This paper presents a facial emotion recognition technique using two newly defined geometric features, landmark curvature and vectorized landmark. These features are extracted from facial landmarks associated with individual components of facial muscle movements. The presented method combines support vector machine (SVM) based classification with a genetic algorithm (GA) for a multi-attribute optimization problem of feature and parameter selection. Experimental evaluations were conducted on the extended Cohn-Kanade dataset (CK+) and the Multimedia Understanding Group (MUG) dataset. For 8-class CK+, 7-class CK+, and 7-class MUG, the validation accuracy was 93.57, 95.58, and 96.29%; and the test accuracy resulted in 95.85, 97.59, and 96.56%, respectively. Overall precision, recall, and F1-score were about 0.97, 0.95, and 0.96. For further evaluation, the presented technique was compared with a convolutional neural network (CNN), one of the widely adopted methods for facial emotion recognition. The presented method showed slightly higher test accuracy than CNN for 8-class CK+ (95.85% (SVM) vs. 95.43% (CNN)) and 7-class CK+ (97.59 vs. 97.34), while the CNN slightly outperformed on the 7-class MUG dataset (96.56 vs. 99.62). Compared to CNN-based approaches, this method employs less complicated models and thus shows potential for real-time machine vision applications in automated systems.
Pre-print author copy (peer-reviewed and accepted): XLiuIEEESensors2020