A new transform for lossy color image coding

Luminance and chrominance components of a color image have redundancies often exploited for coding the image with a reduced number of bits, while maintaining good visual quality for the decoded image. Most existing image-coding algorithms use spatial correlation between pixels in each color-transformed image component via wavelet or other spatial transform, but applied separately to the luminance and to each chrominance component. This paper introduces the eidochromatic transform, a non-linear predictive technique that exploits the interdependence between spatially proximate pixels of different color components. The eidochromatic transform reduces the entropy of a quantized chrominance component up to 40% or more over a wavelet transform alone. Combined with JPEG 2000 coding techniques, the eidochromatic transform reduces the bit-rate for the entire image at a specified mean-square error by up to 10% or more when compared to JPEG 2000 alone.