Multi-Branch Non-Homogeneous Image Dehazing via Concentration Partitioning and Image Fusion

Existing single image dehazing methods falter on non-homogeneous haze—images where haze density varies across regions and transitions abruptly. To address this, researchers from an undisclosed institution (Yingming Zhang, Wuqi Su, Qing Xiao, Yonggang Yang) propose CPIFNet (Concentration Partitioning and Image Fusion Network), detailed in arXiv:2605.00885. The key insight: a non-homogeneous hazy image can be viewed as a composite of local regions each with approximately homogeneous haze. CPIFNet uses a two-stage architecture. Stage one employs multiple Image Enhancement Network (IENet) branches, each independently trained on homogeneous haze datasets of different concentration levels. Each branch specializes in restoring regions matching its haze density. Stage two uses an Image Fusion Network (IFNet) that aggregates advantageous regions from all IENet outputs via deep feature stacking and merging, producing a unified high-quality dehazed result.

The system is supervised by a comprehensive loss function combining reconstruction, perceptual, structural similarity, and color losses to jointly optimize both stages. While no quantitative benchmarks are provided in the abstract, the approach promises to outperform classical methods on challenging real-world haze with uneven density—common in foggy scenes, smoke, or underwater imaging. The paper is listed under Computer Vision and Pattern Recognition and is available on arXiv. This represents a practical step toward robust dehazing in uncontrolled environments, where uniform haze assumptions break down.