Confidence analysis in character recognition

In order to estimate the accuracy of character recognition, the defination of classifier's confidence value and the concept of generalized confidence are given. Based on the confidence value the optimal rejection theorem is proven. The relationship between confidence value and recognition rate is disclosed. The formula for evaluating the generalized confidence values for several commonly used pattern classifiers is presented and an algorithm to convert generalized confidence to confidence value is introduced. Three possible applications of confidence analysis: the selection of rejection area, the estimation of recognition rate and the combination of multiple classifier are proposed. Practice in handwritten numeral recognition and off line handwritten Chinese character recognition strongly supports the ideas and the methods.