Abstract
Recent advances in deep learning have been utilized successively to improve the performance of signature verification (SV) systems. Deep models proposed in the literature are complicated and need to learn many parameters to give acceptable error rates, requiring a lot of training data. On the other hand, those models are designed and hand-crafted specializing in the problem, online or offline SV. In this work, we suggest and show on popular datasets that similar and simple convolutional neural network (CNN) models can achieve state-of-the-art results both for offline and online SV problems. For offline SV, our work outperforms its counterparts with and without data augmentation. We also show that a very similar CNN architecture can be employed for online SV. To the best of our knowledge, this is the first work to show that CNNs can be used to learn online signature representations directly from raw data.