Ask HN: OCR from screenshot returns gibberish

My brother works at a restaurant and his manager sends him screenshots of the schedule via email.

I would like to write a simple OCR app that does the following:

-gets the screenshot from his gmail -finds his name on the schedule -adds the hours from the schedule to his google calendar

This is a fun weekend project. Thinking about building it in a new language I haven't used before.

However, when running the screenshot through the OCR stuff I can find online (before actually writing code) the results are absolutely horrible.

Am I doing this wrong or is OCR just not very good?

4 comments

[ 2.9 ms ] story [ 19.0 ms ] thread

ohiovr 7y ago ↗

I’m not an expert but have you tried tesseract ocr?

raleigh_user 7y ago ↗

yeah. and i get absolutely horrible results. looks like a toddler typed it out

janci 7y ago ↗

You may need to prepare the input - isolate the parts you want read, blank out all other, remove the background color, table borders, graphic elements. Convert to greyscale/BW. Then apply OCR

ColinWright 7y ago ↗

I've generally found that OCR requires high resolution and/or image pre-filtering. With significant pre-filtering I've had some great results.

Tesseract can be very, very good, but also very, very bad. I'd suggest you have a quick hack at writing your own overly simplistic OCR tool and see how well you get on. This will either give you an appreciation of the difficulties and potentially how to do the pre-processing to overcome them, or you will have a tool that is better than the existing ones, and people will love you for it.