Enhanced CNN-Based Model for Facial Emotions Recognition in Smart Car Applications

In smart car applications, monitoring drivers involves assessing drivers' concentration, fatigue, distraction, and emotional state to ensure safe driving and accident prevention. In recent years, several research projects have focused on the impact of emotions on driving safety. Despite the models proposed and the performance achieved, FER2013, one of the well-known datasets, still presents performance challenges. This research paper proposes a CNN-based model to classify seven driver emotions using the FER2013 dataset. After rigorous pre-processing including image label correction, cleaning, balancing with SMOTE-EEN, and data augmentation, our CNN model was trained on the FER2013 dataset. The experiment's outcomes were compared to established facial emotion recognition methods developed over FER2013, revealing that the proposed model achieved an improved accuracy of 98.41%, precision of 98%, F1-score of 94%, and recall of 96%, surpassing the performance of prior research in this field. Moreover, our approach outperforms the most existing CNN pre-trained models, including MobileNet, VGG16, VGG19, ResNET50, DensNet121, DensNet169, MobileNetV3Small, and EfficientNet, all assessed using the FER2013 dataset.