In face recognition, traditional linear dimensionality reduction methods can not be good at keeping the intrinsic distribution of face sample data. While Locally Linear Embedding (LLE) algorithm, which belongs to manifold learning, has the advantage of keeping the intrinsic distribution of face sample data. Principal Component Analysis (PCA) possesses the merits of high recognition efficiency. An improved PCA, in which the formula of PCA is modified, is presented in this paper. This algorithm has the ability of gray normalization and can overcome the influence of light on the target. Then the algorithm is combined with LLE and used in face recognition. In this way, we not only keep the intrinsic distribution of face sample data, but also assure the accuracy of the image characteristics. Experimental results on ORL face database demonstrate that the algorithm is superior to the original LLE.