Unable to detect signature fields in Custom Neural Model

Question

I recently trained a custom neural model using labeled data that included both signature and normal fields. However, when I tested the model on new documents, it only returned the normal fields. I verified in the project settings that I'm using version 4.0 (see attached screenshot), and my understanding is that v4.0 GA supports signature fields. Could you please help clarify whether signature fields should be supported in this release or if I might be missing a configuration step?

Project Settings:

Answer 1

Hi Prashanth Devireddy,
Greetings & Welcome to Microsoft Q&A forum! Thanks for posting your query!

Yes, the Custom Neural Model v4.0 GA does support signature fields. To ensure signature detection works correctly, you must label signatures using the Signature field type and draw the region for the signature, keeping in mind that only one draw region per field is supported. For effective training, at least five labeled samples with signatures must be provided, including variations in signature styles to improve accuracy. If your trained model is not detecting signatures, verify that the signatures were correctly labeled in your training data.
Additionally, check the model's raw JSON output to see if signature fields are present but have low confidence scores. It is also important to test with document samples that closely match the format of your training data.
For more information: Signature detection.
I hope this information helps.