Lightweight Deep Learning Models for Facial Expression Recognition in Inclusive Education

Abstract

Facial expression recognition is an essential component in the development of artificial intelligence-based learning systems, particularly in the context of inclusive education that involves students with special needs. This study aims to evaluate the performance of several lightweight deep learning architectures in detecting facial expressions with high accuracy while maintaining computational efficiency. Facial image data were obtained from both public datasets and newly collected samples, which were preprocessed through face cropping, normalization, and data augmentation. The dataset was split into 70% training, 15% validation, and 15% testing. Four lightweight deep learning architectures: MobileNetV2, MobileNetV3 (Small and Large), and EfficientNetB0, were employed as the primary models using transfer learning and fine-tuning approaches. Evaluation was conducted using accuracy, loss, precision, recall, and F1-score metrics, complemented by visualization through confusion matrices. The results indicate that MobileNetV2 achieved the best performance with a test accuracy of 92%, precision of 93%, recall of 91%, and F1-score of 92%, while maintaining a relatively lightweight parameter size of 2.26 million. EfficientNetB0 ranked second with 83% accuracy, followed by MobileNetV3-Large (77%), whereas MobileNetV3-Small demonstrated the lowest performance (45%). Confusion matrix analysis revealed recurring misclassification patterns for certain expressions, such as Happy often misclassified as Sad, and Neutral overlapping with Angry. This study confirms that MobileNetV2 is the most optimal architecture for implementing facial expression recognition systems in inclusive education environments, as it balances high accuracy with computational efficiency. These findings provide a solid foundation for developing intelligent applications that support adaptive interaction in the learning process..

GS Cited Analysis