Robust, Ambiguity Faithful (RAF) classification

Paul Gader, University of Florida
Robust, Ambiguity Faithful (RAF) classification
Monday, January 26,  2:00– 3:00 PM
3001 Plant and Environmental Sciences

Researchers have demonstrated a variety of classifiers that are able to achieve excellent performance on many different standardized data sets. These classifiers are usually evaluated using the general methodology: Given a set of feature vector samples, X = {x1, …, xN}, that all have known class labels from m classes C1, …, Cm, do the following one or more times: (1) Estimate parameters of a classifier, fi, using a subset of X and (2) Evaluate the classification accuracy of fi using a different, disjoint subset of X by evaluating how often a data sample is assigned to the correct class. However, this evaluation strategy and mode of development ignore very important issues when building classifiers for fielded systems. One issue is that in fielded systems, classifiers are generally part of a larger system and binary decisions are often not required. The “Principle of Least Commitment” espoused by the computer vision pioneer David Marr in 1982 applies. Therefore, classifiers should produce more information than a class label. This information can be represented using probability or possibility distributions. Classifiers should be able to estimate possibility that an input pattern is a sample of one of the classes of interest or is an outlier. This capability is referred to here as robustness. The other issue is that some patterns are truly ambiguous; no distinction can be made between them based on the feature vectors. It is possible that contextual or multi-sensor cues can be used to resolve the ambiguities. Issues involved in designing RAF classifiers are discussed. Examples are given on real-world problems, including handwritten word recognition, landmine detection, and remote sensing.

Paul Gader has been devising pattern recognition algorithms since 1984. He received a Ph.D. in Math in 1986 for parallelizing image processing algorithms. Since then, he has focused on applying theory to real problems. He became a leading figure in the application areas of handwriting recognition and landmine detection and is becoming one in hyperspectral image analysis. He led the development of handwritten character and word recognizers that performed in the top 5 and top 1 in a NIST competition. In 1998, he and H. Frigui devised a real-time Ground Penetrating Radar landmine detection algorithm that was a top performer in blind field testing. He was Technical Director of an Army Demining MURI for 2 years. He and D. Ho developed algorithms for hand-held mine detection system currently in use by the U. S. Army. He participated led many landmine/IED detection projects using data from Acoustic/Seismic, EMI, FLIR, SAR, and LWIR and VISNIR/SWIR sensors. He led teams that studied and implemented Hidden Markov Model and Possibilistic detectors in real-time on a Husky Mine Detection System (HMDS). HMDS was fielded in Afghanistan. The HMDS with his team’s algorithms, is featured in National Geographic Television program: “Bomb Hunters: Afghanistan”.
He has been researching hyperspectral algorithms since 2002. He was general chair of the IEEE Workshop on Hyperspectral Image and Signal Processing in June 2013. Dr. Gader has published 90 journal and over 300 total papers, (was) an Associate Editor of IEEE Geoscience and Remote Sensing Letters (before becoming Chair), led an ad-hoc committee on Standardized Algorithms, Data, and Evaluation (SADE), is a U. of Florida Research Foundation Professor, Chair of the CISE department, and an IEEE Fellow.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s