Real-Time Human Emotion Recognition Using Transfer Learning

Abstract

The research community has been interested in facial emotion recognition (FER) due to its potential uses. The main task of FER in real-time is mapping distinct face expressions to the corresponding states of emotional. The two main phases of the traditional FER are feature extraction and emotion classification. Due to its built-in feature extraction process from images, Deep Neural Networks “DNN”, particularly a Convolutional Neural Network “CNN”, are being extensively employed in FER. This paper proposed a Deep CNN (DCNN) modeling approach based on the Transfer Learning (TL) for developing a highly precise FER system in real-time in which the fully connected layer is fine-tuned to seven emotion states with using k-nearest neighbor (KNN) and support vector machine (SVM) classifiers instead of SoftMax. The suggested FER system has been validated using the ResNet-101 model and SAVEE image datasets. When used with KNN and SVM classifiers for emotion recognition, the suggested ResNet101 model with SVM classifier model obtains 98.41 % accuracy with respect to other performance measures.