Multi-view Representation Learning with Deep Features for Offline Signature Verification

Feature learning is one of the most crucial steps in offline signature verification systems. In this paper, to improve the performance of deep learning-based features for the offline signature verification task, we propose a novel framework to learn the new representations from two views of deep features by Canonical Correlation Analysis-based (CCA-based) multi-view representation learning approaches. Specifically, the features from one view can be extracted from deep learning-based feature extractors and the other view can be generated from the extracted view by adding the noise to another homologous sample. Then, the different CCA-based multi-view representation learning methods are evaluated on these two-view deep features to generate the joint features as the final features for the next verification step. Extensive experiments and discussions on three benchmark offline handwritten signature datasets demonstrate that the proposed framework improves the deep learning-based features and achieves the state-of-the-art results compared with other verification systems.