Generalised approach to the recognition of structurally similar handwritten characters using multiple expert classifiers

It is observed that a particular classifier using a particular set of features will generally exhibit a greater probability of confusion among certain character classes than among others. In general these confusion classes are a substantial source of error in the overall performance of the classifier. A productive way to deal with the problem is to separate these characters and reprocess them further in an independent secondary stage in the framework of a multiple expert configuration. The philosophy is to use multiple classifiers to re-evaluate these relatively difficult characters by treating them as special and specific problem cases. In extending special treatment to these characters, advantage can be taken of distinctive structural features to design tailor-made algorithms suited to a particular problem. Since such classifiers are required to deal only with a limited number of classes, very versatile classifiers can be implemented. The main difficulty of this philosophy is to devise a way to group characters together to make sure that these specialised classifiers receive a stream of input characters which indeed belong to the particular group of characters associated with that particular classifier. The authors present a general philosophy for multi-expert classification and deal with the specific problem of formation of distinctive character streams with a high degree of confidence. It then elaborates on other techniques and variations that can be adopted to make this type of multiple expert configuration more effective