The Research on Multi-Granularity Writer and Writing Style Identification for Interpretability

Writer identification is a significant method in biometric identification, with extensive research conducted by numerous scholars. Currently, the predominant approach involves utilizing deep learning technologies for identification. While these methods achieve high accuracy, their black-box nature presents challenges in fields such as judiciary and finance, where security demands are stringent, leading to issues with the lack of interpretability of identification results. To address this, the following contributions are made in this study: 1. A multi-granularity data augmentation strategy is introduced, ensuring the model’s capability to perceive both global and local features. 2. A multi-task writer identification framework is designed, which can simultaneously identify the writer’s identity and style. 3. An interpretative strategy centered around the consistency of the writer’s identity and style is developed to provide interpretability for the identification results. Experiments conducted on multiple public datasets demonstrate that our approach can accurately identify the writer’s identity and style, with a degree of interpretability in the results.