Position independent neuro pattern matching and its application to handwritten numerical character recognition

A novel one-dimensional pattern-matching neural network which matches an input to multiple candidates of the stored templates in parallel is proposed. It can find the best matching template, whose features are arranged in the same order as those of the input, regardless of positional differences between corresponding features. The proposed network is applied to handwritten numerical character recognition. Six kinds of features are extracted form the input pattern: closure, upward, downward, left and right orientations of line ends, and junction. From the two-dimensional feature distribution pattern, horizontal and vertical projection profiles are made. They are matched in the horizontal and vertical matching networks separately. The results of matching in the two networks are combined and categorized by a recognition network. 600 handwritten characters in the ETL-1 database were used as a learning set, and another 600 characters were used as an open set. The recognition rate for the learning set was 98.2% and that for the open set was 88.2%