I am using a Conv-based architecture to verify two signatures on cheque images. While training, I used the Siamese Network to differentiate between the signatures and the model works pretty well.
I do not have any classes in my verification model. The model outputs logits based on the signature images.
Eg.
(Sig1 ------> [Model] -------> Logits1)
(Sig2 ------> [Model] -------> Logits2)
After that, I calculate the distance between the two logits.
dist = get_distance(logits1, logits2)
Based on a certain threshold and dist I verify whether it is the same signature or not.
As multiple signatures can exist within a single signature field, I am having problems in segmentation of the signatures based on reference signature.
Is it possible to do object detection based on the aforementioned model and generate a bounding box if the signature is matched? If possible how can I do that?
Any help is appreciated. Thank You!