AMSA-YOLO: Real-time object detection with adaptive multi-scale attention mechanism.

Abstract

Object detection, as a fundamental task in computer vision, has extensive applications in autonomous driving, video surveillance, medical imaging, and other domains. The YOLO (You Only Look Once) series of algorithms has become the representative method for single-stage object detection due to their excellent real-time performance. However, existing YOLO algorithms still face challenges in small object detection and dense scene detection. This paper proposes AMSA-YOLO (Adaptive Multi-Scale Attention YOLO), an improved YOLO algorithm based on adaptive multi-scale attention mechanism. By introducing scale-aware modules, adaptive spatial attention, and adaptive channel attention, the proposed method significantly improves detection accuracy, particularly for small object detection. Experimental results demonstrate that AMSA-YOLO achieves a 2.3 percentage point improvement in mAP@0.5:0.95 compared to YOLOv8s on the COCO dataset, with a 3.6 percentage point improvement in small object detection AP, while maintaining inference speed with only a 10.3 % decrease. Significant improvements are also achieved on specialized datasets such as VisDrone and CrowdHuman, proving the effectiveness and practicality of the proposed method.