Shape normalized subspace analysis for underwater mine detection

We present a method for classification and localization of sea mines in 2D side-scan sonar imagery. The approach parallels closely Turk and Pentland's method of eigenfaces and treats mine detection as a two dimensional object recognition and localization problem acknowledging the fact that mine patches are endowed with a measure of regularity in geometry and appearance. We, therefore, set out with the hypothesis that mine patches can be described fairly accurately using low complexity parametric models. We outline a method for constructing a low dimensional non-linear shape plus appearance decomposition model from 2D sonar imaging data. We then characterize the parameter space and formulate a distance metric in this space. The distance metric is used as a mine similarity measure, and thus forms the basis of our mine detector. We show results demonstrating the discriminative power of the mine detector on a benchmarking data created from the mine dataset used in.