A Comparative Analysis of CNNs and ResNet50 for Facial Emotion Recognition

Abstract

Anger, disgust, fear, happiness, sadness, surprise, and neutrality are some of the basic facial emotions that researchers worldwide consider for recognition. Detection of these emotions is important in the present era due to the digital transformation of many processes and human communication. This study analyzes emotion detection methods using the capabilities of deep learning techniques such as a Convolutional Neural Network (CNN) and Residual Network 50 (ResNet-50). The FER2013 benchmark dataset for facial emotion recognition was used for training and testing purposes, along with a few other private images. This study aimed to compare and analyze the performance of the two methods based on several comparative factors, such as architectural differences, feature extraction capability, training dynamics, model performance, computational efficiency, and hardware configuration. The experimental results showed that the ResNet-50 model was significantly more accurate than the CNN, with an accuracy of 85.75% compared to 74%. Although ResNet-50 has higher computational costs, its robustness and accuracy make it the optimal choice for facial emotion recognition tasks. This research provides valuable insight into the capabilities and trade-offs of these models for face emotion recognition techniques.