Comparative Analysis of Optimized Deep Learning Models for Facial Emotion Detection

Abstract

Artificial Intelligence (AI) has been successfully applied across numerous domains, with computer vision being one of its most impactful areas. In this study, a Deep Neural Network (DNN) was utilized for Facial Emotion Recognition (FER), aiming to accurately classify human emotions based on facial expressions. A key objective of this research is to identify and analyze the critical facial features that the DNN prioritizes during the FER process. Understanding these focused regions provides valuable insights into the decision-making patterns of the model and enhances the interpretability of deep learning-based emotion recognition systems. This research proposes a MATLAB-simulation based framework for the comprehensive performance evaluation of deep learning models using multiple optimization techniques. The system facilitates the selection, training, and comparative analysis of widely used convolution neural network (CNN) architectures, including VGG16, InceptionV3, MobileNetV2, GoogLeNet, ResNet50, and DenseNet. It supports both baseline and transfer learning modes, with optimizers such as Stochastic Gradient Descent with Momentum (SGDM), ADAM, and RMSprop, selectable via intuitive push buttons. The system automatically initiates the appropriate training process based on user input and computes essential performance metrics, including accuracy, precision, recall, F1-score, sensitivity, specificity, Jaccard coefficient, Dice coefficient, true positives, true negatives, false positives, and false negatives.. The system consistently identifies ResNet50 as the most effective model across optimizers.