Language-invariant novel feature descriptors for handwritten numeral recognition

Numeral recognition is treated as a benchmark research problem as this is a basic module for designing a comprehensive optical character recognition system. In this context, unconstrained handwritten numeral recognition is still considered as an open research problem. Most of the feature descriptors found in the literature for the said problem, work well for numeral images written in a particular language. To encounter this shortcoming, in this paper, we have proposed two shape-based feature descriptors, namely Point-Light Source-based Shadow (PLSS) and Histogram of Oriented Pixel Positions (HOPP). We have evaluated the proposed feature descriptors on 10 (9 offline and 1 online) publicly available standard handwritten numeral image datasets written in eight different languages. Besides, to prove the usefulness of the descriptors in real-life scenario, we have considered numeral string images also. We have also shown how the proposed feature descriptors are invariant toward broken, noisy and rotated numeral images. Experimental outcomes soundly prove that the proposed feature descriptors have the ability to estimate the shape of a numeral image almost accurately irrespective of the language in which it is written. Comparison of the proposed feature descriptors with other shape-based as well as texture-based features shows that PLSS and HOPP produce the results which are analogous to state of the art. The code of the proposed feature descriptors can be found at-https://github.com/ghoshsoulib/Numeral-Recognition.

Automated summary

This paper introduces two novel shape-based feature descriptors for unconstrained handwritten numeral recognition, evaluates their performance on handwritten numeral image datasets in multiple languages, and demonstrates their invariance to variations in broken, noisy, and rotated numeral images. The proposed feature descriptors, PLSS and HOPP, are comparable to state-of-the-art shape-based and texture-based features.