Adaptive confidence transform based classifier combination for Chinese character recognition

摘要

Classifier combination is an effective way to improve recognition performance. However, in Chinese character recognition the extremely large number of categories results in several difficulties for the combination. In order to overcome these difficulties a novel combination method is presented in this paper. It consists of three main components: adaptive confidence transform (ACT), consensus theoretic combination and reliability based speedup scheme. ACT, which can estimate a posteriori probabilities from raw measurement values, is the focus of this paper. Experimental results show a significant reduction of error rates in both printed (PCCR) and handwritten Chinese character recognition (HCCR).