Abstract: Signature-based writer identification from static images is a critical biometric task, yet it traditionally ignores the rich, dynamic information of the signing process. This paper introduces DynaSigNet (Dynamic Signature Network), a novel deep learning framework that overcomes this limitation by learning to infer the latent generative dynamics—trajectory, velocity, and pressure—from a single offline signature. Our architecture employs a synergistic CNN-Transformer backbone to produce a holistic feature representation that captures both local textures and global structural dependencies. The core of our method is a multi-task learning paradigm where this shared representation is decoded in parallel: a primary head for writer classification, and auxiliary motor-imagination heads for predicting the dynamic properties. By training the model to understand how the signature was formed, we achieve a more discriminative feature space. Validated on a challenging bilingual dataset of Roman and Bengali scripts, DynaSigNet achieves state-of-the-art performance, demonstrating that inferring latent motor processes is a powerful new direction for static biometric analysis. A system graphical user interface (GUI) is also implemented.

Keywords: Signature identification, Deep learning, Visual transformer, Latent Dynamics, Biometrics, Dynamic Signature Network, Multi-script recognition

Pages: 368-381