Improved Sparse Preserving Projections Based Face Recognition Algorithm

In the field of pattern recognition for sample classification criterion, it is used that to keep the sparse reconstruction relationship of the original samples to the sample space after the projection transformation, so as to increase the accuracy of classification. Sparse preserving projection(SPP) is a typical algorithm based on the idea. But the deficiency of the SPP algorithm, which is a global perspective of the original samples to find the best projection transformation.However, the general image datas are presented as global nonlinear and local linear structure, it means that the original image samples are often in a low dimensional manifold structure in a high dimensional space. Based on the deficiencies of the above analysis, the main points of this paper are as follows:The projection transformation vectors obtained by the SPP are not orthogonal, Projection vector orthogonality would be helpful to improve the recognition rate of the algorithm. At the same time, original image sample structure present linear separability has great effect on the improvement of the accuracy of image classification, From these two points, a sparse preserving projection algorithm based on kernel orthogonal algorithm(KOSPP) is proposed.However, the above method is still based on the global perspective of the samples, taking into account the spatial distribution of the original image samples, Under normal circumstances, low dimensional mainfold structure to maintain would be helpful to improve the recognition rate, so the neighbor information matrix is introduced to preserve the neighborhood information, and propose a sparse preserving projection algorithm(LNIFSPP) based on local neighbor information fusion.The above two algorithms are based on non supervision. However, supervised algorithm will be able to improve the rate of recognition, while considering the projection of orthogonal transformation and the original sample separability and retain local neighbor information based on, a kernel orthogonal sparse preserving projection algorithm(LNIKOSPD) based on the local neighbor information of supervised is proposed.