Data Composition Strategy for FFA-Net Image Dehazing
Chae-bin Park, Hyo-jae Lee, Jae‐Ho Nah · Journal of Multimedia Information System · 2026
We propose a generalized training strategy for FFA-Net to achieve robust image restoration in diverse haze conditions. By combining synthetic and real-world datasets (e.g., ITS, OTS, Haze4K, and NH-HAZE), we enhance restoration stability across varying atmospheric scenarios. Furthermore, we leverage performance benchmarks to motivate the choice of FFA-Net alongside detailed dataset overviews. Experimental results demonstrated that the proposed data composition strategy—integrating Haze4K and NH-HAZE at a 1:1 ratio—significantly improves cross-domain generalization and stabilizes dehazing performance under non-homogeneous real-world haze. This provides a practical and generalizable solution for reliable real-world dehazing without requiring manual adjustment according to input haze conditions.