A sparse concept coded spatio-spectral feature representation for handwritten character recognition

Abstract

In this paper, a new sparse concept coding based image representation (SCCST) is proposed which efficiently extracts low dimensional features from the spatio-spectral decomposition of handwritten characters. The multiresolution decomposition is obtained by adopting an octave sampling based non-redundant S-transform. The introduction of sparsity not only reduces feature dimension significantly, but also captures the intrinsic geometric structure of the image space. The robustness of our proposed feature representation is validated by experimenting with the classifiers such as k-NN and SVM. A new public database for handwritten Odia characters is created and reported in order to address the issue of unavailability of open access character database. A five-fold cross validation strategy is adopted on MNIST database, existing handwritten Odia numeral databases and the newly created character database to measure the recognition performance. The proposed approach is found to achieve the best accuracy of 99.28% and 96.35% for unconstrained Odia handwritten numerals and characters respectively. � 2016 IEEE.