Facial expression recognition using convolutional neural network with weighted loss function

Facial expression is an important content in human communication as it is participating in nearly half of the human interaction. The recognition of facial expression has been applied in security, criminal research, entertainment applications, and other human-computer-interaction fields. Facial expression recognition is being extensively studied in recent years and have reached satisfactory results. However, the type of images of expressions in the commonly used datasets are unbalanced and cause the recognition of one to two facial expressions are harder than others, which affects the accuracy of the whole, Therefore, we create a loss function aim to decrease the effect of this unbalance. With the combination of the state-of-art convolutional neural network and our loss function, the accuracy on the FER-2013 dataset has raised about 2%.