[Paper Review] Pyramid Fusion Dark Channel Prior for Single Image Dehazing

Qiyuan Liang, Bin Zhu+1 more|arXiv (Cornell University)|May 21, 2021

Image Enhancement Techniques22 references4 citations

TL;DR

This paper proposes Pyramid Fusion Dark Channel Prior (PF-DCP), an enhanced single image dehazing method that improves the traditional Dark Channel Prior (DCP) by fusing transmission maps across multiple image scales using a weighted fusion strategy. By eliminating the need for manual patch size selection and reducing halo artifacts, PF-DCP achieves state-of-the-art performance among traditional methods and rivals deep learning approaches, with PSNR of 23.07 and SSIM of 0.91 on RESIDE SOTS.

ABSTRACT

In this paper, we propose the pyramid fusion dark channel prior (PF-DCP) for single image dehazing. Based on the well-known Dark Channel Prior (DCP), we introduce an easy yet effective approach PF-DCP by employing the DCP algorithm at a pyramid of multi-scale images to alleviate the problem of patch size selection. In this case, we obtain the final transmission map by fusing transmission maps at each level to recover a high-quality haze-free image. Experiments on RESIDE SOTS show that PF-DCP not only outperforms the traditional prior-based methods with a large margin but also achieves comparable or even better results of state-of-art deep learning approaches. Furthermore, the visual quality is also greatly improved with much fewer color distortions and halo artifacts.

Motivation & Objective

• To address the critical limitation of patch size selection in the Dark Channel Prior (DCP), which affects transmission map accuracy and dehazing quality.
• To improve the robustness and performance of traditional dehazing methods by leveraging multi-scale image pyramids.
• To reduce color distortion and halo artifacts common in DCP-based methods through a weighted fusion strategy.
• To achieve performance comparable to deep learning models without requiring large-scale training or sacrificing interpretability.
• To provide a computationally efficient, explainable, and robust alternative to black-box deep learning models for real-world dehazing applications.

Proposed method

• Construct a multi-scale image pyramid from the input hazy image to capture features at different resolutions.
• Apply the Dark Channel Prior (DCP) independently at each pyramid level to compute transmission maps.
• Use an iterative weighted fusion strategy to combine transmission maps from different scales, with fusion weights optimized per scene type (indoor/outdoor).
• Estimate the atmospheric light and recover the haze-free image using the fused transmission map and the standard atmospheric scattering model.
• Apply a fixed fusion weight ratio of 4:1 for indoor scenes and 80:1 for outdoor scenes, derived from empirical evaluation on RESIDE training sets.
• Maintain image sharpness and reduce artifacts by preserving high-frequency details through multi-scale fusion.

Experimental results

Research questions

• RQ1 Can multi-scale fusion of transmission maps improve the accuracy and robustness of the Dark Channel Prior in single image dehazing?
• RQ2 How does the choice of patch size affect the performance of DCP, and can this dependency be eliminated through multi-scale fusion?
• RQ3 To what extent can a traditional, non-deep learning method like PF-DCP match or exceed the performance of recent deep learning-based dehazing models?
• RQ4 Does the proposed weighted fusion strategy reduce halo artifacts and color distortion compared to the original DCP?
• RQ5 Can PF-DCP achieve high-quality dehazing with minimal computational overhead and without requiring large-scale training data?

Key findings

• PF-DCP achieves a PSNR of 23.07 and SSIM of 0.91 on the RESIDE SOTS benchmark, outperforming traditional prior-based methods such as DCP, BCCR, CAP, and NLD.
• The method significantly reduces halo artifacts and color distortions, resulting in visually more natural and pleasing dehazed images compared to the original DCP.
• PF-DCP achieves performance comparable to state-of-the-art deep learning models like GFN and EPDN, despite being a non-trainable, interpretable method.
• The fusion weight ratio of 4:1 for indoor scenes and 80:1 for outdoor scenes yields optimal performance, demonstrating scene-specific sensitivity in fusion strategy.
• The computational cost of PF-DCP increases by only 30% on average compared to DCP, making it efficient and practical for real-time applications.
• PF-DCP maintains high image sharpness and detail preservation, avoiding the blurring common in many prior-based and deep learning-based dehazing methods.