Hacker News Reader

Comment by Infernal | Hacker News Reader

Their paid features include OCR, both for handwriting and images of printed text. Seems like a textbook machine learning problem.

https://blog.evernote.com/blog/2015/01/23/search-handwriting...

View on HN

And I've been very, very impressed with the handwriting recognition. I discovered it by accident before it was a paid feature when I was searching for something and it brought up an image of some scribbled notes I had taken with the search highlighted. And my handwriting is bad, really bad.

Awhile ago someone asked how a product like ABBYY Finereader, which does great when OCRing typed text, does on handwriting. Not well at all I'm afraid:

https://gist.github.com/dannguyen/50dcb5a8f4230e7a8a40bfe2d0...

Evernote doesn't just OCR handwriting to a single text; it indexes a bunch of possibilities for every handwritten word. This way the search functionality can search through all these variations, not just the most likely one.

I haven't used its handwriting OCR feature so genuinely curious. Does it not show a single piece of text? If not, what value does the OCR handwriting provide? Is it solely for the search feature?

Yep it seems to be, it indexes every possibility that the handwriting might be so that if you search for it then it comes up. You can see this if you dig through the XML it exports.

OCR with Evernote works very well because it's allowed to be fuzzy. One hand written word is allowed to have multiple predicted texts. This is because the predictions are only used for search - you can't copy/paste OCR'd text because Evernote can't guarantee it's correct. But it's good enough for fuzzy search. In fact it's pretty much exactly what you want for search: allowances for misspellings and typos. Very smart.

Reply on news.ycombinator.com