Design of a Dual Attention Mechanism for Small Object Detection

Detecting small objects remains a challenging task in object detection, with even notable models like YOLO struggling to achieve high accuracy in complex backgrounds. A dual attention (DA) mechanism specifically designed for small object detection is proposed in this paper, integrating two attention modules: the Global Grouped Coordinate Attention (GGCA) module, which enhances global information retention in feature maps for improved contextual integration, and the Median Pooling-Enhanced Multi-Scale Attention (MPSA) module, which increases noise resilience and refines multi-scale features. Experimental evaluations on datasets such as SHWD and Hardhat and Safety Vest demonstrate that incorporating the DA mechanism into YOLOv7 yields accuracy improvements of 0.75% and 0.59%, respectively, as well as gains of 0.57% and 0.17% on the Traffic Vehicles Dataset and J-EDI dataset. These results confirm that the proposed DA mechanism significantly enhances small object detection performance in complex scenes. Furthermore, this approach shows potential for application in medical imaging, potentially advancing precision in medical image analysis.