Underwater Target Classification Using A Pose-Invariant Matched Manifold Classifier
Pooria Pakrooh, Colorado State University
Louis L. Scharf, Colorado State University
Mahmood R. Azimi-Sadjadi, Colorado State University

One of the challenges in automatic detection and classification of underwater targets in sonar imagery is variation of the target returns and features with respect to target aspect. This paper adopts a framework for target classification that offers local invariance properties with respect to target aspect. Sonar image snippets of a target type at nearby aspects are related to each other via geometric deformations approximated by locally affine transformations. A transform is then used to map each such image into a low dimensional linear subspace locally invariant to affine geometric transformations of the image. These linear subspaces are subsequently averaged using an extrinsic subspace averaging to yield a subspace corresponding to this set of images. Class label of an unknown image is then decided using a geometrically invariant Matched Manifold Classifier that uses principal angles to measure distance between this average subspace and the subspace associated with the observed image. Results on synthesized sonar images of various underwater targets inserted in real backgrounds show the promise of this new method.