Unmanned aerial vehicle aerial image target detection method and device based on YOLO-GELAN
By optimizing the YOLOv8 network structure and loss function, the ability to detect small targets is enhanced, solving the problem of small target detection in UAV aerial images and achieving efficient and accurate target detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-03
AI Technical Summary
Small target detection in drone aerial images is difficult to perform effectively, and existing deep learning models are difficult to deploy on resource-constrained drone platforms and have insufficient detection accuracy.
The YOLOv8 network structure was optimized by introducing the RepNCSPELAN module and the Wise-IoU loss function to enhance the detection capability of small targets. The number of network parameters and computational complexity were optimized, and the loss function was improved to enhance detection accuracy.
The system enables efficient and accurate small target detection on the drone platform, improving detection accuracy and resource utilization efficiency, and adapting to complex backgrounds and varying lighting conditions.
Smart Images
Figure CN121789090A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of public safety, specifically providing a method and apparatus for target detection in UAV aerial images based on YOLO-GELAN. Background Technology
[0002] With the rapid development and widespread application of unmanned aerial vehicle (UAV) technology, UAV aerial imagery has become an important data source in many fields such as agricultural monitoring, urban planning, environmental protection, and disaster relief. These high-altitude images provide unprecedented perspectives and details, greatly improving the efficiency of data collection and analysis. However, UAV images often cover large geographical areas, making small target detection a particular challenge. These small targets (such as people and vehicles) are small in size and lack detail in the images, making them difficult to detect effectively using traditional target detection algorithms. Therefore, improving the accuracy of small target detection has become a focus of research.
[0003] With the rapid advancement of deep learning technology, the performance of deep learning-based object detection has significantly surpassed that of traditional methods. Deep learning-based object detection algorithms are generally divided into two categories: single-stage methods and two-stage methods. Two-stage methods, such as R-CNN and Faster R-CNN, generate candidate regions first and then extract features for classification and localization. Single-stage algorithms, such as the YOLO series and SSD, directly predict the object category and location using a single network, simplifying the detection process and improving speed. Although two-stage algorithms offer higher detection accuracy, their higher computational complexity makes them unsuitable for scenarios requiring rapid response, such as object detection on fast-moving drones. Therefore, single-stage object detection algorithms, with their faster detection speed, are a better choice.
[0004] Despite significant achievements in deep learning-based object detection technology, its performance in drone object detection applications still needs improvement. One reason is the substantial difference between drone aerial images and general natural landscape images. Drone images typically encompass large-scale scenes, multi-scale targets, complex background interference, and mutual occlusion between targets, significantly increasing the difficulty of detection. Furthermore, drone object detection demands high real-time performance and accuracy, while complex deep learning models are often difficult to deploy directly on resource-constrained drone equipment. Simplified, lightweight models, on the other hand, usually sacrifice some accuracy. Summary of the Invention
[0005] This invention addresses the shortcomings of the prior art by providing a highly practical target detection method for UAV aerial images based on YOLO-GELAN.
[0006] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable target detection device for UAV aerial images based on YOLO-GELAN.
[0007] The technical solution adopted by this invention to solve its technical problem is:
[0008] The target detection method for UAV aerial images based on YOLO-GELAN has the following steps:
[0009] S1. Network structure optimization;
[0010] S2, Added small target detection;
[0011] S3. Improve the loss function.
[0012] Furthermore, in step S1, the traditional C2f module is replaced by the RepNCSPELAN module in the aggregation network in the YOLOv8 network structure. RepNCSPELAN performs feature extraction and fusion through a series of convolutional layers Conv and RepNCSP modules.
[0013] The RepNCSPELAN module includes the number of input channels, the number of output channels, and the number of intermediate channels, which determine the size and complexity of the feature map.
[0014] Furthermore, the Conv and RepNCSP modules enable RepNCSPELAN to effectively fuse local and global features, thereby enhancing the accuracy and robustness of target detection.
[0015] While maintaining feature extraction efficiency, the RepNCSPELAN module reduces the number of network parameters and computational complexity.
[0016] Furthermore, in step S2, a new Detect module with a feature size of 160×160 is added, and the original module with a feature size of 160×160 is deleted. The feature maps of its output layer are modified to be 160×160, 80×80, and 40×40, with the newly added 160×160 feature map.
[0017] Furthermore, the input image is resized to 640×640×3, and three feature maps are generated with sizes of 80×80×128, 40×40×256, and 20×20×512, respectively. An output is taken from the first C2faster module in the Backbone, and upsampling, concat, and C2faster modules are added to the Neck. A Detect module is added to the Head. Then, the upsampling, concat, C2faster, and conv modules connected to SPPF in the Neck and the Detect module in the Head are removed.
[0018] Furthermore, in step S3, Wise-IoU is introduced as a new bounding box loss function to optimize the model's sensitivity to data quality and sample balance. There are three versions of Wise-IoU: WIoUv1, WIoUv2, and WIoUv3.
[0019] Furthermore, WIoUv1 focuses on reducing the gradient gain of high-quality anchor frames and increasing the loss of ordinary-quality anchor frames.
[0020] The WIoUv2, based on WIoUv1, introduces a monotonic focusing mechanism to dynamically update the gradient gain;
[0021] The WIoUv3 uses outlier to evaluate the quality of the anchor frame and employs a dynamic, non-monotonic gradient gain allocation strategy.
[0022] A target detection device for UAV aerial images based on YOLO-GELAN includes: at least one memory and at least one processor;
[0023] The at least one memory is used to store a machine-readable program;
[0024] The at least one processor is used to call the machine-readable program to execute a target detection method for UAV aerial images based on YOLO-GELAN.
[0025] Compared with existing technologies, the target detection method and apparatus for UAV aerial images based on YOLO-GELAN of the present invention have the following outstanding advantages:
[0026] This invention aims to improve the target detection accuracy of UAV aerial images while effectively reducing the number of network parameters by optimizing the network structure, adding a small target detection layer, introducing a generalized efficient aggregation network (GELAN), and using the Wise-IoU loss function. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 This is a schematic diagram of the YOLOv8 network structure in a target detection method for UAV aerial images based on YOLO-GELAN;
[0029] Figure 2 This is a schematic diagram of the YOLO-GELAN structure in a target detection method for UAV aerial images based on YOLO-GELAN.
[0030] Figure 3 This is a distribution map of the VisDrone dataset in a target detection method for UAV aerial images based on YOLO-GELAN;
[0031] Figure 4 This image shows the detection results of the algorithm presented in this paper, which is a target detection method for UAV aerial images based on YOLO-GELAN. Detailed Implementation
[0032] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0033] The following is a preferred embodiment:
[0034] This embodiment presents a YOLO-GELAN-based object detection method for UAV aerial images. YOLOv8, as the latest member of the YOLO family, represents the forefront of object detection. It inherits the advantages of previous YOLO versions and innovates and optimizes upon them to provide superior performance and greater operational flexibility. YOLOv8 is capable of handling various computer vision tasks, including but not limited to object detection, image classification, and instance segmentation; this paper focuses on its application in object detection.
[0035] The YOLOv8 family includes five variants of different sizes: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x, with the number of parameters and computational complexity increasing sequentially. Considering the dual requirements of real-time performance and accuracy in application scenarios, YOLOv8n was chosen as the basic architecture for improvement because it combines a smaller model size, ease of deployment, and high recognition accuracy, making it particularly suitable for tracking fast-moving and feature-blurred low-altitude UAV targets.
[0036] The YOLOv8 model architecture consists of four key parts: the input layer, the backbone network, the neck network, and the head. These parts work together to achieve efficient and accurate object detection.
[0037] The input section employs a mosaic data augmentation technique, which generates more diverse training samples by randomly scaling, cropping, and recombining four images, thereby enhancing the model's generalization ability. Furthermore, YOLOv8 uses an anchor-free mechanism to directly predict the target center point, eliminating the reliance on pre-set anchor boxes. This not only simplifies computation but also improves detection speed and accuracy.
[0038] The backbone network, serving as the core of feature extraction, integrates multiple modules, including convolutional layers (Conv), cross-stage local layers (C2f), and fast spatial pyramid pooling (SPPF). The convolutional modules perform basic feature extraction, while the C2f module borrows design ideas from C3 in YOLOv5, achieving both lightweighting and performance improvement by adding gradient propagation paths. The SPPF module transforms feature maps of different scales into feature vectors of uniform size, enhancing the model's adaptability to multi-scale targets.
[0039] The Neck section focuses on feature fusion and adopts the PANet structure, which combines the Feature Pyramid Network (FPN) and the Path Aggregation Network (PAN). Through top-down feature fusion and bottom-up information transfer, it not only preserves high-level semantic information but also enhances the positional sensitivity of the lower level, thus achieving full utilization of cross-scale features.
[0040] The Head layer, serving as the model's output layer, performs the final prediction based on the aforementioned multi-scale feature maps, including object category and location information. This ensures the model maintains good detection performance on targets of varying sizes. The network structure diagram is as follows: Figure 1 As shown.
[0041] YOLO-GELAN optimizes the original YOLOv8 architecture by reducing the number of model parameters. The improvements include:
[0042] S1. Network structure optimization.
[0043] In the YOLOv8 network architecture, the RepNCSPELAN module from the generalized efficient aggregation network is used to replace the traditional C2f module. While maintaining model compactness and high computational efficiency, the network parameters are reduced and the extraction of key target features is enhanced, thereby improving the target detection effect and further improving the overall performance of YOLOv8-based target detection.
[0044] The RepNCSPELAN module, in order to improve the network's feature extraction capabilities and detection performance, draws inspiration from the design principles of the GELAN network and optimizes the original network structure. GELAN is a lightweight network structure based on gradient path planning, combining the advantages of CSPNet (Cross Stage Partial Network) and ELAN (Efficient Layer Aggregation Network). It optimizes information flow and gradient propagation through cross-layer connections and efficient feature aggregation, maintaining efficient feature extraction and information representation capabilities even with reduced parameter count and computational complexity. This design not only reduces the number of model parameters and computational complexity but also enhances the model's flexibility, making it suitable for various hardware and application scenarios.
[0045] Based on the design principles of the GELAN network, this paper replaces the original C2f module with its key module RepNCSPELAN, further optimizing the network structure. RepNCSPELAN performs feature extraction and fusion through a series of convolutional layers (Conv) and the RepNCSP (Replicated Cross Stage Partial Network) module. The main components of this module include the number of input channels, the number of output channels (c2), and the number of intermediate channels (c3, c4), which determine the size and complexity of the feature map.
[0046] Through the synergistic effect of these convolutional layers and the RepNCSP module, RepNCSPELAN effectively fuses local and global features, thereby enhancing the accuracy and robustness of object detection. While maintaining feature extraction efficiency, the RepNCSPELAN module significantly reduces the number of network parameters and computational complexity, optimizing the use of hardware resources, thus achieving a better balance between performance and resource consumption in the entire detection model.
[0047] S2, Added small target detection layer;
[0048] Based on the original output layer feature maps of 80×80, 40×40, and 20×20, a new Detect module with a feature size of 160×160 is added, and the original module with a feature size of 160×160 is deleted. The output layer feature maps are modified to be 160×160, 80×80, and 40×40. The newly added 160×160 feature map aims to solve the problem of insufficient capture of feature information of small targets in images, thereby further improving the accuracy of small target detection.
[0049] In the YOLOv8s model, the input image is resized to 640×640×3, generating three feature maps with sizes of 80×80×128, 40×40×256, and 20×20×512, respectively. However, when detecting small targets, the minimum size of the feature maps is reduced to 20×20×512, potentially leading to the loss of significant small target feature information. Furthermore, due to deep convolutional operations, the feature information of small targets may be further weakened, thus reducing detection accuracy.
[0050] To address this issue, this paper extracts an output (160×160×64) from the first C2faster module in the Backbone, adds upsampling, Concat, and C2faster modules to the Neck, and adds a Detect module to the Head. Furthermore, the original small target detection layer is improved by removing the upsampling, Concat, C2faster, and conv modules connected to SPPF in the Neck, and removing a Detect module from the Head.
[0051] The original detection layer divided a 640×640 input image into 20×20 detection units, with each unit measuring 32×32. However, for small targets, such large units increase the interference of background information on target features, thus reducing detection accuracy. The improved detection layer, under the same input conditions, uses only 4×4 units per cell, effectively extracting key target features while reducing background interference, thereby significantly improving the detection accuracy of small targets.
[0052] S3. Improve the loss function;
[0053] Using the Wise-IoU loss function reduces the large gradient problem caused by extreme values, thereby improving the overall performance of the network model. The improved network structure diagram is shown below. Figure 2 As shown.
[0054] Wise-IoU is introduced as a new bounding box loss function to optimize the model's sensitivity to data quality and sample balance, thereby further improving the overall performance of object detection.
[0055] Wise-IoU has three versions: WIoUv1, WIoUv2, and WIoUv3.
[0056] WIoUv1 primarily focuses on reducing the gradient gain for high-quality anchor boxes and significantly increasing the loss for ordinary-quality anchor boxes, thereby enhancing the model's attention to ordinary-quality anchor boxes. WIoUv2, building on WIoUv1, introduces a monotonic focusing mechanism, dynamically updating the gradient gain to address the slow convergence issue in the later stages of training. WIoUv3 uses outlier evaluation to assess anchor box quality and employs a dynamic, non-monotonic gradient gain allocation strategy, which more effectively handles low-quality examples in the training data, reducing their negative impact on model performance and improving the overall performance of the detector. Therefore, this paper uses WIoUv3 for experimental research.
[0057] The VisDrone2019 dataset, collected and released by the Machine Learning and Data Mining Laboratory of Tianjin University, contains 8629 images. 6471 images were used for training, 548 for validation, and 1610 for the test set. The dataset covers 10 common everyday scene targets: awning tricycle, tricycle, bus, bicycle, truck, van, people, motor, pedestrian, and car. The class distribution in the dataset is imbalanced, and most targets are small, increasing the difficulty of detection. The class distribution of the VisDrone2019 dataset is as follows: Figure 3 As shown.
[0058] The experimental operating system was Ubuntu 18.04.5, using an NVIDIA GeForce RTX 3090 graphics card with 24GB of video memory. The programming language was Python 3.8.18, CUDA version 11.7, torch version 1.13.1, and YOLOv8 version 8.1.6. All experiments were conducted under the same hyperparameters.
[0059] In the YOLO series models, network performance evaluation metrics include: precision (P), recall (R), and mean precision (mAP). This experiment uses mAP50 as the primary performance reference metric. mAP50 represents the mAP value with a threshold of 0.5; the higher the value, the higher the overall precision of the model. The calculation formulas for each metric are as follows:
[0060]
[0061] When the model correctly predicts the target, Tp represents the number of true positives; when the model incorrectly predicts a non-small target as a small target, represents the number of false positives; when the model misclassifies a small target as another category, represents the number of false negatives; represents the accuracy of a certain class of predictions; represents the average accuracy of all classes, where is the total number of classes.
[0062] The detection results were obtained by analyzing drone aerial images from the VisDrone 2019 test suite. Figure 4 As shown in the figure. The results demonstrate that the proposed algorithm can accurately detect and locate small targets such as cars and pedestrians under different lighting conditions and complex scene backgrounds. Specifically, in different situations such as sparse and dense scenes, day and night, the algorithm successfully suppresses the interference of background noise information such as buildings and trees, achieving accurate target classification and localization. Detection results show that the proposed algorithm performs excellently in real-world scenarios, demonstrating strong adaptability and robustness. This algorithm can not only effectively cope with complex backgrounds and changing lighting conditions, but also accurately detect targets of different sizes, making it highly practical and reliable for target detection tasks in UAV aerial imagery.
[0063] Leveraging the unique properties of UAV aerial images, a YOLO-GELAN object detection network was designed, enabling efficient detection in complex and varied environments and images with densely packed targets. By adjusting the network structure and adding a small target detection layer specifically designed for small targets, the network's ability to detect targets at different scales was improved. To adapt to the UAV platform, the RepNCSPELAN module from the generalized efficient aggregation network was introduced, reducing the number of parameters while improving the detection performance of small targets. This design enhances the network's adaptability to the UAV platform. Finally, by introducing the WIoU loss function, target location information was finely processed, reducing the localization error of small targets and improving detection accuracy. Test results on the VisDrone2019 dataset demonstrate that YOLO-GELAN performs exceptionally well in target detection in UAV aerial images, surpassing existing mainstream detection networks.
[0064] Although YOLO-GELAN already has high detection accuracy, there is still room for improvement, especially at the network level and in detection.
[0065] Regarding accuracy and speed, future research will focus on further optimizing the network architecture, reducing model size, decreasing the number of parameters, and improving real-time detection speed to achieve more efficient target detection.
[0066] Based on the above method, the UAV aerial image target detection device based on YOLO-GELAN in this embodiment includes: at least one memory and at least one processor;
[0067] The at least one memory is used to store a machine-readable program;
[0068] The at least one processor is used to call the machine-readable program to execute a target detection method for UAV aerial images based on YOLO-GELAN.
[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not 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; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A target detection method for UAV aerial images based on YOLO-GELAN, characterized in that, It has the following steps: S1. Network structure optimization; S2, Added small target detection; S3. Improve the loss function.
2. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 1, characterized in that, In step S1, the traditional C2f module is replaced by the RepNCSPELAN module in the aggregation network in the YOLOv8 network structure. RepNCSPELAN performs feature extraction and fusion through a series of convolutional layers Conv and RepNCSP modules. The RepNCSPELAN module includes the number of input channels, the number of output channels, and the number of intermediate channels, which determine the size and complexity of the feature map.
3. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 2, characterized in that, The Conv and RepNCSP modules of the convolutional layer enable RepNCSPELAN to effectively fuse local and global features, thereby enhancing the accuracy and robustness of target detection. While maintaining feature extraction efficiency, the RepNCSPELAN module reduces the number of network parameters and computational complexity.
4. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 3, characterized in that, In step S2, a new Detect module with a feature size of 160×160 is added, and the original module with a feature size of 160×160 is deleted. The feature maps of its output layer are modified to: 160×160, 80×80, 40×40, and a new feature map with a size of 160×160 is added.
5. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 4, characterized in that, The input image is resized to 640×640×3, and three feature maps are generated with sizes of 80×80×128, 40×40×256, and 20×20×512, respectively. An output is taken from the first C2faster module in the Backbone, and upsampling, concat, and C2faster modules are added to the Neck. A Detect module is added to the Head. Then, the upsampling, concat, C2faster, and conv modules connected to SPPF in the Neck and the Detect module in the Head are removed.
6. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 5, characterized in that, In step S3, Wise-IoU is introduced as a new bounding box loss function to optimize the model's sensitivity to data quality and sample balance. There are three versions of Wise-IoU: WIoUv1, WIoUv2, and WIoUv3.
7. The target detection method for UAV aerial images based on YOLO-GELAN according to claim 6, characterized in that, The WIoUv1 focuses on reducing the gradient gain of high-quality anchor frames and increasing the loss of ordinary-quality anchor frames; The WIoUv2, based on WIoUv1, introduces a monotonic focusing mechanism to dynamically update the gradient gain; The WIoUv3 uses outlier to evaluate the quality of the anchor frame and employs a dynamic, non-monotonic gradient gain allocation strategy.
8. A target detection device for UAV aerial images based on YOLO-GELAN, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to perform the method according to any one of claims 1 to 7.