Facial Expression Recognition Based on Optimized ResNet

Facial expression recognition (FER) is a very challenging task for machines to understand the emotional changes in human beings. Recently, using deep learning methods has greatly improved the accuracy of FER. However, considering the requirements of deploying deep learning algorithms in embedded systems, the size of neural networks needs to be further optimized on the basis of ensuring the high accuracy. In order to reduce the model parameters and maintain the high FER accuracy, inspired by the advances that Residual Network (ResNet) have achieved in image recognition and classification, we propose a simplified and efficient neural network based on ResNet and Squeeze-and-Excitation (SENet).Experiments show that the proposed algorithm obtains recognition accuracy of 74.143% and 95.253% on the FFE2013 and CK+ databases, respectively. Compared with state-of-the-art methods using Visual Geometry Group (VGG) or other deep neural networks, the proposed algorithm not only improves the accuracy, but also reduces the size of the model, which is competitive with existing methods in terms of size of the model parameters and recognition accuracy. Furthermore, this work also shows that only increasing the network layers hardly improves the recognition accuracy on the FFE2013 and CK+ databases.