Skip to Main content Skip to Navigation

Approches robustes pour la comparaison d'images et la reconnaissance d'objets

Abstract : The general topic of this dissertation is the comparison of images, which is treated both as a local and a global issue via several applications. We are first interested in object recognition task based on local descriptors. We propose a complete, robust and automatic system for multiple object recognition, which relies on two methodological approaches: the “a contrario” detection theory and the optimal mass transport theory of Monge-Kantorovich. In the latter framework, a dissimilarity measure is introduced for the comparison of SIFT-like descriptors relying on the optimal transportation cost between circular histograms (Circular Earth Mover's Distance). A matching criterion of local descriptors based on the a contrario methodology is then introduced. This criterion, which does not require any detection threshold setting nor than the usual nearest-neighbor restriction, enables multiple correspondences between images. Moreover, we propose an algorithm based on RANSAC strategy (RANdom SAmple Consensus for the grouping of local matches which handles multiple detection. It also enables us to automatically select the best geometrical model describing the object pose modification, due for instance to the viewpoint change between several images. In addition, we study the interest of the EMD (Earth Mover's Distance) for global comparison of images (image retrieval). Eventually, a regularization approach for the transportation map is proposed, inspired from the non-local filters, in the context of characteristics transfer between images by histogram specification. We show the interest of this approach for contrast modification and color transfer applications.
Complete list of metadata

Cited literature [145 references]  Display  Hide  Download
Contributor : Julien Rabin Connect in order to contact the contributor
Submitted on : Monday, April 12, 2010 - 10:33:59 AM
Last modification on : Friday, September 10, 2021 - 3:04:18 PM
Long-term archiving on: : Friday, October 19, 2012 - 1:50:14 PM


  • HAL Id : tel-00472442, version 1


Julien Rabin. Approches robustes pour la comparaison d'images et la reconnaissance d'objets. Traitement du signal et de l'image [eess.SP]. Télécom ParisTech, 2009. Français. ⟨tel-00472442⟩



Record views


Files downloads