A Lightweight Classifier for Facial Expression Recognition based on Evolutionary SVM Ensembles

Evaluation criteria for solutions to facial expression recognition usually bias to classification accuracy. Hence, the utilization of deep neural networks has become a straightforward and popular option in theoretical studies despite the limitations in real usage from data collection, storage space, and power consumption issues. Our work proposes a practical alternative that is consisted of a minimum model configuration and still matches the state-of-the-art performance of deep learning approaches. We establish a conventional two-stage procedure, where feature extraction of a facial subject depends on a universal filter, histogram of oriented gradients (HOG), and classification is implemented through an ensemble learning approach using basic binary classifiers, support vector machines (SVM). Our two designs considerably improve prediction accuracy. One is that we adopt post-hoc statistics, rather than a priori expectations, to interpret the outputs of weak classifiers. The other is we design a genetic algorithm to search for the optimal ensemble of weak classifiers efficiently. Our method demonstrates supreme performance in several benchmark datasets and even outperforms those based on deep learning from big data. Besides, from a practical viewpoint, our model shows the advantage and flexibility of its storage size and power consumption. Lastly, we further display how the evolutionary SVM ensembles in our model contain information about the dependency and similarity among facial expression categories.