TamilSTARNet: segmentation via tri-phase architecture and attention-based recognition for handwritten Tamil characters

Handwriting recognition plays a significant role in preserving cultural heritage, specifically in the context of ancient languages such as Tamil. Traditional Optical Character Recognition (OCR) systems find it challenging to handle these issues and hence report very low accuracies. To address the recognition of the complex structure of Tamil alphabets, this paper presents TamilSTARNet, an improved framework for handwritten Tamil document recognition. The proposed methodology consists of two main stages: (1) a three-stage segmentation process that segments. Tamil handwriting in a document into sentences, words, and characters despite its irregular spacing and intricate layouts and (2) an attention-driven recognition model with the view to enhance accuracy. The performance of the proposed model is evaluated using the uTHCD (90,000+ character samples spanning 156 classes) and HP Labs Offline Tamil (20,000+ handwritten words) datasets. The results show that the proposed model achieves an accuracy of 96.18%, proving its effectiveness in recognizing the intricate structure.