Nondeterminism and Motion Compensation for Weighted Finite Automata

Abstract

Nondeterministic finite automata with states and transitions labeled by real numbers as weights have turned out to be powerful tools for compression of still images and video sequences. These Weighted Finite Automata (WFA) as introduced and studied by Culik II, Kari, Karhumäki and others can exploit self-similarities within single pictures, between colour components and also sequences of pictures. Thus, as a coding-method WFA has much in common with fractal encoding algorithms like Barnsley’s Iterated Function Systems, but it can be viewed as well as a generalization of gain-shape vector quantization. WFA-coding is particularly effective for low bit-rates and our current implementation clearly outperforms the video standard H.263. This paper discusses the control of nondeterminism — i.e. the fan-out of states in the WFA labeled by the same input-symbol — and of motion compensation and shows how both features are handled as two facets of a coding decision taken during WFA-inference.