Fine-tuning AI improves reconstruction of ancient textile fragments

Researchers have developed a new deep learning benchmark to help reconstruct fragmented historical textiles, such as those found in the 834 CE Oseberg ship burial in Norway.

The TEXMET dataset contains 18,644 historical textile images sourced from the Metropolitan Museum of Art Open Access collection. This represents the first textile-specific benchmark for puzzle-solving research in computational archaeology.

By testing different models, researchers found that transformer-based models trained on standard images do not work well on textiles unless they undergo domain-specific fine-tuning. The graph-based DiffAssemble model showed the best results when adjusted for textile data, reaching 85.89% accuracy. When this model was applied to 45 highly degraded fragments from the Oseberg ship burial, the reconstruction accuracy jumped from 26.67% to 64.44%. These findings suggest that specialized training is necessary to use modern artificial intelligence on real archaeological materials.