Harmonic Analysis and Signal Processing Seminar
Unsupervised Regression for Image Denoising
Martin Raphan
CIMS and Laboratory for Computational Vision, Center for Neural Science, NYU
Tuesday, October 30, 2007, 12:30pm, WWH 1302
Abstract
There
are two standard frameworks for describing optimal least squares
estimation of a random quantity from corrupted measurements. The first
technique, Bayesian Least Squares (BLS) estimation, uses explicit
models of both the corruption process and the prior distribution of the
quantity to be estimated in order to formulate an optimal estimator via
Bayes' rule. The second technique, Least Squares regression, uses
supervised training on a data set which has clean samples paired with
corrupted versions of those samples, to choose an optimal estimator
from some family. In many applications, however, one has available
neither a model of the prior distribution, nor uncorrupted measurements
of the variable being estimated. We will describe a framework for
expressing the BLS estimator (regression function) entirely in terms of
a model of the corruption process and the density of the corrupted
measurements. We show a practical implementation of this nonparametric
estimator for additive white gaussian noise (AWGN), and demonstrate the
use of this procedure for denoising photographic images, showing that
it compares favorably with previously published methods which use
explicit prior models. We also describe a dual, prior-free formulation
of the Mean Square Error (MSE) which generalizes Stein's Unbiased Risk
estimator (SURE), and show how this may be used for unsupervised
regression. We then demonstrate the use of this dual formulation in
image denoising. In particular, we use the dual formulation to prove
the empirically observed fact that, despite their suboptimality,
marginal image denoisers chosen to minimize MSE within the subbands of
a redundant multi-scale decomposition will always perform better than
on the orthonormal versions of those bases. We also develop an
extension of SURE that allows minimization of the image-domain MSE for
estimators that operate on subbands of a redundant decomposition, and
show that this gives improvement over methods which optimize MSE within
subbands.