Offline fake signature and real signature classification using deep convolutional networks

In this thesis, signature verification, which has a significant impact on human life today, has been addressed using deep learning methods. After the recent increase in fraudulent activities, the examination of offline signatures has posed a considerable workload for document examiners. With the advancement of artificial intelligence, various methods have been tried to automate signature verification with the progress of deep learning techniques. Convolutional neural networks, a deep learning method, have been employed. Architectures of convolutional neural networks used in image classification, namely AlexNet, DenseNet201, MobileNet, Vgg16, Vgg19, Resnet50, Resnet101, InceptionV1, InceptionV3, have been utilized. The GPDS and ICDAR2011 signature datasets have been used. The highest success rates from the models used for these datasets have been compared. The highest accuracy rate for the GPDS dataset was achieved with the MobileNet architecture, while for the ICDAR2011 dataset, the highest accuracy rate was obtained with the InceptionV3 architecture.