Abstract: Manifold learning algorithms are unsupervised learning methods,the discriminant ability of the low-dimensional feature obtained by these methods are often lower than those obtained by the conventional dimensionality reduction methods.Furthermore,the original feature vectors of face image may include redundancy information such as high-order correlation which cannot be removed by manifold learning methods.To address the two problems,this paper proposes a Log-Gabor based supervised locality preserving projection(LG-SLPP) method for face recognition.We first use Log-Gabor wavelet to extract their corresponding Log-Gabor magnitude features(LGMF) by convolving the normalized face image with multi-scale and multi-orientation Log-Gabor filters.Then,SLPP operates on LGMFs to extract the discriminative submanifolds.Furthermore,the nearest distance classifier is used for classification.The proposed method is robust to illumination and expression variations by combing the Log-Gabor transform and supervised manifold learning.Experimental results on Yale and PIE databases show that the approach is quite effective.