Enhanced Security Screening Using Spectral X-ray Imaging

X-ray image techniques are part of many security-screening systems due to their ability to produce images of the interiors of the object scanned. The classification of threats is carried out via shape recognition and the characterization of materials through the physical properties that are measurable with X-rays. In the case of X-ray Computed Tomography (CT), the measured quantity is the Linear Attenuation Coefficient (LAC), which has energy dependence. The polychromatic nature of laboratory-scale X-ray sources results in energy-dependent distortions in the retrieval of the LAC, when using conventional energy-integrating detectors.
This thesis presents spectral X-ray imaging techniques based on energy-discriminating single-photon counting detectors, with focus on Spectral X-ray CT (SCT). This technique offers the possibility to resolve the energy dependence of the LAC, in a discrete number of energy channels. Thus, it directly enables the estimation of systemindependent properties as the effective atomic number, Z_e, and the electron density, ρ_e.
Energy-dependent effects, such as beam hardening, are mitigated with the proposed method. However, artifacts due to scattering noise and photon starvation by metals remain to be corrected.
This work also introduces a framework for Monte Carlo simulation of SCT measurements used to generate a large amount of training data for machine learningbased correction methods. These posses the advantage of near real-time execution and not requiring a-priori knowledge of the sample. The proposed employs a spectral Convolutional Neural Network architecture, which can learn features from the energy domain. The corrections of real experimental datasets show promising results for both scattering noise and metal artifact removal. The drawback using this method is the introduction of blur, due to the spatial downsampling of the input images and the small size of the network that could fit the computational hardware used for this work. Lastly, this thesis presents a benchmark study of the material classification and threat detection accuracy of 2- and 3-dimensional luggage-type objects. It is found that material features measured with spectral techniques produce better results compared to energy-integrating techniques. The accuracy is improved up to ≈ 35% for the threat detection. In addition, the material classification accuracy is improved up to ≈ 55%.