A Comprehensive Review of Existing CNN Architectures for Handwritten Devanagari Character Recognition

Major advances in deep learning algorithms (Primarily CNN's) for handwritten Devanagari character recognition have driven continued advances over the last ten years. In this review article, we provide an ordered summary of research on Devanagari Character Recognition (DCR) from 2010 through 2024. The methods are summarized and also illustrated with how complex CNN architectures based upon data-driven techniques have replaced simpler manual methods for generating features. The review includes a comparison of the network architectures, data sets and improvements made to hyper-parameter tuning within peer-reviewed articles that were published. Collectively, the results indicate that modern CNN approaches yield higher recognition accuracy than would have been achieved otherwise. Furthermore, issues related to class imbalance and handwriting style variance remain significant barriers. Finally, notable advances and emerging trends within the field include hybrid networks and transfer learning, both of which provide promise for increased accuracy and generalization in future research efforts related to Devanagari OCR. Collectively, these results provide guidance to new researchers as well as identifying remaining questions that should be addressed within the context of this rapidly advancing area.