Scribe Identification for Tang Dynasty Changsha Kiln Poetic Ceramics Using a Dual-Path Multi-Scale Global Attention Model

The poetic ceramics of the Changsha Kiln during the Tang Dynasty are an important part of China’s cultural heritage, known for their special calligraphy styles and abundant historical and cultural connotations. Until now, Studies have mainly concentrated on the linguistic content of these inscriptions, while the identification and study of the individual scribes has not yet been explored. There is also no publicly available dataset of Changsha Kiln calligraphy characters. Experts specializing in ancient handwriting are primarily consulted in conventional methods of investigation. They are time-consuming, difficult to scale, and have no standard, objective metrics due to the inherent differences in expert judgment. To overcome these limitations, we have assembled a dataset of Changsha Kiln calligraphic characters from the Tang dynasty. The dataset contains 1,865 character images (160×150) from 135 different ceramic objects with poetic texts. We suggest a new Dual outwardly dual-path network with Multi-Scale Global Attention (MSGA) module for scribble identification and categorization. The proposed model takes features from character pairs and compares them for scribe attribution. The core framework utilizes two weight-sharing paths. In each path, ResNet-34 is utilized as the backbone feature extraction network. Crucially, we integrated an MSGA module into this backbone. The MSGA module applies convolutional kernels of three different scales simultaneously and fuses multi-scale results. It computes attention distributions at each scale to model long-distance dependencies between each pixel which enhances the model's ability to capture the relationship between local details and global features and enhances the model's sensitivity to subtle handwriting characteristics. Experimental results show that using a standard Non-Local attention module achieved a precision of 93.75%. The proposed MSGA module significantly increased precision to 97.85%. Our application which we use in classifying Changsha Kiln ceramics, further revealed that two inscribed ceramic pieces, housed in separate museums, were written by the same scribe. The proposed technique identifies variations between scribes of poetic inscriptions from the Tang Dynasty Changsha kiln as well as differentiates different calligraphic styles. This paper proposes a new quantitative approach for analysing Changsha Kiln calligraphy styles, which can offer a valid objective way to identify scribes.