Orthogonal Tensor Sparse Neighborhood Preserving Embedding for Two-dimensional Image

Orthogonal Tensor Neighborhood Preserving Embedding (OTNPE) is an efficient dimensionality reduction algorithm for two-dimensional images. However, insufficiencies of the robustness performance and deficiencies of supervised discriminant information are remained. Motivated by the sparse learning, an algorithm called Orthogonal Tensor Sparse Neighborhood Embedding (OTSNPE) for two-dimensional images is proposed in the paper. The algorithm firstly regards two-dimensional images as points in the second-order tensor space, then, the neighborhood reconstruction of samples within the same class is achieved with sparse reconstruction. Finally, projections are gotten to preserve local sparse reconstruction relation and neighborhood relation within the same class and spatial relation of pixels in an image. Experiments operated on Yale, YaleB and AR databases show, in contrast to the existing typical tensor dimensionality reduction algorithms, the algorithm can improve the accuracy rate of classification algorithms based on the shortest Euclidean distance. Key–Words: Dimensionality Reduction, Tensor Image, Neighborhood Preserving Embedding, Sparse Reconstruction, Supervised Discriminant Information, Face Recognition.