Using Multi-inception CNN for Face Emotion Recognition

Abstract

One integral and necessary part of human behavior is emotion, which affects the way people communicate. Although human beings can recognize and interpret facial expressions, the identification of correct facial expressions continues to be a key and challenging task by computer systems. The main issues stem from the face's non-uniform design and variations in conditions such as light, facial structure, and posture. Several Convolutional Neural Network (CNN) approaches have been introduced for Face Emotion Recognition (FER), but these methods cannot completely reflect temporal variations in facial characteristics. In this study, we use the CMU face data collection of four types of emotions to provide a method for the identification of facial emotions. Four classes of distinguished emotions are happy, sad, angry, and neutral. Pixel values are fed into a Neural Network with different architecture, and the accuracy of those methods has been compared. Restricted Boltzmann machine (RBM), Deep Belief Networks (DBN), Convolutional Neural Networks (CNN), and multi-inception ensemble Convolution Neural Networks are different methods that are used in this research. We note the latter has considerably higher accuracy compared to other ones. The results obtained from the proposed methods Multi-inception CNN is slightly more than 87 percent while for the Restricted Boltzmann Machine (RBM) model it is 26.1 percent and for Deep Belief Networks (DBN) results are almost the same and slightly more than 26 percent finally the results for simple CNN model is 55 percent. Keyword: Face Emotion Recognition, FER, Deep Belief, RBM, multi-inception CNN.