An iterative algorithm for context selection in adaptive entropy coders

Context-based adaptive entropy coding is an essential feature of modern image compression algorithms; however, the design of these coders is non-trivial due to the balance that must be struck between the benefits associated with using a large number of conditioning classes, or contexts, and the penalties resulting from data dilution. The problem is especially severe when coding small sub-images where the amount of data available is small. In this paper, we propose an iterative algorithm that begins with a large number of conditioning classes and then uses a clustering procedure to reduce this number to a desired value. This method is in contrast to the more usual approach of defining contexts in an ad-hoc manner. Experiments are conducted on synthetic data sources having varying amounts of memory, as well as on the sub-images resulting from a wavelet decomposition of an image. The results show that our approach to context selection is effective and that the algorithm automatically learns the structure of the data. This technique could be applied to improve the performance of both image and video coders.