GAFSV-Net: A Vision Framework for Online Signature Verification
New method treats signature data as pictures, outperforming all sequence-based models.
GAFSV-Net converts online signature temporal sequences into six-channel asymmetric Gramian Angular Field images, enabling the use of pretrained 2D vision backbones (ConvNeXt-Tiny) instead of 1D architectures. A dual-branch encoder with bidirectional cross-attention processes GASF and GADF matrices representing pen speed, pressure derivative, and direction angle. Trained with semi-hard triplet loss and skilled-forgery hard-negative injection, GAFSV-Net outperformed all sequence-based baselines trained under identical objectives on DeepSignDB and BiosecurID, demonstrating that the representational gain of 2D temporal encoding is consistent and independent of training procedure.
- Converts 1D signature sequences (pen speed, pressure, direction) into 2D GASF and GADF images for vision backbone processing.
- Uses a dual-branch ConvNeXt-Tiny encoder with bidirectional cross-attention to fuse complementary temporal patterns.
- Outperforms all sequence-based baselines on DeepSignDB and BiosecurID, with ablation studies confirming the benefit of the 2D encoding approach.
Why It Matters
Enables high-accuracy signature verification with pretrained vision models, improving security in banking and authentication systems.