Physics-Based Recognition in Outdoor Hyperspectral Images Using Small Image Samples

University Of California-Irvine, Irvine CA

Investigators

Abstract

The project focuses on the problem of recognizing objects in outdoor images acquired under unknown conditions when only a few pixels are available on the object. The development of this capability would enable the recognition of small or partially obscured objects at large distances thereby enhancing the performance of systems for many applications such as wide-area search and image registration. A recognition system that operates in an uncontrolled outdoor environment must overcome several substantial challenges. The appearance of an object in an outdoor scene is highly variable due to spatial and temporal variation in the illumination and atmospheric conditions. Also, in images of distant or obscured objects, modeled object surfaces may appear at subpixel scale therefore reducing the usefulness of geometric features. This project will address several important scientific issues. Methods will be developed to build hyperspectral subspace models for materials of interest and backgrounds using physical models and the underlying image data. Models will also developed that describe the statistical properties of these subspaces. These models will be the basis of statistical algorithms for recognition that are invariant to the scene conditions. The models and algorithms will be evaluated using a range of hyperspectral data acquired under different conditions in different environments. An important goal is to determine the fundamental bounds on recognition performance in unknown environments as spatial resolution degrades.