Hacker News Reader

Comment by dsp_person | Hacker News Reader

I was using something like this for awhile, but I found tesseract did poorly quite often. That resize trick didn't seem to affect much. I'm not sure what pre-processing would make it better.

I'd love to if TextSnatcher does anything to improve on this. The github page is opaque.

View on HN

The source is pretty straightforward - it's calling `scrot -s -o` to a temp file, and then `tessaract` with no further preprocessing.

https://github.com/RajSolai/TextSnatcher/blob/master/src/ser...

> I found tesseract did poorly quite often

The script calls Tesseract in default page segmentation mode (PSM 3). [1]

Depending on the input text, PSM mode 11 for disconnected text would probably work much better. That uses the flag "--psm 11".

[1] From the original repo: string tess_command = "tesseract " + file_path + " " + out_path + @" -l $lang" ;

Having used Tesseract for OCR for other things, getting the right PSM helps but it's still rather terrible, especially for sans-serif fonts, which are common in UIs.

Granted there's a lot of ambiguity in sans serif fonts, lower-case "L", vertical bar, and upper-case "i" can even be pixel-identical, but I've seen tesseract turn

  Chapter III

into

  Chapter |l1

which really surprises me. In fact, for books, I run it through sed to replace vertical bar with upper-case "i" and it significantly improved recognition.

Reply on news.ycombinator.com