Facsimiles-based deep learning for matching relief-printed decorations on medieval ceramic sherds

In this paper, we addressed the problem faced by archaeologists in associating relief-printed decorations on ceramic objects discovered during excavations carried out with the same wheel. This is crucial to understand the trade networks between regions, but highly complex and time-consuming task. We used two approaches: supervised classification or unsupervised clustering of 2D relief views generated from 3D scans of ceramic sherds. Inspired by experimental archaeology, we created wheel facsimiles to supplement significantly the database with numerous plausible and clearly identified samples. Taking advantage of the powerful convolutional neural network EfficientNet to extract reliable discriminating features, experimental results show that the facsimiles significantly improve the networks' training to achieve a classification accuracy exceeding 95% on real sherds. On the other hand, unsupervised spectral clustering from a vector reduced to a few hundred of the most significant features delivered by the network EfficientNet-B5 trained on ImageNet, without any fine-tuning, achieves an accuracy of 77.47% on our database. These results validate the strategy of using facsimiles to supplement a too-small data set and are very promising for the development of a computer-assisted archaeology tool for pattern-wheel association.