Abstract

The advantages of representing a signal by principal components can be briefly summarized as follows, (i) The basis functions of the representation are orthogonal and have uncorrelated coefficients, (ii) The expected square coefficients are minimum compared with other generalized Fourier representations. These properties make principal components attractive for efficient image coding and representation, because uncorrelatedness reduced redundancy and smaller coefficients require smaller dynamic ranges. Principal components computed for natural color images are comprised of basis functions with various spatial profiles and color responses. These principal components are similar to certain visual system image operations in space and in color, including the color opponent transformation and a variety of spatially organized receptive field filters with different color responses. This similarity suggests that units in the visual system are tuned to features which are principal components of natural images, thereby efficiently coding and representing the image.

© 1991 Optical Society of America