Fuzzy Nearest Feature Line-based Manifold Embedding for Facial Expression Recognition

Traditional nearest feature line (NFL) based subspace learning (NFLS) has been successfully applied in face recognition by capturing more variations of face images than original data set. However, it is suboptimal for pattern recognition due to that it only focuses on the within-class information while neglects the interaction between different classes. Besides, NFLS fails to detect the underlying manifold embedded in image space. In this paper, a novel manifold embedded extension of NFLS named fuzzy local nearest feature line (FLNFL) is developed for facial expression recognition. FLNFL achieves its high classification performance by combining graph embedded projection with fuzzy set theory in a Fisher type of learning manner. It not only considers the intrinsic similarity between data points, but also exploits fuzzy assignment technique to investigate inherent uncertainties in expression images arise from variety of illumination, posture, and viewing directions. After this mapping, therefore, the transformed features can reflect both local geometry and intrinsic class assignment of the data set. Experimental results on two widely used facial expression databases suggest that the proposed method provides a better representation of expression features and yields promising performance in facial expression recognition.