A dictionary learning and KPCA-based feature extraction method for off-line handwritten Tibetan character recognition

Kernel principal component analysis (KPCA) is a powerful feature extraction technique. For character recognition, however, the computation cost of KPCA is too high because of much larger sample size of each class. A novel two-stage feature extraction method DL-KPCA that based on dictionary learning and KPCA is proposed for character recognition. In the first stage, with the dictionary learning method K-SVD, a representative sample subset is constructed from the original sample set of each class at first. Then, to the test sample, find its K nearest neighbors from the union of all the constructed sample subsets and consider the classes of their nearest neighbors as the candidate classes. In the second stage, the test sample and the constructed sample subsets of its candidate classes are transformed to the feature space with KPCA, and the test sample is finally classified with K-NN in the feature space. Experimental results on THCDB, a recently developed Tibetan handwritten character sample database, and the reshuffled USPS digit database show that, to character recognition problems, it is feasible to extract the features with the proposed DL-KPCA.