Recognition of Fragmented Characters Using Multiple Feature-Subset Classifiers

This paper addresses the problem of recognizing fragmented characters in printed documents of poor printing quality, which often causes characters to break up. To enhance the recognition accuracy of such characters, most existing approaches attempt to improve the quality of character images by means of some mending techniques. We propose an alternative approach that adopts a bagging-predictor method to build classifiers, using only intact characters as training samples. The resultant classifiers can classify both intact and fragmented characters with a high degree of accuracy. Applying this approach to characters in archived Chinese newspapers, we extract two types of features from character images and form bagging predictors, each of which takes a subset of features as input. As a result, we are able to achieve drastic improvements in the recognition of fragmented characters.