Research map / Detection and segmentation / Open-set and few-shot detection

Dual-view X-ray Detection: Can AI Detect Prohibited Items from Dual-view X-ray Images like Humans?

CVPR 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Industrial anomaly detection · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LDXray, a large-scale dual-view X-ray prohibited-item detection dataset (353,646 instances, 12 categories), and AENet, which uses main-view detection plus saliency-guided auxiliary-view localization and expert models for hard categories.

Problem

Detecting prohibited items in X-ray luggage images is hard for challenging categories under heavy occlusion from a single view, and existing X-ray datasets are single-view or lack sample diversity.

Approach

Mimics human inspectors by pairing a main-view detector with an auxiliary-view pipeline: saliency detection localizes objects in the side view, labels are transferred via horizontal cross-view correspondence, and crops from the main view are refined by expert detectors trained on hard categories.

Key results

• AENet improves umbrella (UM) AP50 from 33.9% to 58.6% (+24.7%) on LDXray with the main-view-only baseline setup described in the paper.
• With Faster R-CNN, AENet raises overall mAP by 1.9%, and with Cascade R-CNN by 2.0%, while keeping simple categories stable.
• Saliency-based auxiliary-view localization raises mAP from 38.2% to 39.1% over the single-view baseline; C-RCNN as auxiliary detector reaches 26.7% mAP.

Measured effects reported in the paper

Surprising findings

• Non-metallic lighters are almost indistinguishable from background under X-ray, so several detectors score near zero AP on them.

Limitations

• Auxiliary view has no ground-truth labels, so localization relies on saliency approximations and horizontal coordinate correspondence assumed from fixed imaging geometry.
• Non-metallic lighters remain nearly undetectable (near-zero AP), so the method does not solve the hardest category.

Experimental setup

Datasets: LDXray

Benchmarks: LDXray

Metrics: AP50 mAP AP75

Baselines: Faster R-CNN Cascade R-CNN Sparse R-CNN RetinaNet CenterNet RepPoints ATSS

Builds on: Faster R-CNN Cascade R-CNN ATSS mmdetection

Closest papers by meaning

• CSPCL: Category Semantic Prior Contrastive Learning for Deformable DETR-Based Prohibited Item Detectors (NeurIPS 2025 · similarity 0.82)
• BenchX: A Unified Benchmark Framework for Medical Vision-Language Pretraining on Chest X-Rays (NeurIPS 2024 · similarity 0.82)
• Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models (NeurIPS 2026 · similarity 0.81)
• SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images (NeurIPS 2026 · similarity 0.81)
• STING-BEE: Towards Vision-Language Model for Real-World X-ray Baggage Security Inspection (CVPR 2025 · similarity 0.81)
• RadZero: Similarity-Based Cross-Attention for Explainable Vision-Language Alignment in Chest X-ray with Zero-Shot Multi-Task Capability (NeurIPS 2025 · similarity 0.80)