An evaluation of deep learning based object detection strategies for threat object detection in baggage security imagery

Abstract

Rising crimes are likely to promote the need for effective security systems for baggage screening at airports. Therefore, technologies enabling public safety are of paramount importance. Out of all the technologies available, X-ray based baggage-screening plays a major role in threat detection. Originally the screening is done manually where a person scrutinizes the X-ray images on a screen to identify potential threat objects. Within this context, with limited dataset availability, we employ an imaging model for a generation of new X-ray images. In this article, an effort is made to perform threat object detection by using deep neural networks based framework. The framework is built upon Convolutional Neural Network (CNN) based techniques such as You Only Look Once (YOLO) and Faster Region based CNN (FRCNN) to perform threat object detection. Apart from this, to improve the model performance with limited original training data the transfer-learning paradigm is also tested out. The performance is studied on 4 classes of threat objects: 1) Gun; 2) Shuriken; 3) Razor-blade; 4) Knife. As compared to traditional Machine Learning (ML) techniques, FRCNN uses region proposal in its first stage to produce better results. On using Faster RCNN with RESNET which was pre-trained on ImageNet dataset, 98.4% accuracy is achieved for 4-class threat recognition requiring 0.16 sec per image. Comparative performance of these threat detection techniques for cluttered X-ray baggage imagery is also presented. We firmly believe that it is possible to fully automate this screening process by using these computer vision techniques