Local similarity and diversity preserving discriminant projection for face and handwriting digits recognition

In this paper, a novel supervised subspace learning algorithm, named local similarity and diversity preserving discriminant projection (LSDDP), is presented. LSDDP defines two weighted adjacency graphs, namely similarity graph and diversity graph. LSDDP constructs the similarity scatter and diversity scatter with the weights, which are adjustable according to the global supervisor and the local semi-supervisor information of the data. Thus LSDDP could utilize both the similarity and diversity information of the data simultaneously for dimensionality reduction. After characterizing the similarity scatter and diversity scatter, a concise feature extraction criterion arised via minimizing the difference between them and the optimal projection is obtained by performing the eigen-decomposition. Thus our method successfully addresses the SSS problem without losing any discriminating information. Finally the proposed model is verified by the face and handwriting digits recognition experiments. The experimental results on Yale, ORL and CMU-PIE face database and the USPS handwriting digits database indicate the effectiveness of our method.