Fuzzy Spatial Relations and Deep Learning for Recognizing Iconographic Elements on Byzantine Seals

While scene recognition on Byzantine seals is a key element of historical interpretation, automating this task is made difficult by seal degradation and the variability of iconographic elements. However, most of the scenes depicted are structured, with a spatial arrangement of objects that is stable across a wide variety of seals. We propose to exploit this structure by modeling the expected spatial relationships between objects in a fuzzy set framework to take into account both their imprecision and variability. This knowledge is then used in a cost function, optimized by a neural network, to recognize each of the objects represented on the seals.