Évènements

Ph.D. Defense of Pierre Blanchart Monday, September 26 at 1:30 pm, Room Grenat Télécom ParisTech -- 46, rue Barrault -- 75013 Paris

Fast learning methods adapted to the user specificities:
application to earth observation image information mining

Author

Pierre Blanchart.

Date and time

Monday, September 26, 2011 at 1:30 pm.

Location

Télécom ParisTech -- Site Barrault -- Room Grenat.

Ph.D. Thesis Directors

• Mihai Datcu (DLR),
• Marin Ferecatu (CNAM).

Thesis board

Reporters

• Constantin Vertan (Politehnica University of Bucharest),
• Chi-Ren Shyu (Missouri University Informatics Institute).

Examiners

• Nozha Boujemaa (INRIA),
• Isabelle Bloch (Télécom ParisTech),
• Arnold Smeulders (University of Amsterdam),
• Alain Giros (CNES).

Abstract

An important emerging topic in satellite image content extraction and classification is building retrieval systems that automatically learn high-level semantic interpretations from images, possibly under the direct supervision of the user. In this thesis, we envisage successively the two very broad categories of auto-annotation systems and interactive image search engine to propose our own solutions to the recurring problem of learning from small and non-exhaustive training datasets and of generalizing over a very high-volume of unlabeled data. In our first contribution, we look into the problem of exploiting the huge volume of unlabeled data to discover "unknown" semantic structures, that is, semantic classes which are not represented in the training dataset. We propose a semi-supervised algorithm able to build an auto-annotation model over non-exhaustive training datasets and to point out to the user new interesting semantic structures in the purpose of guiding him in his database exploration task. In our second contribution, we envisage the problem of speeding up the learning in interactive image search engines. Minimizing the number of iterations in the relevance feedback loop is indeed a crucial issue to build systems which are well-adapted to a human user. With this purpose in mind, we derive a semi-supervised active learning algorithm which exploits the intrinsic data distribution to achieve faster identification of the target category. In our last contribution, we describe a cascaded active learning strategy to retrieve objects in large satellite image scenes. Ensuring fast exchanges between the user and the system is indeed a crucial issue inherent to any interactive image search engine. We propose consequently an active learning method which exploits a coarse-to-fine scheme to avoid the computational overload inherent to multiple evaluations of the decision function of complex classifiers such as needed to retrieve complex object classes. We assess each time our results on Spot5 and QuickBird panchromatic imagery and we show that the methods we propose significantly outperform state-of-the-art techniques while adding interesting new features such as the "unknown" semantic structures discovery feature in the auto-annotation case or the interactive search scheme in the object retrieval part.