Recognition of handwritten characters from Devanagari, Bangla, and Odia languages using transfer-learning-based VGG-16 networks

Abstract

ABSTRACT Despite promising results in character recognition techniques, research on handwritten characters from Indian regional scripts is still limited. In fact, most character recognizers in Indian languages are far less accurate than those built for English alphanumeric characters. Traditional techniques rely on extracted characteristics that require extensive knowledge of the chosen script, which is never practical. In such a circumstance, automatically extracting features may create interest. This study demonstrates how deep CNN VGG-16 networks can be used to increase character recognition accuracy. We have used nine different datasets from three separate Indian regional languages, Devanagari, Bangla, and Odia, to validate the model's efficacy further. When the samples were very noisy, it was revealed that the VGG model performed exceptionally well. The model recognized up to 3% better accuracy when the input samples were noisy and without applying any preprocessing. Furthermore, the model was implemented via transfer learning rather than being trained from scratch. This accomplishment could pave the way for constructing an automatic character recognition system for Indian regional scripts. The model outperforms recognition accuracy by around 1% for two datasets, cMATERdb 3.1.2 Bangla Basic and NITROHCS_V1.0 Odia Basic, compared to the existing approaches in the literature.