Facial Expression Recognition Using Double-Stage Sample-Selected SVM

This paper proposes a double-stage classification model for the classification of six basic facial expressions. Inspired for the fact that an increase in the number of classes brings a drop on the accuracy for facial expression recognition, we use classifiers with fewer classes to improve the performance of a six-class expression recognition classifier. Support Vector Machine (SVM) is adopted as the classifiers due to its excellent performance in small databases. To make SVMs classify samples more precisely, selecting more support vectors trains the model. Active Shape Model (ASM) is used to locate shape points. The shape points are used as features to train the double-stage SVM, which includes a six-class SVM and a following few-class SVM with the classes corresponding to the largest classification probabilities of the former. The approach in this paper achieves an accuracy of 98.25% on the Japanese Female Facial Expression (JAFFE) database, 3.08% and 5.53% higher than those of Local Curvelet Transform method Facial Movement Features method respectively, and besides far better than six other methods.