AMFE-YOLO: A Small Object Detection Model for Drone Images

Abstract

Drones, due to their high efficiency and flexibility, have been widely applied. However, small objects captured by drones are easily affected by various conditions, resulting in suboptimal surveying performance. While the YOLO series has achieved significant success in detecting large targets, it still faces challenges in small target detection. To address this, we propose an innovative model, AMFE-YOLO, aimed at overcoming the bottlenecks in small target detection. Firstly, we introduce the AMFE module to focus on occluded targets, thereby improving detection capabilities in complex environments. Secondly, we design the SFSM module to merge shallow spatial information from the input features with deep semantic information obtained from the neck, enhancing the representation ability of small target features and reducing noise. Additionally, we implement a novel detection strategy that introduces an auxiliary detection head to identify very small targets. Finally, we reconfigured the detection head, effectively addressing the issue of false positives in small-object detection and improving the precision of small object detection. AMFE-YOLO outperforms methods like YOLOv10 and YOLOv11 in terms of mAP on the VisDrone2019 public dataset. Compared to the original YOLOv8s, the average precision improved by 5.5%, while the model parameter size was reduced by 0.7 M.