Information-exchange Enhanced Feature Pyramid Network (IEFPN) for Detecting Prohibited Items in X-ray Security Images

Automatic X-ray security inspection has attracted more and more interest in recent years. Although several methods have been proposed for detecting prohibited items, the accuracy is not high due to scale variation and cluttered backgrounds in X-ray security images. To detect prohibited items with various sizes in cluttered backgrounds, we propose an Information-exchange Enhanced Feature Pyramid Network (IEFPN). Specifically, our IEFPN consists of a Layer-based Recalibration Module (LRM) and a Channel-attention-based Skip-connection Path (CSP). By learning aggregated information weights and re-weighting different level features accordingly, LRM promotes the information exchange of different layers. In this way, features of each level can have access to semantics and details simultaneously. Subsequently, CSP utilizes channel attention to enhance the detailed information in the bottom layer for better localization capacity. Comprehensive experiments on the Security Inspection X-ray (SIXray) and Occluded Prohibited Items X-ray (OPIXray) datasets prove the superior performance of our proposed IEFPN.