Principal component analysis (PCA) can only keep the global structure, while neighborhood preserving embedding (NPE) preserves the similarity between neighbor data, but ignores the difference between them. Focusing on the problems mentioned above, a feature extraction method is proposed by fusing global and local various feature, and is applied to facial expression recognition. PCA is used to preserve global structure and a local diversity scatter and a local similarity scatter is defined by manifold learning algorithms, combining with local maximum scatter difference criterion, the proposed method can efficiently preserve the variety of local manifold. The low dimensional feature is extracted by combining the global feature with local various feature for expression classification. The experiments on JAFFE and Cohn-Kanade facial expression databases indicate that compared with PCA, locality preserving progection (LPP), NPE and other methods, this method not only improves the recognition rate efficiently, but also needs the least dimensions when achieves the highest recognition rate, which demonstrates that this method is superior to others in recognition rate.