SK-PCA: A Novel Feature Extraction Technique for Handwritten Tibetan Letter Recognition

Abstract

Kernel-PCA is a popular technique to extract nonlinear features. For Tibetan letter recognition, however, it is nearly impossible to extract features with Kernel-PCA. The main reason is that the computational cost of Kernel-PCA increases too fast with the increase of sample number. Therefore, based on sparse representation and Kernel-PCA, a novel feature extraction method is proposed for Tibetan letter recognition. There are two stages in the proposed method. In the first stage, a representative sample subset is created for each class by employing a sparse representation technique at first; then, from the union set of all representative subsets, find the K-nearest neighbors of a test sample and consider the classes of the K nearest neighbors as the candidate classes. In the second stage, transform both the test sample and the representative subsets of its candidate classes into feature space with Kernel-PCA, and the test sample is finally recognized with an modified quadratic discriminant function classifier. Experimental results show that it is preferable to extract features with the proposed SK-PCA for Tibetan letter recognition.