Driver Drowsiness Detection Using Swin Transformer and Diffusion Models for Robust Image Denoising

Abstract

With the rapid development of intelligent transportation systems and growing emphasis on driver safety, real-time detection of driver drowsiness has become a critical area of research. This study presents a robust and scalable driver drowsiness detection framework that integrates a Swin Transformer-based deep learning model with a diffusion model for image denoising. While conventional convolutional neural networks (CNNs) are effective in standard vision tasks, they often suffer performance degradation in real-world driving scenarios due to noise, poor lighting, motion blur, and adversarial attacks. To address these challenges, the proposed model focuses on eye-state detection, specifically, prolonged eye closure, as a primary indicator of driver disengagement and fatigue. Our system introduces a novel preprocessing stage using a denoising diffusion model built on a U-Net encoder-decoder architecture, effectively mitigating the impact of Gaussian noise and adversarial perturbations. Additionally, we incorporate adversarial training with Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks, demonstrating significant improvements in classification accuracy and resilience. Evaluations are conducted on two benchmark datasets, Eye-Blink and Closed Eyes in the Wild (CEW), under both clean and noisy conditions. Comparative experiments show that the proposed system outperforms several state-of-the-art models, including ViT, ResNet50V2, InceptionV3, MobileNet, DenseNet169, and VGG19, in terms of accuracy (up to 99.82%), PSNR (up to 41.61 dB), and SSIM (up to 0.984), while maintaining competitive inference times suitable for practical deployment. Moreover, a detailed sensitivity analysis of data augmentation strategies reveals that techniques such as rotation and horizontal flip substantially enhance the model’s generalization across variable visual inputs. The system also demonstrates improved robustness under real-world black-box scenarios and adversarial conditions. While this study primarily targets static image datasets, preliminary evaluations on dynamic video frames suggest potential for real-time monitoring applications. Overall, this research delivers a high-performing driver monitoring system capable of real-time drowsiness detection, even under adverse visual conditions. It lays a strong foundation for future extensions, including temporal modeling, real-time deployment, and multimodal integration (e.g., combining visual input with physiological signals such as EEG and heart rate) to further enhance driver safety and awareness in smart vehicles.