Performance comparison of ensemble models for recognition of offline handwritten Odia numerals

In this paper an ensemble model is proposed for the recognition of Odia handwritten character. The ensemble model is constructed from four base classifiers: Support Vector Machine (SVM), Artificial Neural Network (ANN), C5.0 Decision Tree and Discriminant Analysis (DA). Gradient and curvature based features are extracted from the numerals and a combination of gradient and curvature based features is taken for simulation. The features are reduced by using Principal Component Analysis (PCA). First the numerals are classified by the individual classifiers. Then the ensemble models are formed with all possible combinations. The outputs of the classifiers are combined by using confidential weighted scheme to obtain ensemble models. The classification results obtained from individual classifier and all ensemble models are compared. It is observed that the ensemble of SVM+DA and SVM+C5.0+DA outperformed well each with 97.5% of accuracy on test dataset as compared for other eight ensemble models.