Abstract

In pattern recognition, the reliability and the recognition accuracy of a classification system are of same importance, because even a small percentage of errors could cause a huge loss in real-life handwritten numeral recognition systems, like cheque-reading at financial institutions.
Aiming at improving the reliability of recognition systems, this thesis presents two novel learning-based rejection criteria for single classifiers including SVM-based measurement (SVMM) and Area Under the Curve measurement (AUCM).
Voting based combination methods of multiple classifier system (MCS) are also proposed for rejecting poor handwritten digits. Different rejection criteria (FRM, FTRM and SVMM) are individually combined with MCSs as weight parameters in voting. This method is then evaluated on three renowned databases including MNIST, CENPARMI and USPS. Experimental results indicate that these combinations improve the rejection performances consistently. To further improve the performance of the MCS based rejection method, specialist information has been integrated into the combination process by introducing a new confidence weight parameter. The best result on MNIST is obtained by the simpler one of the two proposed methods of deriving this parameter, which reaches 100% reliability with a rejection rate of only 4.09%, the best value in this field.