Source-linked AI summary
Densely Connected Pyramid Dehazing Network
He Zhang, Vishal M. Patel
TL;DR
Single-image dehazing must recover clear scenes from haze while estimating transmission and atmospheric light, and prior learning-based methods remain limited by inaccurate or separately optimized components. DCPDN embeds the atmospheric scattering model into an end-to-end network, combining dense edge-preserving transmission estimation with atmospheric-light prediction and joint adversarial refinement. The paper reports significant improvements over state-of-the-art methods and uses ablations to assess its modules.
Problem
Learning-based dehazing remains limited by transmission-estimation errors and disjoint optimization of transmission, atmospheric light, and dehazed images.
Method
DCPDN embeds the atmospheric scattering model into an end-to-end network that jointly estimates transmission, atmospheric light, and the dehazed image.
Results
The authors report significant improvements over state-of-the-art methods across experiments and evaluate module contributions through ablation studies.
Takeaways & Limitations
The method provides a physics-model-guided framework for jointly optimizing the principal components of single-image dehazing.
Takeaways & Limitations
The method assumes a homogeneous atmospheric-light map, and one compared implementation could not estimate the transmission map.
Abstract
from arXiv · showhide
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
1. Introduction
DCPDN addresses haze-related image degradation with a jointly optimized network that follows the atmospheric scattering model. It combines transmission estimation, atmospheric-light prediction, physics-based dehazing, edge preservation, dense multilevel features, and joint adversarial learning.
- Severe haze absorbs and scatters light, degrading image quality and potentially affecting classification and detection systems.
- Existing learning-based methods remain suboptimal because transmission errors degrade dehazed images and disjoint optimization misses relationships among key variables.
- DCPDN embeds the atmospheric scattering model into the network to jointly estimate the transmission map, atmospheric light, and dehazed image.
- A stage-wise training strategy progressively optimizes network components before jointly optimizing the entire network.
- The transmission estimator uses an edge-preserving densely connected encoder-decoder with multi-level pooling and a dedicated edge-preserving loss.
- A joint GAN discriminator evaluates paired transmission maps and dehazed images to exploit their correlated structural information.
2. Related Work
Single-image dehazing is highly ill-posed, motivating both handcrafted prior-based and learning-based approaches. Prior work includes scene-albedo, contrast, factorial-MRF, and dark-channel methods, while GANs provide adversarial learning for realistic image synthesis and other vision tasks.
- Single-image dehazing is highly ill-posed, so researchers have developed handcrafted prior-based and learning-based methods.
- Handcrafted methods estimate scene properties using albedo models, patch-based contrast maximization, factorial MRFs, or dark-channel priors.
- GANs learn training-image distributions through adversarial optimization and have been applied to synthesis, image translation, and low-level vision.
3. Proposed Method
DCPDN jointly estimates the transmission map, atmospheric light, and dehazed image within an atmospheric-scattering-model-based network. Its transmission estimator combines dense multi-level features, pyramid pooling, edge-preserving losses, and joint-discriminator refinement, while stage-wise learning supports optimization.
- Architecture: DCPDN contains transmission-map estimation, atmospheric-light estimation, model-based dehazing, and a joint discriminator.The atmospheric-scattering relation is embedded into the optimization framework to connect the three estimation tasks.
- Transmission Map Estimation: The transmission estimator uses a densely connected encoder-decoder to combine features from multiple CNN layers.Dense connections maximize information flow across layers and support convergence.
- Transmission Map Estimation: Multi-level pyramid pooling embeds features from different scales into the final transmission-map estimate, improving global object structure.Features from four levels are upsampled and concatenated with the original feature before final estimation.
- Atmospheric Light Estimation: The atmospheric-light network predicts a spatially uniform atmospheric-light map using an eight-block U-net.The uniform-map assumption assigns the same value to every pixel for a given image.
- Joint Discriminator Learning: A joint discriminator evaluates corresponding transmission-map and dehazed-image pairs to exploit their structural correlation.The predicted transmission map and dehazed image are treated as a pair and compared with corresponding ground truths.
- Edge-preserving Loss: The edge-preserving loss combines L2, horizontal and vertical gradient, and feature losses to retain sharp transmission-map edges.Gradient operators and low-level VGG-16 features are used as edge extractors.
- Stage-wise Learning: Stage-wise learning first optimizes tasks separately and then fine-tunes the full network jointly.This strategy addresses slow convergence attributed to gradient diffusion among different tasks.
4. Experimental Results
Experiments on synthetic and real-world hazy images evaluate DCPDN against prior methods and isolate the contributions of its architectural and loss components. Results report effective module contributions, competitive synthetic performance, and visually improved real-image dehazing, with one implementation-related scope note.
- Experimental setup: The study compares DCPDN with five state-of-the-art methods on two synthetic datasets and real-world hazy images, alongside an ablation study.The synthetic evaluation uses TestA and TestB; real images come from previous authors and Internet sources.
- Implementation: The reported training setup uses ADAM, a 2 × 10^-3 learning rate, batch size 1, 512 × 512 inputs, and 400000 iterations.The parameters were selected through cross-validation.
- Ablation study: The ablation configurations progressively add dense connectivity, multi-level pyramid pooling, gradient loss, edge-preserving loss, and the joint discriminator.DED and DED-MLP use only L2 loss, while DCPDN combines all listed components.
- Ablation study: Multi-level pooling preserves global structure, edge-preserving loss refines transmission-map edges, and the joint discriminator captures fine structural details.The reported fine details include small objects in the table; quantitative results on TestA and TestB also support module effectiveness.
- Synthetic evaluation: On synthetic images, DCPDN preserves sharper contours, reduces color distortion, and appears closer to ground truth than compared methods.Prior methods are described as sometimes over-dehazing or under-dehazing, producing darker results or retaining haze; quantitative results are reported in Tables 2 and 3.
- Evaluation caveat: A comparison-table footnote notes that one code release cannot estimate the transmission map.This is recorded as an N/A condition in the quantitative comparison.
- Real-world evaluation: On real images from previous methods, DCPDN removes haze while maintaining realistic colors, whereas competing methods may retain haze, darken regions, or create color shifts.The comparison covers four sample images in Figure 8 and additional Internet images in Figure 9, where competing outputs show color distortions.
5. Conclusion
The paper presents DCPDN as an end-to-end dehazing method that jointly optimizes transmission map, atmospheric light, and dehazed image. It combines physics-model embedding, a densely connected pyramid estimator with edge-preserving loss, and a joint discriminator, with experiments supporting the proposed modules.
- Conclusion: DCPDN jointly optimizes the transmission map, atmospheric light, and dehazed image within an end-to-end framework.The atmospheric image degradation model is embedded directly into the optimization framework.
- Conclusion: The method estimates transmission maps with a densely connected encoder-decoder and multi-level pooling module optimized by an edge-preserving loss.These components are presented as the efficient transmission-map estimation design.
- Conclusion: A joint-discriminator GAN framework refines details by leveraging structural correlation between the dehazed image and estimated transmission map.The conclusion describes this as an additional component of the proposed method.
- Conclusion: Experiments were conducted to demonstrate the significance of the proposed method.