Inf-Att-OSVNet: information theory based feature selection and deep attention networks for online signature verification

Online Signature Verification (OSV) is a widely adapted biometric characteristics, that objects to determine the signature’s authenticity by computing its unique features. OSV has attracted greater attention in recent years in vital real-time applications like access control, m-commerce, etc. Deep learning (DL)-based Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) are the state-of-the-art methods for many challenging computer vision tasks like pose estimation, scene understanding, semantic segmentation etc. DL based models require huge data samples, but owing to various reasons, it is an intrinsic difficulty in acquiring sufficient number of signature samples per user. Hence, to adopt DL into OSV, deep learning based models learning to generalize from a few signature examples per user, and with fewer features per sample is the critical requirement. We propose a novel Information Theory (IT) based dimensionality reduction technique to meet these critical requirements in which, the relevancy and all possible types of interaction among the features are considered to reduce the dimensionality of a feature set from 100 to 1 regarding MCYT-100 and 47 to 1 in SVC and SUSIG datasets, respectively. Additionally, we present a Deep CNN (DWSCNNA) for signature classification that uses a DepthWise Separable Convolution with Attention mechanism and facilitates one/few-shot learning for test signature verification. Four commonly used datasets, namely MCYT-100, MCYT-330, SUSIG, and SVC are employed in a comprehensive set of experiments to assess the robustness of our proposed dimensionality reduction technique and DWSCNNA framework. The proposed technique with the Information Theory based feature selection & lightweight network with attention produces an EER of 11.56% EER in MCYT dataset, taking into account 35 features per signature. The newest SOTA obtains an EER of 13.38% with 100 features. Similarly, framework results in SOTA EER of 8.71%, 6.46 % and 6.41% in case of MCYT-330, SVC and SUSIG dataset respectively. In comparison to many current and cutting-edge OSV models, the Inf-Att-OSVNet (combination of IT-based feature selection and DWSCNNA) produced a state-of-the-art EER in the majority of test categories.