A Comparison of CNN-based 2D Static Facial Emotion Recognition Techniques

Abstract In the field of computer vision, recognizing expressions in 2D static facial images is a crucial aspect of facial emotion recognition (FER), with broad applications in real-world scenarios such as mental health diagnosis and security monitoring. Despite the various convolutional neural network (CNN) architectures proposed for FER, there is still a lack of systematic comparative studies on the classification capabilities of different CNN architectures. In this paper, we systematically compare the classification performance of four commonly used CNN architectures in FER research for 2D static facial expression recognition on the classic FER2013 dataset. Through experiments, we evaluate the classification accuracy of these architectures and analyze the recognition results for different emotion categories.