A new convolutional map feature extraction for character recognition

This paper presents a novel feature extraction method for off-line recognition of segmented (isolated) handwritten characters. The feature extraction method is discussed in terms of invariance properties, speed and discriminative power with respect to the classification accuracy by benchmarking on a large and realistic handwritten character database. The NIST SD3 is used for training and the SD7 set comprising of 55000 digits is used for testing. The testing set is large so as to present a realistic view of the performance. Neural network classifier is used for aggregating the feature for classification decision making. Accuracy above 97% was achieved with no rejection. With a rejection of just 5% reliability performance can be boosted to 99%