EffKannadaRes-NeXt: An efficient residual network for Kannada numeral recognition

Authors: Saini, A., Daniel, S., Saini, S. and Mittal, A.

Journal: Multimedia Tools and Applications

Volume: 80

Issue: 18

Pages: 28391-28417

eISSN: 1573-7721

ISSN: 1380-7501

DOI: 10.1007/s11042-021-10797-0

Abstract:

In this article a framework regarding correct and efficient identification of Kannada handwritten numerals has been proposed. The EffKannadaRes-NeXt model is based on deep residual network ResNeXt and takes binary as well as gray-scale representations of numeral images into consideration. The study deals with handling numeral images from MNIST-sized Kannada-MNIST dataset and Dig-MNIST dataset, an out-of-domain test dataset. The test datasets derive test sets from two different scenarios and these sets are beneficial for evaluating the robustness of the model. The EffKannadaRes-Next is observed to achieve an accuracy of 97.81% and 97.37% on the Kannada-MNIST test dataset along with 82.08% and 81.67% on Dig-MNIST dataset. A comparison of results with those available in the literature is performed and a close agreement shows the versatility of the present technique.

Source: Scopus