Action Unit recognition in still images using graph-based feature selection

Facial expressions are universal and independent of race, culture, ethnicity, nationality, gender, age, religion, or any other demographic variable. These facts are the main reason for automatic facial expression recognition being one of the hot topics of many research efforts and being useful in so many commercial and scientific fields. The most well-known and probably the most used anatomically based method of defining facial activity is Facial Action Coding System (FACS). In this paper, we propose a Facial Action Unit recognition algorithm using graph-based feature selection in unsupervised and supervised setting. The proposed algorithm is based on a state of the art algorithm for facial key points detection Supervised Gradient Descent method, the classification is carried out using the well know Support Vector Machines classifier. Built this way, the algorithm works on still images where the human expressions are expected to be in their apex phase. Using leave one person out evaluation methodology we achieve average accuracy of 90.1% for unsupervised and 92.7% for supervised feature selection on 12 Action Units.