Characterizing prediction error distributions for lossless image compression

Abstract

In predictive coding for lossless image compression, full knowledge of the prediction error distribution and efficient coding with an arithmetic coding method is the best one can do with the 0-order model assumption. The zero-order error distributions typically are Laplacian with zero mean. Higher-order error distributions are often skewed with a mean that is often positive or negative. Additional compression is achieved by an accurate characterization of context-dependent error distributions. This paper presents the results of a study the different characteristics of the error distributions found in higher-order conditioning contexts of the LOCO and CALIC algorithms. The study includes nonstationary behavior.