Semi-supervised Object-Wise Anomaly Detection for Firearm and Firearm Component Detection in X-ray Security Imagery

Yona Falinie A. Gaus, Brian K.S. Isaac Medina, Neelanjan Bhowmik, Yam T. Lee, Toby Breckon; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 4013-4023

Automatic detection of prohibited items in X-ray imagery plays a vital role in ensuring public safety, particularly in high-throughput venue, transport and postal (border) security. Existing Automatic Prohibited Item Detection Systems (APIDS), based on supervised object detection approaches, are primarily designed for venue and transport screening operations security, where individual baggage items are screened in a controlled manner. However, postal mail screening presents unique challenges due to both of the continuous flow of items, and the desire for high-throughput screening of postal mail items in an unordered and unstructured manner on existing conveyer systems, making the adaptation of current APIDS solutions impractical. To address these challenges, we propose a framework leveraging open-world object detection and semi-supervised anomaly detection as a conduit to effective screening in this context. Our approach jointly uses an open-world object detector to detect generic objects within the cluttered X-ray imagery, followed by a secondary anomaly detection network that identifies outlier objects in a class-agnostic manner. Specifically considering the context of postal screening, experimental results on a UK government evaluation dataset and a locally collected in-house Postal Mail (Parcel) dataset demonstrate the efficacy of our method, achieving high recall (77.76%) and accuracy (75.93%) with low false positive rates (1.98%), thus illustrating future potential in automated postal screening for firearms and firearms components.