Face recognition (FR) offers unmatched advantages as compared to other biometrics, such as easy access or needless explicit cooperation from users, and today, it has attained the reliability and the maturity required by real applications . With recent enormous developments, both academic and industrial research are focusing more and more on unconstrained real-scene face images in order to further extend the application field of FR while keeping its reliability as compared to constrained user cooperative conditions.
Reliable FR in unconstrained conditions needs to handle face images which are taken under various scenarios, notably with respect to uneven illumination environments, large facial expression changes, arbitrary head poses and ageing. In this thesis, the aim is to address one of these issues, i.e., lighting condition changes.
Hypothesis under investigation and main aims
Lighting variations impinge upon many conventional FR algorithms which assume a normalized lighting condition because the lighting conditions may vary across not only the intensity and direction of the lighting sources, but also the color of light. The most challenging part of this thesis, which is also its attractiveness, lies in the realization of an illumination-invariant FR system which will integrate the lighting normalization module.
Specifically, the aim of this thesis is to raise some constraints on the existing 2D FR solutions thanks to 3D, thereby widening their range of application fields. These applications can include for instance FR dealing with non-ideal imaging environment where users may present their face not with a neutral lighting (e.g. side lighting), or even FR using images from video surveillance which can gather all the difficulties such as low resolution images, pose changes, lighting condition variations, occlusions, etc.
In this thesis, the candidate will investigate the possible contribution of 3-D to improve performances of authentication while keeping existing advantages of face recognition from 2-D images. It has been shown that using a 3D model of human face does improve 2D face recognition robustness to illumination and pose variation [2,3,4]. Our goal is to further explore the application and significance of 3D information for illumination processing.
Meanwhile, we propose to prioritize all these difficulties and investigate first the problem of lighting modeling by introducing 3D statistical model as a mathematical/physical explanation, then the focus will be upon the lighting normalization (i.e. delighting), ultimately the research is supposed to be finalized with illumination recovery (i.e. relighting), both delighting and relighting techniques could demonstrate their competitive ability to handle with FR suffering from lighting issues.