Offline Signature Verification using Deep Learning Convolutional Neural Network (CNN) Architectures GoogLeNet Inception-v1 and Inception-v3

Abstract

Biometric systems such as signature verification are highly viable in order to identify individuals in organizations or in finance divisions. Advancement in classification of images using deep learning networks has opened an opportunity for this problem. In this study, the largest available handwritten signature dataset, namely, the GPDS Synthetic Signature Database, was employed for the classification of signatures of 1000 users, each of which having 24 original (or genuine) signatures, and 30 forged (or fake) signatures. Moreover, two popular GoogLeNet architecture versions of CNN, namely, Inception-v1 and Inception-v3, were used. Firstly, algorithms were trained on samples from 20 users, and achieved a validation accuracy of 83% for Inception-v1 and 75% for Inception-v3. In terms of Equal Error Rates (EER), Inception-v1 managed to obtain an EER as low as 17 for 20 users; while EER for Inception-v3 with 20 users obtained 24, which is a good measure compared to prior works in the literature. Although Inception-v3 has performed better in the ImageNet image classification challenge, in the case of 2D images of signatures, Inception-v1 has performed the classification task better than Inception-v3 It is also acknowledged in this study that Inception-v1 spent less time training, as it had a lower number of operations compared to Inception-v3.