Skip to Main content Skip to Navigation
Theses

Towards unconstrained face recognition from one sample

Abstract : Although having been an active research topic for 30 years, recognizing a person from surveillance having seen only one image is unsolved. Within this context, the two greatest challenges are the variations of pose and illumination. Moreover, there are strict constraints upon the complexity in both terms of computational time and stockage requirements. The work developed throughout this dissertation gives several advantages in the context of real-time and unconstrained face recognition. Firstly, an illumination normalization method simulating the performance of human retina is proposed as preprocessing algorithm. Secondly, we propose novel features called POEM (Patterns of Oriented Edge Magnitudes) for representing a local image structure. This descriptor is discriminative and robust to exterior variations (variations of pose, illumination, expression and pose that we always see when dealing with face images). Thirdly, a statistical model for robust face recognition across poses, entered on modeling how facial patch appearance changes as the viewpoint varies, is proposed. Finally, a novel approach modeling the spatial relationships between face components is developed. Except the last algorithm, all proposed methods are very fast and are therefore suitable for the constraints upon real-time of surveillance applications.
Document type :
Theses
Complete list of metadatas

Cited literature [149 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00574547
Contributor : Ngoc Son Vu <>
Submitted on : Tuesday, March 8, 2011 - 11:50:49 AM
Last modification on : Thursday, November 19, 2020 - 12:59:52 PM
Long-term archiving on: : Thursday, June 9, 2011 - 2:36:13 AM

Identifiers

  • HAL Id : tel-00574547, version 1

Collections

Citation

Ngoc-Son Vu. Towards unconstrained face recognition from one sample. Human-Computer Interaction [cs.HC]. Institut National Polytechnique de Grenoble - INPG, 2010. English. ⟨tel-00574547⟩

Share

Metrics

Record views

700

Files downloads

2975