Skip to Main content Skip to Navigation

Nouvelle approche d'identification dans les bases de données biométriques basée sur une classification non supervisée

Abstract : The work done in the framework of this thesis deal with the automatic faces identification in databases of digital images. The goal is to simplify biometric identification process that is seeking the query identity among all identities enrolled in the database, also called gallery. Indeed, the classical identification scheme is complex and requires large computational time especially in the case of large biometric databases. The original process that we propose here aims to reduce the complexity and to improve the computing time and the identification rate performances. In this biometric context, we proposed an unsupervised classification or clustering of facial images in order to partition the enrolled database into several coherent and well discriminated subsets. In fact, the clustering algorithm aims to extract, for each face, a specific set of descriptors, called signature. Three facial representation techniques have been developed in order to extract different and complementary information which describe the human face: two factorial methods of multidimensional analysis and data projection (namely called "Eigenfaces" and "Fisherfaces") and a method of extracting geometric Zernike moments. On the basis of the different signatures obtained for each face, several clustering methods are used in competing way in order to achieve the optimal classification which leads to a greater reduction of the gallery. We used either "mobile centers" methods type such as the K-means algorithm of MacQueen and that of Forgy, and the "agglomerative" method of BIRCH. Based on the dependency of the generated partitions, these different classifying strategies are then combined using a parallel architecture in order to maximize the reduction of the search space to the smallest subset of the database. The retained clusters in fine are those which contain the query identity with an almost certain probability.
Document type :
Complete list of metadatas

Cited literature [221 references]  Display  Hide  Download
Contributor : Frédéric Davesne <>
Submitted on : Tuesday, December 21, 2010 - 7:23:54 PM
Last modification on : Wednesday, July 8, 2020 - 9:59:10 AM
Long-term archiving on: : Tuesday, March 22, 2011 - 3:12:12 AM


  • HAL Id : tel-00549395, version 1



Anis Chaari. Nouvelle approche d'identification dans les bases de données biométriques basée sur une classification non supervisée. Modélisation et simulation. Université d'Evry-Val d'Essonne, 2009. Français. ⟨tel-00549395⟩



Record views


Files downloads