Vision-Language Models for Tamil OCR: Enhancing Image-to-Text Conversion using Contrastive Learning

Tamil has a multi-dimensional script with lots of characters and complex interwoven lines and merged characters, which makes the Optical Character Recognition (OCR) of Tamil faces a very difficult process. Dravidian scripts such as Tamil have complex nature which makes traditional OCR technology problematic specifically for handwritten and historical documents or texts produced from low quality scanning techniques. Based on this, this research develops a new deep learning based image to text conversion solution using Vision Language Models (VLMs) with contrastive learning techniques to improve Tamil text identification. In particular, we adapt a system with a CNN feature extractor and a Transformer decoder with contrastive loss functions that force the system to identify unique characters from similar glyph visual appearances. By utilizing pre-trained multimodal models such as CLIP and BLIP which have their own visual and textual embeddings, the recognition performance improves by handling different Tamil fonts, handwriting styles and degraded documents. The self-made data set of different Tamil writing scripts in different states of print and handwriting and degraded historical document quality was used for proposed system training. The results indicate that not only are we able to provide superior performance for both character accuracy rate (CAR) and word error rate (WER) measurements as well as resistance to signal distortions compared to standard methods, we are also able to employ simple processing. This research work adopts contrastive learning and vision language modeling as a new standard for Tamil OCR, and opens up new opportunities for document digitization and preservation of Tamil heritage with better language processing accessibility.