Harmonic Analysis and Signal Processing Seminar
A Tensor Framework for
Image Analysis (Vision) and Synthesis (Graphics)
M. Alex O. Vasilescu, Media Research Lab, CIMS
Wednesday, December 8, 2004, 2-3:00pm, WWH 1314
Abstract
Natural images are the consequence of multiple factors related
to scene structure, illumination, and imaging. I will present a multilinear
model that computes (nonlinear) manifold representations of image ensembles
in which the multiple constituent factors (or modes) are disentangled, analyzed
explicitly and their second order statistics encoded. This nonlinear model
is computed via a tensor decomposition, known as the N-mode SVD. I
will present TensorFaces, our multilinear representation for automated face
recognition, and TensorTextures, a novel, multilinear approach to image-based
rendering which is an important problem in computer graphics. Next,
I will present Multilinear Independent Components Analysis (MICA) algorithm
which encodes the higher order statistics of different factors associated
with image formation. MICA is a generalization of the conventional (linear)
independent components analysis (ICA) algorithm.