Google Vision API is eons ahead of tessaract in its OCR capabilities. I was quite impressed how good it was when trying to read ID numbers from scans.

I guess my point is that this project is interesting because it doesn't depend so strongly on the OCR being perfect, as he's allowing to prescribe a handwriting style that helps it along. I just think it's a nice compromise, and something that I hadn't thought of previously.

Obviously the tech can always be improved, but I find this to be a secondary concern. Ideas can still be cool without worrying about implementation details. For instance, if the PoC works well enough, but you want better performance, then you can train a new OCR system specifically for it. You can break the project down into pieces, where each piece is an interesting challenge. Meanwhile getting 80% performance out of an existing solution allows you to build up a dataset to be used for improving it.

Ha I was going to say, I was trying to read receipts and Tesseract was having problems, those are machine-printed not hand written. Granted supposed to clean up/up contrast/etc... but yeah. Scanning a screenshot of a website though accuracy is almost 100% every time eg. just black/white text.

It's usually better to segment the text on the image (and twek the perspective if necessary) to help tesseract extract the text. All the "easy" preprocessing that can be done beforehand helps a lot. I understand that Google Vision Services doesn't need these kind of adjustments (or maybe they do them automatically?), which makes for an easier time/less code, but setting up the account sure looks like a lot of work from the article!

That's a good tip about the segmenting text, will have to keep that in mind next time.

I like the local aspect, one project I have in mind is a hand writing to printed letters for note taking/tablets. Not a new concept I think one drive has it, but it would be neat to train it on your own writing.

Why not Zoidberg

If you want to train your own net for handwritten letters, you can start with EMNIST[0] and then move on from there.

[0] https://www.nist.gov/itl/products-and-services/emnist-datase...

Thanks, yeah it's an ambitious project. It's cool but has no real gain immediately. Maybe if I was a student/used a tablet everyday.