Off-line handwritten signature verification using compositional synthetic generation of signatures and Siamese Neural Networks

• Published in: Neurocomputing (Amsterdam) Vol. 374; pp. 30 - 41
• Main Authors: Ruiz, Victoria, Linares, Ismael, Sanchez, Angel, Velez, Jose F.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 21.01.2020
• Subjects: Compositional model of synthetic signature generation; Off-line handwritten signature verification; On-Demand training; Random forgeries; Siamese Neural Networks; Siamese Neural Networks; Compositional model of synthetic signature generation; On-Demand training; Off-line handwritten signature verification; Random forgeries
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2019.09.041

•Siamese Neural Network for random forgery offline signature verification problems.•Method for generating on-demand synthetic signatures based on simple geometric shapes.•Combining synthetic, augmented and original signatures to train Siamese Networks.•Generalization capability when training and testing using different datasets. In this work, we propose the use of Siamese Neural Networks to help solve the off-line handwritten signature verification problem with random forgeries in a writer-independent context. Our proposed solution can be used on new signers without the need for any additional training. Also, we have analyzed three types of synthetic data to increase the amount of samples and the variability needed for training deep neural networks: augmented data samples from GAVAB dataset, a proposal of compositional synthetic signature generation from shape primitives and the GPDSSynthetic dataset. The first two approaches are “on-demand” generators and they can be used during the training stage to produce a potentially infinite number of synthetic signatures. In our approach, we initially trained Siamese Neural Networks using signatures from GAVAB dataset and different combinations of synthetic data. The best verification results were obtained when combining original and synthetic signatures for training. Additionally, we tested our approach on the GPSSynthetic, MCYT, SigComp11 and CEDAR datasets demonstrating the generalization capabilities of our proposal.