A Writer Adaptation Approach to Online Signature Verification Through Feature Clustering and Classifier Selection

We present a novel adaptive approach to online signature verification based on writer specific feature and classifier selection. Data efficiency is a notable advantage of classical pattern recognition methods over deep learning models. Unlike deep learning, which often demands substantial volumes of labelled data to generalize effectively, classical methods can achieve satisfactory performance with smaller datasets. This characteristic proves invaluable in domains like online signature verification, where data availability is scarce or the acquisition process is cost-prohibitive. In this work, we choose feature clustering techniques specifically tailored for both writer authentication and classification based on achieving the lowest possible error rate with training signature samples. This approach diverges from conventional methods in online signature verification, which often adopt writer-independent strategies by clustering a shared set of features for all writers during verification. Similarly, our classifier selection process deviates from the norm by considering writer-dependent characteristics. To rigorously evaluate our model’s performance, we conducted exhaustive testing using the widely recognized MCYT-100 (DB1) dataset, a subset of the MCYT dataset containing signature samples from 100 writers. Our experimental results affirm the efficacy of leveraging writer-dependent features for online signature verification. Notably, our model achieved a lower error rate compared to many state-of-the-art signature verification models.