Bio-Inspired Feature Selection and Fusion for Facial Emotion Recognition

This work presents an approach to improve emotion recognition systems by using two approaches: selection of feature subset using swarm intelligence based bio-inspired algorithms, and fusion of features from different regions of the face. For feature subset selection, Particle Swarm Optimisation (PSO) and Cuckoo Search algorithms are used. Classification is based on the fusion of features from three regions - entire face, mouth and eyes - and the predicted emotion (class label) is decided as the one given by two or more regions (Majority Vote). The ensemble of classifiers includes Random Forest and k-NN classifiers. It can be inferred that the ensemble model on an average yields 92% accuracy for Random Forest and 90% for k-NN, with a selected subset of almost half, or even less than half of the total features extracted. Moreover, not just in the majority region alone, the ensemble model outperformed the models in the mouth region and the face region. Though this work does not intend to show significant improvement in classification accuracy, the hybrid approaches and feature subsets presented in this work can significantly reduce computations with large datasets along with promising classification performance; moreover, some general observations and outcomes would help follow-ups in this area of research.