BL-YOLO:无人机航拍图像目标检测算法

BL-YOLO: A Target Detection Algorithm for Aerial Images Captured by Drones

摘要:

针对传统无人机航拍图像中目标尺度多样、背景复杂、小目标密集等特点，当前目标检测算法存在漏检、误检等问题，提出一种BL-YOLO（BiFPN-LSKA YOLO）目标检测算法。首先，在检测头前增加LSKA（Large Separable Kernel Attention）注意力机制，使模型在检测前能将注意力集中于最关键的特征。然后，分析了无人机航拍图像的各尺寸目标特征，发现在实际检测中，分辨率在20x20及以下的小尺寸目标很难有效地检测到，针对性的在YOLO v10n模型的基础上，添加一个大尺寸160x160目标检测头以提升对航拍图像小目标的检测能力。最后，设计使用BiFPN（Bidirectional Feature Pyramid Network）重构颈部结构，使得网络实现了多层特征融合的同时使网络结构更加轻量化。实验结果表明，基于BL-YOLO网络结构的目标检测的mAP@0.5为37%，模型较YOLO v8n提升了4.0%，较YOLO v10n提升4.6%，较YOLO v11n提升4.4%。BL-YOLO的参数量较YOLO v8n下降了55W，较YOLO v10n下降了23W，较YOLO v11n下降了12W，通过在Dronevehicle红外数据集上的泛化实验也验证了模型的有效性，BL-YOLO算法能够有效地实现无人机航拍图像的目标检测。

Abstract:

In response to the characteristics such as diverse target scales, complex backgrounds, and dense small targets in traditional unmanned aerial vehicle (UAV) aerial images, the current target detection algorithms have issues like missed detections and false detections. Thus, a BL-YOLO (BiFPN-LSKA YOLO) target detection algorithm is proposed. Firstly, an LSKA (Large Separable Kernel Attention) attention mechanism is added before the detection head, enabling the model to concentrate on the most critical features before detection. Subsequently, the characteristics of targets of various sizes in UAV aerial images are analyzed. It is discovered that small-sized targets with a resolution of 20x20 or lower are challenging to be effectively detected in actual detection. Specifically, a large-sized 160x160 target detection head is added on the basis of the YOLO v10n model to enhance the detection capability for small targets in aerial images. Finally, the BiFPN (Bidirectional Feature Pyramid Network) is designed to reconfigure the neck structure, allowing the network to achieve multi-level feature fusion while making the network structure more lightweight. Experimental results indicate that the mAP@0.5 of the target detection based on the BL-YOLO network structure is 37%, which is 4.0% higher than that of YOLO v8n, 4.6% higher than that of YOLO v10n, and 4.4% higher than that of YOLO v11n. The parameter quantity of BL-YOLO is 55W less than that of YOLO v8n, 23W less than that of YOLO v10n, and 12W less than that of YOLO v11n. The generalization experiment on the Dronevehicle infrared dataset also validates the effectiveness of the model, the BL-YOLO algorithm can effectively implement the target detection of UAV aerial images.