CNN-swarm intelligence hybrid model for facial expression recognition

Emotion is essential to our social interactions, as it allows us to communicate with others effectively. Emotion is expressed through facial expressions, voice prosody, and body language. As technology advances rapidly to facilitate human connection, the capac- ity to discern emotions through facial expressions, voice prosody, and body language has markedly improved. Facial expressions serve as non-verbal indicators for interpreting emotions. Recent research indicates that models utilising extensive datasets and deep learning architectures, such as Convolutional Neural Networks (CNN) and Vision Transformers, have strong performance on benchmarks for face expression datasets. Nevertheless, significant constraints exist when representing emotions through facial expressions (i.e., image or video-based). The variance and illumination significantly impact the model’s performance. This study seeks to examine the effectiveness of combining a conventional optimisation technique (i.e., gradient-based optimisation) with a metaheuristic search approach (i.e., swarm intelligence) to improve model performance. An inception-based architecture is utilised to model emotion recognition from facial cues. A hybrid optimisation approach that integrates gradient-based and swarm intelli- gence techniques is employed to improve the architectures. All the offered models demonstrate significantly enhanced performance relative to the baseline. Model B (20-5) attained a training accuracy of 99.15%, a validation accuracy of 100%, a training loss of 0.0934, and a validation loss of 0.0402.