Signature Verification Using CNN Deep Learning-Based Approach

A signature is a representation of a person's name affixed to papers to verify their identity. Every person's signature is different, making it crucial to be recognized. Human verification might be imperfect and uncertain at times. In some cases, signatures on papers are occluded by document texts or rubber stamps, which makes them difficult to verify. The same might be said for antique documents, in which the signatures may not always be visible. Others, when a client's signature has slight alterations in the pattern, some financial transaction requests cannot be validated precisely. Handwritten signature verification is a challenging task that has been widely studied in recent decades to solve all the problems. Despite the progress that has been made, it remains an active research, whereby new approaches are being developed all the time. Researchers are experimenting variety of methods for distinguishing between genuine and forged signatures. Motivation to these problems, a novel method is proposed for reliably recognizing signatures in documents and performing identification checking using a CNN-based deep learning algorithm. The proposed algorithm's validity is determined by the matching sign with the online signature database. An autoencoder is utilized to create random distortions in genuine photos from the database, which were then given to the classifier during training to create false signatures. The proposed algorithm is based on Siamese Network, which works on two inputs from VGG-16 with the same weight and same structure and produces two features. The classification results of the proposed algorithm are about 99% accurate.