Frame Rate Hyperspectral Target Segmentation

Practical use of hyperspectral imagery for target recognition relies on hardware speed and the ability of algorithms to properly separate (segment) targets from image background in the presence of changing illumination. Recent hardware advances have put frame-rate hyperspectral target segmentation on the horizon. Such applications are computationally intense -- requiring over 200 billion arithmetic operations per second. Yet, sensor and ASIC technology will soon be small and fast enough for practical sensing and processing. Algorithm technology for hyperspectral target segmentation is also mostly "there". Basic algorithms descend from developments in earth resource monitoring and a smaller body of military target recognition work. However, key challenges related to illumination must be solved to make the technology robust. JDLL understands these very real practical problems and proposes tasks and a program that will deepen our understanding and address these problems directly and thoroughly. Using simulation, we will probe the physical processes that undermine segmentation and develop algorithms that respond to these processes. Our preliminary research points toward a key innovation of this proposal - an adaptive model for continuously tracking and compensating the effects that shifting illumination hyperspectra have on segmentation algorithms. We propose an adaptive segmentation algorithm based on this model.