This paper proposes a novel algorithm for face recognition, which is based on curvelet transform and (2D)2PCA. Contrast to traditional tools such as wavelet transform, curvelet transform has better directional and edge representation abilities. Inspired by these attractive attributes, we decompose face images to get low frequency coefficients by curvelet transform. (2D)2PCA with an exponential decay factor is applied on these selected coefficients to extract feature vectors, which will achieve dimension reduction as well. The nearest neighbor classifier is adopted for classification. Extensive comparison experiments on different data sets are carried out on ORL and Yale face database. Results prove that the proposed algorithm has high recognition accuracy and short recognition time, and it is also robust to changes in pose, expression and illumination.