YOLO-GML: An object edge enhancement detection model for UAV aerial images in complex environments

Uav target detection is a key technology in low altitude security, disaster relief and other fields. However, in practical application scenarios, there are many complex and highly uncertain factors, such as extreme weather changes, large scale and span of the target, complex background interference, motion ambiguity, etc., which makes accurate and real-time UAV target detection still a great challenge. In order to reduce the interference of these situations in real detection scenes and improve the accuracy of UAV detection, a Global Edge Information Enhance (GEIE)module is proposed in this paper, which enables edge information to be fused into features extracted at various scales. It can improve the attention of the network to the edge information of the object. In addition, special weather conditions can greatly reduce the detection accuracy of the target, this paper proposes a Multiscale Edge Feature Enhance(MEFE) module to extract features from different scales and highlight edge information, which can improve the model's perception of multi-scale features. Finally, we propose a Lightweight layered Shared Convolutional BN(LLSCB) Detection Head based on LSCD, so that the detection heads share the convolutional layer, and the BN is calculated independently, which improves the detection accuracy and reduces the number of parameters. A high performance YOLO detector (YOLO-GML) based on YOLO11 model is proposed. Experimental results show that Compared with YOLO11s, YOLO-GML can improve AP50 by 2.3% to 73.6% on the challenging UAV detection dataset HazyDet, achieving a better balance between accuracy and inference efficiency compared to the most advanced detection algorithms. YOLO-GML also showed good performance improvement in the SODA-A and VisDrone-2019 datasets, demonstrating the generalization of the model.