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0041] In this embodiment, a UAV detection method based on blended wing-body features is described, such as... Figure 1 As shown, it includes the following steps:

[0042] Step 1: Collect and construct a UAV sample dataset that conforms to the YOLO annotation format, preprocess the dataset, convert the UAV sample dataset format to the YOLO-trainable and recognizable .txt format, and construct the fuselage dataset and wing dataset respectively.

[0043] The collection of datasets mainly involves searching for drone images or videos on domestic and international websites, or taking drone videos and images with a mobile phone. These images are then filtered, integrated, and processed to create drone sample images, forming a drone sample dataset. First, the overall fuselage data of the drone is labeled to form a fuselage dataset. Then, images of each rotor of the drone are cropped out and labeled to form a wing dataset.

[0044] The annotation of the drone sample dataset uses XML files. For each image, the detected object is labeled with category information and detection region BoundingBox data. The data format is VOC format. The YOLO dataset recognition format is Label, X_center / Width, Y_center / High, Bbox / Width, and Bbox / High.

[0045] The labeled drone sample dataset is preprocessed by reshaping the images and performing data augmentation operations, namely rotation, translation, and cropping, to increase the diversity of the dataset. The transformed data information is stored in .txt format files, with one .txt file for each image sample. The number of annotations in the .txt file is the number of target samples in the image. Each target sample occupies one line and includes category information, the ratio of the x-coordinate of the target sample center point to the overall width of the image, the ratio of the y-coordinate of the target sample center point to the overall width of the image, the ratio of the width of the recognition region to the overall width of the image, and the ratio of the height of the recognition region to the overall width of the image.

[0046] Step 2: Train the UAV's fuselage dataset using the YOLOv5 model, and optimize and adjust the YOLOv5 model structure and weight parameters;

[0047] The fuselage dataset is divided into training, validation, and test sets. The images in the fuselage dataset are preprocessed to meet the input requirements of the YOLOv5 model. The YOLOv5 model is then used to train the fuselage dataset. The preprocessing of the fuselage dataset images includes operations such as resizing, cropping, scaling, and normalization.

[0048] After preprocessing, the fuselage images in the fuselage dataset are used for feature extraction through the backbone network structure of the YOLOv5 model. Then, the neck network enhancement model is used to detect drones of different sizes, thus enabling the identification of drones of different sizes. Finally, the bounding box prediction results are obtained in the detection head part.

[0049] After multiple iterations, the YOLOv5 model network parameters are continuously updated until the predetermined requirements are met;

[0050] This embodiment utilizes a server to train the device using the YOLOv5 network. The training environment uses Python 3.8, the deep learning framework is PyTorch 1.8.1, CDUA is 11.1, the operating system is Windows 11, the processor is a 16-core Intel i9-12900KF, the memory is 32GB, and the graphics card is an NVIDIA RTX 3090 Ti with 24GB of video memory. The specific steps include:

[0051] S1. Obtain the training set of the drone body, select training images from the training set, and obtain the center point p and the true bounding box width and height [W,H] of the objects in the training images; input the training drone images based on... Figure 2 The YOLOv5 object detection model shown first processes the drone dataset using the Mosaic method, which flips, scales, and changes the color gamut of four images, then stitches them together after random cropping to form a new drone image as the dataset for model training.

[0052] S2. The preprocessed drone dataset extracts features through a cross-stage partial network (CSP) backbone network, and then downsamples the drone images using a focus structure. A Neck network is then used to further enrich the feature diversity to improve model robustness. This part performs feature fusion, obtaining training labels corresponding to three output feature maps: a first output feature map, a second output feature map, and a third output feature map. The information in the output feature maps includes: presence or absence of class information, class information, center point information, and width and height information. Specifically, the width and height of the first output feature map are preferably 20×20, the second output feature map is 40×40, and the third output feature map is 80×80.

[0053] S3. Scale the object's center point and ground truth bounding box dimensions proportionally to the dimensions of the first, second, and third output feature maps to obtain the training labels corresponding to each output feature map. This yields the object's reference bounding box and anchor points in each output feature map. By calculating the mathematical relationship between each anchor point and the corresponding output feature map regarding the object's center point and dimensions, the predicted center point and predicted dimensions of the object in each output feature map are obtained.

[0054] S4. Calculate the predicted center point and predicted width and height of the object, and the loss between these values ​​and the predicted center point and width / height of the object to obtain the first loss value of the model. Add the first loss value to the loss value of the object category to obtain the overall loss value of the model. The loss function used here is the CIoU loss function, as shown in the following formula:

[0055]

[0056]

[0057]

[0058] In the CIoU formula, ρ 2 (b,b g t ) represents the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box, and c represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box. IoU is a standard for measuring the accuracy of detecting corresponding objects in a specific dataset.

[0059] S5. Input the overall loss value into the Adam optimizer for optimization. After multiple iterations, continuously update the network parameters until the predetermined requirements are met. Obtain the trained UAV prediction model based on YOLOv5.

[0060] Step 3: Train the UAV wing dataset using the YOLOv7 model, and optimize and adjust the YOLOv7 model structure and weight parameters to make the YOLOv7 model network structure better adapt to learning small sample datasets.

[0061] The wing dataset is divided into a training set, a validation set, and a test set;

[0062] Since the wing is a small target to be detected, the YOLOv7 model was selected for training. First, the image was preprocessed through a series of operations such as input data augmentation, and then fed into the backbone network of the YOLOv7 model to extract features from the processed image. Subsequently, the extracted features were fused by the neck module to obtain features of three sizes: large, medium, and small. Finally, the fused features were fed into the detection head, and the results were output after detection.

[0063] After multiple iterations, the YOLOv7 model network parameters are continuously updated until the predetermined requirements are met. This process includes the following steps:

[0064] A1. First, preprocess the input image, including image scaling, cropping, normalization, etc., and align it into a 640×640 RGB image.

[0065] A2. Input the preprocessed image into, for example... Figure 3 In the YOLOv7 model shown, the backbone network is used to extract features from the preprocessed image. First, after four convolutional layers, the drone feature map becomes 160×160×128 pixels. Then, the drone feature map passes through the ELAN module. Following this are the outputs of three MP+ELAN modules, corresponding to the drone feature map outputs C3, C4, and C5, with sizes of 80×80×512, 40×40×1024, and 20×20×1024 respectively. The feature map is then input into the Neck network, which uses a traditional PAFPN structure. The three effective feature layers obtained in the backbone are fused here, while the effective feature layers already obtained in the FPN section are used for further feature extraction. Then, a Panet structure is used to upsample and downsample the features to achieve feature fusion. Finally, a RepConv structure is used to convert the complex residual structure into a simple 3×3 convolution, reducing the network's complexity without compromising predictive ability. By using Backbone and FPN, three enhanced effective feature layers can be obtained. These feature maps contain feature information at different levels and scales. Each feature layer has width, height, and number of channels. We can then view the feature map as a collection of feature points, each with three prior boxes, and each prior box has a number of channels and features. This allows us to obtain the wing prediction information.

[0066] A3. The loss function consists of three parts: coordinate loss, target confidence loss, and classification loss. The target confidence loss and coordinate loss use the CIoU loss function, and the classification loss uses BCEWithLogitsLoss.

[0067] A4. Select the Adam optimizer with default parameters and set the initial learning rate to 1e-3, decreasing it by one-tenth as training progresses. Use the output of the original image set (region coordinates, confidence score, and classification label) as inputs to the three loss functions. Calculate the loss function results, perform backpropagation, and update the network parameters. After a preset training period, obtain the trained convolutional neural network.

[0068] Step 4: Use the trained YOLOv5 and YOLOv7 models to detect the drone's fuselage and wings respectively; use the drone fuselage and wing detection results to obtain the final drone detection result through distance and confidence level determination, such as... Figure 4 As shown;

[0069] Step 4.1: Perform distance determination on the drone fuselage and wings identified by the YOLOv5 and YOLOv7 models;

[0070] The Euclidean distance is used to determine the distance between the fuselage and wings identified by the YOLOv5 and YOLOv7 models. The center positions of the UAV fuselage and wings are represented as two-dimensional coordinates (x1, y1) and (x2, y2), respectively. The Euclidean distance between the UAV fuselage and wings is calculated as follows:

[0071] d = sqrt((x2-x1)) 2 +(y2+y1) 2 )

[0072] Where d represents the Euclidean distance between the fuselage and the wing;

[0073] If the Euclidean distance between the fuselage and wings is within the set range, the fuselage and wings are determined to be the same drone; otherwise, proceed to step 4.2 and use the confidence level formula to determine whether the identified fuselage and wings belong to the same drone.

[0074] Step 4.2: Use the confidence level formula to associate the fuselage and wings with the target and determine whether the identified fuselage and wings belong to the same UAV.

[0075] The training weights obtained in steps 2 and 3 are used to detect the drone image to obtain the drone's position, as well as the confidence scores for the fuselage and wings. The following confidence score determination formula is then used to perform target association between the fuselage and wings:

[0076] p = w1*X1 + w2*X2

[0077] Where w1 is the weight of the wing, X1 is the confidence level of the wing; w2 is the weight of the fuselage, X2 is the confidence level of the fuselage, and p is the confidence level of the UAV.

[0078] Wing weight Where n is the number of drone wings, α is a constant set according to the actual situation of the drone, 0 < α < 1; the weight of the fuselage w2 = 1 - w1; and the confidence level of the wings. That is, the average confidence level of each wing of the drone; therefore, the final confidence level determination formula for the drone is:

[0079]

[0080] Finally, the calculated final confidence score of the drone is compared with the confidence score threshold. If the final confidence score of the drone is higher than the confidence score threshold, the identified fuselage and wings belong to the same drone, and the detection result is considered to be a drone with high credibility.

[0081] First, the images to be detected are preprocessed, including image scaling, normalization, and channel conversion, to adapt to the model's input requirements. The drone images are then input into the network, where a pre-trained weighted model detects information such as the confidence scores of the drone's fuselage and wings. Post-processing of the detection results includes Non-Maximum Suppression (NMS) and confidence filtering. NMS aims to eliminate overlapping detection boxes, resulting in more accurate detections. Confidence filtering removes detections with low confidence scores. The output detection results include fuselage and wing categories, locations, and confidence scores. Then, information fusion is performed, and the resulting information is used to calculate the confidence score of the drone in the image using a custom confidence level.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for detecting unmanned aerial vehicles (UAVs) based on blended wing-body features, characterized in that: Includes the following steps: Step 1: Collect and construct a UAV sample dataset that conforms to the YOLO annotation format, preprocess the dataset, convert the UAV sample dataset format into a YOLO-trainable and recognizable format, and construct the fuselage dataset and wing dataset respectively. Step 2: Train the UAV's fuselage dataset using the YOLOv5 model, and optimize and adjust the YOLOv5 model structure and weight parameters; Step 3: Train the YOLOv7 model on the drone wing dataset, and optimize and adjust the YOLOv7 model structure and weight parameters to make the YOLOv7 model network structure better adapt to learning small sample datasets. Step 4: Use the trained YOLOv5 and YOLOv7 models to detect the UAV fuselage and wings respectively; use the UAV fuselage and wing detection results to obtain the final UAV detection result through distance and confidence level. Step 4.1: Perform distance determination on the drone fuselage and wings identified by the YOLOv5 and YOLOv7 models; Euclidean distance is used to determine the distance between the fuselage and wings identified by the YOLOv5 and YOLOv7 models; the center positions of the UAV fuselage and wings are represented as two-dimensional coordinates. and The Euclidean distance between the fuselage and wings of the drone is given by the following formula: in, d This represents the Euclidean distance between the fuselage and the wing; If the Euclidean distance between the fuselage and wings is within the set range, the fuselage and wings are determined to be the same drone; otherwise, proceed to step 4.2 and use the confidence level formula to determine whether the identified fuselage and wings belong to the same drone. Step 4.2: Use the confidence level formula to associate the fuselage and wings with the target and determine whether the identified fuselage and wings belong to the same UAV. The training weights obtained in steps 2 and 3 are used to detect the drone image to obtain the drone's position, as well as the confidence scores for the fuselage and wings. The following confidence score determination formula is then used to perform target association between the fuselage and wings: in, For the weight of the wing, The confidence level of the wing; For the weight of the fuselage, denoted as , where p represents the confidence level of the airframe and p represents the confidence level of the drone. Wing weight , where n is the number of wings of the drone. These are constants set according to the actual conditions of the drone. Weight of the fuselage Confidence level of the wing This refers to the average confidence level of each wing of the drone; therefore, the final confidence level determination formula for the drone is: Finally, the calculated final confidence score of the drone is compared with the confidence score threshold. If the final confidence score of the drone is higher than the confidence score threshold, the identified fuselage and wings belong to the same drone, and the detection result is considered to be a drone with high credibility.

2. The UAV detection method based on blended wing-body features according to claim 1, characterized in that: Step 1 involves collecting drone images or videos, then filtering and integrating them to create a drone sample image dataset. First, the overall fuselage data of the drone is labeled to form a fuselage dataset. Then, images of each rotor of the drone are cropped out and labeled to form a wing dataset.

3. The UAV detection method based on blended wing-body features according to claim 2, characterized in that: Step 1 uses XML files to annotate the UAV sample dataset, annotating the category information and detection region BoundingBox data for the detected objects in each image, with the data format being VOC format; the YOLO dataset recognition format is Label, X_center / Width, Y_center / High, Bbox / Width, and Bbox / High; The labeled drone sample dataset is preprocessed by reshaping the images and performing data augmentation operations, namely rotation, translation, and cropping, to increase the diversity of the dataset. The transformed data information is stored in .txt format files, with one .txt file corresponding to each image sample. The number of annotations in the .txt file is the number of target samples in the image. Each target sample occupies one line and includes category information, the ratio of the x-coordinate of the target sample's center point to the overall width of the image, the ratio of the y-coordinate of the target sample's center point to the overall width of the image, the ratio of the width of the recognition region to the overall width of the image, and the ratio of the height of the recognition region to the overall width of the image.

4. The UAV detection method based on blended wing-body features according to claim 3, characterized in that: The specific method for step 2 is as follows: The fuselage dataset is divided into training, validation, and test sets. The images in the fuselage dataset are preprocessed to meet the input requirements of the YOLOv5 model. The fuselage is then trained using the YOLOv5 model. After preprocessing, the fuselage images in the fuselage dataset are used for feature extraction through the backbone network structure of the YOLOv5 model, and then the neck network enhancement model is used to detect drones of different sizes, thus enabling the identification of drones of different sizes. Finally, the bounding box prediction results are obtained in the detection head part; Through multiple iterations, the YOLOv5 model network parameters are continuously updated until the predetermined requirements are met.

5. The UAV detection method based on blended wing-body features according to claim 4, characterized in that: The specific method for step 3 is as follows: The wing dataset is divided into a training set, a validation set, and a test set; Since the wing is a small target to be detected, the YOLOv7 model was selected to train the wing. First, the image was preprocessed by a series of operations such as input data augmentation, and then fed into the backbone network of the YOLOv7 model to extract features from the processed image. Subsequently, the extracted features are fused by the neck module to obtain features of three sizes: large, medium, and small. Finally, the fused features are fed into the detection head, and the results are output after detection. Through multiple iterations, the YOLOv7 model network parameters are continuously updated until the predetermined requirements are met.

Images