Automatic recognition of facial expressions

Accuracy, Robustness, Lightweight, Speed. Can we have the best of everything? Yes! We present a novel Facial Expression Recognition network, designed to address the two real-life key challenges of existing models. First, we observed that existing models consistently outperform on certain facial expressions than others, due to these models being trained on class-imbalanced datasets with long-tailed distributions, and possibly due to some expressions being more complex and therefore more challenging to recognize than others. To address this, we adopt the Sharing Affinity block that splits facial features into two components for easier learning, and utilise Additive Angular Margin Loss for enhanced discriminatory power. A second issue is many existing models contain unnecessarily excessive number of parameters and require tremendous amount of computing power to train and deploy, making them uneconomical for practical applications. For this challenge, our model relies on a ResNet-18 backbone to keep complexity low, and adopts the Feature Integration block to further abstract high-level features before passing them to the fully connected layer. Experiments on RAF-DB prove that our model boosts FER performance remarkably, with an overall accuracy of 92.24% and average accuracy of 86.69%, exceeding all current state-ofthe- art methods. With 11,178,309 parameters and 221,493,504 estimated floatingpoint operations, our model is quite possibly also the most lightweight and most efficient model of its class.Bachelor of Engineering (Computer Science