The goal of the research is to identify efficient (nonredundant) representations for correspondence problems in vision. This research investigates the efficacy of new multiscale representations obtained from both classical wavelet and best-adapted wave packet bases applied to three matching problems: stereo matching, texture discrimination and pattern recognition. Multiresolution and orientation selectivity, known biological mechanisms in primate vision, are ingrained in wavelet representations and inspire the techniques presented in this proposal. The multiresolution decomposition of wavelet transforms provides a natural hierarchy in which to embed an incremental stereo matching strategy. In place of traditional hierarchical (coarse-fine) matching, a novel incremental strategy for stereo matching strategy is proposed, where matching is accomplished by selecting bands of wavelet channels, constrained by orientation, scale and geometry (space). The approach to texture discrimination investigated the method of best-adapted wavelet packets for computing textural signatures. The scale, frequency and location (spacial) indices of the most energetic wave packet coefficients will be analyzed. Unique signatures shall be characterized by their links through 3-D scale space using the method of best-adapted wavelet packets. Finding practical solutions to such fundamental problems will have significant impact in the field of machine intelligence and foster the development of applications in the areas of navigation systems, robot vision, machine inspection, and image understanding.
