Sparsity Based Framework for Discrimination between Confusion Pairs

Sparse Representation (SR) has successful applications in vision and classification problems. SR methods work by transforming the data into a low-rank latent space to find sparse representations for the data. Handwritten Character Recognition (HWCR) is a challenging classification task. Its difficulty arises from the high intraclass variances of the writing. HWCR for Latin scripts is better studied, but a few challenges remain. For Indic scripts, HWCR is not very mature and much research is needed. Indic scripts reflect the shallow orthography and are quite syllabic. Telugu is one such Indic script with more than 80 million native speakers. Literature available on Telugu HWCR is less. In this paper, we propose a framework for Telugu HWCR using sparse representation techniques focusing on confusion pairs. The proposed method is evaluated on confusion pairs from the standard HPL isolated Telugu character data set.