Deep Neural Networks for Robust Handwritten Signature Verification: A Comparative Study

Abstract

This work is another step in the direction of state-of-the-art deep neural networks-based robust handwritten signature verification. In the field of fraud detection and identity verification, there is increasing need for robust authentication techniques, and so signature verification becomes a key problem . In this paper, three of the most widely used deep learning architectures are objectively analyzed on a well-known dataset to measure their performance; ResNet 50,Resnet50_Augmentation, Inception V3,MobileNet V2,MobileNetV2_Augmentation.A systematic training and testing of models with performance evaluation using metrics such as accuracy. The models were trained under a common training procedure so that all results could be comparable. And after that we have applied Data Augmentation in data set and trained these model again.The results show that of the three models tested, ResNet 50 gives the best accuracy and claim the first place with an accuracy of 98.21%, while Resnet50_Augmentation at second place give us a result with an accuracy of 89.48% Inception V3 is 79.47%Inception_V3_Augmentation is 47.71 and finally MobileNet V2 is in third place captured an new accuracies as well which was about 75.22%,MobileNetV2_Augmentation 67.29%. These results demonstrate the potential to outperform the others, further illustrating ResNet 50 architecture's capability of tackling handwriting signature complexity. Papers with code Significance The paper explains effectiveness of different neural net architectures on a signature verification task. Based on our conclusions we find that ResNet 50 is the highest performing in terms of Signature Authentication and additionally believe this leaves potential future work for optimization and use cases in realistic implementations