Representation-Aligned Atmospheric Light Decoupling for Stable Single-Image Dehazing

초록

Single-image dehazing remains a challenging inverse imaging problem, particularly under colored haze caused by industrial pollution and complex outdoor environments. Many existing methods rely on fixed color representations or data-driven learning, which often leads to unstable reconstruction, color distortion, or limited generalization. This paper presents a model-based dehazing framework that reformulates the atmospheric scattering inversion through a haze-aware coordinate system aligned with the physical image formation model. We introduce the FLθ representation, in which one axis is explicitly aligned with the atmospheric light vector, while the orthogonal subspace captures object-dependent optical information. This construction enables an explicit decoupling of air light and scene radiance components, simplifying the structure of the inverse problem. Within this representation, atmospheric light estimation, depth estimation, and transmission recovery become mutually consistent, leading to stable reconstruction without reliance on large-scale training data or image-specific parameter tuning. Extensive experiments on synthetic datasets and real-world hazy images demonstrate that the proposed framework achieves robust dehazing performance across diverse haze colors and scene types, with improved color consistency and structural preservation compared to representative prior-based and learning-based methods. Additional evaluations on downstream segmentation tasks further show that the recovered images enhance perceptual reliability, highlighting the practical relevance of the proposed approach for real-time outdoor vision systems. © 2015 IEEE.

키워드