Deep fusion driven semantic segmentation for the automatic recognition of concealed contraband items

Abstract

Automatic detection of prohibited items in passenger baggage is a challenging task, especially in cluttered and occluded concealment scenarios. In this paper, we present a deep fusion driven semantic segmentation network that leverages multi-scale feature representations (extracted via CNN backbone) to generate highly accurate segmentation masks of the suspicious items irrespective of the clutter and concealment. Assessed with the public GDXray, SIXray and OPIXray datasets our proposed architecture reached a mean IoU performance of 0.7768, 0.6263, and 0.6713 respectively, outperforming the leading frameworks. Keywords: Aviation security, Convolutional neural networks, Object recognition, Semantic segmentation, X-ray baggage imagery