A new feature for Chinese handwriting writer identification: Directional pixel angle

Abstract

In the task of writer identification, although deep learning methods have provided good recognition performance, the black-box nature of these methods also limits their application scope, especially in scenarios that require more stable, reliable, and interpretable algorithms. Focusing on the writing characteristics of Chinese handwriting, we have designed a new edge description feature called directional pixel angle (abbreviated as DPA). The motivation for this feature comes from forensic handwriting analysis, where the long basic strokes in Chinese characters typically have a habitual inclination angle that is highly correlated with the writer’s identity. This paper focuses on four basic strokes in Chinese characters—horizontal, vertical, left-falling, and right-falling—and calculates the relative positional relationships of the edge pixels of these strokes. This paper design and extract first-order and second-order DPA features and apply them to writer identification tasks. Experimental results show that the writer identification performance based on this feature is comparable to that of mainstream CNN models and offers better interpretability.