Invariant 3D recognition in hyperspectral images

Machine vision has been very effective in recognizing subjects under controlled environments but unsuccessful dealing with illumination and pose variations. Representing illumination using a low-dimensional linear subspace is frequently used to conduct illumination-invariant recognition. Using a set of 7,258 global spectral irradiance functions measured at Boulder, Colorado, the largest known so far, linear models are built to compare with the seminar study of Judd et al. over visible range and with MODTRAN generated global irradiance spectra over near-infrared range. A database of 223 materials is also used to analyze the impact of the Boulder spectral variability on material discriminability.; Algorithms to recognize 3-D objects in airborne 0.4–2.5 micron hyperspectral images acquired under unknown conditions are also presented. The Digital Image and Remote Sensing Scene Generation program is used to build spatial/spectral subspace models for the objects that capture a range of atmospheric and illumination conditions and viewing geometries. The study shows that the 3D objects with single or multiple surface materials could be identified in a couple of pixels even subpixels.; Face recognition methods in hyperspectral images are investigated for its invariance to unknown illumination, face orientation and facial expression. A hyperspectral face image database of 200 human subjects is collected for this study. Recognition is achieved by combining local spectral measurements for different tissue types. One front view image of each subject is stored in the gallery dataset with known identity. Other images of different expressions, face orientations and imaging times are recognized using this gallery. The spectral reflectance of rotated faces are adjusted to extract pose-invariant information. 10 subjects have been imaged outdoor under unknown illumination. The Boulder illumination dataset is used to synthesize radiance spectra and construct linear subspace for each tissue type of each subject. Both outdoor and synthesized radiance images are recognized by the projection of radiance spectra onto the linear subspace of gallery images. Specular reflection on human face is considered in the radiance model. The hyperspectral face image dataset is also evaluated using eigenface method based on both spatial and spectral features.