header image
Home arrow Topic arrow Ph.D and HDR theses
Ph.D and HDR theses PDF Print E-mail
Written by Romain Raveaux   
Jul 18, 2007 at 11:21 PM

Habilitation (HDR) thesis in computer science:

Title: Contributions and perspectives on combinatorial optimization and machine learning for graph matching and classification.

Draft version of my manuscrit (PDF file)

HDR defense. The presentation slides (PDF file)

Ph.D thesis in computer science:

Graph Mining and Graph Classification : Application to cadastral map analysis.

This thesis tackles the problem of technical document interpretation applied to ancient and colored cadastral maps. This subject is on the crossroad of different fields like signal or image processing, pattern recognition, artificial intelligence, man-machine interaction and knowledge engineering. Indeed, each of these different fields can contribute to build a reliable and efficient document interpretation device. This thesis points out the necessities and importance of dedicated services oriented to historical documents and a related project named ALPAGE. Subsequently, the main focus of this work: Content-Based Map Retrieval within an ancient collection of color cadastral maps is introduced. The organization of this thesis paper is in five chapters.

My Phd thesis presentation (Video)

PhD defense. The presentation slides (PDF file)

My Phd thesis in short (PDF file)

Draft version of my manuscrit (PDF file)

Abstract: ALPAGE Project

Technical documents have a strategic role in numerous organisations, composing somehow a graphic representation of their heritage. In the context of a project named “ALPAGE”, a closer look is given to ancient French cadastral maps related to the Parisian urban space during the 19th century. Hence, the data collection is made up of 1100 images issued from the digitalization of Atlas books. Each image contains a vast number of domain dependent objects, ie. parcels, water collection points, stairs, windows/doors… From a computer sciences point of view, the challenge consists in the extraction of information from colour images in the objective of providing a vector layer to be inserted in a Geographical Information System (GIS). This raster to vector conversion requires a priori knowledge to control the quality of the vectorization and to adjust algorithm parameters. The constraints given by historians are translated into a computer sciences paradigm called Attributed Relational Graphs (ARG). From a raw colour image segmentation, the system iterates to fit at best a semantic graph. Such a minimization process involves three types of algorithms: a classification stage where image regions are categorized into nominal variables; a graph matching method in order to compare a user-defined model and computer generated graphs; and finally, a relevancy feedback scheme to infer local changes into the segmentation process. Our contribution can be easily identified as a model driven image segmentation.

Alpage Project : http://alpageproject.free.fr
Last Updated ( Jul 17, 2019 at 03:58 PM )