GS-YOLO: A lightweight high-accuracy model for small target detection in drone aerial images

Abstract

For the problems of weak feature representation, significant scale variation, and background interference in small target features of unmanned aerial vehicle (UAV) aerial images, existing detection methods struggle to achieve both lightweight deployment and detection accuracy. Therefore, this paper proposes an extremely lightweight and accurate small target detection architecture named GS-YOLO. Through modular innovation, it achieves extreme lightness and improved detection performance. The core innovations include: 1) Design of a lightweight small target perception attention fusion module C2FGhostLight, using proportionally optimized GhostConv to replace traditional convolution, combined with a dual-path lightweight attention mechanism, which significantly reduces parameters while dynamically suppressing background interference; 2) Proposal of a lightweight channel attention module for small target perception (SOLCA), through a “channel focusing-local enhancement” dual-branch compact structure and adaptive weighted fusion, to strengthen weak feature representation. Experimental results show that on the VisDrone public dataset, GS-YOLO improves mAP50 by 0.9% compared to YOLOv8n, with a model parameter size of only 0.84M. It maintains lightweight characteristics and provides a solution for engineering applications in UAV aerial photography scenarios.