Applications of hyperspectral imaging techniques to forensic image analysis

• Main Author: Dowler, Shaun Wallace
• Other Authors: Andrews, Mark
• Published: ResearchSpace@Auckland 2011
• Online Access: http://hdl.handle.net/2292/9604

Hyperspectral imaging is a form of imaging spectroscopy developed for remote sensing. Hyperspectral algorithms have many useful properties: particularly robustness to scene conditions and the versatility to analyse a wide variety of scene compositions. Hyperspectral techniques are, however, computationally expensive. Imaging spectroscopy has been applied to the analysis of forensic crime scenes in the recent past with some success. The relative simplicity of the techniques used in these studies, however, has created an opportunity to apply hyperspectral techniques to forensic scenes. This work focused on the development of analysis techniques for camera systems suitable for imaging forensic scenes in the field. Hyperspectral unmixing allows for a scene to be decomposed into a list of material signatures and maps of the abundances of those materials. Winter’s N-FINDR was selected as a suitable unmixing technique for examination due to its popularity, performance and well-understood operation. Analyses of the operation, complexity and performance on simulated and real remote sensing scenes of N-FINDR were conducted to establish a baseline against the body of remote sensing literature. N-FINDR was shown to be an effective, albeit computationally costly, algorithm for analysing hyperspectral data. Two complementary means for reducing the complexity of the N-FINDR algorithm were considered. The algorithm was restructured and the use of an LDU decomposition allowed for redundancies in the computations to be removed. Secondly, a means for reducing the search space of the algorithm was examined and shown to have a favourable complexity-accuracy trade-off. These modifications allow for N-FINDR to form the basis of a hyperspectral still camera system. A new algorithm, Abundance Guided Endmember Selection (AGES), was developed with the property that iterations have low complexity and produce intermediate material maps. A modified version of AGES was used to develop a framework for a video camera system that made use of between-frame redundancy. Both N-FINDR and AGES were compared to more traditional techniques from forensic literature in their performance on blood shoemarks and treated fingermarks. On these scenes, NFINDR and AGES were shown to equal or outperform traditional techniques. The work constitutes major progress towards a system capable of field deployment.