MS-YOLO: An Efficient Detection Algorithm for Multi-size Prohibited Objects
Pinxuan Zhang, Jiansen Qiao, Defeng Du · 2024
Utilizing security inspection images for prohibited object detection is vital for public safety. Current research primarily addresses occlusion issues in X-ray images but overlooks multi-size object detection, especially for small targets like bullets and lighters. To tackle this problem, we introduce MS-YOLO (YOLO designed for multi-size objects), an improved YOLOv8-based algorithm. MS-YOLO includes an additional small object detection head and uses the BiFPN structure for balanced multi-scale feature fusion. Shape-IoU loss is introduced to focus on object size information, and the DySample method preserves more edges and contours. Experiments on PIDray and CLCXray datasets demonstrate MS-YOLO's SOTA(State of the Art) performance, achieving 88.3%, 84.1%, and 70.8% mAP on PIDray's easy, hard, and hidden sets, respectively, and 72.1% mAP for small targets. On the CLCxray test set, MS-YOLO achieved 65% mAP.