This paper examines benefits of using concavity-based structural features in recognition?of handwritten digits. An overview of existing concavity features is presented?and a new method is introduced. These features are used as complementary features?to gradient and chaincode features, both among the best performing features?in handwritten digit recognition. Two support vector classifiers (SVCs) are chosen?for classification task as the top performers in previous works; SVC with radial basis?function (RBF) kernel and the SVC with polynomial kernel. For reference, we also?used the k-nearest neighbor (k-NN) classifier. Results are obtained on MNIST, USPS?and DIGITS datasets. We also tested dataset independency of various feature vectors?by combining different datasets. The introduced feature extraction method gives the?best results in majority of tests.
展开▼
