Public Dataset Test of Custom Lightweight CNN Model for Facial Emotion Recognition

Abstract

Facial emotion recognition (FER) remains a challenging topic even though it is crucial for many applications. This study focuses on enhancing FER utilizing a customized lightweight Convolutional Neural Network (CNN) model in order to get around the computational expense often associated with traditional AI approaches. We evaluated our model on many publicly available datasets, including FER2013, RAF-DB, Young AffectNet HQ, and CK-Dataset, to evaluate its performance in both classification and detection tasks. We employed a range of classification techniques, including Xception, ShuffleNet, MobileNetV2, and a modified version of MobileNetV2, in order to achieve a decent balance between accuracy and processing efficiency. Additionally, we used state-of-the-art detection algorithms, namely the YOLO family (YOLOv5x6, YOLOv8, and YOLOv9), to accurately detect emotions across the datasets. Furthermore, we developed techniques to calculate F1 Score, Precision, and Recall, enabling a comprehensive evaluation of the model's functionality. Our findings advance the field of emotion recognition in real-time applications by showing how effectively the YOLO techniques and the customized lightweight CNN model cooperate to improve FER accuracy while maintaining computational viability.