Transform coding of stereo image residuals.

Abstract

Stereo image compression is of growing interest because of new display technologies and the needs of telepresence systems. Compared to monoscopic image compression, stereo image compression has received much less attention. A variety of algorithms have appeared in the literature that make use of the cross-view redundancy in the stereo pair. Many of these use the framework of disparity-compensated residual coding, but concentrate on the disparity compensation process rather than the post compensation coding process. This paper studies specialized coding methods for the residual image produced by disparity compensation. The algorithms make use of theoretically expected and experimentally observed characteristics of the disparity-compensated stereo residual to select transforms and quantization methods. Performance is evaluated on mean squared error (MSE) and a stereo-unique metric based on image registration. Exploiting the directional characteristics in a discrete cosine transform (DCT) framework provides its best performance below 0.75 b/pixel for 8-b gray-scale imagery and below 2 b/pixel for 24-b color imagery, In the wavelet algorithm, roughly a 50% reduction in bit rate is possible by encoding only the vertical channel, where much of the stereo information is contained. The proposed algorithms do not incur substantial computational burden beyond that needed for any disparity-compensated residual algorithm.