An Approach for Rejoining Ancient Dunhuang Manuscripts Automatically Based on Self-Supervised Learning and Encoding with Azimuths Information

Ancient manuscripts, which document daily life, religious history, social development, and historical progress, embrace immeasurable value. However, these manuscripts often suffer damage and fragmentation due to natural degradation and improper preservation methods. Traditionally, archaeologists dedicated several years to manually reassembling these fragments, a process that is not only labor-intensive but also unsuccessful. Based on previous scholars’ work, this paper presents a new approach comprising feature extraction, data augmentation, and a self-supervised Siamese network that successfully speeds up the process of scrutinizing original fragments. Specifically, we have compiled a well-classified and fully annotated dataset of ancient Dunhuang manuscripts. Among these manuscripts, we extracted the edges of the fragments and encoded them as one-dimensional ordinal sequences using a novel method we named AZfreeman. For the first time, we propose AZfreeman, an improved Freeman encode that fuses azimuths information and can be effectively used to rejoin ancient Dunhuang fragments, which minimizes spatial overhead and enhances operability. Furthermore, we develop a data augmentation strategy capable of learning the features of fragments, which performs well with unbalanced and dynamically expanding datasets. It also brings convenience to subsequent dataset usage. Finally, we employ a self-supervised framework based on a lightweight backbone network, which incorporates attention mechanisms to extract features within sequences. Experimental results show that our approach can quickly and accurately automate the scrutinizing process of fragments, which significantly facilitates subsequent reconstruction. Our finding demonstrates the promising prospect of computer-assisted artifact restoration technology in the field of archaeology.