Previous work has demonstrated that manifold learning can effectively preserve the local geometry among nearby data, and has become an active topic in pattern recognition and machine learning. However, it ignores or even impairs the local diversity of data, which will impair the recognition accuracy and lead to unstable local geometrical structure representation. In this paper, a novel approach, namely two-dimensional diversity preserving projection (2DDPP), is proposed for dimensionality reduction. 2DDPP constructs an adjacency graph to model the variation of data and measures the variation among nearby data by the diversity scatter, on the basis of which a concise criterion is raised by maximizing the diversity scatter. Moreover, 2DDPP directly calculates the diversity scatter matrix from the image matrix, which effectively avoids the small sample size problem. Experiments on Yale, UMIST, and AR databases show the effecitveness of the proposed method.
