ESI-YOLO: Enhancing YOLOv8 with Efficient Multi-Scale Attention and Wise-IoU for X-Ray Security Inspection

Security inspection is a priority for preventing threats and criminal activities in public places. X-ray imaging can help with the closed luggages checking process. However, interpreting X-ray images is challenging due to the complexity and diversity of prohibited items. This paper proposes ESI-YOLO, an enhanced YOLOv8-based model for prohibited item detection in X-ray security inspection. The model integrates Efficient Multi-Scale Attention (EMA) and Wise-IoU (WIoU) loss function to improve multi-scale feature representation and detection accuracy. EMA improves multi-scale feature representation, while WIoU enhances bounding box regression, particularly in cluttered and overlapping scenarios. Comprehensive experiments on the CLCXray and PIDray datasets validate the effectiveness of ESI-YOLO. A systematic exploration for the optimal placement of EMA integration on YOLOv8 architecture reveals that the scenario with direct integration in both backbone and neck sections emerges as the most effective configuration without introducing significant computational complexity. Ablation experiments demonstrate the synergistic effect of combining EMA and WIoU in ESI-YOLO, outperforming individual component additions. ESI-YOLO demonstrates notable advancements over the baseline YOLOv8 model, achieving mAP50 improvements of 0.9% on CLCXray and 3.5% on the challenging hidden subset of PIDray, with a computational cost of 8.4 GFLOPs. Compared to other nano-sized models, ESI-YOLO exhibits enhanced accuracy while maintaining computational efficiency, making it a promising solution for practical X-ray security inspection systems.