DAM-YOLO: A Real-Time Detection Strategy for Non-Salient Objects Derived from Feature Enhancement Modeling

The task of real-time object detection aims to identify targets from video streams or continuous image sequences with minimal latency. However, detecting small and easily occluded non-salient objects remains challenging. To confront this challenge, this work introduces an improved object detection method based on YOLOv8, termed DAM-YOLO. Specifically, we first refine the feature fusion mechanism of YOLOv8 by means of implementing a Multi-Scale Aggregation Module (MAM) to assist the backbone in extracting more fine-grained features related to non-salient objects. In the subsequent step, the feature processing method of YOLOv8 is advanced through the integration of a Dual-Stream Attention Mechanism (DAM) module to further strengthen the contextual feature representation linked to nonsalient objects. Compared to other popular real-time detection models in recent years, DAM-YOLO achieves competitive results on two large-scale public datasets, MS COCO 2017 and PASCAL VOC.