Optimizing facial emotion recognition using custom light weight CNN across public datasets

Abstract

FER is a necessary component in numerous applications, as a healthcare, human-computer interaction, and mobile technology. Nonetheless, FER continues to be a formidable challenge because to the intricacies of facial emotions and the computing expenses linked to conventional AI methodologies. This paper addresses these issues through the development of a custom lightweight CNN model that enhances FER and is computationally efficient. Various publicly available datasets, including FER2013, RAF-DB, Young AffectNet HQ, and CK+ Dataset, were used to evaluate the effectiveness of the model in classification and detection. We explored the methods of categorization, such as MobileNetV2 which is an adapted version of MobileNetV2, ShuffleNet as well as Xception to trade off accuracy and efficiency. Moreover, we used the latest detection algorithms belonging to the YOLO family which are YOLOv5x6, YOLOv8, and YOLOv9 to enhance precision of the detection in different datasets. Performance metrics like Recall, Precision, and F1 Score were used to comprehensively evaluate the effectiveness of the models. The findings suggest that the proposed lightweight CNN and YOLO models significantly improve the accuracy of FER with the least processing requirements, thus, developing realtime emotion recognition systems.