Windmill Graph based Feature Descriptors for Facial Expression Recognition

Abstract

Automatic Facial Expression Recognition (FER) is an important research area in computer vision because it has many real-time applications. However, the main issue lies in the design of a feature descriptor that could effectively capture the appearance changes in the facial images. For FER systems, graph based methods have not been much explored for feature extraction. Hence, this paper introduces novel graph based methods named Windmill Graph based Feature Descriptors (WGFD) for feature extraction. Two variants of WGFD’s (WGFDh and WGFDv) have been proposed in this work to effectively encode the relationship among neighboring pixels in a local neighborhood. For expression classification, multi-class Support Vector Machine (SVM) is utilized. The performance of the proposed feature descriptors have been evaluated on six benchmark FER datasets namely JAFFE, KDEF, MUG, CK+, TFEID and FERG with respect to seven expressions classification. The experimental results showed promising results when compared to the recent FER methods.