Handwritten digit recognition based on prototypes created by Euclidean distance

Handwritten digits are recognized using prototypes created by a training algorithm based on the Euclidean distance. The subsequent classification of a handwritten digit is based on criteria considering the Euclidean distance to the prototypes. A training set of 2361 patterns is used to create the prototypes and a separate set of 1320 patterns is used to test the proposed method. The system performance is compared to two other known classification algorithms: a MLP (multilayer perceptron network), and SOM (self-organizing map) plus LVQ1 (a linear vector quantization algorithm). The proposed method reached a recognition rate of 93.5% when using the nearest-prototype criterion, and raised to 94.8% when using a nearest-prototype-voting criterion. It compared favorably with the MLP (91.8%) and SOM+LVQ1 (91.5%).