Robust Handwritten Signature Representation with Continual Learning of Synthetic Data over Predefined Real Feature Space

Talles Brito Viana, Victor L. F. Souza, Adriano L. I. Oliveira, Rafael M. O. Cruz, Robert Sabourin

Abstract: Deep learning methods have emerged as state-of-the-art techniques for learning handwritten signature feature representations. However, successful results in deep learning require a significant amount of training data. The GPDS-960 dataset used to be the largest publicly available dataset of offline handwritten signatures for training deep models. However, due to data protection regulatory issues, the GPDS-960 dataset is no longer publicly available. This way, new investigators starting research in this field have suffered from the absence of a large-scale real signature dataset and have resorted to adopt the GPDSsynthetic dataset, a large-scale set of synthetically generated signatures for training models. Nevertheless, we have found a difference in verification performance between models trained using real and synthetic signature data. To deal with this problem, we apply a method based on data-free knowledge transfer learning. Firstly, we generate inverted examples with the same distribution as the real examples. Then, we complement a feature space based on real data using synthetic data while minimizing the divergence in distribution between the representations provided by these two different types of data sources. This is achieved through continual learning based on knowledge distillation. We evaluated models obtained with the proposed method in terms of the equal error rate on GPDSsynthetic, GPDS-300, CEDAR, and MCYT-75 datasets in a writer-dependent verification approach. Experiments demonstrated that the proposed method provides a more robust model for writer-dependent verification when considering real and synthetic signature datasets. Inverted data is available for download at https://github.com/tallesbrito/continual_sigver.