Light Weight Online Signature Verification Framework by Compound Feature Selection and Few-shot Separable Convolution Based Deep Learning

Online Signature Verification (OSV) is an extensively used biometric trait aims to verify genuineness of a test signature by computing unique features of the signature. The advancements in mobile and communication technologies resulted in usage of computationally sparse mobile devices in critical applications like m-commerce etc., demands for OSV frameworks which are able to classify the dynamic test signature with fewer number of training signature samples and lesser number of features. The recent advancements in Deep Learning (DL) technologies, resulted in exponential improvements of accuracy in traditional tasks like Object Detection, Scene Text Detection etc. The main disrupt in usage of DL based frameworks for OSV is the requirement of extensive number of training samples and larger number of parameters to learn. To overcome the above pitfalls, we propose a novel dimensionality reduction technique which reduces the dimensionality of a feature set from 100 to 3 in case of MCYT-100 and 47 to 3 in case of SVC, SUSIG datasets respectively. In addition to it, we propose a depth wise separable (DWS) convolution based OSV framework which enables one/few shot learning for test signature verification. To inspect the robustness of our proposed dimensionality reduction technique and DWS OSV framework, exhaustive experiments are conducted with three widely used datasets i.e. MCYT-100, SUSIG and SVC. We have attained state of the art EER in majority of experimentation categories compared to many recent and state-of-the art OSV models.