Beyond relighting: RTI for clustering fragmented heritage textiles using deep learning

Abstract This paper introduces a computational framework for clustering and visualizing textile fragment relationships using Reflectance Transformation Imaging (RTI). Our approach leverages deep learning models and utilizes the Hemispherical Harmonics (HSH) to extract discriminative features. The RTI data is first modeled using HSH before feature extraction by the deep feature extractor. The feature vectors are visualized through dimensionality reduction techniques, which help reveal the relationship between fragments by clustering them. We tested the proposed framework on the Oseberg textile collection (an open archeological artifact assembly problem) and a control dataset of Polish Dragoons textiles, demonstrating that the algorithm achieves good intra-class similarity and inter-class separation, distinguishing different textiles. RTI-based framework achieves higher clustering and dimensionality reduction evaluation scores between related fragments than RGB photography. The results confirm RTI’s potential as a data-rich, non-destructive imaging technique for supporting archeological reconstruction.