Any way to use surrounding context to improve recognition especially of similar glyphs (e.g. capital i vs number 1)

[ConcaveTrillion]

I've been training my model very successfully, and getting great results so far, with one (common in OCR) exception.

The capital letter i (I), the number 1, and the lowercase letter L all look very similar especially in some older public domain books (which is the data set I am using and training for). Generally lowercase L and capital i are well-recognized within words (probably thanks to context within the words), but capital i by itself is very often confused for the number 1, since the recognition task seems to take place 'in a vacuum'.

Example:
Image Shows:
in different parts of the world. In the following pages I
OCR reads:
in different parts of the world. In the following pages 1

The 1 and capital I look very similar in many of the old fonts I've fine-tuned the data on.

Any ideas for how to improve this type of recognition, especially given that the OCR engine is built around the idea of "single word detection/recognition"...

[felixdittrich92]

Hi @ConcaveTrillion 👋,

The docTR result.export() contains also the line predictions (built from the word predictions), but I don't think that's really helpful here.

Sounds more like a post-processing with an LM or LLM.
Maybe one of the following models could help with this: https://huggingface.co/datasets/PleIAs/Post-OCR-Correction

Best,
Felix

[ConcaveTrillion]

Thanks Felix. I'll take a look at that. I'm hesitant to use LLM-based correction due to its probabilistic nature and the possible introduction of characters that don't exist into the text. I was also considering some of the rules-based NLP engines which could provide suggestions and auto-fix a few things.