COMPARATIVE ANALYSIS OF CNN, RESNET50, EFFICIENTNETB0, AND DENSENET121 FOR FACIAL EMOTION RECOGNITION USING THE RAF-DB DATASET

Dr. Abdul Aziz; Dr. Shahid Khan Yusufzai; Naveed Ali Rajper

Authors

Dr. Abdul Aziz MS Scholar in Data Science, Department of Robotics and AI, SZABIST University, Karachi
Dr. Shahid Khan Yusufzai Adjunct Faculty Member, Department of Robotics and AI, SZABIST University Karachi
Naveed Ali Rajper MS Scholar in Data Science, Department of Robotics and AI, SZABIST University, Karachi

Abstract

Facial emotion recognition has got significant attention in the field of computer vision. It has several applications such as healthcare, surveillance, human-computer interaction, and intelligent system. The aim of this study is to compare four deep learning models CNN, ResNet50, EfficientNetB0, and DenseNet121. All four models have been trained on RAF DB dataset to recognize seven human facial emotion classes Angry, Disgust, Fear, Happy, Neutral, Sad, and Surprise. All models have been trained and evaluated on the basis of different parameters such as accuracy, precision, recall, F1-score, confusion matrix, and ROC curve. The results suggest that the CNN model achieved the highest test accuracy of 77.36%, followed closely by ResNet50 with 77.57%, on the other hand DenseNet121 and EfficientNetB0 achieved 74.66% and 73.53% respectively. All models performed really well on the dominant classes like Happy and Surprise but poorly performed on minority classes like Disgust and Fear. The overall results suggest that pre-trained models have strong capability to extract robust feature and better generalization capabilities to recognize human facial emotions. This study confirms the usefulness of deep learning models and their capabilities to recognize human emotions.