Handwriting Trajectory Recovery Via Trajectory Transformer With Global Radical Context-Aware Module

Junxiang Lin, Zhounan Chen, Lingyu Liang, Wenjie Peng, Shuangping Huang

Abstract: Handwriting trajectory recovery aims to reconstruct the writing trajectories from images of handwritten characters, holding applications in various areas such as text recognition, signature authentication, and forensic handwriting analysis. Current methods commonly used CNNs for feature extraction from character images, excelling in local feature identification but lacking in capturing global handwriting structures. Subsequent trajectory generation is typically handled by RNNs, which model the temporal sequence of writing but are hindered by gradient vanishing issues, leading to accuracy reduction in longer sequences. In this paper, we propose a Trajectory Transformer with a Global Radical Context-Aware (GRCA) module to realize precise trajectory recovery by analyzing intricate structural relationships in handwriting characters and modeling contextual correlations within trajectory sequences. Concretely, the GRCA module utilizes dilated convolutions to extract character radical features across various scales and perceives the structural associations embedded within the multi-scale features. Additionally, we introduce a Transformer to capture the contextual correlations among trajectory sequences, thus alleviating the issue of trajectory drift. Experiment results show that our proposed Trajectory Transformer achieves state-of-the-art performance on four benchmark datasets.