Automatic Signature Verification

less than 1 minute read

Signature verification involves distinguishing between authentic signatures and forgeries. Inspired from facial recognition networks, a one-shot learning convolutional neural network was designed to detect forgeries of a given signature.

A Siamese architecture was developed using PyTorch and was trained using triplet loss. A dataset containing several authentic signatures & forged signatures of around 2600 people (around 75,000 samples) was used. The final network converts the input signatures to a lower-dimensional embedding vector for comparison. It is able to identify forgeries even with only one or two true signatures for reference.

placeholder image 1
Training: Training the Siamese architecture using triplet loss, using a pivot image, true image and false image as inputs.

This project was done in Summer 2019 as an internship project (as a requirement of the ‘Practice School’ programme of BITS Goa) for Bank of Maharashtra.

[GitHub Repo]