Multimedia Systems Computer Networks and Communications
Recognizing ancient Chinese handwriting of limited samples and tiny variations by data augmentation and contrastive learning
Abstract
Recognizing ancient Chinese handwriting helps people understand historical texts. However, recognizing ancient Chinese handwriting is challenging due to limited samples, subtle differences between distinct characters, and the lack of a standard form. To address these challenges, we propose a few-shot learning framework called ACHR for ancient Chinese handwriting recognition. In particular, to address limited sample availability, we combine few-shot learning with an augmentation strategy designed to increase sample diversity while preserving character structure. To distinguish visually similar characters, we use contrastive learning and self-attention to enhance feature discrimination. To address the lack of a standard form, we combine global and local features to capture character structure and consider changes in component positions associated with character evolution. We construct the Chinese Character Evolution Dataset (CCED), which includes different historical forms of the same character within each class. Experiments on CCED and the publicly available Oracle-241 dataset demonstrate the effectiveness of our method. In particular, our method achieves accuracy improvements of 1.2% and 3.3% over the strongest baselines on CCED for the 1-shot and 5-shot tasks, respectively.