CEDP-YOLO: UAV Object Detection Based on Context Enhancement and Dynamic Perception

Zhenhui Ding, Zengbin Zhang, Mao Yuan, Guangxiao Ma, Guohua Lv

Abstract: The object detection of unmanned aerial vehicle (UAV) images has widespread applications in numerous fields. However, the dense occlusion, scale changes, and small scale of objects in UAV images make object detection a challenging task. Based on the complex background, high resolution and large object scale differences in UAV images, we propose an object detection network based on context enhancement and dynamic perception (CEDP-YOLO). Firstly, by combining the small object feature layer, Content-Aware ReAssembly of FEatures (CARAFE) and Separated and Enhancement Attention Module (SEAM), the Context Enhanced Pyramid (CEPAN) network is designed to enhance the model’s feature perception and feature fusion capabilities for small objects. Secondly, deformable convolution, global perception module and semi-decoupled head are used to dynamically perceive scale changes and enhance classification and positioning capabilities. Finally, a bounding box loss function Inner-Minimum Point Distance Intersection over Union (IMPDIoU) is proposed to improve the bounding box regression rate and positioning accuracy, which combines bounding box similarity with minimum point distance and auxiliary bounding boxes to optimize bounding box detection. Extensive experiments on the benchmark dataset show that the proposed method has achieved the highest score of 34.0% on mAP@.5, which is 5.3% higher than the baseline method. The code of the proposed method is publicly available on https://github.com/GODFArher/CEDP-yolo.