BYOLSiam: Self-Supervised Pre-training for Offline Signature Verification
Abundance Udo, Jianhui Guo, Aloysious Barlea, Samia Ben Amarat · 2024
We introduce BYOLSiam, a novel approach to offline signature verification that utilizes self-supervised learning techniques to address the challenges of cross-domain generalization and the scarcity of well-annotated signature datasets. Specifically, we adapt the Bootstrap Your Own Latent (BYOL) framework to learn discriminative and generalizable representations suitable for OSV. Our approach comprises two key stages: self-supervised pre-training using BYOL and implementing a Siamese architecture dubbed BYOLSiam for the subsequent downstream OSV task. The BYOLSiam model incorporates our pre-trained BYOL encoder with a multi-layer perceptron, a layer normalization, and a flexible prediction layer which allows for both metric learning and binary classification tasks. We conducted experiments across multiple signature benchmark datasets to evaluate our model’s performance in within-domain and cross-domain scenarios. Our results demonstrate that BYOLSiam achieves state-of-the-art performance, particularly excelling in challenging cross-domain verification tasks.