Adaptive Incremental Learning for Robust X-Ray Threat Detection in Dynamic Operational Environments

Prohibited item detection in X-ray imagery requires rapid adaptation to newly emerging datasets driven by evolving threats, scanner variability, and limited labeled data. In such dynamic environments, vision–language models like Grounding DINO offer advantages over closed-set detectors by leveraging semantic flexibility for unseen categories. However, incremental fine-tuning of vision–language models can lead to knowledge degradation, reducing performance on previously learned security-critical items. To address these challenges, we propose SA-GDINO, an incremental learning framework tailored for vision–language grounding models in the X-ray prohibited-item detection domain. Unlike conventional incremental learning approaches that heavily emphasize stability, SA-GDINO prioritizes plasticity for newly emerging threats while maintaining stability to preserve previously acquired X-ray knowledge. Our method integrates two key components: (1) a phrase prompt strategy that prevents semantic structure collapse by unifying semantically equivalent but lexically distinct tokens into coherent contextual phrases, and (2) a projection-based semantic anchor that preserves inter-class topology in a shared embedding space, achieving a balanced trade-off between plasticity and stability. Evaluated on PIDray, SIXray, and HiXray datasets, SA-GDINO improves Average Precision at 0.5 IoU by 11.6 points over naive fine-tuning and 48.9 points over ZiRa, demonstrating superior adaptability with minimal forgetting.