oa Enhancement of multispectral face recognition in unconstrained environment using regularized linear discriminant analysis (LDA)

In this paper, Face recognition in unconstrained illumination conditions is investigated. A two manifold contribution is proposed: 1) Firstly, Three sate of the art algorithms, namely Multiblock local Binary pattern (MBLBP), Histogram of Gabor Phase Patterns (HGPP) and Local Gabor binary pattern histogram sequence (LGBPHS) are challenged against The IRIS-M3 multispectral face data base. 2) Secondly, The Performance of the three mentioned algorithms, being drastically decreased due to the non monotonic illumination variation that distinguish the IRIS-M3 face database, This performance is enhanced using multispectral images (MI) captured in the visible spectrum. The use of MI images like near infrared images(NIR), short wave infrared images images (SWIR) or even visible images captured at different wavelengths rather then the usual RGB spectrum, is getting more and more the trust of researcher to solve problems related to uncontrolled imaging conditions that usually affect real world application like securing areas with valuable assets, controlling high traffic borders or law enforcement. However, one weakness of MI is that they may significantly increase the system processing time due to the huge quantity of data to mine (in some cases thousands of MI are captured for each subject). To solve this issue, we proposed to select the optimal spectral bands (channels) for face recognition. Best spectral bands selection will be achieved using linear discriminant analysis (LDA) to increase data variance between images of different subjects(between class variance) while decreasing the variance between images of the same subject(within class variance). To avoid the problem of data overfitting that generally characterize LDA technique, we propose to include a regularization constraints that reduce the solution space of the chosen best spectral bands. Obtained results highlighted further the still challenging problem of face recognition in conditions with high illumination variation, as well as proven the effectiveness of our multispectral images based approach to increase the accuracy of the studied algorithms namely MBLBP, HGPP and LGBPHS of at least 9% upon the proposed database.