Matching ostraca fragments using a siamese neural network

Abstract As part of sociological studies, artifacts such as pottery ostraca from Upper Egypt are studied by egyptologists. These pottery pieces are covered with textual inscriptions many of which concern accounting and other economic and administrative matters. Thus, these writings contain pertinent information for understanding ancient Egyptian civilizations. The ostraca fragments retrieved from excavations are available in large quantities, but most of them are not yet analyzed, sorted, and reassembled by the egyptologists. We present a fragment matching approach based on pairwise local assemblies, using a 2D Siamese Neural Network to evaluate matching probabilities. This network is designed to predict simultaneously the existence or absence of a match, and the spatial relationship of one fragment in relation to the other (up, down, left, or right). We trained our deep learning model on a dataset of 6000 patches extracted from ostraca images, and achieved 96% accuracy on a validation dataset of 1000 patches. Then we propose a pipeline to reconstruct larger ostraca images using step by step pairwise matching, and able to produce a series of image reconstruction proposals. This method is based on the construction of a graph through iterative addition of small fragments. This work is intended as a proof of concept that archaeologists can benefit from automatic processing of their ostraca dataset.