Georgia Institute of Technology College of Sciences

There are several definitions of the word shape; of these, the most important to this research is “the external form or appearance of someone or something as produced by its outline.” Shape Analysis in this context focuses specifically on the mathematical study of explicit, parameterized curves in 2D obtained from the boundaries of targets of interest in Synthetic Aperture Sonar (SAS) imagery. We represent these curves with a special “square-root velocity function,” whereby the space of all such functions is a nonlinear Riemannian manifold under the standard L^2 metric. With this curve representation, we form the mathematical space called “shape space” where a shape is considered to be the orbit of an equivalence class under the group actions of scaling, translation, rotation, and re-parameterization. It is in this quotient space that we can quantify the distance between two shapes, cluster similar shapes into classes, and form means and covariances of shape classes for statistical inferences. In this particular research application, I use this shape analysis framework to form probability density functions on sonar target shape classes for use as a shape prior energy term in a Bayesian Active Contour model for boundary extraction in SAS images. Boundary detection algorithms generally perform poorly on sonar imagery due to their typically low signal to noise ratio, high speckle noise, and muddled or occluded target edges; thus, it is necessary that we use prior shape information in the evolution of an active contour to achieve convergence to a meaningful target boundary.