Towards interpretable Chinese writer identification: analysis and enhancement via SHAP

View at publisher Summarise with VIRAT
Authors: Zhenjiang Li, Miao Zhang
Cited by: 0
References: 47

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Text-independent offline writer identification is challenged by subtle inter-writer 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 group-level 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.

Fields of Study & Subjects

Handwritten Text Recognition Techniques Advanced Neural Network Applications User Authentication and Security Systems

Publication details

Journal / source
Engineering Research Express
Volume
8
Issue
7
Pages
075241–075241
Publication date
1 April 2026
Language
EN