| Summary: | Customs and border officers are overwhelmed by the hundreds of millions of cargo containers that constitute the backbone of the global supply chain, any one of which could contain a security- or customs-related threat. Searching for these threats is akin to searching for needles in an ever-growing field of haystacks. This thesis considers novel automated image analysis methods to automate or assist elements of cargo inspection. The four main contributions of this thesis are as follows. Methods are proposed for the measurement and correction of detector wobble in large-scale transmission radiography using Beam Position Detectors (BPDs). Wobble is estimated from BPD measurements using a Random Regression Forest (RRF) model, Bayesian fused with a prior estimate from an Auto-Regression (AR). Next, a series of image corrections are derived, and it is shown that 87% of image error due to wobble can be corrected. This is the first proposed method for correction of wobble in large-scale transmission radiography. A Threat Image Projection (TIP) framework is proposed, for training, probing and evaluating Automated Threat Detection (ATD) algorithms. The TIP method is validated experimentally, and a method is proposed to test whether algorithms can learn to exploit TIP artefacts. A system for Empty Container Verification (ECV) is proposed. The system, trained using TIP, is based on Random Forest (RF) classification of image patches according to fixed geometric features and container location. The method outperforms previous reported results, and is able to detect very small amounts of synthetically concealed smuggled contraband. Finally, a method for ATD is proposed, based on a deep Convolutional Neural Network (CNN), trained from scratch using TIP, and exploits the material information encoded within dual-energy X-ray images to suppress false alarms. The system offers a 100-fold improvement in the false positive rate over prior work.
|
|---|
