Removing Background from Noisy Handwritten Signatures on Banking Documents Using GANs

In this study, the CycleGAN (Cycle-Consistent Gen-erative Adversarial Network) architecture is used to remove background noise such as stamps, seals, handwriting, or marks behind signatures to increase the accuracy of our pre-existing signature verification system. In Turkey, aforementioned back-ground noise types are frequent on official documents, resulting in lower signature verification ac curacy than expected. We train a model from scratch using the CycleGAN architecture to remove such background noise, which learns not only to remove background noise but also to preserve signatures without background noise as is. We evaluate the performance of our model and observe a significant increase in s ignature verification accuracy pipeline. Further tests show that using our trained model on clean signatures enhances visual clarity; positively affecting the signature verification pipeline.