Recognition of Handwritten Chinese Characters by Combining Regularization, Fisher's Discriminant and Distorted Sample Generation

The problem of offline handwritten Chinese character recognition has been extensively studied by many researchers and very high recognition rates have been reported. In this paper, we propose to further boost the recognition rate by incorporating a distortion model that artificially generates a huge number of virtual training samples from existing ones. We achieve a record high recognition rate of 99.46% on the ETL-9B database. Traditionally, when the dimension of the feature vector is high and the number of training samples is not sufficient, the remedies are to (i) regularize the class covariance matrices in the discriminant functions, (ii) employ Fisher's dimension reduction technique to reduce the feature dimension, and (iii) generate a huge number of virtual training samples from existing ones. The second contribution of this paper is the investigation of the relative effectiveness of these three methods for boosting the recognition rate.