Multi-phase Offline Signature Verification System Using Deep Convolutional Generative Adversarial Networks

We present an offline signature verification system with new architecture, using Deep Convolutional Generative Adversarial Networks (DCGANs) to learn features unsupervised instead of using hand-crafted features. The advantage of this architecture is that system has a robust generalization ability. Besides, aiming at the conflict between convenience and accuracy, a hybrid Writer-Independent-Writer-Dependent classifier is used, which is an approach compromise two different kind of classifier. We test our method in GPDSsyntheticSignature database, which is an updated version of GPDS-960. In order to make it's convenient to compare with other works, we also use GPDS-960 to test our method. Our experimental results show that, the accuracy of proposed method is satisfactory. Due to this kind of architecture, the more query samples are tested and enrolled into the system, the more accurate our system will be conceptually.