Offline Efficient Signature Authentication Using Octave Convolution Neural Network

Handwritten signature verification aims to separate the genuine signature of an individual from forgeries for security reasons. Handwritten signature verification is the process of comparing the image of the questioned signature to the reference signatures of an individual with a predefined threshold to decide if it is considered genuine or forged signature. The task could be done in two ways: online and offline systems. Offline signature verification is more challenging than online systems due to being not easy to distinguish in skilled forgeries processes, as offline confirmation involve fewer features to aid its classification model. Modern approaches include deep learning (DL) algorithms as specified machine learning is trailed to be efficient in the signature image classification task. Octave convolution (OctConv) overcame limitations and achieved better results in terms of accuracy compared to the well-known convolution neural network (CNN). In this paper, we propose to utilize OctConv to build classification model with the purpose of separating genuine and forged signatures. The study revised different architectures using CNN adjusted as baseline models to be used for fair comparisons of OctConv to others. The work different popular signature datasets used have been compared and evaluated, i.e., consisting standard CEDAR, UTSig, BHSig260-Hindi, and BHSig260-Bengali. Interestingly, our OctConv-based model achieved the best competitive results compared to all four baseline model datasets, remarking competitive results among state-of-the-art models including CNN and capsule network with accuracy of high attractive percentages. The work showed attractive modeling strategy to verify signatures needed for today's cybersecurity authentication applications.