A New AdaBoost Algorithm for Large Scale Classification And Its Application to Chinese Handwritten Character Recognition

The present multiclass boosting algorithms are hard to deal with Chinese handwritten character recognition for the large amount of classes. Most of them are based on schemes of converting multiclass classification to multiple binary classifications and have high training complexity. The proposed multiclass boosting algorithm adopts the descriptive model based multiclass classifiers (Modified Quadratic Discriminant Function, MQDF) as the element classifiers, which perform multiclass classifications directly. The proposed boosting algorithm does not need to convert multiclass classifications to multiple binary classifications, and has lower training complexity than most of present multiclass boosting algorithms. So it is more suitable for dealing with large scale classification problems. The algorithm updates samples' weights according to the generalized confidence which is simple and effective. Further, in order to reduce the recognition complexity, the pruning method was performed to pick out only one best element classifier from all boosted classifiers to do the classification. Applying the proposed algorithm to Chinese handwritten character recognition on the different datasets, the recognition rate is significantly improved; meanwhile the recognition complexity is the same as the traditional MQDF classifier.