Hacker News new | ask | show | jobs
by YZF 4275 days ago
It has OCR but wasn't working so great. It uses Tesseract. I'm not absolutely sure why it wasn't working well in the past, possibly something to do with different fonts/display rendering (e.g. ClearType and such). It "almost" worked so maybe it got better or maybe there's some tuning you can do. Didn't spend too much time on it.
1 comments
drothlis 4275 days ago
OCR is never perfect. I do a lot of automated UI testing in a way similar to Sikuli, and while we do rely on OCR a lot, you have to use certain workarounds (like fuzzy matching instead of looking for a perfect match of your expected text).

Ultimately Tesseract was primarily designed to operate on text which had been printed and then scanned, whereas the text on screen is lower resolution, anti-aliased, on a coloured background, etc etc.

Some further details of our OCR investigations here: http://stb-tester.com/blog/2014/04/14/improving-ocr-accuracy...

The TLDR version is: Training Tesseract on your font doesn't help; scaling up the text 3x before passing it to tesseract gives a massive improvement (I don't know if Sikuli does this); normalising ligatures & punctuation gives an additional slight improvement.

link