Abstract

The objective is to remove from a given image all parts that do not belong to some logically defined class of objects using both spectral and spatial information. As the spectral information (real spectra in the world or data from a hyperspectral camera) usually have many more than the two spatial components, the first step taken in animal color vision and here is projection of that spectral information onto two or more spectrally overlapping color classes. Most humans, for instance, have three overlapping cone cell spectral sensitivities. Then, again following the biological model, the spectral information can be processed independently using statistical pattern recognition. This removes much of the input image from consideration. Then spatial recognition can be used. We have previously shown that multiple Fourier correlation properly and nonlinearly combined can effect a powerful nonlinear discrimination while preserving the object-locating power of Fourier correlators. We show using an otherwise-difficult task how ell this two step process performs.

© 2006 Optical Society of America

PDF Article