Staticand Dynamic Facial Emotion Recognition Using Neural Network Models

Abstract

Emotion recognition is the process of identifying human emotions. It is made possible by processing various modalities including facial expressions, speech signals, biometricsignals,etc. Withtheadvancementsincomputingtechnologies,FacialEmo tion Recognition (FER) became important for several applications in which the user’s emotional state is required, such as emotional training for autistic children. The recent years witnessed a major leap in Artificial Intelligence(AI),specially neural networks for computer vision applications. In this thesis, we investigate the application of AI algo rithms for FER from static and dynamic data. Our experiments address the limitations and challenges of previous works such as limited generalizability due to the datasets. We compare the performance of machine learning classifiers and convolution neural networks (CNNs) for FER from static data (images). Moreover, we study the perfor mance of the proposed CNN for dynamic FER(videos),in addition to Long-ShortTerm Memory(LSTM)inaCNN-LSTM hybrid approach to utilize the temporal information in the videos. The proposed CNN architecture out performed the other classifiers with an accuracy of 86.5%. It also outperformed the hybrid approach for dynamic FER which achievedanaccuracyof74.6%