Place: Large Lecture Room
Affiliation: Université de Rouen, France
Graphs are data structures which allow an analytic description of complex data. A scene or an object can be described by a graph whose vertices are devoted to the description of parts (themselves can be described by subdivisions…) and whose edges are used to represent relationships between subparts at different depths. Edges may include, as is the case for the adjacency graphs, the description of neighboring relations, or inclusion/composition relations. Vertices and edges can carry complex labels (nominal labels, numerical vector) describing the intrinsic characteristics of the subpart described in the case of vertices, or the relation in the case of eges. Because of their representation power, graphs are often used to represent complex data (social networks, chemical molecules, 2D or 3Dscenes… ). In pattern recognition, the use of such data structures has long been hampered by the algorithmic complexity error-tolerant methods imposed by the presence of noise in the data. In recent years, the increasing computing power of machines combined with the proposal of new robust methods tends to break down these barriers. In this presentation, we present an overview of the work carried out in recent years at the LITIS on the topics related to structural methods for pattern recognition . After a general introduction of the main issues, we will detail some approaches developed for graph classification, graph indexing and subgraph isomorphism in document image analysis. We will also present the tools and databases that we make available to the community.