Multi-Scale Single Image Dehazing Using Laplacian and Gaussian Pyramids

Zhengguo Li, Haiyan Shu, Chaobing Zheng

TL;DR

This paper introduces a multi-scale dehazing algorithm that tackles two persistent "white whales" of image restoration: the ambiguity between object color and haze, and the ugly noise amplification typically seen in dehazed sky regions. By combining a new Dark Direct Attenuation Prior (DDAP) with Laplacian and Gaussian pyramid decomposition, the authors achieve state-of-the-art results that preserve fine details while keeping the sky clean.

What Problem Does This Paper Address?

Single image dehazing is an "ill-posed" problem. Mathematically, we follow Koschmieder’s law, which models a hazy image as a combination of the true scene radiance (the "haze-free" image) and the atmospheric light (the "haze"). Because we only have one input image but multiple unknowns per pixel, we have to rely on "priors"—statistical assumptions about what natural, haze-free images look like.

The authors identify two major gaps in existing model-based methods:

  1. Ambiguity: It is often difficult to distinguish between a bright/white object and the haze itself. Popular priors like the Dark Channel Prior (DCP) often fail in bright regions or sky areas, leading to "morphological artifacts" (weird halos or blocks).
  2. Noise Amplification: In regions with heavy haze (like the sky), the transmission signal is very weak. When you try to "undo" the haze, you effectively divide by a very small number, which acts as a massive volume knob for sensor noise, resulting in grainy, unappealing skies.

Methods

The researchers propose a three-pronged strategy to solve these issues:

  • Dark Direct Attenuation Prior (DDAP): Instead of assuming a local patch of the final image has a dark pixel (as DCP does), DDAP assumes the direct attenuation (the part of the light that actually reaches the camera from the object) is dark. This is a more flexible constraint that holds true even in sky regions where DCP typically breaks down.
  • Haze Line Averaging: To fix the blocky artifacts that priors often create, the authors use "haze lines." This concept groups pixels that share the same intrinsic color but exist at different depths. By performing non-local averaging across these lines, they can refine the transmission map without the massive "blurring" radius required by older filters, allowing them to keep fine details like hair or distant branches sharp.
  • Multi-Scale Pyramid Decomposition: This is the "divide and conquer" step. The image is decomposed into a Gaussian pyramid (low-frequency, smooth components) and a Laplacian pyramid (high-frequency, edgy components).
  • The low-pass components (where the haze signal is strongest) are dehazed.
  • The high-pass components (where the noise lives) are treated with a specialized noise-reduction approach.
  • Finally, the levels are "collapsed" back together to form the final image.

Datasets

The authors evaluated their method using:

  • A 29-image real-world dataset: Used for psychophysical (human preference) testing with 12 students.
  • A 79-image composite dataset: Including 31 images from the RESIDE dataset, plus 48 images from other benchmarks and the internet.
  • Vehicle Detection: They also tested the algorithm's utility as a pre-processing step for computer vision tasks using a vehicle detection benchmark.

Key Insights and Ideas

  • The Sky is the Limit: Most dehazing papers treat the sky as a nuisance to be masked out. This paper treats it as a signal-to-noise ratio problem. By recognizing that noise is a high-frequency phenomenon, they can suppress it in the Laplacian pyramid while still removing haze from the Gaussian base.
  • Smaller is Better for Filters: Traditional dehazing uses a Guided Image Filter (GIF) with a large radius (e.g., 60 pixels) to smooth out artifacts. The authors show that by using haze line averaging first, they can use a much smaller radius (25 pixels), which prevents the "over-smoothing" that usually kills fine textures.
  • Physics-Based vs. Data-Driven: While deep learning (CNNs/GANs) is popular, the authors argue that model-based approaches are more "photo-realistic" because they respect the physics of light scattering, whereas GANs might hallucinate textures that weren't there.

Key Results

  • Quantitative: The proposed method achieved the highest average DHQI (Dehazing Quality Index) score of 60.97, outperforming SOTA models like FFA-Net (55.33) and MSBDN (54.32).
  • Qualitative: In side-by-side comparisons (see Fig. 6 in the paper), the proposed method shows significantly less grain in the sky and sharper edges on foreground objects compared to DehazeNet and DCP.
  • Downstream Impact: Using this dehazing method as a pre-processor significantly improved the accuracy of a Faster R-CNN vehicle detector on hazy highway footage.

Limitations and Future Work

The primary limitation is the parameter $\eta$, which controls the "strength" of the dehazing. The authors note that $\eta$ is adaptive to the haze degree (e.g., 1/4 for normal haze, 1/8 for heavy haze). However, in the current version, this still requires some level of manual selection or prior knowledge of the haze density.

Future work could involve automatically estimating the optimal $\eta$ value from the image statistics to make the algorithm fully autonomous for mobile or real-time applications. Additionally, while the pyramid approach is efficient, the authors suggest it could be further optimized for low-power mobile hardware.

Summary updated

Abstract

Model-based single image dehazing was widely studied due to its extensive applications. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ambiguity between object radiance and haze</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">noise amplification in sky regions</i> are two inherent problems of model-based single image dehazing. In this paper, a dark direct attenuation prior (DDAP) is proposed to address the former problem. A novel haze line averaging is proposed to reduce the morphological artifacts caused by the DDAP which enables a weighted guided image filter with a smaller radius to further reduce the morphological artifacts while preserve the fine structure in the image. A multi-scale dehazing algorithm is then proposed to address the latter problem by adopting Laplacian and Gaussian pyramids to decompose the hazy image into different levels and applying different haze removal and noise reduction approaches to restore the scene radiance at the different levels. The resultant pyramid is collapsed to restore a haze-free image. Experiment results demonstrate that the proposed algorithm outperforms state-of-the-art dehazing algorithms.
BibTeX
@article{li2021multi,
  author = {Zhengguo Li and Haiyan Shu and Chaobing Zheng},
  title = {{Multi-Scale Single Image Dehazing Using Laplacian and Gaussian Pyramids}},
  journal = {IEEE Trans. Image Process.},
  year = {2021},
  url = {https://openreview.net/forum?id=nc129qmHyW},
}

Loading related content...

Figure