Évènements

Soutenance de thèse de Payam Birjandi Vendredi 16 septembre à 10H30, Amphi B310 Télécom ParisTech -- 46, rue Barrault -- 75013 Paris

Modélisation et description des images satellite haute résolution :
Des algorithmes a la base de l'analyse en composantes indépendantes

Auteur

Payam Birjandi.

Date

Vendredi 16 septembre 2011 à 10H30.

Lieu

Télécom ParisTech -- Site Barrault -- Amphi B310.

Directeur(s) de thèse

• Mihai Datcu.

Membres du jury

Rapporteurs

• Inge Gavat,
• Philippe Bolon.

Examinateurs

• Mohammad Ali Djafari,
• Tullio Tanzi,
• Michel Roux,
• Alain Giros.

Abstract

The main purpose of the thesis is to define descriptors for Sub-meter resolution satellite images especially for those who contain geometrical or man-made structures. Independent Component Analysis (ICA) is a good candidate for this purpose, since previous studies demonstrated that the resulted basis vectors contain some small lines and edges.

As a basic analysis, a study about scale size and dimensionality behavior of ICA components for satellite image indexing is presented and the optimum dimensionality and scale size are found.

There are two points of view for feature extraction using ICA. The usual approach is to use the ICA coefficients (ICA sources) and the other is to use the ICA basis vectors related to every image. From the first point of view, an ordinary ICA source based approach is proposed for feature extraction. Then this approach is developed and generalized as Topographic ICA method to extract middle level features which leads to a significant improvement in the results.

From the other point of view, two methods are proposed. One of them uses the Bag of words idea which considers the basis vectors as the visual word. Second method uses the line properties inside the basis vectors to extract features. Also, using the line properties idea, another method is developed which detect the line segments directly in the images.

In addition, the capabilities of proposed descriptors are compared through a supervised classification which is based on the Super Vector Machine (SVM).