search

card-io / card.io-Android-SDK

card.io provides fast, easy credit card scanning in mobile apps
Other
1.99k stars 535 forks source link

How to retrain card.io model to include non-embossed digits? #217

Open sprtkd opened 6 years ago

sprtkd commented 6 years ago

As mentioned in all issues, card.io fails in detecting non-embossed digits on credit cards. There are few questions regarding this:

  1. Is the model not trained on OCR-B digits, or is this problem related to non-embossed only?
  2. Is there any hacky way to implement this using Tesseract, or a new model using Tensorflow?
  3. Why are non-embossed cards a problem?
jeffrwatts commented 6 years ago

+1 on this.

Would be great to have ability to get non-embossed... whether that is an update to this model, or ability to insert a new model using TF.

@sprtkd did you get an answer offline that you wouldn't mind sharing?

Thanks in advance.

sprtkd commented 6 years ago

So I tried training on a TF model with computer generated digits, which gave little accuracy. nearly 50%

Then I found out a repo which does it.

https://github.com/SmartEngines

This works as expected.

I would be happy to train a tf model on non embossed cards if anyone provides me with a dataset, since computer generated datasets are futile here.

jabransiddiqui commented 5 years ago

does anyone got solution of non-embossed card digits scan ?