A novel deep learning model for classifıcation of handwritten signature images

Signatures are used as a means of identification for banking, security controls, certificates and contracts. It also plays an active role in legal issues that need to be examined. From this point of view, in this thesis, it is aimed to design network architectures that work very fast in areas that require signature images. For this purpose, a new Si-CNN network architecture was designed with existing layers. A new classification layer called "Si-CL" was developed to improve the classification performance of the model. "Si-CNN+YS" network architecture was also developed by replacing the classification layer used in the "Si-CNN" network design with the newly created "Si-CL" layer. In addition, the "Si-CNN+SVM" model was developed by adding the SVM (Support Vector Machine) layer to the "Si-CNN" model. Si-CNN, Si-CNN+YS and Si-CNN+SVM models were trained with three datasets. For the training of signature images, experiments were conducted using "Si-Signatures", "Si-Signatures-Fake" and "Cedar" datasets. GoogleNet, DenseNet201, Inceptionv3 and ResNet50 were used to compare the performance of the developed networks. For "Si-Signatures", "Si-Signatures-Fake" and "Cedar" datasets, 98.64%, 95.91% and 99.09% accuracy rates were obtained respectively and the most successful model was found to be Si-CNN+YS. The findings of the study show that the proposed network models can learn features from three different handwritten signature images and achieve higher accuracy than other benchmark models. The test performance of the trained networks shows that the Si-CNN+YS network performs better in terms of both accuracy and speed. Due to its superior performance, Si-CNN, Si-CNN+YS and Si-CNN+SVM can be used by signature experts as an aid in various applications including crime detection and forgery.