An illumination insensitive framework using robust illumination normalization and spectral regression kernel discriminant analysis for face recognition

Recognition under uncontrolled lighting conditions remains one of the major challenge for practical face recognition systems. In this work, we present an efficient and effective framework to improve the recognition performance from two aspects: image preprocessing and subspace representation. The step of image preprocessing is mainly used to eliminate the effects of illumination. The step of subspace representation is used for dimension reduction and further removing the effects of illumination and expression. First, a novel and efficient image preprocessing method based on rotation invariant LBP and gradient direction, which we name “GDP-face”, is proposed to extract the illumination insensitive face appearance, it can extract more discrimination information by adjusting the parameters. Experimental results on Extended Yale B and PIE data sets show that the GDP-face outperforms some compared state-of-the-art image preprocessing methods. Second, we apply the Spectral Regression Kernel Discriminant Analysis (SRKDA) which is an effective and efficient subspace learning algorithm to get a more compact, robust, and discriminative feature descriptor. Experimental results show that our framework is more efficient than some popular methods based on Gabor features or Local Binary Patterns Histogram features, and achieve a better performance than some state-of-the-art methods.