Abstract
Abstract
Reconstructing fragmented historical textiles is challenging because of irregular shapes, erosion, and missing boundary cues. We present a graph-based deep learning framework that predicts likely adjacencies between fragments, functioning as a recommendation system for virtual reconstruction. The method learns a latent graph structure from visual features and refines it using a graph autoencoder. To evaluate it, we introduce a benchmark created by virtually fragmenting a high-resolution Bayeux tapestry scan. The model performs well across puzzle sizes and remains robust under boundary erosion. Ablation studies confirm the complementary roles of the graph structure predictor and the autoencoder, and the importance of balanced negative sampling. We further show that the pipeline can adapt to a second historical textile, the Skog tapestry, through fine-tuning, giving initial evidence of cross-textile transfer. Overall, the framework offers an effective tool for experts and a foundation for learned graph reasoning in cultural heritage reconstruction.