Feature selection in the recognition of handwritten Chinese characters

Abstract

A method is proposed here to extract appropriate features for the recognition of handwritten Chinese characters. The features represent the lengths, positions and directions of the character strokes. In addition, two approaches (Karhunen-Loeve and stroke density analyses) have been used to analyze the information content of Chinese characters, from which a cost-effective, non-uniform, and two-dimensional sampling scheme for feature extraction has been derived. The resulting scheme extracts more samples from regions of higher information content. A recognition system that combines all the proposed approaches was built, and experiments were performed on the 500 most frequently used Chinese character classes, with 20,000 handwritten samples. Results indicated that the proposed methods are useful. © 1997 Elsevier Science Ltd. All rights reserved.