Authors: Anouar Mellakh
Keywords: Biometrics, Face Recognition, Illumination, Photometric Normalisation, Gabor filters, Quality.
Abstract: Nowadays, the algorithms of face recognition, proposed in the literature, reached a correct performance level when the acquisition's conditions for the tested images are controlled, but this performances fall when these conditions degraded. The controlled conditions of acquisition correspond to a good balance of illumination, as well as a high-resolution and a maximum sharpness of the face image.
In order to determine the problem of degradation of performances under difficult capture's conditions and also to propose adapted solutions, we carried out several studies at various levels of the recognition's chain. These studies relate to the behavior of the algorithms based on global approaches. They also relate to the various methods of photometric standardization as well as strategies of recognition based on the quality of the face images.
The solutions suggested on each level of this chain resulted in a significant improvement of the performances compared to the traditional approaches. For the recognition algorithms, we proposed to fuse the phase and magnitude of Gabor's representations of the face as a new representation, in the place of the raster image. Although the Gabor representations were largely used, particularly in the algorithms based on global approaches, the Gabor phase was never exploited. We explain in this thesis the problems involved in the use of this phase and we propose a solution to solve this problem. Various methods of photometric normalization for face were studied and compared. We, thereafter, proposed a new approach of normalization based on the correction of the brightness component. Lastly, we presented a strategy of recognition based on the quality measure of face. This measurement is a fusion of several quality standards and according to our experiments ; this strategy offers an improvement of the verification rate compared to the classical methods.
The various studies, the validation of our quality measurements as well as the validation of the recognition strategy were carried out on the two public and largely used databases of FRGCv2 face and BANCA.