Emotion Recognition from Masked Faces using Inception-v3

Abstract

The identification of human emotions from facial expressions is intriguing and challenging research given the subtle differences between certain emotions. Face masks are nowadays strongly recommended to minimize infection transmission due to Covid-19. Successful emotion identification from masked faces is challenging since the lower part of the face contributes significant cues for emotion identification. In this work, we investigate transfer learning using deep pre-trained networks for emotion recognition from masked faces. Specifically, we fine-tune the pre-trained models: - EfficientNet-BO, ResNet-50, Inception-v3, Xception and AlexNet, on the benchmark Facial Expression Recognition (FER) 2013 dataset containing seven categories of emotions, namely, angry, disgust, fear, happy, sad, surprise and neutral. The experiments reveal that the Inception-v3 model outperformed all other deep learning models and the machine learning models Support Vector Machine (SVM) and Artificial Neural Network (ANN), for facial emotion recognition from masked faces.