Abstract
This paper presents a novel manifold learning method, namely Directional two-dimensional neighborhood preserving embedding (Dir-2DNPE), for feature extraction. In contrast to standard NPE, Dir-2DNPE directly seeks the optimal projective vectors from the directional images without image-to-vector transformation. Moreover, Dir-2DNPE can well reserve the spatial correlations between variations of rows and those of columns of images. Experiments on the ORL and Yale databases show the effectiveness of the proposed method.