An Occlusion-Representation and Background-Redundancy-Suppression Network for Prohibited Item Detection with X-Ray Images

Abstract

Prohibited item detection in baggage X-ray images remains challenging because object overlap, partial occlusion, and cluttered backgrounds often obscure target boundaries. These factors make discriminative structural cues difficult to preserve in complex luggage scenes. This study proposes YOLO11-ORS, an occlusion representation and background redundancy suppression network built upon YOLO11n. The proposed model improves feature representation in two ways. First, a Background Redundancy Suppression Pyramid (BRSP) module is designed after semantic feature aggregation to refine contextual features at multiple scales and suppress irrelevant responses from cluttered baggage contents. Second, an Occluded Structure Enhancement Module (OSEM) is constructed at the attention-based feature transformation stage to preserve weak edge cues and incomplete local structures of partially occluded prohibited items. Experiments on the SIXray dataset show that YOLO11-ORS achieves 94.1% mAP@50 and 72.7% mAP@50:95, improving the YOLO11n baseline by 1.2 and 0.9 percentage points. On the OPIXray dataset, YOLO11-ORS improves mAP@50 and mAP@50:95 over YOLO11n by 1.2 and 0.8 percentage points, respectively, while the mAP@50:95 gain reaches 1.5 percentage points under severe occlusion (OL3). Evaluations on difficult test subsets further demonstrate its robustness under cluttered backgrounds, object overlap, and weak target visibility. Additional validation on a self-collected grayscale X-ray dataset confirms its stable performance under different imaging conditions. Overall, YOLO11-ORS improves detection reliability in complex X-ray security inspection scenarios while maintaining practical inference efficiency.