Dynamic Noise Injection for Facial Expression Recognition In-the-Wild

Abstract

Facial expression-based emotion analysis is one of the most important artificial intelligence research fields. However, a lot of works still suffer from the low classification/regression performance caused by overfitting. Therefore, we propose new noise injection techniques to alleviate the overfitting problem on the task of facial expression recognition in the wild. Specifically, both techniques are based on the ResNet-18 architecture, and we periodically or dynamically add feature-level noise into the BN+ReLU unit to learn more robust features. The periodic method needs to probe the optimal hyperparameter with respect to the interval for the noise injection through trials and errors. Therefore, we propose the second method in order to make a dynamic noise injection mechanism work without a non-trivial time-consuming hyperparameter search process. Finally, the performance of the two methods is reported in the experiment. Our experiments on facial expression classification with the AffectNet dataset demonstrated the usefulness of the proposed approach.