Real-World Nighttime Dehazing via Score-Guided Multi-Scale Fusion and Dual-Channel Enhancement

Nighttime dehazing is a challenging image restoration task owing to the presence of non-uniform illumination, artificial light sources, and haze. Existing physics-based methods struggle to effectively adapt to complex real-world nighttime haze scenarios, and learning-based approaches also fail to generalize well in such environments. In this paper, we propose a novel two-stage real-world nighttime image dehazing framework using Score-guided Multi-scale Fusion and Dual-channel Enhancement, called SMFDE. In the first stage, we first apply gamma correction and dark channel prior-based operations to generate a series of intermediate improved images. Four haze-related features are then utilized to construct a score mechanism, and binary weight maps are derived by selecting the optimal score for each pixel. Subsequently, a multi-scale fusion strategy is employed to integrate all the intermediate images based on the binary weights to yield an initial result that effectively captures rich and informative features across multiple scales. In the dual-channel enhancement stage, two primary channels of the initial result, namely brightness and saturation, are further refined to enhance details and correct colors, respectively. For brightness enhancement, we first fuse the high-frequency texture component of the input image and the initial result to provide texture-enhanced brightness. Next, the edge component is extracted by calculating the difference between the sharpened and smoothed versions of the input image. The texture-enhanced brightness and the noise-free edge component are then fused to yield the enhanced brightness. For saturation adjustment, a non-local adaptive saturation adjustment algorithm, considering color similarity, is developed to enhance color vibrancy. Experiments on real-world datasets prove that SMFDE achieves superior performance in both visual quality and objective assessments. Our code is available at https://github.com/TaoLi-TL/SMFDE.