Video compression with weighted finite automata

Abstract

Summary form only given. Weighted finite automata (WFA) exploit self-similarities within single images and also video streams to remove spatial and temporal redundancies. The WFA image codec combines techniques from fractal image compression and vector-quantization to achieve performance results for low bit-rates which can be put on a par with state-of-the-art codecs like embedded zerotree wavelet coding. Moreover, frame regeneration of WFA encoded video streams is faster than that of wavelet coded video streams due to the simple mathematical structure of WFA. Therefore, WFA were chosen as a starting point for a fractal-like video compression with hierarchical motion-compensation. Video streams are structured as proposed by the MPEG standards: the entire video is subdivided into several groups of pictures which are made up of one I-frame and a given number of predicted B- or P-frames. The macro block concept of the MPEG standard is replaced by a hierarchical and adaptive image partitioning. We integrated motion compensation with variable block sizes into the WFA coder to exploit the inter-frame redundancy. The general concept of the WFA compression was retained since it already provides a hierarchical subdivision of the image. The video stream is encoded frame by frame with an improved version of the WFA inference algorithm.