Abstract
Facial expression recognition is a classical problem in computer vision. With recent success in applying deep learning to a number of computer vision tasks, we propose a deep learning based facial expression recognition system. The system is composed of face region detection and convolutional neural network (CNN) classifier. In the module of face region detection, we decide the best face region by facial landmark points. We design our own CNN architecture which is more suitable for the expression recognition task. The expression recognition accuracy of our CNN model could be further improved by using the recurrent neural network module. Experimental results on some standard datasets show that our framework is superior to or comparable to the state-of-the-art methods. We also collect three datasets from different domain to further investigate the generalization of the CNN model. We propose an illumination augmentation scheme which effectively reduce the overfitting issue while training with different domain types. Moreover, because our system did not rely on complicated pre-processing or rectification, our system is very efficient and it could run at about 15 FPS on notebook with GPU.