fringe marsh
Hello,
We recently started using your text recognition plugin, particularly for MRZ code reading, and encountered some issues. Firstly, the model sometimes confuses the '<' symbol with the letters 'C' or 'K'. There are also problems with recognizing the zero ('0'), which is often read as the letter 'O'. Is there an option to retrain the model to improve recognition of these symbols, or any other solutions you might suggest?
Thank you for your assistance!
stable ivy
This is actually a sample which uses Text OCR from Google's ML Kit.
You need to have some custom processing to parse the text.
If it's a document number (based on the position of the characters), you need to consider it as a number instead.
Unfortunately, currently, there is no option to retrain the model and use as it's not yet offered by google's ML Kit.
But, there is an option via TFLite library which I need to further explore.
Do you have any existing model which works well for your domain?
We currently support custom models for Object detection alone. But once I explore futher on tflite, may be I can see if we can support for ocr too. But would like to know if you have a model for your domain already.
fringe marsh
Yes, I'm already using post-processing of the recognized text. However, there are situations where a character could be either a number or a letter. Also K at the end of a surname, it could be part of the surname or a wrongly recognized '<' sign.
I don't have a model that works 100%, but this solution provides a decent result: https://github.com/SerdarHelli/MRZ_Passport_Reader_From_Image
stable ivy
The shared project uses Tesseract and as of now we haven't thought to provide as an alternative to the existing solution as it holds good in most of the cases.
Did you try https://github.com/alimertozdemir/EPassportNFCReader as it uses ml kit too? However, you may see the same limitations as it uses the same under the hood.
