Improved Facial Emotion Detection Using CoAtNet and Temperature Scaling

Facial emotion detection is a pivotal AI technology with broad applications in healthcare and entertainment, enhancing natural human-machine interactions. However, existing methods, based on traditional machine learning and deep learning, often suffer from accuracy and robustness limitations. This research introduces an emotion detection method based on the CoAtNet model. We compare the efficacy of CNN and transformer backbones and investigate the role of temperature scaling in emotion classification. Our approach retains pixel-level information by optimising downsampling layers, preserving critical facial nuances, and enhances training stability through temperature scaling. Experimental results, based on the RAF-DB dataset, reveal a 9% and 8% accuracy improvement over the ResNet baseline and the original CoAtNet method.