Unmanned aerial vehicle aerial photography target detection method based on improved YOLOv8

Abstract

The invention discloses an unmanned aerial vehicle aerial photography target detection method based on improved YOLOv8, and the method comprises the steps: constructing and training a target detection network model of the improved YOLOv8, extracting first to fourth scale feature maps of which the scales are sequentially reduced through a backbone network, inputting the feature maps of the four scales into a neck network, and carrying out the detection of the target detection network model of the improved YOLOv8. Context information of different levels is effectively integrated through the improved neck network, feature expression is enriched, and the sensing precision of small targets is improved. And in the head network, enabling the four scale feature maps output by the neck network to respectively pass through respective detection heads to obtain detection results. According to the method, the adaptability of the model to the form change of the target is enhanced, and the influence of complex scenes on detection is reduced. The overall performance is superior to that of other algorithms in the aspect of small target detection, and the performance is more accurate and stable in a complex unmanned aerial vehicle application scene.

Description

Technical Field

[0001] The present application belongs to the field of target detection technology, and in particular to a method for detecting targets in drone aerial photography based on improved YOLOv8. Background Art

[0002] The research and application of combining drones and deep learning technology is developing rapidly and has become an important trend in realizing smart city management. This integration not only provides a new perspective for improving urban management efficiency, but also lays the foundation for future innovations in urban governance models.

[0003] Deep learning technology, especially convolutional neural networks, provides efficient solutions for urban management due to its powerful feature extraction capabilities. Object detection algorithms based on convolutional neural networks can be divided into two categories: two-stage methods (such as Faster-RCNN) improve detection accuracy by selecting candidate regions, but the processing speed is slow; one-stage methods (such as YOLO series, SSD) directly perform object classification and bounding box adjustment, which is more real-time and therefore more common in practical applications.

[0004] Although deep learning has achieved remarkable results in urban management, current methods still have certain limitations in many existing application scenarios. For example, the complexity of the urban environment and the limited viewing angle of fixed cameras make some areas impossible to cover, affecting the effective detection of targets. To overcome these challenges, drones, with their flexible viewing angles and maneuverability, can achieve all-round monitoring and rapid response, thus providing new solutions for blind spot monitoring in urban management.

[0005] Although drone technology has many advantages in urban management, it still faces many challenges in practical applications. First, the shooting angle and height of drones are unstable, resulting in large changes in the size of targets in the image; second, occlusion and background interference in complex environments affect the detection effect of targets; finally, the aggregation of small targets increases the risk of missed detection and false detection. Summary of the invention

[0006] The purpose of this application is to provide a drone aerial target detection method based on improved YOLOv8, aiming to solve the problem of low detection accuracy of drone images in urban management scenarios due to complex background, small target scale, occlusion, etc.

[0007] To solve the above problems, the technical solution of this application is as follows: A drone aerial target detection method based on improved YOLOv8, including: Build and train an improved YOLOv8 target detection network model, including the backbone network, neck network, and head network; The drone aerial image to be identified is input into the backbone network, and the first to fourth scale feature maps with decreasing scales are extracted; The feature maps of the four scales are input into the neck network. The neck network first preprocesses each scale feature map of the input by convolution operation. After the preprocessing, the fourth scale feature map is upsampled and fused with the preprocessed third scale feature map through the feature fusion structure to obtain the first fused feature map. Then, after the upsampled feature map is further fused upward through the feature fusion structure to obtain the second fused feature map and the third fused feature map in turn. The third fused feature map, the second fused feature map and the preprocessed second scale feature map are Figure 1 The feature fusion structure is used to fuse the second-scale output feature map, and then the second-scale output feature map, the first fusion feature map and the preprocessed third-scale feature map are combined. Figure 1 The third-scale output feature map is fused with the fourth feature map after preprocessing through the feature fusion structure to obtain a third-scale output feature map; the third-scale output feature map and the fourth feature map after preprocessing are fused through the feature fusion structure to obtain a fourth-scale output feature map; In the head network, the third fusion feature map, the second scale output feature map, the third scale output feature map and the fourth scale output feature map are respectively passed through their respective detection heads to obtain the detection results.

[0008] Furthermore, a multi-scale global branch coordinate attention module is provided after the spatial pyramid pooling layer of the backbone network, and the output of the spatial pyramid pooling layer is passed through the multi-scale global branch coordinate attention module to obtain a fourth-scale feature map.

[0009] Furthermore, all C2f modules of the backbone network are replaced by C2f-RFMGCAConv modules; Or, all C2f modules of the neck network are replaced by C2f-RFMGCAConv modules; Alternatively, all C2f modules in the backbone network and the neck network are replaced with C2f-RFMGCAConv modules.

[0010] Furthermore, the C2f-RFMGCAConv module replaces cv2 in the Bottleneck structure of C2f with the RFMGCAConv module, and the RFMGCAConv module adds a multi-scale global branch coordinate attention module between the Adjust Shape module and the Re-weight module of the original RFAConv module.

[0011] Furthermore, the multi-scale global branch coordinate attention module performs the following operations: Obtain multi-scale features of the input feature map through global average pooling operation; The height direction features and width direction features of the input feature map are obtained by performing average pooling operations independently along the height and width directions; After concatenating the height-direction features and the width-direction features, the spatial aggregation features are obtained through convolution operation; The multi-scale features are concatenated and then convolved to obtain multi-scale aggregated features; The spatial aggregation features are divided into height and width directions, and then the weight matrices in height and width directions are obtained by convolution operation and activation function respectively; The spatial aggregation features are interactively fused with the multi-scale aggregation features after convolution operation and activation function, and then passed through an activation function to obtain the channel direction weight matrix; Multiply the height and width weight matrices, the channel weight matrix, and the input feature map element-wise to obtain the enhanced feature map.

[0012] This application proposes a method for detecting aerial targets of drones based on improved YOLOv8. Compared with the prior art, the beneficial effects are as follows: the MGCA attention mechanism is proposed to improve the ability to capture features of different scales through multi-scale branches, and to enhance the local and global feature capture of small targets. The neck network is reconstructed into a Bi-MSFPN structure, which effectively integrates context information at different levels, enriches the feature expression, and improves the perception accuracy of small targets. The C2f-RFMGCAConv module is designed to enhance the model's adaptability to target morphological changes, while reducing the impact of complex scenes on detection. The overall performance is superior to other algorithms in small target detection, and is more accurate and stable in complex drone application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of the drone aerial target detection method based on improved YOLOv8 in this application.

[0014] Figure 2 Schematic diagram of the target detection network model in Examples 1 and 2.

[0015] Figure 3 Schematic diagram of the target detection network model in Example 3.

[0016] Figure 4 Schematic diagram of the MGCA structure.

[0017] Figure 5 This is a schematic diagram of the RFMGCAConv structure.

[0018] Figure 6 Schematic diagram of the C2f-RFMGCAConv structure.

[0019] Figure 7This is a comparison chart of evaluation indicators. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0021] Example 1: Figure 1 As shown, a method for detecting drone aerial photography targets based on improved YOLOv8 is provided, including: Step S1: construct and train an improved YOLOv8 target detection network model, including a backbone network, a neck network, and a head network.

[0022] The target detection network model constructed in this application is improved on the basis of YOLOv8. YOLOv8 includes a backbone network (Backbone), a neck network (Neck) and a head network (Head). The backbone network has 10 layers, from layer 0 to layer 9, layers 0, 1, 3, 5, and 7 are convolutional layers, layers 2, 4, 6, and 8 are C2f modules, and the 9th layer is a spatial pyramid pooling layer SPPF; the neck network is responsible for multi-scale feature fusion, and the feature representation capability is enhanced by fusing feature maps from different stages of the backbone network; the head network is responsible for the final target detection.

[0023] It should be noted that YOLOv8 is a relatively mature technology in this field. Its backbone network and bottleneck network include multiple convolution and C2f modules respectively, which are not described here one by one.

[0024] Step S2: input the drone aerial image to be identified into the backbone network, and extract the first to fourth scale feature maps with decreasing scales.

[0025] The YOLOv8 backbone network has four stages (Stage Layer). The traditional YOLOv8 backbone network outputs the scale feature maps extracted in the last three stages to the subsequent neck network. For example, the input image of the YOLOv8 model is 640×640×3 pixels, and the sizes of the three scale feature maps finally generated are 80×80×256 pixels (second scale), 40×40×512 pixels (third scale), and 20×20×1024 pixels (fourth scale).

[0026] For the drone aerial images, this embodiment outputs the scale feature map extracted in the first stage to the subsequent neck network based on the feature maps of the three scales extracted above. Referring to the above example, the scale is 160×160×128 pixels (first scale).

[0027] This embodiment adds a feature map of 160×160×128 pixels, which can more effectively capture the features of small targets. The use of higher resolution feature maps provides more detailed spatial information, allowing the network to detect and locate small targets more sensitively, thereby significantly improving the detection performance of small targets.

[0028] Step S3, input the feature maps of the four scales into the neck network, the neck network first preprocesses each scale feature map of the input by convolution operation, the fourth scale feature map after preprocessing is upsampled and fused with the third scale feature map after preprocessing through the feature fusion structure to obtain the first fused feature map, and then upsample it and continue to fuse upward through the feature fusion structure to obtain the second fused feature map and the third fused feature map in turn, the third fused feature map, the second fused feature map and the second scale feature map after preprocessing are Figure 1 The feature fusion structure is used to fuse the second-scale output feature map, and then the second-scale output feature map, the first fusion feature map and the preprocessed third-scale feature map are combined. Figure 1 The third-scale output feature map is fused with the fourth feature map after preprocessing through the feature fusion structure to obtain a third-scale output feature map; the third-scale output feature map and the fourth feature map after preprocessing are fused through the feature fusion structure to obtain a fourth-scale output feature map.

[0029] This embodiment improves the original neck network of YOLOv8, such as Figure 2 As shown, compared with the original neck network of YOLOv8, the neck network of this embodiment adds a convolution module for preprocessing for each input scale feature map, which can unify the number of channels of the feature map and optimize the fusion between subsequent feature maps. After preprocessing, feature maps of four scales after preprocessing are obtained.

[0030] After upsampling the fourth scale feature map after preprocessing, it is fused with the third scale feature map after preprocessing through the feature fusion structure to obtain the first fused feature map. Then, after upsampling the first fused feature map, it continues to be fused upward through the feature fusion structure, that is, it is fused with the second scale feature map after preprocessing to obtain the second fused feature map. After upsampling the second fused feature map, it continues to be fused upward through the feature fusion structure, that is, it is fused with the first scale feature map after preprocessing to obtain the third fused feature map.

[0031] It should be noted that in the original neck network of YOLOv8, the feature fusion structure includes the Concat module and the C2f module. Taking the first fused feature map as an example, the fourth scale feature map is preprocessed and upsampled to be consistent with the third scale, and then fused with the preprocessed third scale feature map, that is, through the Concat module and the C2f module, to obtain the first fused feature map. The acquisition process of other fused feature maps is similar to the above example, please refer to the attached Figure 2 It is very clear that I will not elaborate on it here.

[0032] In this embodiment, since there are four scale feature maps to be processed, the neck network is improved to achieve weighted bidirectional fusion of features of different scales.

[0033] Specifically, the third fused feature map is directly output to the head network, so the third fused feature map is also called the first-scale output feature map.

[0034] After convolution (Conv), the third fusion feature map is combined with the preprocessed second scale feature map and the second fusion feature map. Figure 1 The input feature fusion structure is used for fusion to obtain the second-scale output feature map.

[0035] The second-scale output feature map is convolved (Conv) and combined with the preprocessed third-scale feature map and the first fusion feature map. Figure 1 The input feature fusion structure is used for fusion to obtain the third-scale output feature map.

[0036] After convolution (Conv), the third-scale output feature map is fused with the preprocessed fourth-scale feature map input feature fusion structure to obtain the fourth-scale output feature map.

[0037] The model constructed in this embodiment can better integrate features from various scales.

[0038] Step S4: In the head network, the third fusion feature map, the second scale output feature map, the third scale output feature map and the fourth scale output feature map are respectively passed through respective detection heads to obtain detection results.

[0039] The head network of this embodiment adds a detection head for the third fusion feature map (that is, the first scale output feature map), which improves the network model's ability to detect small targets.

[0040] It should be noted that the backbone network of this embodiment is Figure 2 The backbone network is slightly different, which is reflected in that there is no MGCA module after the spatial pyramid pooling layer SPPF module, but the fourth-scale feature map is directly output by the SPPF module. That is, the backbone network of this embodiment has the same structure as the original YOLOv8 backbone network, but it outputs the first-scale feature map.

[0041] Embodiment 2: This embodiment is improved on the basis of Embodiment 1, that is, a multi-scale global branch coordinate attention module (Multi-scale Global-branch Coordinate Attention, MGCA) is added after the spatial pyramid pooling layer SPPF module of the backbone network.

[0042] See also Figure 2 ,The MGCA module is added after the spatial pyramid pooling layer SPPF module, and the output of the SPPF module is processed by the MGCA module to obtain the fourth-scale feature map.

[0043] The MGCA module structure is as follows Figure 4 As shown, perform the following operations: Step F1, obtain the multi-scale features of the input feature map through 1x1, 2x2 and 3x3 global average pooling (G Avg Pool) operations; obtain the height direction features and width direction features of the input feature map through independent average pooling (X Avg Pool / Y Avg Pool) operations along the height and width directions.

[0044] This processing method not only captures multi-scale features, but also extracts global information in height and width directions, providing a richer feature basis for subsequent stages.

[0045] Step F2: After concatenating the height direction features and the width direction features, a convolution operation is performed to obtain a spatial aggregation feature; after concatenating the multi-scale features, a convolution operation is performed to obtain a multi-scale aggregation feature.

[0046] The height directional features and width directional features are also called spatial directional features. In this step, the spatial directional features and multi-scale features are aggregated separately.

[0047] Specifically, the features in the height and width directions are concatenated and then convolved to obtain spatial aggregation features to enhance spatial correlation. Multi-scale features are also concatenated and convolved to obtain multi-scale aggregation features to adapt to input images of different scales.

[0048] Step F3: Split the spatial aggregation features according to the height and width directions, and then perform convolution operations and activation functions to obtain the height and width direction weight matrices respectively; perform convolution operations and activation functions on the spatial aggregation features, and then interactively fuse them with the multi-scale aggregation features, and then perform an activation function to obtain the channel direction weight matrix.

[0049] Specifically, the spatially aggregated features are segmented according to the height and width directions, and then convolution operations are performed separately to generate height and width weight matrices to differentially weight the feature responses in different directions.

[0050] Next, the spatial aggregation features are interactively fused with the multi-scale aggregation features after convolution operations and activation functions. The multi-scale features are shared with spatial information by calculating the average value of the processed spatial aggregation features and multiplying them with the multi-scale aggregation feature elements, thereby improving the expression ability of the model in complex scenes. Then, an activation function is passed to obtain the channel direction weight matrix.

[0051] The weight matrix generated by this process can adaptively adjust the importance of each channel in different directions and scales, thereby improving the model's feature resolution and recognition of subtle structures.

[0052] Step F4: multiply the height and width direction weight matrices, the channel direction weight matrix and the input feature map element-wise to obtain an enhanced feature map.

[0053] This step multiplies the previously obtained weight matrices by the input feature map element by element to obtain the enhanced feature output, which combines the weighting of height, width and channel direction, so that the output feature can adaptively adjust the feature strength at each scale and direction, thereby achieving accurate and comprehensive feature expression in the final output, significantly improving the performance of the model in segmentation or detection tasks. Finally, the enhanced feature map is consistent with the input feature map in scale.

[0054] It should be noted that Figure 3 The scales marked in the average pooling module and convolution module indicate the scale of the feature map obtained after the operation. Sigmoid is the activation function.

[0055] Among them, after the height direction features and the width direction features are spliced, the spatial aggregation features are obtained through the convolution operation. The scale of the feature map obtained after the convolution operation is , W represents the width of the input feature map, H represents the height of the input feature map, C represents the number of channels of the input feature map, and r represents the scaling factor, which is used to reduce the computational complexity without excessively losing feature information. After the spatial aggregation features are divided according to the height and width directions, they are convolved respectively, and the feature map scales obtained are and , it can be seen that it is the result of segmentation and reconvolution in the height and width directions respectively.

[0056] Embodiment 3: This embodiment further improves on the basis of Embodiment 2, and replaces all C2f modules in the backbone network and / or the neck network with C2f-RFMGCAConv modules.

[0057] In this embodiment, the original C2f module is replaced by the C2f-RFMGCAConv module. Figure 3 As shown, the model's ability to fuse multi-scale features is improved, thereby reducing the problem of missed detection in occlusion situations.

[0058] Receptive field attention convolution RFAConv is a new type of convolution operation. In this embodiment, the multi-scale global branch coordinate attention module (MGCA) is added between the Adjust Shape module and the Re-weight module of the original RFAConv. Its structure is as follows: Figure 5 As shown in the figure, it is called RFMGCAConv. The adaptive receptive field and efficient feature extraction capabilities of RFAConv complement the multi-scale information fusion capabilities of MGCA, forming a more powerful feature extraction system by focusing on the spatial features of the receptive field. RFMGCAConv combines the advantages of both to improve target detection performance, which can not only strengthen the interaction between features, but also improve the ability to capture spatial and channel information.

[0059] Finally, if Figure 6 As shown in the figure, C2f-RFMGCAConv replaces cv2 in the Bottleneck structure of C2f with the RFMGCAConv module, concentrates the enhancement on the later feature processing, and retains cv1 as a normal convolution, avoiding unnecessary computational burden in the early stage while ensuring the simplicity and efficiency of the initial feature extraction; and can better utilize its receptive field adaptive characteristics to enhance the capture of spatial and channel information, especially when processing multi-scale targets, improving the model's ability to distinguish subtle differences.

[0060] It should be noted that Figure 6 In the , after replacing cv2 in the Bottleneck structure with the RFMGCAConv module, it is called Bottleneck-RFMGCAConv.

[0061] Example 4: Train an improved YOLOv8 target detection network model.

[0062] This example selects and filters the VisDrone2019 dataset, uses the LabelImg tool to additionally annotate the filtered images, and divides them into training sets, test sets, and validation sets. The Mosaic data enhancement method is used to enhance the dataset and establish a drone image dataset.

[0063] In the specific embodiment, 1352 images are selected as the training set, 248 images as the validation set, and 200 images as the test set. The dataset covers 15 categories, including pedestrians, people, bicycles, cars, vans, trucks, tricycles, sunshade tricycles, buses, motorcycles, tables, billboards, vendors, balloon arches, and other obstacles.

[0064] Training parameter configuration, set the initial learning rate to 0.01, the cosine annealing parameter to 0.01, the image size to 640, the learning rate momentum to 0.937, the optimizer to SGD, the batch size to 8, the number of training times to 500, and the weight decay to 0.0005.

[0065] Finally, the training set is sent to the target detection network model for training to obtain the trained model.

[0066] This application will also construct a test set in the dataset and send it to the trained model for testing, evaluate it with indicators such as precision and recall, and compare and analyze it with other target detection methods.

[0067] Specifically, the test set is fed into the trained model for testing, and evaluation indicators such as precision, recall, average precision, model parameter count, billion floating-point operations per second, and number of detection frames per second are obtained. The results are compared and analyzed, and the inference effects of the input images are compared.

[0068] Table 1 algorithm MGCA C2f-RFMGCAConv Bi-MSFPN P / % R / % mAP50 / % mAP50:95 / % YOLOv8 × × × 51.7 34.4 36.2 20.7 A √ × × 52.7 35.4 37.3 21.1 B × √ × 49.4 36.6 37.6 20.9 C × × √ 54.4 40.4 42.7 24.7 D √ √ × 52.3 37.7 39.5 22.1 E √ √ √ 57.5 44.1 47.4 28.0 Table 1 and Figure 7 The ablation experiment results are given, and AE represent different ablation schemes, A is a technical scheme that only uses the MGCA module, B is a technical scheme that only replaces C2f, C is the technical scheme of Example 1 (the improved structure of the neck network is called Bi-MSFPN in this application), D is the A+B scheme, and E is a technical scheme that includes all improvements. It can be seen that the MGCA attention mechanism improves mAP50 and mAP50:95 by 1.1 and 0.4 percentage points respectively. After replacing the C2f module with the C2f-RFMGCAConv module, the recall rate increased by 2.2%, indicating that the module helps to improve the model and the recognition performance of small targets. After introducing the Bi-MSFPN structure, all evaluation indicators showed significant improvements, especially mAP50 and mAP50:95, which increased by 6.5% and 4.0% respectively. When the three improvement methods are applied in combination, the algorithm performance is comprehensively improved. Compared with the baseline model YOLOv8, the precision and recall rates are improved by 5.8% and 9.7% respectively, and the mAP50 and mAP50:95 are improved by 11.2% and 7.3% respectively.

[0069] Table 2 algorithm P / % R / % mAP50 / % mAP50:95 / % Params / M GFLOPs FPS Faster-RCNN 51.0 35.4 40.4 24.8 41.42 208.0 25.4 RetinaNet 50.2 32.1 35.4 21.6 36.32 211.8 28.7 SSD 51.0 34.6 37.3 19.0 26.45 89.5 55.4 YOLOX 56.3 37.2 39.8 22.1 8.9 13.3 79.1 YOLOv5 49.0 31.9 34.1 19.2 9.12 23.9 247.8 YOLOv6 49.3 31.2 33.4 19.0 16.30 44.0 193.1 YOLOv8 51.7 34.4 36.2 20.7 11.13 28.5 208.6 YOLOv9 52.8 36.0 37.4 21.6 7.17 26.8 172.7 YOLOv10 49.4 32.7 33.8 19.2 7.22 21.4 214.7 YOLOv11 47.7 34.6 35.7 20.1 9.43 21.6 210.3 This application 57.5 44.1 47.4 28.0 10.29 57.9 73.6 As shown in Table 2, the proposed method has achieved the best performance in core indicators such as P, R, mAP50 and mAP50:95, which are 57.5%, 44.1%, 47.4% and 28.0% respectively, which is significantly better than mainstream detection algorithms such as Faster-RCNN and SSD. In contrast, the mAP50 of Faster-RCNN is only 40.4%, and the mAP50 of SSD is 37.3%, which is far behind the proposed method. This shows that the proposed method has significant advantages in the accuracy of target detection, and can identify targets more accurately, especially in target detection tasks in complex environments. In terms of model complexity, the number of parameters of the proposed method is 10.29M, which is 8.2% lower than that of YOLOv8. Compared with other real-time algorithms (such as YOLOX-s and YOLOv11s, etc.), its number of parameters remains at the same level. Although the GFLOPs of the method in this application is 57.9, slightly higher than YOLOv8's 28.5, this complexity is still controlled within a reasonable range while ensuring accuracy, demonstrating its ability to efficiently utilize computing resources in a multi-target detection environment. In terms of reasoning speed, the FPS of the method in this application reaches 73.6. Although it is lower than the same YOLO series algorithm, it has significant advantages over Faster-RCNN's 25.4 and RetinaNet's 28.7. This means that while ensuring high detection accuracy, the method in this application still has good real-time performance and is suitable for use in application scenarios with speed requirements. In addition, compared with mainstream real-time algorithms such as YOLOv10s and YOLOv11s, although the method in this application is slightly lower in FPS, it is significantly superior in accuracy, achieving a good balance between speed and accuracy.

[0070] In addition, the experimental results also show that compared with YOLOv8, the method of this application effectively reduces the missed detection rate, can more accurately identify small targets that are far away and obscured, and the detection effect is significantly improved. When the target size difference is large, YOLOv8 has obvious missed detection problems when dealing with near and far targets, while the method of this application can not only accurately identify nearby motorcycles and pedestrians, but also detect more distant cars. In scenes of strong light and darkness, the method of this application demonstrates superior adaptability in extreme lighting environments, which not only enhances the detection capability in strong light scenes, but also significantly reduces the false detection rate in dark environments.

[0071] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A drone aerial target detection method based on improved YOLOv8, characterized in that: The UAV aerial photography target detection method based on improved YOLOv8 includes: Build and train an improved YOLOv8 target detection network model, including the backbone network, neck network, and head network; The drone aerial image to be identified is input into the backbone network, and the first to fourth scale feature maps with decreasing scales are extracted; The feature maps of four scales are input into the neck network, and the neck network first preprocesses each input scale feature map through a convolution operation, and the fourth scale feature map after preprocessing is upsampled and fused with the third scale feature map after preprocessing through a feature fusion structure to obtain a first fused feature map, and then the fourth scale feature map is upsampled and further fused upward through the feature fusion structure to obtain a second fused feature map and a third fused feature map in sequence, and the third fused feature map, the second fused feature map and the second scale feature map after preprocessing are fused together through the feature fusion structure to obtain a second scale output feature map, and then the second scale output feature map, the first fused feature map and the third scale feature map after preprocessing are fused together through the feature fusion structure to obtain a third scale output feature map; the third scale output feature map and the fourth feature map after preprocessing are fused through the feature fusion structure to obtain a fourth scale output feature map; In the head network, the third fusion feature map, the second scale output feature map, the third scale output feature map and the fourth scale output feature map are respectively passed through their respective detection heads to obtain the detection results.

2. The unmanned aerial vehicle target detection method based on improved YOLOv8 according to claim 1, characterized in that: After the spatial pyramid pooling layer of the backbone network, a multi-scale global branch coordinate attention module is further provided, and the output of the spatial pyramid pooling layer is passed through the multi-scale global branch coordinate attention module to obtain a fourth-scale feature map.

3. The unmanned aerial vehicle target detection method based on improved YOLOv8 according to claim 1, characterized in that: All C2f modules in the backbone network are replaced with C2f-RFMGCAConv modules; Or, all C2f modules of the neck network are replaced by C2f-RFMGCAConv modules; Alternatively, all C2f modules in the backbone network and the neck network are replaced with C2f-RFMGCAConv modules.

4. The unmanned aerial vehicle target detection method based on improved YOLOv8 according to claim 3 is characterized in that: The C2f-RFMGCAConv module replaces cv2 in the Bottleneck structure of C2f with the RFMGCAConv module. The RFMGCAConv module adds a multi-scale global branch coordinate attention module between the Adjust Shape module and the Re-weight module of the original RFAConv module.

5. The unmanned aerial vehicle target detection method based on improved YOLOv8 according to claim 2 or 4, characterized in that: The multi-scale global branch coordinate attention module performs the following operations: Obtain multi-scale features of the input feature map through global average pooling operation; The height direction features and width direction features of the input feature map are obtained by performing average pooling operations independently along the height and width directions; After concatenating the height-direction features and the width-direction features, the spatial aggregation features are obtained through convolution operation; The multi-scale features are concatenated and then convolved to obtain multi-scale aggregated features; The spatial aggregation features are divided into height and width directions, and then the weight matrices in height and width directions are obtained by convolution operation and activation function respectively; The spatial aggregation features are interactively fused with the multi-scale aggregation features after convolution operation and activation function, and then passed through an activation function to obtain the channel direction weight matrix; Multiply the height and width weight matrices, the channel weight matrix, and the input feature map element-wise to obtain the enhanced feature map.

Images