Image Region Selection Based GLRR for Face Recognition

Local ridge regression classifier (LRR) is an effective local face recognition method. It suppresses the influence of local changes by setting a voting RR classifier for each image region, thus has partial robustness to local changes caused by lighting, occlusions and poses. LRR uses the concatenated vector of a sub-image as its input feature, such a feature is still not sufficient to represent an image, thus leading to possibly imprecise voting and limited increase in recognition rate. In order to boost its recognition rate, we first develop a novel classifier GLRR which combines LRR classifier and Gabor-LBP features which can improve the feature representation greatly. Experiments on AR database demonstrate that GLRR is superior to LRR and other local methods such as Aw-SpPCA and SpCCA. When just fewer classifiers can be available and some occlusion regions exist, majority-voting recognition rate will still be imprecise. To remedy this, in this paper, we add an occlusion detection step before classification using GLRR for which we call it S-GLRR. In this way, we can purposely shield locally-occluded regions using the detection step, thus get better performance for face recognition. Experiments show that S-GLRR achieves better recognition rate than GLRR, especially when only a few sub-classifiers are provided.