Deep Feature Selection Using Moth-Flame Optimization for Facial Expression Recognition from Thermal Images

Abstract

Computer-based automated facial expression recognition has a wide variety of applications that include interactive games, data-driven animation, sociable robotics, and many other human–computer interaction systems. The recognition of facial expressions is a challenging research problem as people have different ways to show their expressions at different occasions. This paper proposes a model for the recognition of facial expressions using thermal infrared (IR) images which has two distinct stages. In the first stage, we have used the concept of snapshot ensembling to combine the features of a customized Convolutional Neural Network (CNN) at different epochs. Though CNN models can learn differently at different epochs from the inputs, there may be some redundancy if the feature maps of different epochs are combined straightforwardly. Hence, in the next stage, we have applied a nature-inspired meta-heuristic called Moth-Flame Optimization algorithm to reduce the feature dimension of the feature vector generated by the snapshot ensembling-based CNN model. The optimized feature vector is fed to the Support Vector Machine for classification of the facial images based on its expression. The proposed model is evaluated on a publicly available facial image database, called IR database, and it successfully classifies 99.63% of the images using only 45% features of the original feature vector. The model outperforms many state-of-the-art methods developed for facial expression recognition which are experimented on the same IR database considered here.