Fire Segmentation in Complex Environments: Learning Representations for Reflection, Haze, and Day-Night Adaptation

This paper presents an efficient flame segmentation framework using RGB images and a lightweight U-Net architecture. The method incorporates color correction, haze removal, reflection elimination, and day–night adaptation to enhance flame features under complex conditions such as smoke and lighting variations. Unlike prior RGB-D or thermal-based approaches, our model operates on single RGB input and achieves real-time performance. Experimental results on real-world fire datasets show that the proposed method outperforms existing models, achieving up to 22.8% higher IoU and 24.2% higher Boundary IoU, demonstrating strong accuracy–speed balance for practical applications such as CCTV fire detection and drone-based monitoring.