Static signature verification using NASNet deep learning architectures

• Main Authors: Arulanandar, Daniel Raj, Padmanabhan, Venugopal, Nammalwar, Padmapriya, Govindan, Sakthi Priya
• Format: Conference Proceeding
• Language: English
• Subjects: Datasets; Deep learning; Documents; Signatures; Verification
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0156725

Static signatures are the signatures in the offline documents. The genuineness of these signatures is very important to verify the originality of the documents. There are many methods available to check the authentication of the static signatures. In this paper, NASNetMobile and NASNetLarge deep learning architectures are proposed for static signature verification (SSV). These architectures were built with ImageNet weights and then fine-tuned to distinguish real and fake signatures. Initially, the architectures have been trained using signatures from the ICDAR signature dataset, both original and augmented. The transfer learning approach was used to train all layers of the architectures. These architectures were evaluated on signature datasets with a multitude of genuine signature samples using the NASNetMobile and NASNetLarge for the both datasets. This comparison shows that the proposed strategy improves accuracy in verifying static signatures. The results achieved using the transfer learning process were compared to those obtained using deep learning architectures like CNN-trained architectures.