Independent Component Analysis: Embedded LTSA

In order to solve the adaptability problem of local tangent space alignment (LTSA) with potential higher-order information loss in manifolds, a novel algorithm is proposed to optimize the extraction of local neighborhood information. The algorithm is based on LTSA and ICA algorithms, which is called the IELTSA algorithm. By optimizing the extraction of the tangent vectors, the algorithm can improve dimension reduction in high-dimensional and unevenly distributed manifolds. The proposed algorithm is feasible in carrying out manifold learning, and the reconstruction error is no more than that of LTSA. Experiments show that IELTSA can be applied to changed-density 3D curves and face images targeted at lower dimensions, showing highest performance over LTSA and other improved methods based on LTSA, and achieving the highest recognition rate in lower-dimensional embedding. The algorithm can effectively reconstruct low-dimensional coordination in curves with changed-density and also shows adaptability in high-dimensional images targeted at lower dimensions.