Writer independent handwritten signature verification on multi-scripted signatures using hybrid CNN-BiLSTM: A novel approach

Abstract

Authenticating important documents by identifying individuals using handwritten signatures make signature verification a critical task. Interpersonal similarity and intrapersonal variation of individuals along with high skilled imitation of signature structure make automatic signature verification a challenging task. In such scenarios, a signature verification system should detect the small differences between genuine and forged signatures with high efficacy. This paper proposed a novel approach towards offline signature verification where a hybrid deep learning network, consisting of a Convolutional Neural Network and a Bidirectional Long Short Term Memory network is used. Signature written with freehand and an imitation of it are almost identical structurally. Deep Neural Network is used to recognize skilled forgery from genuine signatures because of its capability to learn critical details and subtle patterns from the image pixels. A Convolutional Neural Network is trained, and the trained network is then used to extract diverse features of the signature images. The generated feature vectors are then used for classification using Bidirectional Long Short Term Memory. The hybrid deep learning network classifies the input signature as skilled forgery or genuine with high accuracy. For this work, state-of-the-art datasets, such as, GPDS-300, GPDS-Bengali, GPDS-Devanagari, CEDAR, BHSig260-Bengali, BHSig260-Hindi, and a local dataset, Meitei Mayek signature are used. In order to verify the robustness of the system, different multi-scripted offline signatures belonging to multi-lingual Indian society, are used for evaluation. The experimental results determined from multi-scripted signatures, exhibit that the proposed system is found comparable to many state-of-the-art systems and in some specific cases, it out performs some of the existing systems.