CNN-based multilingual handwritten numeral recognition: A fusion-free approach

Abstract

Numeral recognition plays a vital role in making automated systems like postal address sorting and license plate recognition. In a multilingual country like India, more often, the multiple languages are mixed while writing. Numeral recognizer systems which can handle more than one language are very much useful to recognize numerals of various scripts. Handwritten numeral recognition is much more complicated than the printed one because of the different writing styles. Few multilingual works have been reported earlier, but these systems are either trained for each language individually or fusion of similar shaped classes of performed. In either case, multiple classes exist for a single digit. In this work, we have developed a script independent numeral recognition system for multilingual handwritten digits which is independent of fusion and has only 10 classes corresponding to every single numeric digit. This work is first of its kind in which we have addressed the problems of multilingual numeral recognition systems. Exhaustive experiments are done with numeral datasets of 8 Indic and non-Indic scripts. We have obtained the accuracy of 96.23% collectively for all the eight scripts. This obtained high accuracy is promising and demonstrates the hypothesis that multilingual handwritten numeral recognition is void with CNN.