Facial Expression Recognition On Static Images

Facial expression recognition (FER) is currently one of the most attractive and also the most challenging topics in the computer vision and artificial fields. FER applications are ranging from medical treatment, virtual reality, to driver fatigue surveillance, and many other human-machine interaction systems. Benefit from the recent success of deep learning techniques, especially the invention of convolution neural networks (CNN), various end-to-end deep learning-based FER systems have been proposed in the past few years. However, overfitting caused by a lack of training data is still the big challenge that almost all deep FER systems have to put into a concern to achieve high-performance accuracy. In this paper, we are going to build a FER model to recognize eight commons emotions: neutral, happiness, sadness, surprise, fear, disgust, anger, and contempt on the AffectNet dataset. In order to mitigate the effect of small training data, which is prone to overfitting, we proposed a thoughtful transfer learning framework. Specifically, we fine-tuning ResNet-50 model, which is pre-trained on ImageNet dataset for object detection task, on the AffectNet dataset to recognize eight above mentioned face emotions. Experiment results demonstrate the effectiveness of our proposed FER model.