Combinatory Connection Network for Facial Expression Classification

Abstract

Facial expression classification is important to the fields of computer vision, knowledge discovery, and human-computer interaction. Improving the accuracy of facial expression classification contributes to the development of safety, traffic, and recommendation. In this study, we proposed a Combinatory Connection Network, which called ComNet. In each layer, we divide the feature maps into different areas and then use the intra combinatory connection to learn more local information. Not only improving the accuracy of expression classification, but ComNet also has better performance in the case of less labeled samples. Between each layer, we use the inter combinatory connection to optimize the propagation of the gradient, which improves the accuracy of the network and reduces the generalization error. To verify the accuracy of the network, we performed experiments on the CK+ dataset to present the performance of the ComNet on facial expression classification tasks. We also experimented on other datasets to prove that ComNet is not only effective on specific datasets. A similar phenomenon was obtained.