Robust affect analysis using committee of deep convolutional neural networks

Human emotion has attracted researcher's attention as it finds potential applications in identifying consumer's mood and interest towards their product, assessments of learner emotional states, manufacturing smart cars, automotive industry and detecting mental states of the person in health care applications. In this paper, a well-designed committee network that focuses on the applicability of deep features for human emotion recognition from facial expressions is proposed. This architecture has the advantage of multi-level feature extraction using multiple filters that improve the performance of the network. The designed variant of inception-residual structure helps in the flow of input data through multiple paths, thus explicitly captures emotion variation from multi-path sibling layers and concatenated for recognition. The proposed algorithm is experimented with eNTERFACE, SAVEE and AFEW databases and the accuracy of 94.76%, 98.67% and 66.84%, respectively, is obtained.