Stabilization of probability estimates by conditioning on prior

Abstract

A system and method involving a statistical conditioning technique that improves the coding efficiency in compression systems which have unstable statistical properties by conditioning the probability estimate for a given model context on prior decisions for that context thus enlarging the conditioning decision set. Instead of extending the model context by increasing the range of prior decisions in the pixel neighborhood, each model context is expanded into a set of two contexts which are the actual coding contexts. For a given probability estimation model context, the selection of a coding context index is done on the basis of the previous coding decision (1 or 0) for that model context. Thus, if a model context is assigned an index A, the coding context would be assigned an index 2*A +D(A)', where D(A)' is the previous (immediately preceding) binary decision for model context A; the decision of which coding context to use turns on whether the model context was used most recently for coding a 1 or a 0. More generally, each model context index A is expanded into a set of 2**N coding context indices, where N is the number of previous decisions for model context index A used in the conditioning of, for example, an arithmetic coding decision. The addressing of the context expansion can be done in any way which expands the total number of contexts by a factor of 2**N. The index for a case where the expansion is by one decision bit may be 2*A+D(A)', whereas in the case where the expansion is by two decision bits the index may be 4*A+2*D(A)'+D(A)'.