Automatic Mosaic Digitalization: a Deep Learning approach to tessera segmentation

A mosaic, made of small coloured tiles, called tesserae or tessellas, is a form of decorative art to create images and patterns on a surface. One of the first steps to obtain digital mosaics is the segmentation process to separate (identify) tesserae. Then, various tools are necessary to catalog the digitalized scenes, to extract the figures, such as animal or human beings, from the puzzle of small pieces, to geolocalize the segmented objects and to assign a semantic meaning to them. While some mosaic segmentation approaches have already been reported in the literature, currently, mosaic segmentation is still done by human operators, resulting time consuming and error prone. We propose in this paper an automatic approach, based on Deep Learning, to segment the mosaic floor tesserae of the church of S. Stephen in Umm ar Rasas. Our approach allows to obtain automatically and with good reliability the description of the main elements of a mosaic (the tesserae) that are not homogeneous. The experiments performed on the collected tesserae dataset yield high accuracies and demonstrate the effectiveness and suitability of our approach.