A Self-Supervised Learning Approach for Pairwise Matching of Ancient Dunhuang Fragments

In the field of ancient manuscript rejoining, particularly for Dunhuang manuscripts, traditional approaches often focus on complete manuscripts or real paired fragments while overlooking orphan fragments, which are equally critical for restoration. Due to natural degradation and improper preservation methods, most Dunhuang manuscripts survive only as fragmented remnants. Scholars have traditionally relied on prior rejoining experience to guide the restoration of orphan fragments, a process that often takes years and achieves limited success. In this study, we propose a novel self-supervised pairwise matching approach designed to assist scholars in digital fragment rejoining. We construct a dataset comprising complete manuscripts, real paired fragments, and orphan fragments and apply a detailed preprocessing scheme that integrates expert knowledge as implicit constraints without requiring manual annotations. A novel feature encoding—BFAz encoding (Binary Freeman Code with Azimuth Information)—is introduced to provide robust matching cues. To address the scarcity of positive samples and the difficulty of annotating fragment similarity, we develop a self-augmentation strategy that generates supervisory signals by learning global fragment features, forming fragment groups to establish positive and negative sample relationships. Pairwise similarity is then quantified using a Siamese network. Experimental results demonstrate that our approach efficiently prioritizes correct rejoining candidates, highlighting the potential of computer-assisted restoration technologies in archaeological research.