Semi-supervised learning framework based on statistical analysis for image set classification

Highlights：

• We propose a semi-supervised learning method to solve image set classification based on statistical Gaussian manifold from the perspective of Lie Group.

• We derive two new positive definite manifold kernels to capture the structure information of Gaussians based on Lie group isomorphisms.

• We adopt manifold distance metric to construct a “fully trusted” graph structure.

• We derive data dependent probabilistic manifold kernel to strongly reflect the underlying connection relationships between the training and query Gaussian manifold components.

• We propose a new kernel fuzzy discriminant framework to facilitate robust classification.