Atmospheric scattering by fine particulates degrades visibility and impairs object detection. We address this by introducing a DCP-based lightweight U-Net as a preprocessing module cascaded with YOLOv8x. Using the NH-Haze dataset, we evaluate performa...
Atmospheric scattering by fine particulates degrades visibility and impairs object detection. We address this by introducing a DCP-based lightweight U-Net as a preprocessing module cascaded with YOLOv8x. Using the NH-Haze dataset, we evaluate performance at confidence thresholds of conf = {0.25, 0.50, 0.75} At conf=0.25, true positives increased by 27.9% and false detections decreased by 41.2% relative to the Noisy input, demonstrating that the DCP-based lightweight U-Net preprocessing improves detection robustness in foggy conditions.