An image coding algorithm based on a class of doubly stochastic image models

The authors present a novel image coding algorithm based on a class of image models known as doubly stochastic Gaussian models (DSGM). They exploit the nonhomogeneous nature of images by a space-variant autoregressive representation that switches in a set of linear predictive submodels. The switch is controlled by a 2-D Markov chain. The coder is a DPCM (differential pulse code modulation) system that is given the submodel (predictor) coefficients. In order to obtain these predictors, a recursive state estimation of the underlying Markov chain is done. This combination provides the doubly recursive prediction nature of this algorithm. Two coding schemes based on this structure are introduced. The first is a backward adaptation DPCM coder that needs very few side information bits. The other transmits codes for both model indices and quantized prediction residuals. Experimental results for different bit rates are presented