SignForensics: A Robust Framework for Forensic Offline Signature Verification With Enhanced Detection, Noise Removal, and Multi-Stage Authentication

Reliable signature authentication is essential across fields where identity verification is critical. Despite advances in machine learning, validating handwritten signatures remains challenging due to inherent variability in individual writing styles. Most current methods focus exclusively on verification, assuming signatures are already isolated and noise-free—an assumption rarely met in real-world scenarios. In this paper, we present a comprehensive framework called SignForensics for robust signature authentication, directly extracting and refining signatures from real-world documents. Our system employs YOLOv10 for signature detection and extraction, and leverages CycleGAN to effectively remove residual visual noise, producing noise-free samples optimized for downstream analysis. To evaluate authenticity, we developed specialized SigNet and CapsNet neural networks, achieving strong generalization and accuracy. Our method outperforms current state-of-the-art techniques, reaching 100% accuracy on the CEDAR dataset, 94.79% on the HINDI dataset, and 96.52% on the BENGALI dataset, remaining robust across writer-independent settings and previously unseen samples.