Kernel maximum likelihood scaled locally linear embedding for night vision images

•A kernel manifold algorithm is designed to analyze the complex night vision data.•An outlier-probability is derived by solving maximum likelihood in kernel space.•A robust embedding is completed by scaling the kernel LLE with outlier-probability.

This paper proposes a robust method to analyze night vision data. A new kernel manifold algorithm is designed to match an ideal distribution with a complex one in natural data. First, an outlier-probability based on similarity metric is derived by solving the maximum likelihood in kernel space, which is corresponding with classification property for considering the statistical information on manifold. Then a robust nonlinear mapping is completed by scaling the embedding process of kernel LLE with the outlier-probability. In the simulations of artificial manifolds, real low-light-level (LLL) and infrared image sets, the proposed method show remarkable performances in dimension reduction and classification.