Learning Manifold from Incomplete Image Set

Recently, there have been several advances for developing manifold learning algorithms to learn the nonlinear manifold of collected data. To our best knowledge, however, learning manifold from incomplete data set, wherein some features of samples are missing, is still an untouched problem so far. In context of incomplete image set, an improved LTSA algorithm is proposed to learn manifold from corrupted image set. The proposed algorithm exploits an extended EM-based PCA algorithm, which can learn the principal components of incomplete image set only using the known pixels, to obtain the local tangent space coordinates instead of standard SVD technique. Experiments on benchmark data sets demonstrate the effectiveness of the proposed approach.