The goal of this research is to develop novel techniques for the analysis of filtered discrete self-similar signals and investigate their application in speech modeling and recognition, image texture segmentation and object recognition based on contour information. Specifically, filtered discrete self-similar signals are used as tools for modeling both the long term and short term characteristics of signals of interest and novel multiscale algorithms for estimating their parameters are constructed. A multiscale solution to the problem of finding an iterated function system whose attractor best approximates a given M-D curve in the Hausdorff measure is also being developed. Particular attention will be given to parallel implementation of the algorithms. Filtered discrete self-similar signals are also studied as models for speech signals with particular emphasis on the naturalness of the produced speech. Furthermore, their parameters and the parameters of iterated function systems approximations to speech demisyllables are tested for use as feature vectors in speech recognition systems. M-D experiments are also being performed to investigate the usefulness of M-D filtered discrete self-similar signals and iterated function systems in texture segmentation and object recognition tasks.
