Dynamic best spectral bands selection for face recognition

Abstract

In this paper, face recognition in uncontrolled illumination conditions is investigated. A twofold contribution is proposed. First, three state-of-art algorithms, namely Multiblock Local Binary Pattern (MBLBP), Histogram of Gabor Phase Patterns (HGPP) and Local Gabor Binary Pattern Histogram Sequence (LGBPHS) are evaluated upon the IRIS-M3 face database to study their robustness against a high illumination variation conditions. Second, we propose to use visible multispectral images, provided by the same face database, to enhance the performance of the three mentioned algorithms. To reduce the high data dimensionality introduced by the use of multispectral images, we have designed a system to dynamically select the best spectral bands for each new subject. Our semi-supervised system for best spectral bands selection learn the relation between the recognition performance of each spectral band and its intrinsic quality using techniques of transfer learning and finite mixture of Gaussian for data distribution estimation. The obtained model is function of the image quality, and for each new spectral band, the likelihood ratio test is used to determine if the former belongs to either the set of good spectral bands or bad spectral bands. To the best of our knowledge, this is the first system proposed to dynamically select the best visible spectral bands for face recognition. Our results highlighted further the still challenging problem of face recognition in conditions with high illumination variation, as well as the effectiveness of our subspectral images based approach to increase the accuracy of the studied algorithms by at least 21.66 % upon the proposed database. Finally, our dynamic system has shown a superiority of performance over non-dynamic systems developed for the same face database.