In order to extract the facial feature s from face images effectively,a novel supervised linear method of reducing dime nsionality is proposed for face recognition.In this study,the concept of image distance is first introduced to measure the similarity between face samples,which enhances the robustness to the translation and deformation of the face image.And then the prior class label information of train samples is i ncorporated into the criterial equation of neighborhood preserving embedding (NP E) algorithm which is a manifold learning method developing from the classical a lgorithm of locality linear embedding (LLE).After optimizing the criterial equat ion,the distribution of the reduced subspace is made to be the structure of mult i-manifold,which not only optimally preserves the local geometry of the origina l space,but also minimizes the intra-class scatter while maximizes the between -class scatter of the projected data.Thus the discrimination of the embedding i s enhanced,and then the recognition rate of the proposed algorithm is improved o bviously.Experiments are conduced on the two open face databases,the Extended Ya le Band CMU PIE face databases,and the results show that the proposed method ca n effectively find the key facial features form face images and can achieve bett er recognition rate compared with other existing ones.