Joshua Ash (Advisor), Steve Gorman (Committee Member), Arnab Shaw (Committee Member)
Master of Science in Electrical Engineering (MSEE)
Point set classification methods are used to identify targets described by a spatial collection of points, each represented by a set of attributes. Relative to traditional classification methods based on fixed and ordered feature vectors, point set methods require additional robustness to obscured and missing features, thus necessitating a complex correspondence process between testing and training data. The correspondence problem is efficiently solved via spatial pyramid histograms and associated matching algorithms, however the storage requirements and classification complexity grow linearly with the number of training data points. In this thesis, we develop optimal methods of identifying salient point-features that are most discriminative in a given classification problem. We build upon a logistic regression framework and incorporate a sparsifying prior to both prune non-salient features and prevent overfitting. We present results on synthetic data and measured data from a fingerprint database where point-features are identified with minutia locations. We demonstrate that by identifying salient minutia, the training database may be reduced by 94\% without sacrificing fingerprint identification performance. additionally, we demonstrate that the regularization provided by saliency determination provides improved robustness over traditional pyramid histogram methods in the presence of point migration in noisy data.
Department or Program
Department of Electrical Engineering
Year Degree Awarded
Copyright 2017, all rights reserved. My ETD will be available under the "Fair Use" terms of copyright law.