Towards Real-Time Anomaly Detection within X-ray Security Imagery: Self-Supervised Adversarial Training Approach

Abstract

Automatic threat detection is an increasingly important area in X-ray security imaging since it is critical to aid screening operators to identify concealed threats. Due to the cluttered and occluded nature of X-ray baggage imagery and limited dataset availability, few studies in the literature have systematically evaluated the automated X-ray security screening. This thesis provides an exhaustive evaluation of the use of deep Convolutional Neural Networks (CNN) for the image classification and detection problems posed within the field. The use of transfer learning overcomes the limited availability of the object of interest data examples. A thorough evaluation reveals the superiority of the CNN features over conventional hand-crafted features. Further experimentation also demonstrates the capability of the supervised deep object detection techniques as object localization strategies within cluttered X-ray security imagery. By addressing the limitations of the current X-ray datasets such as annotation and class-imbalance, the thesis subsequently transitions the scope to- wards deep unsupervised techniques for the detection of anomalies based on the training on normal (benign) X-ray samples only. The proposed anomaly detection models within the thesis employ a conditional encoder-decoder generative adversarial network that jointly learns the generation of high-dimensional image space and the inference of latent space — minimizing the distance between these images and the latent vectors during training aids in learning the data distribution for the normal samples. As a result, a larger distance metric from this learned data distribution at inference time is indicative of an outlier from that distribution — an anomaly. Experimentation over several benchmark datasets, from varying domains, shows the model efficacy and superiority over previous state-of-the-art approaches. Based on the current approaches and open problems in deep learning, the thesis finally provides discussion and future directions for X-ray security imagery.