Wavelet Transform Methods for Acoustic and Image Pattern Recognition

ORINCON proposes to investigate real-time pattern recognition techniques that exploit the multiresolution properties of the discrete wavelet transform (DWT). We will process one-dimensional acoustic signals and two-dimensional visual and thermal images of vehicles provided by the Army to generate several different wavelet-based signal representations (including the DWT, multiscale-edge representations, and best-basis expansions). We will design classifiers to recognize patterns of interest in the data. Classification schemes to be investigated include hierarchical decision trees and multifeature-fusion classifiers. Basic architectures will be based on neural network paradigms; ORINCON has several fully implemented architectures available for use, including multilayer perceptrons, radial (or ellipsoidal) basis function networks, and fuzzy min-max networks. Classifiers will be trained and tested using a jackknife approach. As an optional task, we will undertake a preliminary design study for a real-time hardware system on which to implement the pattern recognition algorithms developed. All implementation and testing will be performed on Silicon Graphics workstations using ORINCON's RIPPEN graphical programming software package, which combines a graphical user interface and a software utility library to facilitate rapid prototyping of signal and image processing algorithms. All software developed for this contract will be included as deliverables.