Page de Frédéric SUARD

20e colloque GRETSI sur le traitement du signal et des images 8-11 septembre 2009, Dijon, France. http://www.gretsi2009.org

Abstract This paper presents an extension of multiple kernels like Kernel Basis to the Relevance Vector Machine algorithm. The framework of kernel machines has been a source of many works concerning the merge of various kernels to build the solution. Within these approaches, Kernel Basis is able to combine both local and global kernels. The interest of such approach resides in the ability to deal with a large kind of tasks in the field of model selection, for example the feature selection. We propose here an application of RVM-KB to a feature selection problem, for which all data are decomposed into a set of kernels so that all points of the learning set correspond to a single feature of one data. The final result is the selection of the main features through the relevance vectors selection.

Résumé – Nous présentons ici une adaptation des noyaux multiples de type Kernel Basis à l'algorithme Relevance Vector Machine. L'intérêt du Kernel Basis réside dans la capacité d'adapter des noyaux globaux et locaux dans une même solution. La finalité consiste en effet à affecter aux vecteurs de la solution un ensemble de noyaux qui soit spécifique à chaque vecteur. Nous proposons d'utiliser cette approche pour un problème de sélection de variables, afin de choisir pour chaque vecteur les variables les plus pertinentes. Les performances obtenues sur les résultats préliminaires et comparées avec une approche de noyau multiple composite, sont très prometteuses et ouvrent de nouvelles perspectives.

@inproceedings{fsuardgretsi09,
author = {Frédéric Suard and David Mercier},
title = {Application des noyaux multiples de type Kernel Basis à la méthode Relevance Vector Machine pour la sélection de modèles},
booktitle = {GRETSI09, Dijon, France},
year = {2009},
pages = {4p},}

21e colloque GRETSI sur le traitement du signal et des images 11-14 septembre 2007, Troyes, France www.gretsi2007.org

Abstract This paper presents a pedestrian detection method based on the multiple kernel framework. One main problematic of pattern recognition resides in the pertinent characterization of the data. Depending on the descriptor, we sometimes have to tune the descriptor in order to be more efficient. Instead of accomplishing this tuning manually by testing and comparing all possible values we propose here to use the multiple kernel framework. The aim is to use a kernel as a linear combination of different kernels in order to combine and select automatically the best kernels within a set of kernels. This can be assimilated as model selection, where one kernel of the set corresponds to one model. We first introduce the MKL framework and finally apply this approach for a parameter tuning task and a feature selection problem.

Intelligent Vehicle Symposium 13-15 June 2007, Istanbul, Turkey http://www.iv2007.itu.edu.tr/

Abstract This paper presents a pedestrian detection method based on the multiple kernel framework. This approach enables us to select and combine different kinds of image representations. The combination is done through a linear combination of kernels, weighted according to the relevance of kernels. After having presented some descriptors and detailed the multiple kernel framework, we propose three different applications concerning combination of representations, automatic parameters setting and feature selection. We then show that the MKL framework enable us to apply a model selection and improve the performance.

International Conference on Image Analysis and Recognition 18-20 septembre 2006, Povoa de Varzim, Portugal [http://www.iciar.uwaterloo.ca/iciar06/index.php |http://www.iciar.uwaterloo.ca/iciar06/index.php|fr] Abstract : This paper presents a method for object categorization. This problem is difficult and can be solved by combining different information sources such as shape or appearance. In this paper, we aim at performing object recognition by mixing kernels obtained from different cues. Our method is based on two complementary descriptions of an object. First, we describe its shape thanks to labeled graphs. This graph is obtained from morphological skeleton, extracted from the binary mask of the object image. The second description uses histograms of oriented gradients which aim at capturing objects appearance. The histogram descriptor is obtained by computing local histograms over the complete image of the object. These two descriptions are combined using a kernel product. Our approach has been validated on the ETH80 database which is composed of 3280 images gathered in 8 classes. The results we achieved show that this method can be very efficient.

@inproceedings{fsuardiciar06,
author = {Frédéric Suard and Alain Rakotomamonjy and Abdelaziz Bensrhair},
title = {Object Categorization using Kernels combining Graphs and Histogram of Gradients},
booktitle = {International Conference on Image Analysis and Recognition, Póvoa de Varzim, Portugal},
year = {2006},
month = {September},
pages = {23--34},}

20e colloque GRETSI sur le traitement du signal et des images 6-9 septembre 2005, Louvain-La-Neuve, Belgique. http://www.gretsi2005.org

Abstract This article presents a novel method concerning pedestrian detection, thanks to graph kernels. Nowadays, the pedestrian detection is a hard task, due to the variability of its shape : size and posture. To address this problem, we choose to transform a pedestrian into a graph. The aim of this method consists of extracting a graph from each object (pedestrian or non-pedestrian), contained in a database. We compute the kernel with the inner product between graphs in order to apply a supervised classifier, here the SVMs (Support Vector Machine). We applied this method on a real images database in order to test its efficiency, particularly for scale invariance, and we obtained a good classification rate.

Résumé – Cet article présente une méthode concernant la reconnaissance de piétons à l’aide de graphes et de méthodes à noyaux. La détection du piéton est limitée à cause de la grande variabilité de la forme du piéton : taille, posture. Pour surmonter ce problème, nous avons choisi de le représenter à l’aide d’un graphe. Le but de la méthode est d’extraire le graphe de chaque objet (piétons ou non-piétons) présent dans une base d’images et de calculer un noyau à partir de ces graphes afin d’effectuer un apprentissage supervisé basé sur les SVMs (Séparateurs à Vaste Marge). L’application sur une base d’images réelles nous permet de démontrer l’efficacité de cette méthode, au niveau des invariances en échelle, avec un bon taux de reconnaissance.

@inproceedings{fsuardgretsi05,
author = {Frédéric Suard and Alain Rakotomamonjy and Abdelaziz Bensrhair},
title = {Détection de piétons par stéréovision et noyaux de graphes},
booktitle = {GRETSI05, Louvain-la-Neuve, Belgique},
year = {2005},
pages = {686-686},}

Congré Francophone de Reconnaissance de Formes et d'Intelligence Artificielle 2004 28 - 30 Janvier 2004, Toulouse, France http://www.laas.fr/rfia2004/

@inproceedings{rakotoRfia04,
author = {Alain Rakotomamonjy and Frédéric Suard},
title = {Sélection de variables par SVM: application à la détection de piétons},
year = {2004},
booktitle = {RFIA04},}'