A Semi-supervised Nighttime Dehazing Baseline with Spatial-Frequency Aware and Realistic Brightness Constraint

TL;DR

This work tackles nighttime single-image dehazing by addressing localized glow, haze, and frequency-inconsistent distortions as well as brightness realism gaps between synthetic and real data. It introduces SFSNiD, a semi-supervised baseline that fuses spatial and frequency information through a Spatial-Frequency Information Interaction (SFII) module and leverages pseudo-label-based retraining along with a local brightness loss to enforce realistic brightness. The SFII framework uses components such as Fourier-domain spectral filtering, Density/Attention-inspired Bidomain modules, and dual-domain losses ($L_G$ in the spatial domain and $L_F$ in the Fourier domain), plus a local brightness constraint $L_B$ with a brightness coefficient $$ to align dehazed outputs with real-world brightness. Experimental results on GTA5, UNREAL-NH, NightHaze, and RWNH demonstrate superior quantitative performance and improved brightness realism, with code released for reproducibility.

Abstract

Existing research based on deep learning has extensively explored the problem of daytime image dehazing. However, few studies have considered the characteristics of nighttime hazy scenes. There are two distinctions between nighttime and daytime haze. First, there may be multiple active colored light sources with lower illumination intensity in nighttime scenes, which may cause haze, glow and noise with localized, coupled and frequency inconsistent characteristics. Second, due to the domain discrepancy between simulated and real-world data, unrealistic brightness may occur when applying a dehazing model trained on simulated data to real-world data. To address the above two issues, we propose a semi-supervised model for real-world nighttime dehazing. First, the spatial attention and frequency spectrum filtering are implemented as a spatial-frequency domain information interaction module to handle the first issue. Second, a pseudo-label-based retraining strategy and a local window-based brightness loss for semi-supervised training process is designed to suppress haze and glow while achieving realistic brightness. Experiments on public benchmarks validate the effectiveness of the proposed method and its superiority over state-of-the-art methods. The source code and Supplementary Materials are placed in the https://github.com/Xiaofeng-life/SFSNiD.

Figures (9)

• Figure 1: Visualization of real-world dehazed images, where the "IM-" and "GE-" denote the dehazed results obtained by training on imaging model (IM) and game engine (GE) simulated datasets, respectively. The curve figure represents the pixel histogram, where the $x$ and $y$ coordinates represent the pixel values and corresponding numbers, respectively. The $x$ and $y$ coordinates of the bar figure represent the color channel and the corresponding average pixel value, respectively.
• Figure 2: The overall pipeline of the proposed SFSNiD.
• Figure 3: The sub-modules of the proposed SFII.
• Figure 4: The overall architecture of the proposed SFII.
• Figure 5: Visual results on synthetic dataset liu2023nighthazeformer.