Fatigue Detection Using an Enhanced and Scalable MobileNetV2 Model

Fatigue is a psychological and physiological condition that occurs due to prolonged activity, usually resulting in decreased performance and possible safety risks. This paper introduces a fatigue detection approach using a deep learning-based model built on an enhanced MobileNetV2 model. An extensive dataset of facial expressions such as anger, fear, happiness, sadness, neutrality, and surprise is used to train the model. The proposed model performs outstandingly well in the identification of fatigue expressions. MobileNetV2 has the advantage of being lightweight in architecture, being applicable in real-time, and having high accuracy while using fewer layers. For the betterment of the model, Adam optimizer, categorical cross-entropy loss, dropout layers, batch normalization, L2 regularization, and data augmentation are employed here, all serving to better feature extraction, guard against overfitting, and guarantee stable convergence. The proposed model provides an accuracy of 76.61% and is proving itself with a macro-and weighted average F1 score of 0.77 for a balanced classification output. Confusion matrix analysis asserts high recall towards fatigue detection as 0.80 and significant precision in classifying non-fatigue as 0.78. This research identifies the effectiveness of the model in providing safety, scalability, productivity, and real-time monitoring.