A HYBRID CNN TRANSFORMER FRAMEWORK FOR OFFLINE HANDWRITTEN SIGNATURE VERIFICATION

ABSTRACT

Aim: Offline handwritten signature verification remains a widely used biometric authentication method, yet Convolutional Neural Network (CNN)-based approaches, while effective at extracting local features, are limited in modelling global structural dependencies, which constrains their resilience to skilled forgery. This study proposes and evaluates HMT-Net, a hybrid CNN Transformer architecture that combines CNN-based local feature extraction with a Transformer's capacity for long-range dependency modelling, with the aim of improving verification accuracy and forgery resistance relative to single-architecture baselines.

Methods: HMT-Net uses a ResNet-18 backbone for spatial feature extraction; the resulting feature maps are flattened into patch embeddings, combined with positional encoding, and passed through a stack of Transformer encoder blocks employing multi-head self-attention. Classification is performed via global average pooling followed by fully connected layers trained with binary cross-entropy loss. The model was trained and evaluated on the GPDS-960 offline signature benchmark (24 genuine and 30 forged signatures per individual) using an 80:10:10 training validation test split, and benchmarked against a CNN-only model, a Transformer-only model, and a previously reported hybrid architecture

Keywords: Offline Signature Verification; Hybrid Deep Learning; Convolutional Neural Network; Transformer; Biometric Authentication.