ALERT | Awareness and Localization of Explosive-Related Threats

Detection of Explosives using Hyperspectral Imaging
F2-A

Download the 2012 Project Report

The focus of this project was to develop and implement detection algorithms for imag­ing and non-imaging spectroscopy sensing modalities to best detect and identify explosive-related threats. The proposed approach included modeling of spectral variability, signal and image enhance­ment, and feature extraction for target detection, and computational implementation using GPUs. The developed algorithms were tested and validated using standard data sets used in hyperspectral target detection and phantom data generated at UPRM-LARSIP. We demonstrated the use of Graphic Processing Units (GPUs) to implement hyperspectral target detection algorithms. Significant speed­up was demonstrated which is needed for real-time systems such as those used in explosive detection portal systems. We also developed an understanding and show with examples how factors such as memory management and data transfer between CPU and GPU affect algorithm performance. An in­ternal REU effort was developed with students from the UPRM Computer Engineering program to de­velop a library of GPU routines for hyperspectral target detection, focusing on the portability aspects of the library. Cmake and Ctest and software engineering practices and tools were used to develop a cross platform library called Libdect. The use of a library can encourage new developers in the field of hyperspectral image application development. The efforts concentrated on the design of the Lib­dect library prototype structure, coding guidelines, its build system, and tools. Two target detection algorithms were ported to the library: RX and Matched Filter (MF). One additional algorithm, AMSD is still under development. A new hyperspectral image enhancement algorithm was developed using tensor anisotropic nonlinear diffusion (TAND). The approach is based on a new structure tensor that better incorporates spectral/spatial information and serves as a basis for the diffusion tensor. The TAND approach resulted in improved contrast enhancement between background and small spatial features such as traces of explosives.