Multiple Filter Mask Learning of Feature Extraction in Character Recognition

An incremental learning for the feature extraction of OCR was applied to the learning of multiple filter mask in the feature extraction. The multiple filter mask means plural masks which are used for one local region of an input pattern. The output of the filter is provided to a recognition unit as a feature vector. In the recognition unit, distance is calculated by a reference vector and the feature vector. Both the reference vectors and the feature extraction filter masks are gradually changed by recurrence equations. The purpose of this learning was similar character discrimination. The effectiveness of the learned masks was checked as an accuracy improvement. The learned masks were visually displayed and they were checked by human observation in the experiment. It was found that the masks had good performance in recognition accuracy, and that the masks were grown at the important area of a pattern for discrimination through the learning.