A Unified ResNet18-Based Approach for Offline Signature Classification and Verification Across Multilingual Datasets

This paper introduces a novel ResNet18-based unified framework capable of performing both offline handwritten signature classification and writer-independent verification. Unlike traditional approaches relying on metric learning or Siamese architectures, our model uses a simple ResNet18 classifier trained with only three genuine samples per user. The model’s embeddings from the classification head are repurposed for verification through cosine similarity. Evaluated on a multilingual dataset comprising 430 users and over 17,000 unseen signatures, the proposed system achieves 93.85% classification accuracy and 90.07% verification accuracy. The approach demonstrates competitive performance relative to specialized methods, offering substantial simplicity, scalability, and robust generalization to new users and diverse scripts.