Tuesday, December 3, 1996
4:30-5:30 pm
1200 EECS
Abstract -
This is the first in a series of talks that I plan on giving on robust target
detection and recognition for statistical imaging applications. In this
talk, we specialize to binary images arising in document processing,
character recognition, and speckle images. An asymptotic spatial point
process representation for the binary image is proposed. Under this
representation multiplicative "bit flip" noise in the bitmap domain is
rendered additive in the point process intensity domain. In particular, we
show that a spatially inhomogeneous noise translates into the statistical
model as an epsilon-contaminated mixture of spatially varying object and
noise intensities. Using this representation we develop a pattern recognition
algorithm which is capable of identifying very weak signatures in noise
contaminated images. This algorithm is based on constructing a higher order
moment matrix formed from the bitmaped image, performing noise subspace
processing, and finally deploying optimal detection. We illustrate the
methodology in the context of word spotting and character recognition for
optical document processing.
For your information the following are titles and abstracts for two other talks that I plan to give after the holiday break.
This is the second of a series of talks on robust target detection and recognition statistical imaging applications. As in the previous talk (Part I), we specialize to binary images. This talk presents methods of detection based on pruning a spanning tree which passes through the points of a binary image. This non-parametric method is based on constructing Renyi-entropy estimates derived from the span lengths of a sequence of D-minimal spanning trees. We will illustrate these approaches for problems including word spotting, optical character recognition, and target detection in speckle images.
This is the last of series of talks on robust target detection and recognition in statistical imaging applications. In this talk we review methodologies for robust detection under nuisance parameters (local UMP, min-max, MP invariant), apply these methods to target detection in images with unknown clutter, show that these methods are equivalent to well known CFAR methods for the case of homogeneous clutter, and derive novel CFAR tests for structured inhomogeneous clutter backgrounds. The novel tests are shown to outperform other CFAR methods over important ranges of SNR. We conclude with examples in IR/SAR imaging when the target signature is known up to a scale factor and when the target signature lies in a set of known target signatures.