My Prior OCR Work

[driscoll42]

Hello! Following up on my post on reddit for my own OCR work. I was mostly working with documents from the Ivan Allen Jr. Mayoral Records - https://ivanallen.iac.gatech.edu/mayoral-records/traditional/ - as a research project in grad school. They're all 1960s era documents, all in English, about 80% typed, 20% handwritten, so my results obviously won't generalize to everything.

Here's a summary (from about 2022/4)

• Google Vision/Textract are the best, particularly if you need handwriting, but you're paying/uploading to Google/Amazon
• surya is the new OCR that's surprisingly good, had not heard of it before this year. Works great, just depends on whether the commercial license works for you. On my 3080, took 188 seconds to run 88 pages
• Tesseract is still really good for what it is, the real advantage of Tesseract is it's speed. It's 100% CPU bound and for processing 88 pages, it took:
  • tessdata-fast: 121 seconds
  • tessdata-best: 201 seconds
  • tessdata-default: 150 seconds
    • (note in my results, fast was sometimes better than best
  • And that was just leaving it single threaded, I could easily parallelize that, from memory it took roughly a 20th of the time.

However the other issue with tesseract or some of the others is if you need to keep things like tables, headers, etc... for formatting. Or if you need non-latin characters, I expect the results could be quite different (like paddleocr might be much better, I think it's more optimized for Chinese)

I've pushed all my code to GitHub, forgive the very dirty nature of it, I had long meant to make it cleaner but never got around to it. Still I hope it should give you and idea of what I've been up to:

https://github.com/driscoll42/VIP_OCR

Everything probably won't work perfectly out of the box, but should be straightforward fixes I could work on if you wanted the help here.

Happy to try to answer any more questions you have or help in ways you'd find useful. It would be really useful to have an OCR benchmark database with sample code and a test set of files for everyone to compare.

[driscoll42]

Also thinking a bit, to make this database even more useful, comparing the results from the GUI/standard tools liek Adobe Scan, ABBYY, Adobe Pro, and the built in Android/Apple OCRs would be great as well.

[GiftMungmeeprued]

Thank you so much! That's very helpful. I'll let you know if I have any questions. Cheers!

[driscoll42]

Just wondering if you're still working on this? I've been sending it to a few people for OCR suggestions

[GiftMungmeeprued]

Unfortunately, something came up and I've been busy since. I plan to come back to this later. What kind of stuff should I prioritise first?

[driscoll42]

Life comes up! No worries at all.

This is your project, so I defer to you on what you get the most value/interests out of. That being said, I would suggest making some kind of list of possible next steps and then also what the ultimate outcome would look like so if there's anything that needs to be done for all parsers, it's done first.

For myself, the thing I would find most useful/interesting to add on are some more columns with quality of output, time to run, and cost (if applicable), and then for each parser the datetime run, the version run, and system specs run on. When it comes to document parsers and OCR, there's really not a Works/Not Works, it's a percentage. And also really would need more than one document for each type, probably want a hundred or more. Also for the prompt based ones, what the prompt used.

It would likely be helpful for each parser to have a notebook with the code to run each one so someone can copy/paste and see the example, and then also provide a consistent setup so that if you were measuring time to run, could run it on the same machine for consistency.

I know I'm suggesting... a lot, but really it's what you want from this project. What I'm suggesting is a part time job worth of work for a few weeks (though with AI coding... maybe not so much anymore, this is probably something also Claude Code could just churn through).