A computational model of spatiochromatic image coding in early vision

We applied decorrelation and redundancy reduction methods to image signals in space and color to derive an image coding operation that simultaneously reduces correlation in both these dimensions. This method, which we apply to the “red” and “green” planes of a color image, is based on an eigenvector (Karhunen-Loeve) transformation in color in conjunction with linear predictive coding in space. The resulting image transformation has a distinct major component associated with an achromatic response, and a second component associated with a chromatic response. The achromatic component has a high spatial frequency bandpass form and contains a large portion of the signal energy. The chromatic component is spatially lowpass and contains chromatic differences which represent only a small portion of the signal energy. A color image is thereby transformed into two distinct components that contain primarily spatial information and primarily color information, respectively. These image components correlate with a high spatial frequency cutoff luminance mechanism and low spatial frequency cutoff chromatic mechanism in the visual system and can be implemented by a simple combination of retinal color-opponent receptive fields.