Abstract
Security inspection often deals with a piece of baggage or suitcase whereobjects are heavily overlapped with each other, resulting in an unsatisfactoryperformance for prohibited items detection in X-ray images. In this work,first, we contribute a high-quality dataset named OPIXray, each of which isannotated manually by professional inspectors from an international airport. Tothe best of our knowledge, this is the first dataset specifically designed forobject detection in security inspection. Second, we propose a De-occlusionAttention Module (DOAM) that can be inserted as a plug-and-play module intomost detectors, aiming at detecting occluded prohibited items in X-ray images.Central to DOAM are EIEM, RIAM and HAG. EIEM and RIAM capture edge informationand region information of the prohibited item respectively, and HAG generatesthe attention map by hybridizing the two feature maps generated by EIEM andRIAM to serve a refined feature map to a general detector. We evaluate ourmethod on the OPIXray dataset and compare it to several baselines, includingpopular methods for detection and attention mechanisms. As is shown from theresults, our proposed method significantly outperforms existing models. Thedata and code are released at https://github.com/OPIXray-author/OPIXray