Nonlinear spectral techniques for signal/image recognition, classification and estimation
Abstract
Nonlinear spectral techniques for signal/image recognition, classification and estimation are presented. The two techniques are nonlinear matched filters (NMF's) and neural networks with spectral probabilistic feature extraction. The NMF applies a nonlinearity to the spectrum of the input as well as the spectrum of the reference signal. The nonlinearity can be adapted to achieve varying degrees of discrimination when cross-correlating two vectors. These systems have very high power of discrimination and lack of false correlation signals and artifacts than previously known phase-only, binary phase-only and other partial information filters for correlations. Because of the nonlinearities, the analysis of nonlinear matched filters requires new approaches different from techniques valid in linear systems. With care, these networks are also suitable for optical implementations. The applications of the NMF to machine vision, image recognition, pattern recognition and phase estimation are discussed. Also the relation between Hamming distance, maximum likelihood and the NMF with the hardlimiter nonlinearity is described. A neural network architecture with multichannel probabilistic spectral feature extraction (PSFNN) and its application to the classification of remote-sensing data are described. The PSFNN consists of preprocessing and postprocessing. The postprocessing can be done by any neural network. The preprocessing involves estimation of probabilistic spectral features, special ways to generate spectral information, choice of best features and conversion of inputs considered the probability density function and the class separability in the spectral domain. The preprocessing is performed in order to obtain a best classification. In this thesis, the preprocessing steps is described in detail. The PSFNN is applied to spectral feature vectors. Studies of the PSFNN in comparison to two other neural network architectures in classification of remote-sensing data show that the PSFNN holds promise for high classification accuracy.
Degree
Ph.D.
Advisors
Ersoy, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.