Optimized Ensemble Framework Using Levy-Enhanced Particle Swarm Optimization for Facial Emotion Recognition

Facial Emotion Recognition (FER) remains a significant challenging task due to the complexity of classifying emotions and the variation in facial expressions. To address this issue, we propose an ensemble learning approach aimed at improving the accuracy and efficiency of FER systems. To solve the problem of effectively capturing diverse features from facial expressions, we integrate three high-performance Convolutional Neural Networks (CNNs) including DenseNet121, MobileNetV1, and VGG16, into a unified model. To enhance the feature extraction process, we incorporate Naïve Inception blocks with each pre-trained model. This is critical because these blocks enable multi-scale feature extraction by allowing the model to capture both fine-grained and global patterns in facial expressions, which is essential for accurate emotion classification under varying conditions. Additionally, we optimize the ensemble model using a Particle Swarm Optimization (PSO) algorithm, enhanced by Levy Flight, which intelligently assigns weights to each base model within the ensemble by ensuring optimal and improved classification performance. The fitness function evaluates the most relevant features for recognizing seven distinct emotions by leading to more accurate predictions. Experimental results on the JAFFE dataset demonstrate that the proposed approach achieves an accuracy of 97.62%, showcasing the potential of combining optimization techniques with CNNs to build robust and scalable FER systems for real-world applications.