Signature handwriting identification based on generative adversarial networks

Abstract

Handwritten signature has been an important identity-verification method since ancient times. Compared with manual handwriting verification, the use of computer image recognition technology for handwriting verification is faster and avoids subjectivity. However, there are still some challenges in traditional image recognition methods, such as feature selection, lack of a standard basis, and low accuracy. For the first time, generative adversarial nets (GAN) technology is adopted to study the task of handwritten signature identification. A special network SIGAN (Signature Identification GAN, SIGAN)is proposed based on the idea of dual learning. The loss value of the trained discriminator of SIGAN is used as the identification threshold. The authenticity of the test handwritten signature is determined by comparing the threshold and loss value of the test image obtained through the network. The experimental data set in this study consists of five hard pen-type signatures, which include some real signatures and some deliberate imitations. The experimental results show that the average accuracy of the SIGAN-based signature identification model is 91.2%, which is 3.6% higher than that of the traditional image classification method.