Towards Interpretable Chinese Writer Identification: Analysis and Enhancement via SHAP

Abstract Text-independent offline writer identification is challenged by subtle interwriter variations and diverse handwriting styles. While Convolutional Neural Networks (CNNs) improve performance by learning hierarchical features, they often lack interpretability and robustness. We analyze a ResNet-based baseline using SHapley Additive exPlanations (SHAP), revealing its over-reliance on limited distinctive characters and insufficient modeling of local and global patterns. To overcome these limitations, we propose TAP-ResNet, which integrates Triple Attention and Pyramid Pooling modules to refine multi-dimensional feature representations and aggregate multi-scale contextual information. Experiments on HWDB series demonstrate that TAP-ResNet outperforms existing methods in Top-1 accuracy, F1 score, and grouplevel consistency. SHAP visualizations confirm enhanced interpretability, showing the model captures both fine-grained strokes and global handwriting traits. Our framework provides a robust and transparent approach for writer identification, advancing both accuracy and explainability in practical applications.