Geometric decision trees for optical character recognition

Abstract

A fundamental problem in computer vision is identifying which of a given set of geometric models is present in an image. We consider an approach to model recognition based on computing efficient strategies (decision trees) for `probing' a scanned image of a typeset document, in order to perform fast and effective optical character recognition (OCR). We consider a `probe' to be a simply computed local operator that can be applied to discriminate between two sets of possible models. By carefully constructing effective probes, and assembling them into a geometric decision tree, we have devised, implemented, and compared a variety of methods to perform OCR. In this paper, we present algorithms for probing strategies and decision tree construction, and we report experimental results on the effectiveness of these algorithms in identifying English characters and numerals in scanned images of printed pages of text. These algorithms are implemented as part of a system used by a document processing company (Syngen Corp.).