Geometric Methods for Optical Character Recognition a Dissertation Presented by Abstract of the Dissertation Geometric Methods for Optical Character Recognition

of the Dissertation Geometric Methods for Optical Character Recognition by George N. Sazaklis Doctor of Philosophy in Computer Science State University of New York at Stony Brook Advisor: Joseph S. B. Mitchell 1997 Abstract Optical Character Recognition (OCR) is an important problem having both theoretical and practical interest. In this dissertation, we present solutions to three problems within the area of OCR. A di culty encountered by many OCR systems is confusions between similar shapes, when exible matching is employed as a primary recognition mechanism. Our solution, constrained matching as a second stage classi cation technique, can discriminate between similar shapes, using shape geometric attributes; thus the system is enabled to reach a nal decision on the character identity. Another important problem in OCR is the fast and reliable xed-font recognition. We present a hierarchical classi cation technique that utilizes the concept of geometric probe trees from [4]. At each node of the probe tree, a geometric probe collects information from the shape at hand, and makes a partial decision about its identity, eliminating certain candidates from further consideration. The probe tree can be constructed o -line in a preprocessing step and can provide us with high speed recognition for a xed font. ii We also present an extension of geometric probes, the pre x probing technique that solves the important practical problem of touching characters for a xed font. Pre x probing is based on a forest of probe trees that succeeds in identi ing each member from a sequence of touching characters independently of its neighbors. This avoids the excessive demands of a straightforward solution. To substantiate our methods, we have developed several tools, and we have done experiments with scanned documents, so we also present our experimental results. iii