Machine learning approach towards automatic target recognition.

摘要

The overall goal of this research project is to develop online algorithms for detecting military targets in radar return images. In this dissertation, Automatic Target Recognition is taken as a multi-stage generalized pattern recognition problem. Three problems are addressed: feature extraction, feature selection and inductive learning.; Four mathematical transforms are utilized to extract features from raw image chips. They are Fourier Transform, Principle Component Analysis, Singular Value Decomposition and Radon Transform. Feature selection is carried out to identify important features and to reduce the dimensionality of feature space. In this work, feature selection is considered as a stochastic combinatorial optimization problem. The estimated error rate of a naive Bayesian classifier is used as performance metric to evaluate the goodness of a given subset while forward multi-selection algorithm is proposed to improve the efficiency of searching in large search space. For inductive learning, a decision fusion scheme is proposed to improve the predictive accuracy of weak learners. Different subsets of features are utilized to train weak learners and several combining algorithms are compared here. Finally, all the techniques developed are put together to solve the Automatic Target Recognition problem.