Abstract Objective The purpose of this research study is to establish a practical independent drowsy eye detection method with deep learning (DL) techniques. This detection technique will assist in determining the drowsiness state of drivers by visually detecting their drowsy eye condition to provide for early warning systems and improve the safety of the vehicle and passengers in areas such as driver monitoring and the assessment of fatigue. Materials A publicly available dataset is employed in this study, which includes 49,793 images of different eye conditions captured in various settings. In preparation for model training, the dataset must be processed through initial preprocessing steps, which include image size reduction, image normalization, and augmentation to help increase the model's ability to generalize and reduce the risk of the model becoming too specialized or over-fitting. Once the dataset has been processed, it will then be divided into 70% training, 15% validation, and 15% testing subsets. Additionally, a tenfold cross-validation approach will also be used to evaluate the model and ensure its reliability. Methods The experimental framework includes the evaluation of five different DL architectures: ConvNeXt-Base, ResNeSt101, CaiT-S36, Twins-SVT-Base, and the developed architecture—DrowSFormer. In each of these architectures, the models were trained for 50 epochs with a batch size of 64 on the pre-processed dataset. The experimental framework consists of three primary stages in order to train and test the models; training phase, validation phase, and testing phase. The performance of the models was evaluated using various classification metrics to assess their effectiveness. Finally, the use of Grad-CAM (Gradient-weighted Class Activation Mapping) has been implemented as an XAI tool to provide explanations of the contribution of specific areas to the prediction made by the models. Results The experimental evaluation reveals that the proposed DrowSFormer model achieves the highest testing performance among all models, with an obtained accuracy of 99.58%. This result significantly surpasses the baseline and state-of-the-art (SOTA) models used in the study, including transformer-based and convolutional neural networks (ConvNet) based models. Conclusion This study successfully demonstrates the effectiveness of transformer-based architectures for standalone drowsy eye detection. The proposed DrowSFormer model, supported by thorough training and evaluation procedures, exhibits outstanding performance across all evaluation metrics. With the integration of Grad-CAM visualizations, the system not only delivers high accuracy but also provides interpretability, reinforcing user trust and offering insights into the model’s decision-making process.
Rajalingam et al. (Mon,) studied this question.