Enhancing Occluded X-Ray Security Screening via Collaborative Optimization and Representation Learning

Image-level multi-label recognition in dual-view X-ray luggage screening faces challenges due to occlusioninduced visibility loss and uneven category difficulty. We propose MCB-U, a lightweight dual-view collaborative framework, integrating a closed-loop coupling between optimization and representation. Our approach includes the Meta Class Balancing-Convex loss function (MCB-Convex) for dynamic category reweighting and CoordAtt-U for task-aware coordinate attention at high-resolution layers. Evaluated on the DvXray classification protocol, MCB-U achieves an mAP of 0.9334, outperforming the Base-BCEwmin baseline by +2.52 points, with notable improvements on heavily occluded and under-learned categories. The model’s attention mechanism remains fixed at inference, introducing no extra computational overhead. The source code is available at https://github.com/DCY512/MCB-U-code.