Heuristic objective for facial expression recognition

Facial expression recognition has been widely used in lots of fields such as health care and intelligent robot systems. However, recognizing facial expression in the wild is still very challenging due to variations, light intensity, occlusions and the ambiguity of human emotion. When training samples cannot include all these environments, the classification can easily lead to errors. Therefore, this paper proposes a new heuristic objective function based on the domain knowledge so as to better optimize deep neural networks for facial expression recognition. Moreover, we take the specific relationship between the facial expression and facial action units as the domain knowledge. By analyzing the mixing relationship between different expression categories and then enlarging the distance of easily confused categories, we define a new heuristic objective function which can guide deep neural network to learn better features and then improve the accuracy of facial expression recognition. The experimental results verify the effectiveness, universality and the superior performance of our methods.