Skip to Main Content
Mobile Menu

Character Recognition Using Fourier Descriptors

Masters Thesis, Boise State University, August 2006

By Jared A. Hopkins


A set of features for performing optical character recognition of bi-tonal images is implemented within the Gamera framework. The features are based on the Fourier Descriptor but include a method which allows classification of images which contain multiple boundaries. This is accomplished by assigning to each character image a signature which encodes the boundary types that are present in the image as well as the positional relationships that exist between them. This allows a boundary-wise comparison of different images to be accomplished using their Fourier Descriptors. Under this approach, only images having the same signature are comparable. This fact in turn affects the architecture of exemplar-based classifiers which use these features. Effectively, a meta-classifier is used which first computes the signature of an input image and dispatches the image to a classifier which is trained to recognize images having that signature. The implementation is carried out by extending the functionality of Gamera, an existing open-source framework for building document analysis applications. The features are then tested by implementing a fuzzy-knn and a neural-network classifier based on them. The fuzzy-knn classifier achieves an estimated generalization accuracy of 94%, which is the best rate yet achieved on the particular data set used. The neural-network classifier achieves an estimated generalization accuracy of 91%.