Leveraging ResNet and label distribution in advanced intelligent systems for facial expression recognition.

Abstract

With the development of AI (Artificial Intelligence), facial expression recognition (FER) is a hot topic in computer vision tasks. Many existing works employ a single label for FER. Therefore, the label distribution problem has not been considered for FER. In addition, some discriminative features can not be captured well. To overcome these problems, we propose a novel framework, ResFace, for FER. It has the following modules: 1) a local feature extraction module in which ResNet-18 and ResNet-50 are used to extract the local features for the following feature aggregation; 2) a channel feature aggregation module, in which a channel-spatial feature aggregation method is adopted to learn the high-level features for FER; 3) a compact feature aggregation module, in which several convolutional operations are used to learn the label distributions to interact with the softmax layer. Extensive experiments conducted on the FER+ and Real-world Affective Faces databases demonstrate that the proposed approach obtains comparable performances: 89.87% and 88.38%, respectively.