Threat detection in X-ray images using CNN

Abstract

Automated X-Ray image inspection makes inspection fast and efficient by focusing inspector’s attention only on the images that are likely to be anomalous. Classification of X-Ray images into threat and non-threat classes was done using Convolutional Neural Network (CNN). CNN architectures AlexNet and VGGNet were trained on a database of 1.2 × 105 X-ray images containing threat and non-threat images, generated using ‘Threat Image Projection’ technique by infusing X-Ray scan of threat with realistic transformations, placed at random positions within the bounds of the baggage. Four channel inputs consisting of log transformation, difference and summation of high and low energy were selected to improve the detection of occluded threats. Ensemble Network, designed to improve threat detection rate by combining and optimizing AlexNet and VGG-19 models, trained using adam optimizer, achieved 99.37% classification accuracy with 99.06% threat detection rate. To get insight into operation of the classifier and features detected by intermediate layers, CNN visualization technique, deconvolutional network was used. Effect of variation in number of layers of VGG-19 on accuracy was also studied. The paper discusses automated threat detection of a prohibited item in X-Ray baggage images.