Entropy based merging of context models for efficient arithmetic coding

The contextual coding of data requires in general a step which reduces the vast variety of possible contexts down to a feasible number. This paper presents a new method for non-uniform quantisation of contexts, which adaptively merges adjacent intervals as long as the increase of the contextual entropy is negligible. This method is incorporated in a framework for lossless image compression. In combination with an automatic determination of model sizes for histogram-tail truncation, the proposed approach leads to a significant gain in compression performance for a wide range of different natural images.