Probabilistic feature extraction for object recognition
Abstract
In most object recognition systems, features are extracted from an image, grouped into hypotheses, and matched against a model database. The constructed hypotheses are usually formed in a multi-level hierarchy of features from the lower level pixels into higher level geometric features. Many existing systems include a statistical or heuristic measure of the hypothesized features formed at a given level but do not consider the measure of the underlying features of lower levels in the hierarchy. By incorporating a statistical likelihood measure at each level of the hierarchy from the lower level to higher level features, the overall accuracy and speed of the system can be improved. The system can also provide a final measure representing the strength of a match between the image and a given model object that cannot be calculated accurately in other systems. The feature extraction task has been divided into several steps. The first step extracts a set of edge pixel hypotheses from an image using a statistical model of edges. The edge pixel hypotheses are then grouped into straight line edge hypotheses, or curves represented by splines using a unified statistical model. Using the edge model, the features can not only be evaluated but can also be improved to increase the accuracy of the system. The benefits of statistically measuring the features at each step is demonstrated with a OCR program for Korean Hangul.
Degree
Ph.D.
Advisors
Chelberg, Purdue University.
Subject Area
Electrical engineering|Computer science|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.