A local linear discriminant analysis method for handwritten Chinese character recognition

LDA is one of popular dimension reduction techniques in existing handwritten Chinese characters (HCC) recognition systems. To deal with the class separation problem and the multimodal samples in tradition LDA method, we proposed a new local linear discriminant analysis (LLDA) method for handwritten Chinese character recognition in this paper. The algorithm operates as follows: (1) Using the clustering algorithm to find clusters for each class. (2) Finding the nearest neighbor clusters for each cluster and using cluster means in the computation of the between-class scatter in LDA while keeping the within-class scatter unchanged. (3) Finally vector normalization is applied to further improve the class separation problem. A series of experiments on HCL2000 have indicated that our method can effectively improve the recognition, the error rate reduction reaches 14.8% comparing to the traditional LDA method, showing effectiveness of the proposed approach.