Combiner apprentissage profond et connaissances sur la forme et les relations spatiales pour l'interprétation de sceaux byzantins

Byzantine seals are small circular objects attached to documents, bearing religious iconography and inscriptions that allow the identification of the senders. These seals are one of the primary sources of information for the study of the Byzantine Empire. However, since the majority of surviving seals are made of lead, they are often oxidized and corroded. Many still retain visual elements that make them interpretable, but others are so severely damaged that only experts in sigillography can interpret them. Currently, there is no computational tool to assist historians in interpreting these seals. In this thesis, we propose hybrid approaches combining multiple disciplines such as computer vision, artificial intelligence, fuzzy logic, spatial relationship modeling, and knowledge-based reasoning. The objective is to develop tools to assist historians in interpreting Byzantine seals. Three main axes, inspired by the methodological approaches of historians, are explored. 1-Automatic characterization of border deterioration. Assessing the physical condition of seals through visual examination is one of the first steps sigillographers take in their interpretive work. In this context, we explore a new characterization technique aligned with traditional analytical practices used by historians, with a particular focus on border deterioration, notably the loss of material. The proposed method includes seal segmentation, feature extraction based on the geometric properties of the contour, and a classification step using machine learning algorithms. 2- Automatic delineation of seal content. Delineating the seal's content is a preliminary step that helps focus the analysis on existing and missing design elements (due to seal deterioration). This task can be challenging, particularly when the seal is broken. To address this issue, we propose a hybrid approach combining deep learning and the shape model of a seal design pattern. This involves detecting remaining fragments of a circular border pattern that typically surrounds the seal's content and using an ellipse estimation method based on prior knowledge of the pattern's shape to delimit it.