Expression Recognition Using Region Features And Facial Action Units

Expression recognition has been a popular research field for decades, since its wide applications such as human-computer interaction. In this area, with the rapid development of deep learning, Convolutional Neural Networks (CNNs) have shown good performance. In this work, a deep learning model based on CNNs is constructed, using not only traditional global features, but also region features. Further more, facial action units are used to improve it, and a Bayesian Network model is built to analyze the probabilities of action units (AUs). At the final stage of our method, ensemble learning is utilized to classify expressions. Experiments on CK+ and JAFFE datasets show accuracies of about 96.76% and 96.2% respectively.