Kannada Character Recognition From Ancient Epigraphical Inscription Using OCR

Abstract: The digitization and recognition of regional and ancient Indian languages have gained significant attention in recent years, driven by the need to preserve linguistic heritage and improve accessibility. This literature survey consolidates recent research focusing on handwriting recognition, translation, and transliteration techniques applied to scripts such as Kannada, Halegannada, Tulu, and Brahmi. Jayanna et al. (2024) explored deep learning models-CNN, RNN-LSTM, and Vision Transformers-for Kannada handwriting recognition, achieving a maximum accuracy of 99.7% with Vision Transformers, although with higher computational costs. Harsha A C et al. (2024) tackled the low-resource Halegannada-to-Hosagannada translation using a hybrid LSTM and dictionary-based approach, achieving 77.92% accuracy while facing challenges in word sense disambiguation. Prathwini et al. (2024) addressed the recognition and translation of the Tulu language using CNNs and encoder-decoder models, reaching up to 92% recognition accuracy and BLEU scores of 0.83 for translation, but struggling with dialect variations and limited datasets. Mubarakkaa et al. (2024) developed an OCR system for Brahmi-to-Tamil transliteration using CNNs and Tesseract OCR, showing promise but limited by degraded character forms and restricted script coverage. Lastly, Bhumika Purant and Mallamma Reddy (2024) proposed a VGG19-based model for converting Kannada inscriptions into modern Hosagannada via OCR, deployed as a web application with 80-90% accuracy, though constrained by data scarcity and non-standard inscription structures.

Keywords: Epigraphical Inscription Recognition