Graph Convolutional Teacher-Student Framework for Writer Inspection from Intra-variable Handwritten Words

Handwriting exhibits distinctive individual characteristics, making it a crucial biometric modality for forensic document examination and legal authentication. In this paper, we propose a graph convolutional teacher-student framework for writer inspection using single handwritten word samples. Our approach integrates both writer identification and verification networks to enhance reliability and robustness. We first train a teacher model for writer identification and then distill its knowledge into a student model for writer verification. Both models leverage Graph Convolutional Networks (GCNs) to learn discriminative writer-specific representations from handwriting graph structures. To evaluate our approach, we introduce a new dataset comprising intra-variable handwritten word samples collected intermittently over several months. This dataset contains 79 distinct English words, each written 6 times by 100 different writers, resulting in a total of 47400 handwritten word images. Additionally, we tested our model on some benchmark datasets to ensure its robustness and generalizability. Experimental results demonstrate that our method achieved promising performances in both identification and verification tasks, outperforming state-of-the-art approaches. These findings highlight the effectiveness of graph-based handwriting analysis in capturing intra-writer variations and the benefits of knowledge distillation for efficient writer inspection.