Signature Verification using Convolutional Neural Network and Autoencoder

Abstract

Signature has been one of the widely used verification biometrics out there. Handwritten signatures are used in cheques, forms, letters, applications, minutes, etc. The Signature of every individual is unique in nature, that is why it is essential that a person’s handwritten signature be uniquely identified. Signature Verification is a widely used method for authenticating any individual during absence. Human verification is prone to inaccuracy and sometimes indecisiveness. This paper presents an investigation of using Convolutional Neural Network (CNN) for Writer-Dependent models in signature verification. Random distortions were generated in genuine images using an autoencoder to get forged signatures, which were passed to the classifier during training. The paper details all the pre-processing steps carried out on the image and shows various test results for changing the number of training sets of images. The average test accuracy for Persian dataset is 83% when the system was trained with 22 genuine images. There was a decrease of 9.4% in accuracy when the model was trained with 9 genuine images.

Subscripto multiplex: A Riemannian symmetric positive definite strategy for offline signature verification

“…We also provide a comparative summary with characteristic methods found in the literature given the fact that a more thorough comparison is not feasible given the numerous and diverse nature in the design and implementation of the system's stages. We additionally report here other augmentation research efforts i) the work of Yapici et al [57] which utilized a novel cycle-GAN topology for generating synthetic signature images but did not evaluate the used datasets, ii) the work of Prajapati et al [69] which exploited autoencoders for synthetic signatures; but it was restricted on a private Persian dataset and iii) the work of Ruiz et al [58] tested only with random forgeries. We feel that that the proposed method is definitely worthy a new line of research for OSV.…”