To even have a chance at doing it you'd need to start the training from scratch with _huge_ penalties for filling in missing information and a _much_ larger vision component to the model.
See an old post I made on what you need to get above sota OCR that works today: https://news.ycombinator.com/item?id=42952605#42955414
To even have a chance at doing it you'd need to start the training from scratch with _huge_ penalties for filling in missing information and a _much_ larger vision component to the model.
See an old post I made on what you need to get above sota OCR that works today: https://news.ycombinator.com/item?id=42952605#42955414