Facial Emotion Recognition using CNN, VGGNet, and ResNet

Abstract

Facial Emotion Recognition (FER) has emerged as an important field of study in the field of computer vision with various applications in human-computer interaction, affective computing, healthcare monitoring, intelligent surveillance systems, etc. This work provides a comparative analysis of three types of Deep Learning architectures, i.e., a custom Convolutional Neural Network (CNN), a VGGNet-inspired architecture, and a Residual Network (ResNet), for the seven-class emotion classification task in FER2013 benchmark dataset. To make sure a fair comparison, all models were trained in identical experimental conditions under the same optimizer, learning rate, batch size and preprocessing strategy and callback configuration, so that performance differences could be attributed to different architecture variations only. The FER 2013 is a set of 35887 grayscale facial images with resolution 48 × 48 and are labeled in Angry, Disgust, Fear, Happy, Sad, Surprise and Neutral categories. Experimental evaluation shows that the model based on VGGNet has achieved the highest test accuracy of 81.23% with a loss of 1.67, 7.18 million parameters. The ResNet model achieved an accuracy of 78.45% with only 3.05 million parameters, which is a reduction of 57% in the model size with a negligible performance drop. The accuracy of the baseline CNN was 59.71% accuracy. In addition, a Streamlit-based real time web application to perform ensemble inference among the three models was created. The results provide practical advice for the choice of suitable architectures for FER, especially in cases where either computational or latency constraints are prevalent.