Combining Physical and Statistical Models for Recognition in Hyperspectral Images.

摘要

We have completed work in the areas of physics-based illumination modeling, invariant 3D object recognition, and spectral/spatial modeling. The illumination models consider over 7,000 measured visible through short-wave infrared spectra I irradiance functions. We developed compact representations for the spectra, and used the representations to establish new results for invariant material discriminability. We have also developed models and algorithms for the recognition of 3D objects in unknown illumination conditions. The DIRSIG image generation code was used to build invariant spectral/spatial 3D object models. The algorithms have been applied to a series of hyperspectral images with varying spatial resolution. We have also developed a multi- scale opponent representation to hyperspectral texture based on Gabor filter outputs. We have applied this representation to hyperspectral texture classification in AVIRIS images. We have also developed a more detailed hyperspectral spatial structure model using multiband correlation functions.