Dual Task Training for Face Expression Recognition

Due to the increased interest in human-computer interaction, Facial Expression Recognition (FER) is among the most popular research fields in computer vision and state of the solutions are based on machine learning techniques. In this paper, we propose the use of a deep neural network trained with a joint loss for better discrimination of action units (AUs). The main idea is to label the action units images with what we have entitled pseudo-expressions. In addition to the classic Mean Squared Error (MSE) loss used for a multi-class, multiple instance leaning problem, we added a Cross Entropy term to gain advantage of pseudo-emotions to better learn the AUs. Last part of the experiment consisted of using the center loss in order to better cluster the embeddings harvested on fully-connected layers.