ADAPTIVE CONFIDENCE TRANSFORM BASED CLASSIFIER COMBINATION FOR CHINESE CHARACTER-RECOGNITION

Classifier combination is an effective way to improve recognition perf ormance. However, in Chinese character recognition the extremely large number of categories results in several difficulties for the combinat ion. In order to overcome these difficulties a novel combination metho d is presented in this paper. It consists of three main components: ad aptive confidence transform (ACT), consensus theoretic combination and reliability based speedup scheme. ACT, which can estimate a posterior i probabilities from raw measurement values, is the focus of this pape r. Experimental results show a significant reduction of error rates in both printed (PCCR) and handwritten Chinese character recognition (HC CR). (C) 1998 Elsevier Science B.V. All rights reserved.