Hierarchical offline-to-online Chinese handwriting trajectory reconstruction via dual-domain priors

Offline-to-online (O2O) handwriting trajectory reconstruction aims to recover online trajectories from static glyph images, supporting applications such as document recognition and human-computer interaction. Existing methods often model the input image holistically and rely mainly on geometric cues, leading to structurally incorrect strokes, implausible motion patterns, and limited generalization. These limitations are particularly pronounced for Chinese characters with complex structures, dense strokes, and background interference. To address these issues, we propose a knowledge-guided hierarchical trajectory reconstruction framework that reconstructs online trajectories that are visually consistent and aligned with human writing patterns. The framework consists of three parts. The structure-aware encoder (SAE) integrates frequency-domain enhancement with dual-branch deformable spatial attention to extract stroke-structure representations that are robust to background interference. The spatial layout network (SLN) predicts stroke regions and writing order to provide explicit layout constraints, while the hierarchical vector decoder (H-VD) generates continuous, smooth, and editable trajectories in a parametric space. In addition, dual-domain contrastive priors align generated outputs with image and writing regularities under limited supervision, thereby improving generation stability. Comprehensive experiments and user evaluations show that our method achieves competitive or superior performance over representative baselines across quantitative and perceptual metrics. The reconstructed trajectories can further support downstream tasks such as data augmentation, handwriting imitation, and style transfer.

Text 128116.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (2MB) Official URL or Download Paper: https://link.springer.com/10.1007/s44443-026-01050...

Actions (login required)

View Item