[Paper Review] Densely Connected Pyramid Dehazing Network

He Zhang, Vishal M. Patel|arXiv (Cornell University)|Mar 22, 2018

Image Enhancement TechniquesComputer Science48 references78 citations

TL;DR

DCPDN jointly estimates transmission map, atmospheric light, and dehazed image in an end-to-end framework by embedding the atmospheric scattering model and using a densely connected encoder-decoder for transmission, a U-net for air-light, and a joint GAN discriminator.

ABSTRACT

We propose a new end-to-end single image dehazing method, called Densely Connected Pyramid Dehazing Network (DCPDN), which can jointly learn the transmission map, atmospheric light and dehazing all together. The end-to-end learning is achieved by directly embedding the atmospheric scattering model into the network, thereby ensuring that the proposed method strictly follows the physics-driven scattering model for dehazing. Inspired by the dense network that can maximize the information flow along features from different levels, we propose a new edge-preserving densely connected encoder-decoder structure with multi-level pyramid pooling module for estimating the transmission map. This network is optimized using a newly introduced edge-preserving loss function. To further incorporate the mutual structural information between the estimated transmission map and the dehazed result, we propose a joint-discriminator based on generative adversarial network framework to decide whether the corresponding dehazed image and the estimated transmission map are real or fake. An ablation study is conducted to demonstrate the effectiveness of each module evaluated at both estimated transmission map and dehazed result. Extensive experiments demonstrate that the proposed method achieves significant improvements over the state-of-the-art methods. Code will be made available at: https://github.com/hezhangsprinter

Motivation & Objective

• Motivate robust single-image dehazing under complex atmospheric scattering by jointly estimating transmission, atmospheric light, and dehazed image.
• Introduce an end-to-end architecture that embeds the physical hazing model into learning for physically-consistent results.
• Develop a densely connected encoder-decoder transmission map estimator with multi-level pyramid pooling.
• Incorporate an edge-preserving loss to maintain sharp edges and reduce halo artifacts.
• Leverage a joint GAN discriminator to exploit mutual structural information between transmission and dehazed image.

Proposed method

• Propose a pyramid densely connected transmission map estimation network to extract multi-scale features for transmission estimation.
• Estimate atmospheric light with an 8-block U-net structure.
• Recover the dehazed image using the physical model J = (I - Â(1 - t))/t.
• Introduce a joint discriminator that assesses the pair (transmission map, dehazed image) to enforce realism.
• Employ an edge-preserving loss comprising L2, two-directional gradient loss, and a feature-edge loss based on early CNN/VGG features.
• Train with a stage-wise learning strategy to stabilize optimization before joint end-to-end training.

Experimental results

Research questions

• RQ1 Can a single architecture jointly learn transmission, atmospheric light, and dehazing while strictly adhering to the haze imaging model?
• RQ2 Does a densely connected pyramid encoder-decoder improve transmission map accuracy and edge preservation?
• RQ3 Does a joint discriminator leveraging the correlation between transmission maps and dehazed images improve results?
• RQ4 Does an edge-preserving loss help maintain sharp edges and reduce halo artifacts in transmission maps?

Key findings

• The proposed DCPDN achieves significant improvements over state-of-the-art methods on synthetic datasets TestA and TestB and real-world images.
• A densely connected encoder-decoder with a multi-level pyramid pooling yields superior transmission map estimation.
• The edge-preserving loss preserves sharp edges and reduces halo artifacts in the estimated transmission map.
• A joint discriminator further refines the transmission map and dehazed image by leveraging structural correlations.
• Stage-wise training stabilizes optimization before full end-to-end training.
• Quantitative results show high SSIM gains for transmission and dehazed image on the synthetic datasets (e.g., Transition values up to 0.9776 and Image values up to 0.9560 on TestA).