FWB-Net: Front White Balance Network for Color Shift Correction in Single Image Dehazing Via Atmospheric Light Estimation

Cong Wang, Yan Huang, Yuexian Zou, Yong XuarXiv 2021 · DOI · arXiv

TL;DR

This paper introduces FWB-Net, a novel deep learning approach for single image dehazing that tackles the persistent problem of color shift. By proposing a non-homogeneous atmospheric scattering model and a "front white balance" module, FWB-Net effectively corrects color distortions before generating the final dehazed image, leading to more visually accurate and natural results.

What Problem Does This Paper Address?

Single image dehazing aims to remove atmospheric haze from a photograph to reveal the underlying scene. Many state-of-the-art methods rely on the Atmospheric Scattering Model (ASM), which typically assumes a uniform atmospheric light factor across the entire image. However, real-world scenes often have non-uniform illumination (e.g., shadows, multiple light sources), causing this assumption to break down. This mismatch leads to a common and visually unappealing artifact: color shift in the dehazed output. Existing methods struggle to fully eliminate this color distortion, especially in complex real-world scenarios.

Methods

The authors propose a multi-pronged approach:

  • Non-Homogeneous Atmospheric Scattering Model (NH-ASM): They extend the traditional ASM by making the atmospheric light factor ($A$) position-dependent ($A(x, y)$). This allows for a more accurate modeling of complex and non-uniform lighting conditions.
  • Front White Balance Module (FWB-Module): This is the core innovation. Instead of performing white balance after dehazing (which is standard practice), this U-Net based module estimates the pixel-wise atmospheric light map and uses it to correct color shifts before the final dehazing step. This "front" approach aims to prevent color distortions from propagating through the dehazing process.
  • FWB-Loss: A novel loss function is introduced specifically for training the FWB-Module. It operates in the perceptually uniform CIELAB color space and penalizes deviations between the estimated and ground truth atmospheric light, directly targeting color shift.
  • FWB-Net: An end-to-end CNN architecture that integrates the NH-ASM, FWB-Module, and an existing transmission estimation module (ETM). The network is trained to produce haze-free images while simultaneously restraining color shifts.

Datasets

  • NYU-Depth-v2: Used to synthesize a dataset of hazy images by applying the NH-ASM with varying atmospheric light factors and transmission maps derived from depth information.
  • O-Haze: A real-world hazy image dataset used for evaluating the method on authentic challenging scenarios.

Key Insights and Ideas

  • Color Shift as an Illumination Problem: The paper frames color shift not just as a dehazing artifact, but as a direct consequence of inaccurate atmospheric light estimation, particularly under non-uniform lighting.
  • "Front" White Balance is Key: Performing color correction before the main dehazing process, rather than after, is shown to be more effective in preventing color distortions from becoming ingrained in the final output.
  • CIELAB for Perceptual Accuracy: Utilizing the CIELAB color space for the FWB-Loss function is crucial because it better reflects human perception of color differences, leading to more visually pleasing results.
  • Modularity and Portability: The FWB-Module is demonstrated to be effective not only within the proposed FWB-Net but also when applied to other classical dehazing methods, highlighting its versatility.

Key Results

  • Superior Performance: FWB-Net achieves state-of-the-art results on both synthetic (NYU-Depth) and real-world (O-Haze) datasets, outperforming several established dehazing methods in terms of PSNR, SSIM, and CIEDE2000.
  • Effective Color Shift Restraint: Qualitative results show FWB-Net producing clean, haze-free images with significantly less color shift compared to other methods, which often exhibit blueish or unnatural color casts.
  • Improved Existing Methods: Applying the FWB-Module to classical dehazing algorithms like DCP, Berman, and DehazeNet also leads to noticeable improvements in color shift correction.

Limitations and Future Work

  • Dataset Synthesis: While the synthesized dataset uses depth information, it's still a synthetic representation of real-world haze. The authors acknowledge the scarcity of real-world paired hazy/clean datasets.
  • Computational Cost: While not explicitly detailed as a limitation, deep learning models, especially those with U-Net architectures, can be computationally intensive.
  • Future Work: The paper suggests that the FWB-Module's portability makes it a valuable component for improving other dehazing algorithms. Further exploration of its integration into more complex dehazing pipelines could be a future direction.

Summary updated

Abstract

In recent years, single image dehazing deep models based on Atmospheric Scattering Model (ASM) have achieved remarkable results. But the dehazing outputs of those models suffer from color shift. Analyzing the ASM model shows that the atmospheric light factor (ALF) is set as a scalar which indicates ALF is constant for whole image. However, for images taken in real-world, the illumination is not uniformly distributed over whole image which brings model mismatch and possibly results in color shift of the deep models using ASM. Bearing this in mind, in this study, first, a new non-homogeneous atmospheric scattering model (NH-ASM) is proposed for improving image modeling of hazy images taken under complex illumination conditions. Second, a new U-Net based front white balance module (FWB-Module) is dedicatedly designed to correct color shift before generating dehazing result via atmospheric light estimation. Third, a new FWB loss is innovatively developed for training FWB-Module, which imposes penalty on color shift. In the end, based on NH-ASM and front white balance technology, an end-to-end CNN-based color-shift-restraining dehazing network is developed, termed as FWB-Net. Experimental results demonstrate the effectiveness and superiority of our proposed FWB-Net for dehazing on both synthetic and real-world images.
BibTeX
@misc{wang2021fwb,
  author = {Cong Wang and Yan Huang and Yuexian Zou and Yong Xu},
  title = {{FWB-Net: Front White Balance Network for Color Shift Correction in Single Image Dehazing Via Atmospheric Light Estimation}},
  year = {2021},
  eprint = {2101.08465},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2101.08465},
}
Figure