Task-Oriented and Fog-Severity-Stratified Evaluation of Image Dehazing for UAV Object Detection

Abstract

Image dehazing is commonly placed before object detection under reduced visibility, yet an image that is visually closer to a clear reference is not necessarily more useful to a detector. This study presents a task-oriented and fog-severity-stratified evaluation of ten dehazing methods for unmanned aerial vehicle (UAV) object detection. One hundred held-out VisDrone scenes were transformed into ten ordered fog levels using an atmospheric scattering model and normalized relative depth, yielding 1,000 annotation-preserving foggy images. DCP, FFA-Net, DehazeNet, GCA-Net, AOD-Net, DehazeFormer, DEA-Net, BiLaLoRA, DeHamer, and CLAHE were evaluated with fixed YOLOv11m and YOLOv11n detectors trained exclusively on 5,432 clear images. Fog reduced mAP@0.5 by 41.25% for YOLOv11m and 48.15% for YOLOv11n. Eight methods improved both reported mAP metrics on both detectors, but the preferred method changed with severity and detector capacity. BiLaLoRA achieved the best dehazed YOLOv11m result (0.4156 mAP@0.5 and 0.2583 mAP@0.5:0.95) and led under heavy fog on both detectors, with mAP@0.5 gains of 0.2139 and 0.2022 over direct foggy-image detection. Method-gain rankings were strongly reproducible across detector capacities ( $\rho =0.891$ for mAP@0.5), whereas visual-quality criteria emphasized different restoration properties. FOM showed the strongest overall association with detection gain, and all four quality metrics became significant predictors under heavy fog after Holm correction. The results establish that dehazing utility is severity-, detector-, and category-dependent, and that downstream detection metrics should complement image-quality metrics when selecting preprocessing for UAV perception.