AI-Based Driver Drowsiness Detection Using Image Enhancement and Swin Transformer

Abstract

This study introduces a smart system designed to detect driver drowsiness in real time, focusing mainly on checking whether a driver's eyes are closed for an unusually long period. While conventional AI image detection methods (like standard CNNs) often fail when faced with poor lighting, motion blur, noisy images, or deliberate visual distortions (adversarial attacks), this solution solves these issues by combining two advanced techniques: Diffusion-Based Denoising (U-Net Architecture): Acts as a preprocessing cleanup layer. Before analyzing the image, it strips away visual noise, blur, and distortion, turning corrupted camera frames into clear, high-quality images. Swin Transformer Model: Serves as the primary classification engine. It uses a multi-scale, window-based attention mechanism to accurately detect eye states (open vs. closed) across both fullface images and close-up eye patches. To make the AI more resilient to unpredictable driving conditions, the researchers trained it using artificial image modifications (like rotation, zooming, and shifting) along with simulated visual attacks.