Abstract Tang Dynasty Changsha Kiln ceramics are iconic cultural heritage featuring unique calligraphy. Current research focuses on textual analysis, neglecting scribe identification due to the absence of public datasets and limitations of expert-dependent methods. We address this gap by constructing the first comprehensive dataset of 1865 character images from 135 artifacts. To enable automated scribe attribution, we develop a dual-path convolutional neural network integrating a multi-scale global attention (MSGA) module, enhancing feature perception via multi-scale fusion and attention. Experiments show MSGA achieves 97.85% precision, significantly outperforming non-local attention baselines (93.75%). Applying our model to museum collections revealed two ceramics in separate institutions that originate from the same scribe—a finding undetected conventionally. This work establishes a quantitative framework for calligraphic style discrimination and ancient handwriting attribution, transforming historical calligraphy study through deep learning methods.
Jiang et al. (Wed,) studied this question.