Face Emotion Recognition From Static Image Based on Convolution Neural Networks

Abstract

Human-Machine Interaction systems have not yet reached all the emotional and social capacities. In this paper, we propose a face emotion recognition system from static image based on the Xception convolution neural network architecture and the K-fold-cross-validation strategy. The proposed system was improved using the fine-tuning method. The Xception model pre-trained on ImageNet database for objects recognition was fine-tuned to recognize seven emotional states. The proposed system is evaluated on the database recorded during the Empathic project and the AffectNet database. Our experimental results achieve an accuracy of 62%, 69% on Empathic and AffectNet databases respectively using the fine-tuning strategy. Combined the AffectNet and Empathic databases to train our proposed model, show significant improvement in the emotion recognition that achieves an accuracy of 91.2% on Empathic database.