XrayCLIP: A VLM-Based X-Ray Security Inspection System for Railway Safety

Abstract

Railway safety is an important component of public security. Ensuring the safety of high-speed rail systems and passengers is also crucial for railway transportation enterprises. The performance of X-ray security inspection systems is one of the key factors in improving intelligent railway security. Previous studies have shown that, due to the complexity of X-ray images, both traditional vision methods and deep learning approaches struggle to meet the requirements of real-world railway security inspection. With the development of vision-language models, this paper proposes XrayCLIP, a CLIP-based method designed to improve the accuracy and robustness of computerized X-ray security inspection. XrayCLIP adapts CLIP to the semantic and imaging characteristics of prohibited-item inspection through security-oriented prompts, texture-aware visual representations, and global–local supervision. The key component of the model is a set of learnable prompt templates, which guide the model to learn generic features of prohibited objects in complex environments. Multi-level global–local (glocal) features enable the model to focus on both global context and local details, while text space optimization, texture enhancement, text–image fusion, and inference enhancement further improve model performance. XrayCLIP is evaluated on the PIDray and derived HiXray(seg) benchmarks against eight conventional and recent single-view baselines under a common evaluation protocol. A railway-station study is additionally conducted using operational X-ray data, including a same-set missed-detection comparison with a commercial system for knives and power banks. The results show that XrayCLIP achieves the best performance among the evaluated methods on PIDray and HiXray(seg), while reducing the missed-detection rate relative to the commercial system on the two categories examined.