What would be the advantage compared to something like Tesseract ?
It means that it performs not-so-good when for example image contains black text and white text on green background since this is not "normalized" through image preparation steps and it cannot detect white text on green background (but you can do it yourself)
This does involve creating your own labeled data.
I got this 99% accuracy by performing incremental training using latest Manheim model as a base. I added about 20k lines which is not really that much. https://github.com/tesseract-ocr/tesseract/wiki
The hard part was crowd sourcing those 20k lines :)
Tesseract might not be best for photos as you said but I did not have major problems.
Of course some documents the source is so bad that a human can't achieve 99%.
Tesseract used to be quite average before they moved onto LTSM models a few years ago.
Also this: https://github.com/UB-Mannheim/tesseract/wiki
The original data was here: https://github.com/tesseract-ocr/langdata_lstm
I did use another data source from Manheim but can't locate it right now.
Using vanilla Ubuntu 18.04
I looked at the example training files and made a small script to convert my own labeled data to fit the format that tesseract requires.
I did do a bit of pre-processing adjusting contrast.
All the data munging was done on Python (Pillow for image processing, Flask for collecting data into a simple SQLite DB before converting back to format that Tesseract requires).
Python was not necessary just something that felt most comfortable to me. I am sure someone could do it using bash scripts or node.js or anything else.
EDIT: To make life easier for my curators I did run Tesseract first to generate prelabeled data for my training set. It was about 90% accurate to start with.
So the process was: Tesseract OCR on some documents to be trained -> hand curation (2 months)-> train (took about 12 hours) -> 99% (on completely separate test set)