Time: Tuesday, November 10, 2015

Speaker: Professor Nathan Cahill (Rochester Institute of Technology)

Title: Semi-Supervised Graph-Based Methods for Image Segmentation and Classification

Abstract: Laplacian Eigenmaps (LE) is a powerful graph-based technique for manifold learning and recovery. It solves the same generalized eigenvector problem that arises in the continuous relaxation of Normalized Cut (NCut)-based graph partitioning. Schroedinger Eigenmaps (SE) has emerged as a powerful semi-supervised extension of LE that uses barrier and/or cluster potentials to encode expert/labeled information provided at a subset of graph vertices. In this talk, we first explore how SE can be used for hyperspectral image analysis, including in the fusion of spatial/spectral data for classification, the exploitation of user-provided "paintbrush" strokes for segmentation, and the inclusion of target spectra for target detection. Then, we show how the idea of cluster potentials in SE can be used to construct an analogous Semi-Supervised version of the Normalized Cuts algorithm, which we call SSNCuts, and we show how SSNCuts improves RGB image segmentation of complicated scenes.

Back to seminar homepage