A generalization of Linear Active Discriminant Functions named as Nonlinear Active Discriminant Functions (nonlinear ADF) to deal with nonlinear deformations of handwritten character is proposed. In Nonlinear ADF

Kernel PCA is applied to capture and represent the nonlinear deformations. Input space is mapped to feature space through nonlinear mapping. Then an optimal active prototype model is produced in principal subspace of the feature space and the distance between it and the projection of character feature vector in the principal subspace is defined as Nonlinear ADF. In addition

the Nonlinear ADF is further optimized using Minimum Classification Error criterion. Experimental results demonstrated that Nonlinear ADF has achieved a higher recognition rate than that of Linear ADF.