Key points are not available for this paper at this time.
Character recognition of historical handwritten stone inscription documents is a significant challenge for machine learning algorithms due to the following inherent complexities of these texts: (1) Since these documents are aged and heavily degraded, characters are superimposed with unwanted artifacts such as stains, deformations, and clutter. (2) Characters are not very legible, as they contain spurious strokes due to the extension of ascenders and descenders of adjacent overlapping characters. (3) Owing to the irregularity and variability of the script, multiple variants of the same character are present, which can lead to ambiguity in interpretation. This paper presents a use case of historical handwritten character recognition on a newly compiled EHHKSI (Estampages of Historical Handwritten Kannada Stone Inscriptions) character dataset. The following experiments were conducted on this dataset: (a) transfer learning of various convolutional neural networks (CNNs) and (b) data augmentation. The best recognition accuracy was achieved using the Xception CNN model coupled with data augmentation.
HT et al. (Tue,) studied this question.