FACIAL EXPRESSION RECOGNITION USING ENSEMBLE DEEP LEARNING TECHNIQUES

Abstract

Human facial expressions are the most popular way to express emotions. In the eld of Human-machine interaction, Facial expression recognition acts as a signi cant part of the area of human-machine intercommunication. In most of the papers, there are two most frequent methods employed for automatic facial emotion recognition, Geometric based and Appearance- based. Facial Expression Recognition usually involves four steps: pre-processing the image, face detection, feature extraction, and expression classi cation. In this report, I have analyzed several traditional and deep learning methods to identify basic human emotions, i.e., anger, happiness, disgust, fear, surprise, sadness, and neutrality. Training di erent models from the start require high computational power and a tremendous amount of time. Therefore, I have used transfer learning and ne-tuned some of the best models for face detection. I have also performed di erent experiments to improve the accuracy, like balancing the class or adding auxiliary data, which has improved the accuracy but not more than the state-of-art accuracy. Therefore I have performed an ensembling of the best of these models and achieved an accuracy of 75.7 percent.