In this thesis, we present an off-line numeral recognition algorithm using an Error Back Propagation(EBP) Neural Network. For printed numeral recognition, we extract conventional feature vectors and recognize using simple EBP Neural Network. For unconstrained hand written numeral recognition, we propose a new feature vector and a new EBP Neural Network structure which is composed of multiple sub-networks. The proposed feature vectors give high recognition rate, because these feature vectors include enough information of hand written numeral. These vectors are gradient of longest line which crosses the hand written numeral pixels on the image plane. Since the proposed structure uses pre-learned weight of Sub-EBP Neural Network, it achieves much higher learning speed and recognition rate than conventional networks. If there are additional input patterns, conventional network-based methods need to repeat the entire learning process using the new input pattern as well as the previously trained pattern. However, the proposed method only learns with the new input pattern. The proposed structure can use pre-learned weights, instead of learning again previous input patterns.
In order to verify the performance of the network, we experiment with 7 fonts of Microsoft Word 5.0 numeral for printed numeral recognition, and with hand written numeral database of Concodia University of Canada for hand printed numeral. For printed numeral, correct recognition rate is 100% for different scale of printed numeral, and for hand written numeral, 93.9% correct recognition rate is achieved for Concodia Numeral Data base.