A New Divide and Conquer Based Classification for OCR
Hamïd Parvïn, Hosein Alizadeh, Behrouz Minaei Bidgoli · InTech eBooks · 2010
We proposed a new method for improving the performance of multiclass recognition system. This method uses population based approaches, in special GA, to obtain best way to conquer classification the classes from each other. Also the method is based minimizing error on evaluation set without learning the evaluation. This is because of necessitated speed in evaluating the chromosome. Applying the proposed approach leads to more accurate logical results than the simple classification, on handwritten digits dataset. We used this method on two classifier ensemble systems. Also, we used one kind of base classifier per experiment. It yields to a better result in both cases. We divided the classes as balanced as possible, in each level. As future work, it can be worked on as unbalanced division. However, hereby, our optimization has been done globally; we can further, focus on the level based optimization as well on future.