Enhancing Deep Learning-Based Human Detection in Flooded Areas Via Dynamic Lighting-Adaptive Image Enhancement

Flood zones pose significant challenges for search and rescue (SAR) operations due to poor visibility, noise, reflections, and occlusions caused by adverse weather and low-light conditions. These factors degrade object detection performance. To address this, the proposed system integrates YOLOv11 with a Retinex-based image enhancement method. The Retinex theory separates illumination from reflectance, improving detail extraction in low-light images. Two publicly available datasets, the Roadway Flooding Image Dataset and the Human Detection in Floods Dataset covering varied flood scenarios, lighting, and occlusions, are used. Results show that Retinex enhancement significantly improves detection accuracy: YOLOv11 reached 95.72%, outperforming baselines by up to 9%–10% in recall and 5%–6% in F1-score. Enhanced models also showed 7%–10% robustness gains on occluded and reflection-heavy images. This demonstrates that the Retinex-enhanced YOLOv11 model provides reliable, real-time human detection in flood environments, offering valuable tools for life-saving SAR efforts.