Dual-threshold Based Local Patch Construction Method for Manifold Approximation And Its Application to Facial Expression Analysis

In this paper, we propose a manifold based facial expression recognition framework which utilizes the intrinsic structure of the data distribution to accurately classify the expression categories. Specifically, we model the expressive faces as the points on linear subspaces embedded in a Grassmannian manifold, also called as expression manifold. We propose the dual-threshold based local patch (DTLP) extraction method for constructing the local subspaces, which in turn approximates the expression manifold. Further, we use the affinity of the face points from the subspaces for classifying them into different expression classes. Our method is evaluated on four publicly available databases with two well known feature extraction techniques. It is evident from the results that the proposed method efficiently models the expression manifold and improves the recognition accuracy in spite of the simplicity of the facial representatives.