Byzantine seals are small circular objects, attached to documents, that bear religious iconography and inscriptions identifying the senders. They are one of the main sources of information for the study of the Byzantine Empire. However, since most surviving seals are made of lead, they are often oxidized and corroded. Many still preserve visual elements that make them interpretable, but others are so severely damaged that only experts in sigillography can interpret them. Currently, no computational tool exists to assist historians in the interpretation of seals.
In this thesis, we propose hybrid approaches combining several disciplines, such as computer vision, artificial intelligence, fuzzy logic, modeling of the spatial relationships, and knowledge-based reasoning. The overall goal is to develop tools capable of assisting historians in the interpretation of Byzantine seals. Three main directions, inspired by historians’ methodological approaches, are explored.
Automatic characterization of border deterioration:
Assessing the physical condition of a seal through visual inspection is one of the first steps sigillographers take when interpreting it. In this context, we explore a new characterization technique aligned with traditional practices used by historians, with a particular focus on border deterioration, especially material loss. The proposed method includes seal segmentation, feature extraction based on geometric properties of the contour, and a classification step using machine-learning algorithms.
Automatic delimitation of the seal’s area of interest:
Delineating the seal’s area of interest is a useful preliminary step that allows the analysis to focus on both existing design elements and those that have disappeared due to deterioration. This task can be challenging, especially when the seal is deteriorated. To address this problem, we propose a hybrid approach combining deep learning and a shape model of the seal’s border motif. This involves detecting the remaining fragments of the circular border motif that typically surrounds the seal’s content and using ellipse-estimation methods based on prior knowledge of this motif’s shape to delineate it.
Analysis of the iconographic scenes on Byzantine seals Byzantine seals display a significant number of religious iconographic representations. Recognizing these scenes is central to historical interpretation. However, automating this task is difficult due to the frequent deterioration of seals. To overcome this challenge, we leverage the structured nature of these iconographic scenes and propose a loss function that incorporates structural constraints—based on expected spatial relations between iconographic elements—into the training of a neural network designed to recognize the objects depicted on the seals. These relations are modeled in a fuzzy manner to account for their imprecision and variability.