Can we train custom model on handwritten signatures not recognized by OCR?

Hello @Bogdan Pechounov

Thanks for reaching out to us, yes, if you are training the custom model only for signature, I would highly recommend you trying signature field to see if the accuracy is good enough for you.

For your question "Do custom models perform token classification on the text detected by OCR? Or do they use techniques such as YOLO for object detection to first detect any arbitrary object and then classify it?" I think this information is not disclose at this moment, from technical side, both of the ways are possible.

Please try recognize the signature fields and see how it works, if it is not good enough for you, we can discuss further to see any other ways we can accomplish it.

@YutongTie-MSFT "both of the ways are possible". The reason I ask is because if OCR is used, then the signature wouldn't even be detected. However, it might be detected if a different algorithm is used, such as one for object detection.

Are signature fields only for template models? If I understood correctly, they use the text detected by OCR in a specific position.

Here is an example where signatures aren't detected:

Thank you,