Creating custom OCR from known fonts

dairy

I'm looking to create an OCR model that will read numerical digits only (no other characters) from a few known specific fonts. This will read a known number of digits in a string e.g. 8 in a row. What's the best way to go about this please? I feel like it should be reasonably efficient process to give the font to a pre-existing model design? Is it possible to train on several fonts at the same time so it will read several fonts particularly to cope for example if the 4 is looped at the top or more like a fork shape, or if the 7 is crossed/not crossed.

Thank you for taking the time to look at this, I'm just starting out…

jakaskerl

Hi dairy
It should be relatively straightforward to train a model on your own dataset. You will need some labeled examples of these characters in various scales and orientations. Then you train the model on a shuffled dataset and the model you get should be able to do OCR on any character from that dataset you created.

I assume you want to run this on an OAK device?

Thanks,
Jaka

dairy

Hi jakaskerl

Thank you for coming back to me, what's the best route to go about this? Indeed I want to run this on an OAK 1 we have a few and testing it using the demo OCR model we're not getting accurate results. Mostly seeing numbers as characters and misreading font variations hence I thought it would be worth training something up.

Thanks, John

jakaskerl

Hi dairy
You will likely need to retrain both stages.

text detection - if you are detecting numbers when trying to detect characters, that would signify the problem if that the model is unable to distinguish between a number and a character.
character recognition - if you are getting bad accuracy when trying to infer character class, you will need to retrain this stage as well, especially if you want it to classify characters it hasn't seen before (different fonts).

You can do this by applying transfer learning to the already trained models. openvinotoolkit/training_extensions should work for the second stage i think. Try that first, then if problems persist you might need to update the first stage as well. I also recommend you follow online guides to avoid confusion.

Thanks,
Jaka