Fusion of Conventional and Deep Learning Methods for Offline Signature Verification

Abstract

Offline signature verification remains a critical yet challenging task in biometrics and forensic document analysis due to the complete lack of dynamic behavioral trajectories such as velocity, acceleration, and pen pressure. This paper presents a comprehensive study on the architectural paradigm that fuses conventional handcrafted feature-extraction techsniques with modern deep learning representation learning models. While conventional techniques like Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) robustly preserve exact geometric proportions and micro-textures, deep learning models like Convolutional Neural Networks (CNNs) capture highly complex, abstract structural representations. We systematically explore early feature-level fusion, late decision-level fusion, and hybrid metric learning pipelines. Our critical evaluation across benchmarks demonstrates that hybrid models dramatically mitigate the threat of skilled forgeries and generalize exceptionally well under constrained reference environments with limited training templates.