Abstract

Most color image coding techniques transform the primary color vectors, at each pixel, independently of their spatial arrangement. Since the spatial and chromatic dimensions of natural color images are not independent, efficient coding requires elimination of both spatial and chromatic correlation across chromatic bands. Global spatiochromatic decorrelation has not been utilized because of its enormous computational load.¹ A more feasible local approach is presented which partitions a large image into subimages, where the computational load is reasonable. The subimages are transformed using the most significant eigenvectors generated by diagonalizing a model subimage covariance matrix found empirically by averaging subimage covariance matrices in a natural color image. These eigenvectors generate transformation masks that separate space and color information and form color image basis functions similar to mechanisms in the visual system. The masks are (i) achromatic, spatially oriented passing high spatial frequencies and (ii) color opponent, unoriented passing low spatial frequencies. By utilizing the cross correlation between space and color, this method achieves a high compression ratio (16:1) with good quality reconstruction without elaborate quantization techniques or variable word length coding.

© 1989 Optical Society of America