Unmanned aerial vehicle aerial image small target detection method, system and device and medium

Abstract

The invention provides an unmanned aerial vehicle aerial image small target detection method, system and device and a medium. The method comprises the implementation steps that an unmanned aerial vehicle aerial image data set is acquired and preprocessed; yOLOv8s is used as a basic network, and SPD convolution is introduced into a backbone network; the neck structure is improved, and the new structure is named as a fused context feature alignment pyramid network (FCAFPN); a feature fusion module SIAF is introduced into the feature fusion module; then, a dynamic up-sampling module (DySample) is used for replacing nearest neighbor sampling (NN); an improved YOLOv8 network is obtained; the improved YOLOv8 network is applied to unmanned aerial vehicle aerial image recognition, and a high-precision small target detection deep learning model is constructed through training; inputting a to-be-detected image into the model for detection to obtain a target detection result; the system, the equipment and the medium are used for implementing the method. According to the invention, accurate identification of the small target of the unmanned aerial vehicle is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning target detection, and in particular to a method for detecting small targets in unmanned aerial vehicle aerial images based on improved YOLOv8. Background Art

[0002] Unmanned aerial vehicles (UAVs) are aircraft that can perform various tasks in the air without the need for personnel. In recent years, with the development of science and technology, drones have been widely used in the fields of aerial photography, traffic flow control, military reconnaissance, agriculture, industrial target detection, etc. Due to the low resolution of images taken by drones, the features of small targets in the images are not obvious enough, making it difficult for target recognition algorithms to accurately identify them.

[0003] Object detection is a key task in computer vision, which can be divided into single-stage and two-stage methods. Two-stage methods (such as Faster R-CNN) first generate candidate regions and then classify them, which is highly accurate but slow. Single-stage methods (such as YOLO and SSD) directly predict categories and locations, which is fast but slightly less accurate. Both methods have their own advantages and disadvantages and are suitable for different scenarios.

[0004] In the field of target detection, the YOLO (You Only Look Once) series of algorithms are widely used due to their superior performance. The algorithm can achieve efficient real-time target recognition, with the advantages of high accuracy and easy training. However, although the YOLO series of algorithms perform well in many application scenarios, they still have certain limitations in small target detection. Specifically, the YOLO algorithm often has difficulty capturing its detailed information when dealing with small targets. This is mainly because small targets occupy relatively few pixels in the image, resulting in the loss of important detail information during feature extraction. In addition, background noise and occlusion can also have a negative impact on the detection accuracy of small targets. These problems pose a major challenge to the UAV target recognition task. Therefore, how to effectively improve the performance of the YOLO series of algorithms in small target detection has become one of the hot spots and difficulties in current research. In response to these problems, developing improved detection algorithms and technical solutions to improve the detection capability of small targets is of great significance for achieving efficient and accurate UAV target recognition. Summary of the invention

[0005] In order to solve the problem of low accuracy in drone target detection, the present invention provides a method for detecting small targets in drone aerial images.

[0006] In order to solve the above technical problems, the specific implementation steps of a method for detecting small targets in drone aerial images of the present invention are as follows:

[0007] A method for detecting small targets in drone aerial images.

[0008] S1. Obtain real aerial images taken by drones, convert the obtained images into different formats, and finally obtain dataset images in YOLO format; divide the dataset images in YOLO format into training set, validation set, and test set; enhance the training set using Mosaic data enhancement method to obtain enhanced training set images;

[0009] S2. Using YOLOv8s as the basic network, the improved YOLOv8 network obtains fine-grained image information by introducing SPD convolution in the backbone of the basic network; and introduces Adown and CBS to construct a new feature fusion module SIAF; the information of adjacent detection layers is fused in the neck structure through the module to form a new feature pyramid network FCAFPN; then the dynamic upsampling module (DySample) is used to replace the nearest neighbor sampling (NN) to reduce the loss of feature information;

[0010] S3, inputting the enhanced data set obtained in step S1 into the improved YOLOv8 network described in step S2 for use in the drone aerial image recognition network; setting initial training parameters; and obtaining a deep learning model for drone small target detection through training;

[0011] S4: Input the drone aerial image to be detected into the model obtained in step S3 for detection to obtain the target detection result.

[0012] The Mosaic data enhancement method in step S1 enriches the background and scene complexity of the training data by randomly cropping and splicing four different pictures in the data set into a new picture.

[0013] The specific implementation method of step S2 includes the following steps:

[0014] S21, replace the ordinary convolution of the adjacent layers of the backbone network P2 detection layer with SPD convolution, which divides the input feature map into multiple sub-feature maps and splices them in the depth dimension to obtain a drone aerial photography feature map with half the spatial dimension;

[0015] S22, constructing a feature fusion module SIAF, wherein the feature fusion module includes Adown, CBS and ordinary convolution;

[0016] S23, adding a small target information flow transmission channel in the P2 and P3 detection layers to the top-down path of the original PAFPN, and combining the original network connection to perform multi-scale feature fusion through the SIAF module to form a new feature pyramid network FCAFPN;

[0017] S24, replace nearest neighbor sampling with dynamic upsampling. Dynamic upsampling can better capture details in images or data and generate rich information density feature maps by adjusting sampling strategies in different contexts;

[0018] S25. Build an improved YOLOv8 network.

[0019] In the step S21: the P2 detection layer is the detection layer where the input feature map is located after being downsampled four times, and a 2-fold downsampling is performed again on the basis of the P2 detection layer to obtain the P3 detection layer.

[0020] The step S22 of constructing the feature fusion module SIAF includes the following steps:

[0021] The present invention includes two feature fusion modules SIAF, which are used to fuse the feature maps of the two layers of the current layer.

[0022]

[0023] Among them, ADown contains an average pooling layer, a maximum pooling layer and two CBS modules (the convolution kernel sizes are 1×1 and 3×3 respectively), which downsample the input feature map by two times. This module divides the feature map into two paths for independent processing, making full use of the diversity of input features and extracting feature information of different scales respectively. Downsample the input feature map by two times;

[0024]

[0025] Among them, the CBS module consists of a cascade of convolutional layers, batch normalization layers, and SiLU activation functions to perform channel matching on the input;

[0026]

[0027] This stage directly receives input that is dynamically upsampled by a factor of two;

[0028] Finally, the Contact operation is used to perform feature fusion in the channel dimension, and a 1×1 convolution is used to adjust the number of output channels.

[0029]

[0030] The specific implementation method of step S3 includes the following network initial training parameters: the training cycle is 300, the initial learning rate is 0.01, the weight decay is set to 0.0005, the optimizer selects SGD, the batch size is 16, the input image is resized to 640×640, the momentum is set to 0.932, and the number of epochs required to stop training early is set to 50.

[0031] In step S4: use the improved YOLOv8 network model to perform target detection on the test set, generate an image containing a detection frame, compare the image with the original image, and obtain the detection effect of each image.

[0032] The present invention also provides a UAV image small target recognition system, which is characterized by comprising:

[0033] Dataset processing module: obtains public drone aerial photography datasets, converts the dataset format, and performs data enhancement and division on the converted dataset to obtain an enhanced dataset;

[0034] Network construction module: used to build a small target detection network for drone images based on improved YOLOv8;

[0035] Network training module: used to train the drone image small target detection network based on improved YOLOv8 using the training set according to the set network training parameters, and obtain the deep learning model of drone small target detection after the training;

[0036] Target recognition module: used to input the test set and the obtained model weights into the trained UAV image small target detection network based on improved YOLOv8 to obtain the target recognition results.

[0037] The present invention also provides a small target recognition device for unmanned aerial vehicle aerial images, which is characterized by comprising:

[0038] Memory: a computer program storing the above-mentioned method for detecting small targets in drone images, which is a computer-readable device;

[0039] Processor: used to implement a method for detecting small targets in drone images when executing the computer program.

[0040] The present invention also provides a computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the method for detecting small targets in drone images.

[0041] The beneficial effects of adopting the above technical solution are: the improved YOLOv8 network adopted by the present invention. First, SPD convolution is added to the P2 and P3 target detection layers in the backbone network to replace the traditional strided convolution. The traditional strided convolution often has poor feature extraction effect on small targets due to its large receptive field in target detection, especially when processing small objects, which easily leads to information loss. SPD convolution can process feature maps in parallel at multiple scales by constructing a spatial pyramid structure, thereby enhancing the feature extraction capability of small targets; secondly, a feature fusion module that can fuse key information of small targets is designed, which can effectively reduce feature loss during information transmission by integrating multiple inputs from adjacent levels. This multi-level feature fusion not only improves the detection capability of small targets, but also increases the model's understanding of target relationships in complex scenes. Finally, a feature fusion network is designed. This network architecture ensures that the model can use shallow feature information to capture more details when detecting small targets by strengthening feature transmission. The network enhances the detection capability of small targets in drone aerial images. The accuracy of small target classification and recognition is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the overall architecture of the present invention;

[0043] Figure 2 This is the overall architecture diagram of the improved YOLOv8 of the present invention;

[0044] Figure 3 Schematic diagram of the SPD convolution structure of the present invention;

[0045] Figure 4 This is a schematic diagram of the structure of the SIAF feature fusion module of the present invention;

[0046] Figure 5 This is a comparison chart of index changes before and after the improvement of the present invention;

[0047] Figure 6 This is a diagram showing the detection effect of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The embodiment of the present invention discloses a method for detecting small targets in unmanned aerial images. Figure 1 As shown, the method comprises the following steps:

[0050] S1. Obtain real aerial images taken by drones, convert the obtained images into different formats, and finally obtain dataset images in YOLO format; divide the dataset images in YOLO format into training set, validation set, and test set; enhance the training set using Mosaic data enhancement method to obtain enhanced training set images;

[0051] Specifically, in the present invention, the obtained drone aerial photography public data set refers to image data obtained from the drone aerial photography public data set VisDrone2019, which contains ten objects to be detected, namely pedestrians, people, bicycles, cars, vans, trucks, tricycles, awning tricycles, buses, and motorcycles, with a total of 8269 pictures; the data set is processed, firstly the data set is formatted to obtain a data set in YOLO format; then the data set images in YOLO format are divided into a ratio of 7:1:2, wherein the training set includes 6471 pictures, the validation set includes 548 pictures, and the test set includes 1610 pictures; finally, the image is enhanced by using the Mosaic data enhancement method of randomly cropping and splicing four different pictures in the training set into a new picture to obtain an enhanced training set image;

[0052] S2. Using YOLOv8s as the basic network, the improved YOLOv8 network obtains fine-grained image information by introducing SPD convolution in the backbone of the basic network; and introduces Adown and CBS to construct a new feature fusion module SIAF; the information of adjacent detection layers is fused in the neck structure through the module to form a new feature pyramid network FCAFPN; then the dynamic upsampling module (DySample) is used to replace the nearest neighbor sampling (NN) to reduce the loss of feature information;

[0053] S3, inputting the enhanced data set obtained in step S1 into the improved YOLOv8 network described in step S2 for use in the drone aerial image recognition network; setting initial training parameters; and obtaining a deep learning model for drone small target detection through training;

[0054] S4: Input the drone aerial image to be detected into the model obtained in step S3 for detection to obtain the target detection result.

[0055] The following is a detailed description of the above steps based on the specific content:

[0056] The specific implementation method of improving the YOLOv8 network described in step 2 includes the following steps:

[0057] S21, replace the ordinary convolution of the adjacent layers of the backbone network P2 detection layer with SPD convolution, which divides the input feature map into multiple sub-feature maps and splices them in the depth dimension to obtain a drone aerial photography feature map with half the spatial dimension;

[0058] S22, constructing a feature fusion module SIAF, wherein the feature fusion module includes Adown, CBS and ordinary convolution;

[0059] S23, adding a small target information flow transmission channel in the P2 and P3 detection layers to the top-down path of the original PAFPN, and combining the original network connection to perform multi-scale feature fusion through the SIAF module to form a new feature pyramid network FCAFPN;

[0060] S24, replace nearest neighbor sampling with dynamic upsampling. Dynamic upsampling can better capture details in images or data and generate rich information density feature maps by adjusting sampling strategies in different contexts;

[0061] S25. Build an improved YOLOv8 network. The specific structure is as follows Figure 2 shown.

[0062] Furthermore, the P2 detection layer in step S21 is the detection layer where the input feature map is located after four times downsampling. On the basis of the P2 detection layer, a 2-fold downsampling is performed to obtain the P3 detection layer, and so on to obtain the P4 detection layer. The SPD convolution module is represented as

[0063]

[0064] Among them, given the feature map X, sub-graph f x,y It consists of all entries X(i,j), where i+x and j+y are divisible by the ratio. Therefore, each subgraph will downsample X in proportion. The present invention specifies the ratio as 2, and four subgraphs f can be obtained. 0,0 , f 1,0 , f 0,1 , f 1,1 By connecting the sub-graphs by channel, the size of the feature graph is reduced by two times, and the channel dimension is increased by four times. Figure 3 As shown. The depth dimension splicing operation is expressed as:

[0065] Given two feature maps A and B, with shapes of (H, W, C1) and (H, W, C2) respectively, the shape of the feature map obtained after the depth dimension splicing operation is (H, W, C1+C2). The height and width of the new feature map remain unchanged, and the number of channels increases.

[0066] Furthermore, constructing the feature fusion module SIAF in step S22 includes the following steps:

[0067] The fusion module used in this invention is SIAF (Spatial Information Alignment and Fusion), which fuses features of multiple scales from different levels. The original Contact operation only uses simple channel splicing. In order to introduce global information and reduce information loss during sampling, we designed a feature fusion module SIAF. For each feature of scale n in level k,

[0068]

[0069] Among them, ADown contains an average pooling layer, a maximum pooling layer and two CBS modules (the convolution kernel sizes are 1×1 and 3×3 respectively), which downsample the input feature map by two times. This module divides the feature map into two paths for independent processing, fully utilizing the diversity of the input features and extracting feature information of different scales respectively. This dual-path structure not only enhances the expressiveness of the features, but also reduces information loss and ensures the retention of details during the downsampling process. By combining average pooling and maximum pooling, the module realizes a variety of downsampling methods, which can not only retain important global features but also enhance the significance of local features. At the same time, through the combination of 1×1 and 3×3 convolutions, the computational complexity is further optimized, and the feature extraction efficiency is improved while keeping the network lightweight. The module schematic diagram is shown as follows Figure 4 shown.

[0070]

[0071] Among them, the CBS module consists of a cascaded convolutional layer, a batch normalization layer and a SiLU activation function to perform channel matching on the input.

[0072]

[0073] This stage directly accepts input that is dynamically upsampled by a factor of two.

[0074] Finally, the Contact operation is used to perform feature fusion in the channel dimension, and a 1×1 convolution is used to adjust the number of output channels.

[0075]

[0076] The SIAF module effectively integrates feature information of different scales by downsampling and convolution processing multiple input features of different resolutions and fusing them at the end. This multi-scale fusion enhances the model's ability to recognize objects of different sizes and can more accurately capture small objects and details. The structural diagram of this module is shown in the figure below. Figure 4shown.

[0077] Furthermore, when the PAFPN described in step S23 needs to integrate information across layers, the structure cannot transmit information losslessly, which prevents YOLO from better performing information fusion. Based on the concept of global information fusion, the present invention improves the network structure in the low-to-up feature fusion process, adds the fusion of the features of the previous layer, and retains more small target information. This structure can accept information from multiple dimensions, realize multi-scale feature fusion, and inject global information into a higher level. In the low-to-up process of the feature pyramid, two upsampling operations are included. After each upsampling, two feature maps with different spatial dimensions are obtained. We use the output feature maps of the P3 and P4 layers as the reference feature maps, and use the feature fusion module SIAF to fuse the information of adjacent levels, while introducing the small target information of the P2 and P3 layers, reducing the increase in parameters and calculations. The improved feature pyramid network FCAFPN is shown as follows Figure 2 shown.

[0078] Further, the dynamic upsampling module in step S24 is expressed as:

[0079] X' = ​​grid_sample(X,S)

[0080] Where X is a feature map of size C×H×W combined with an upsampling factor s, and then a linear layer is used with input and output channels of C and 2s 2 , used to generate a size of 2s 2 ×H×W, and then reshape it to 2×sH×sW by pixel shuffling. The sampling set S is the sum of the offset O and the original sampling grid G, that is,

[0081] O=linear(X), S=G+O

[0082] Finally, the feature map X' is regenerated by combining the original feature map and the sampled set through grid sampling.

[0083] Furthermore, the network training in step S3 includes the following steps:

[0084] Set the initial training parameters of the model: the training cycle is 300, the initial learning rate is 0.01, the weight decay is set to 0.0005, the optimizer selects SGD, the batch size is 16, the input image is resized to 640×640, the momentum is set to 0.932, and the number of epochs required to stop training early is set to 50.

[0085] The enhanced dataset image obtained in step S1 is input into the improved YOLOv8 network described in step S2 for training, and CIoU is used to calculate the bounding box regression loss, the binary cross entropy loss is used to calculate the target confidence loss, and the multi-classification cross entropy loss is used to calculate the classification loss. Back propagation is performed to update the weight parameters, and a deep learning model for drone small target detection is obtained through training.

[0086] In order to evaluate the performance of the model, the evaluation index mean average precision (mAP) is introduced, and the calculation method is as follows:

[0087]

[0088] Among them, n represents the number of categories, AP i represents the AP of the i-th category, P represents the precision, and R represents the recall. The calculation formulas of P and R are as follows:

[0089]

[0090] Among them, T / F stands for true / false, indicating whether the prediction is correct, and P / N stands for positive / negative, indicating whether the prediction is positive or negative.

[0091] Furthermore, in step S4:

[0092] Use the improved YOLOv8 network model to detect targets on the test set, generate an image containing a detection box, compare the image with the original image, and obtain the detection effect of each image. The detection effect is as follows: Figure 6 shown.

[0093] Comparing the experimental results with the benchmark model, the detection accuracy of the present invention increases by 2.8% and 6.5% in the threshold range of 50%-95% and below 50%, respectively. Figure 5 shown.

[0094] The present invention also provides a UAV image small target recognition system, comprising:

[0095] Dataset processing module: used to obtain the drone aerial image dataset in step S1; convert the acquired image format to finally obtain the dataset image in YOLO format; divide the dataset image in YOLO format into training set, verification set and test set; enhance the training set using Mosaic data enhancement method to obtain the enhanced training set image.

[0096] Network construction module: used to implement the construction of an improved UAV aerial image small target recognition network based on YOLOv8 in step S2, and add a small target detection head to the detection head;

[0097] Network training module: used to set the network training parameters in step S3, use the training set to train the drone image small target detection network based on improved YOLOv8, and obtain a deep learning model for drone small target detection after the training is completed;

[0098] Target recognition module: used to implement step S4 in which the test set and the obtained model weights are input into the trained UAV image small target detection network based on improved YOLOv8, generate an image containing a detection frame, compare the image with the original image, and obtain the target recognition result.

[0099] The present invention also provides a small target recognition device for unmanned aerial vehicle aerial images, which is characterized by comprising:

[0100] Memory: a computer program for storing the above-mentioned method for detecting small targets in drone images, which is a computer-readable device;

[0101] Processor: used to implement the method for detecting small targets in drone images when executing the computer program.

[0102] The present invention also provides a computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the method for detecting small targets in drone images.

[0103] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

[0104] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for detecting small targets in drone aerial images, characterized in that: include: S1. Obtain real aerial images taken by drones, convert the obtained images into different formats, and finally obtain dataset images in YOLO format; divide the dataset images in YOLO format into training set, validation set, and test set; enhance the training set using Mosaic data enhancement method to obtain enhanced training set images; S2. Using YOLOv8s as the basic network, the improved YOLOv8 network obtains fine-grained image information by introducing SPD convolution into the backbone of the basic network; Adown and CBS are introduced to construct a new feature fusion module SIAF; the information of adjacent detection layers is fused in the neck structure through this module to form a new feature pyramid network FCAFPN; then the dynamic upsampling module (DySample) is used to replace the nearest neighbor sampling (NN) to reduce the loss of feature information; S3, inputting the enhanced data set obtained in step S1 into the improved YOLOv8 network in step S2 for use in the drone aerial image recognition network; Set initial training parameters; A deep learning model for small target detection of drones is obtained through training; S4: Input the drone aerial image to be detected into the model obtained in step S3 for detection to obtain the target detection result.

2. The method for detecting small targets in drone aerial images according to claim 1, characterized in that: The Mosaic data enhancement method in step S1 enriches the background and scene complexity of the training data by randomly cropping and splicing four different pictures in the data set into a new picture.

3. The method for detecting small targets in drone aerial images according to claim 1, characterized in that: The specific implementation method of step 2 includes the following steps: S21, replace the ordinary convolution of the adjacent layers of the backbone network P2 detection layer with SPD convolution, which divides the input feature map into multiple sub-feature maps and splices them in the depth dimension to obtain a drone aerial photography feature map with half the spatial dimension; S22, constructing a feature fusion module SIAF, wherein the feature fusion module includes ADown, CBS and ordinary convolution; S23, adding a small target information flow transmission channel in the P2 and P3 detection layers to the top-down path of the original PAFPN, and combining the original network connection to perform multi-scale feature fusion through the SIAF module to form a new feature pyramid network FCAFPN; S24, replace nearest neighbor sampling with dynamic upsampling. Dynamic upsampling can better capture details in images or data and generate rich information density feature maps by adjusting sampling strategies in different contexts; S25. Build an improved YOLOv8 network.

4. The method for detecting small targets in drone aerial images according to claim 3, characterized in that: In the step S21: the P2 detection layer is the detection layer where the input feature map is located after being downsampled four times, and a 2-fold downsampling is performed again on the basis of the P2 detection layer to obtain the P3 detection layer.

5. The method for detecting small targets in drone aerial images according to claim 3, characterized in that: The step S22 of constructing the feature fusion module SIAF includes the following steps: The present invention includes two feature fusion modules SIAF, which are used to fuse the feature maps of the two layers of the current layer. Among them, ADown contains an average pooling layer, a maximum pooling layer and two CBS modules (the convolution kernel sizes are 1×1 and 3×3 respectively), which downsample the input feature map by two times. This module divides the feature map into two paths for independent processing, making full use of the diversity of input features and extracting feature information of different scales respectively. Downsample the input feature map by two times; Among them, the CBS module consists of a cascade of convolutional layers, batch normalization layers, and SiLU activation functions to perform channel matching on the input; This stage directly receives input that is dynamically upsampled by a factor of two; Finally, the Contact operation is used to perform feature fusion in the channel dimension, and a 1×1 convolution is used to adjust the number of output channels.

6. The method for detecting small targets in drone aerial images according to claim 1, characterized in that: The specific implementation method of step S3 includes the following network initial training parameters: the training cycle is 300, the initial learning rate is 0.01, the weight decay is set to 0.0005, the optimizer selects SGD, the batch size is 16, the input image is resized to 640×640, the momentum is set to 0.932, and the number of epochs required to stop training early is set to 50.

7. The method for detecting small targets in drone aerial images according to claim 1, characterized in that: In step S4: use the improved YOLOv8 network model to perform target detection on the test set, generate an image containing a detection frame, compare the image with the original image, and obtain the detection effect of each image.

8. A small target recognition system for drone aerial images, characterized by: include: Dataset processing module: obtains public drone aerial photography datasets, converts the dataset format, and performs data enhancement and division on the converted dataset to obtain an enhanced dataset; Network construction module: used to build a small target detection network for drone images based on improved YOLOv8; Network training module: used to train the drone image small target detection network based on improved YOLOv8 using the training set according to the set network training parameters, and obtain the deep learning model of drone small target detection after the training; Target recognition module: used to input the test set and the obtained model weights into the trained UAV image small target detection network based on improved YOLOv8 to obtain the target recognition results.

9. A small target recognition device for drone aerial images, characterized by: include: Memory: a computer program storing a method for detecting small targets in drone images as described in any one of claims 1 to 8, which is a computer-readable device; Processor: used to implement the method for detecting small targets in drone images as described in any one of claims 1-8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the method for detecting small targets in drone images as described in any one of claims 1 to 8.

Images