One-Class and Bi-Class SVM Classifier Comparison for Automatic Facial Expression Recognition

Facial expressions might be seen as a relevant and useful source of information. Indeed, they allow understanding and even identifying people behavior based on the emotional changes. Therefore, automatic facial expression recognition has been widely solicited in the context of smart cities and homes. However, recognizing human emotion automatically through facial expressions remains challenging. Moreover, multi-class Support Vector Machine classifiers have been widely employed and in most cases, the proposed architectures are based on the use of bi-class classifiers. In this paper, we propose an approach that exploits selected geometric-based features using the Extremely Randomized Trees method while the recognition is handled by three distinct multi-class Support Vector Machine architectures namely bi-class (One-against-One and One-against-All) and one-class classifiers. We also investigate the performance of the three different architectures by performing a comparison in terms of accuracy and computation time. The carried experiment on three benchmark datasets attests to the efficiency of the one-class classifier since the proposed approach yields 92.68%, 85.83% and 93. 33% with the JAFFE, RaFD and KDEF datasets, respectively.