Overview and Significance
In CT-based security screening, a challenging problem is to correctly identify and label objects in a scene from X-ray projection data. Conventionally, material parameter reconstruction and labelling are performed as two decoupled steps. Image artifacts induced from metal and other clutter cause variations in apparent material density as well as streaking that can break up homogeneous objects, making their correct identification and assessment challenging. In this project, methods are developed that incorporate the tools of machine learning, physical modeling and Bayesian inference into a unified framework for direct material identification and labeling. Reliable material labelling is critical to the efficient operation of the checkpoint, which is made extremely challenging due to the large range of objects that can appear in baggage, the presence of high clutter, and metal induced image artifacts. The new approach that has been developed can mitigate image artifacts and robustly label materials, thus reducing the number of corner cases, which can in turn reduce false alarms and the need for On-Screen Alarm Resolution Protocol (OSARP) and manual inspection.
We have developed and tested our new methods for robust material identification from multi-energy X-ray sensed data. Using LOIS, we have developed a robust and efficient learning-based method for direct material labeling from multi-energy data sets.Phase 2 Year 2 Annual Report
Faculty and Staff Currently Involved in Project
Students Currently Involved in Project
- Ahmet Tuysuzoglu