OneFocus: Enabling Real-World X-ray Security Screening with a Unified Vision-Language Model

2026-06-14 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Data Science & Analytics · Depth: Expert, quick

Summary

OneFocus is a unified Vision-Language Model (VLM) designed to enhance real-world X-ray security screening by improving contraband detection and visual understanding. Addressing the limitations of conventional detectors and the scarcity of high-quality X-ray image-caption data for VLMs, the system introduces several key components. MMXray, a meticulously curated benchmark, comprises 52,124 image-caption pairs across 28 fine-grained contraband classes. To simulate realistic occlusion, CleanDET, a synthesis dataset with 28 contraband categories, and AnyContraSyn, a controllable synthesis method, were developed. An extensible data curation pipeline, OnePipe, supports this effort. Built upon MMXray, OneFocus supports four core tasks: visual question answering, contraband localization, classification, and image understanding, achieving leading performance and robust cross-domain generalization in X-ray contraband understanding.

Key takeaway

For security screening professionals evaluating next-generation contraband detection, OneFocus demonstrates a robust approach to overcoming data limitations. You should consider integrating unified vision-language models, especially those trained on extensive, curated X-ray datasets like MMXray, to improve adaptability to new threats. This approach offers leading performance and strong cross-domain generalization, enhancing overall security posture.

Key insights

OneFocus, a unified VLM, utilizes a new large-scale X-ray dataset and synthesis methods for leading contraband detection.

Principles

• Data scarcity for VLMs in X-ray security is a critical gap.
• Synthetic data can enrich real-world occlusion patterns.
• Unified VLMs can perform multiple vision-language tasks.

Method

OneFocus is built on MMXray, a 52,124 image-caption pair benchmark. It uses CleanDET and AnyContraSyn for synthetic occlusion data, curated via OnePipe, to train a VLM for four core tasks.

In practice

• Curate large-scale, fine-grained image-caption datasets for VLMs.
• Synthesize occlusion patterns to enhance model robustness.
• Develop unified VLMs for multi-task security screening.

Topics

• X-ray Security Screening
• Vision-Language Models
• Contraband Detection
• MMXray Dataset
• Data Synthesis
• OneFocus Model

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis — DF3DV-1K is a new dataset for benchmarking distractor-free novel view synthesis methods.
• Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models — MLLMs can achieve full-spectrum forensic signal perception by integrating low-level artifact signals via residual injection without losing semantic understanding.
• MMBU: A Massive Multi-modal Biomedical Understanding Benchmark to Probe the Perception Capabilities of Vision-Language Models — The MMBU benchmark reveals current VLMs lack robust visual perception and domain generalization in biomedical contexts.
• DUALVISION: RGB-Infrared Multimodal Large Language Models for Robust Visual Reasoning — Integrating infrared and RGB data via localized cross-attention significantly boosts MLLM robustness in degraded visual conditions.
• CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning — CAVE improves VLM reasoning by assigning structured, action-level credits for visual evidence exploration.
• Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents — Granite 4.0 3B Vision offers compact, modular multimodal intelligence for enterprise document understanding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.