Swintrocr: A Transformer-Based Approach for High-Accuracy Tamil Text Recognition

Tamil is a complex script with many dialects and handwriting styles to this end, handwritten Tamil text Rerecognition is a challenging task to solve and even golder not much annotated datasets are available for Tamil handwritten text recognition. We present TrOCR-Swin, a novel Transformer-based optical character recognition (OCR) model that employs the Swin Transformer encoder instead of the vanilla Vision Transformer (ViT) to improve feature extraction via hierarchical representation learning and shifted-window self-attention. The model is trained using a sequence-tosequence learning approach on a handpicked corpus of labelled Tamil text images. Our evaluation results show that TrOCRSwin significantly outperforms conventional OCR models, with Word Error Rate (WER) of 0.3200 and Character Error Rate (CER) of 0.0544, and facilitates Tamil text recognition, even in terms of noisy and different handwriting styles. With this, it provides a strong and accurate OCR solution for Tamil handwritten and printed text along with serving as the new state of the art (SoTA) for Tamil OCR Task till date.