Classification with spatial-temporal context and design of partially supervised classifiers

Byeungwoo Jeon, Purdue University

Abstract

Pattern recognition technology has had a very important role in many fields of application including image processing, computer vision, remote sensing, etc. The advent of more powerful sensor systems should enable one to extract far more detailed information than ever before from observed data, but to realize this goal requires the development of concomitant data analysis techniques which can utilize the full potential of the observed data. In this dissertation, two different issues are investigated. One involves classification using spatial and/or temporal contextual information. Although contextual information has been an important and powerful data analysis clue for the human-analyst, the lack of a good contextual classification scheme especially which can both use spatial and temporal context has not allowed its usefulness to be put to full use. Two different approaches to spatio-temporal contextual classification are investigated. One is based on statistical spatio-temporal contextual classification, and the other is based on decision fusion of temporal data sets which are classified individually with spatial contexts. The second part of this dissertation addresses a partially supervised classification problem, especially when the class definition and corresponding training samples are provided a priori only for just one particular class. In practical applications of pattern classification techniques, a frequently observed characteristic is the heavy, often nearly impossible requirements on representative prior statistical class characteristics of all classes in a given data set. Considering the effort in both time and man-power required to have a well-defined, exhaustive list of classes with a corresponding representative set of training samples, this "partially" supervised capability would be very desirable, assuming adequate classifier performance can be obtained. Two different classification algorithms are developed to achieve simplicity in classifier design by reducing the requirement of prior statistical information without sacrificing significant classifying capability. The first one is based on optimal significance testing, where the optimal acceptance probability is estimated directly from the data set. In the second approach, this partially supervised classification is considered as a problem of unsupervised clustering with initially one known cluster or class. A weighted unsupervised clustering procedure is developed to automatically define other classes and estimate their class statistics. The operational simplicity thus realized should makes these partially supervised classification schemes very viable tools in pattern classification.

Degree

Ph.D.

Advisors

Landgrebe, Purdue University.

Subject Area

Electrical engineering

COinS