Global-Local Context Enhanced YOLO for Small Object Detection in UAV Images

Object detection using Unmanned Aerial Vehicles (UAVs) has emerged as a crit- ical application across diverse domains. However, the wide-angle views of drones often result in images containing a high density of small objects, posing chal- lenges for object detection such as few learnable features, significant occlusion, and an imbalanced distribution of positive and negative samples. To address these issues, this paper introduces AGLC-YOLO, an enhanced version of the YOLOv7 architecture specifically designed for detecting small objects in UAV images. AGLC-YOLO integrates global and local context information through a Attention guide Global-Local Context Information Extraction (AGLC) module. This module employs parallel dilated convolutions to capture local context infor- mation and a transformer-based structure to extract global dependencies, which are then fused using an improved attention mechanism. The network also adds an additional small object detection head to enrich the small object informa- tion in the model. Additionally, AGLC-YOLO utilizes an auxiliary bounding box in conjunction with the Inner-Wise Intersection over Union (Inner-WIoU) loss function to accelerate the bounding box regression process and improve detec- tion accuracy. Experimental results on the VisDrone and ManipalUav datasets demonstrate that AGLC-YOLO achieves significant improvements over the base- line YOLOv7 model, with an increase of 3% in AP50 and 2.7% in AP95 on the VisDrone dataset, and 1.9% in AP50 and 2% in AP95 on the ManipalUav dataset. Source code is released in https://github.com/hanks124/aglc.