Source-linked AI summary
Gated Context Aggregation Network for Image Dehazing and Deraining
Dongdong Chen, Mingming He, Qingnan Fan, Jing Liao, Liheng Zhang, Dongdong Hou, Lu Yuan, Gang Hua
TL;DR
Image dehazing must recover clean content despite unknown atmospheric transmission and light, which makes the task ill-posed. The paper proposes GCANet, an end-to-end gated context aggregation network using smoothed dilation and gated feature fusion. It reports large-margin improvements over prior dehazing methods and state-of-the-art performance on deraining, while noting that better losses remain future work.
Problem
Image dehazing is an ill-posed recovery problem because atmospheric transmission and global atmospheric light are unknown.
Method
GCANet directly restores haze-free content with smoothed dilated convolutions and a gated subnetwork that fuses features from different levels.
Results
GCANet outperforms previous state-of-the-art image dehazing methods by a large margin and also achieves state-of-the-art performance on image deraining.
Takeaways & Limitations
The same end-to-end network demonstrates reported effectiveness across both image dehazing and image deraining tasks.
Takeaways & Limitations
The paper identifies exploring more advanced losses as future work, despite state-of-the-art performance with its simple loss.
Abstract
from arXiv · showhide
Image dehazing aims to recover the uncorrupted content from a hazy image. Instead of leveraging traditional low-level or handcrafted image priors as the restoration constraints, e.g., dark channels and increased contrast, we propose an end-to-end gated context aggregation network to directly restore the final haze-free image. In this network, we adopt the latest smoothed dilation technique to help remove the gridding artifacts caused by the widely-used dilated convolution with negligible extra parameters, and leverage a gated sub-network to fuse the features from different levels. Extensive experiments demonstrate that our method can surpass previous state-of-the-art methods by a large margin both quantitatively and qualitatively. In addition, to demonstrate the generality of the proposed method, we further apply it to the image deraining task, which also achieves the state-of-the-art performance. Code has been made available at https://github.com/cddlyf/GCANet.
1. Introduction
Image dehazing restores clean content lost to atmospheric degradation, but estimating physical corruption factors is difficult. GCANet addresses this with smoothed dilation and gated multilevel fusion, achieving strong dehazing and deraining results.
- Motivation: Atmospheric turbidity degrades image contrast and saturation, disrupting vision systems designed for clean capture conditions.Dehazing is studied as preprocessing to recover clean images from corrupted inputs.
- Problem formulation: The physical model expresses a hazy image through clean radiance, atmospheric light, and a spatially varying transmission map.Transmission depends on unknown depth information.
- Existing approaches: Traditional methods estimate transmission or atmospheric light using handcrafted image priors such as local contrast and dark channels.These approaches compensate for information lost during atmospheric corruption.
- Proposed method: GCANet is an end-to-end dehazing network using smoothed dilated convolution to reduce gridding artifacts and gated fusion to combine feature levels.The gated subnetwork assigns importance weights to features from different levels.
- Experiments: GCANet outperforms previous state-of-the-art dehazing methods qualitatively and quantitatively, with ablations evaluating component importance.Evaluation uses the RESIDE benchmark and compares methods using PSNR and SSIM.
- Experiments: Applying GCANet to image deraining also outperforms previous state-of-the-art methods, supporting the network’s generality.The paper presents deraining as a second restoration task for the proposed architecture.
2. Related Work
Dehazing research has progressed from handcrafted-prior methods to learning-based networks that learn image statistics from data. GCANet differs by directly regressing the residual between hazy and clean images with a lightweight architecture.
- Background: Single-image dehazing is an ill-posed inverse problem because transmission and global atmospheric light are unknown.Prior-based and learning-based methods differ mainly in whether image priors are handcrafted or learned automatically.
- Prior-based methods: Traditional methods impose image-statistics priors as constraints to compensate for information lost during atmospheric corruption.Examples include albedo estimation, dark channel priors, and local-contrast optimization.
- Learning-based methods: Learning-based methods use large datasets and neural networks to learn priors, with differences in targets and network structures.Existing networks may predict transmission, combine physical-model variables, or directly produce clean images.
- GCANet: GCANet directly regresses the residual between hazy and clean images rather than predicting transmission or atmospheric light.The paper describes its architecture as lightweight while reporting better results than previous methods.
3. Method
GCANet uses smoothed dilated residual blocks and gated multi-level feature fusion within an auto-encoder to restore haze residue without gridding artifacts. It predicts the residue end-to-end and adds it to the hazy input to obtain the haze-free image.
- Network Structure: GCANet encodes hazy images, aggregates context with smoothed dilated convolutions, fuses multi-level features through a gate, and decodes the haze residue.The residue is added to the hazy input to produce the final haze-free image.
- Smoothed Dilated Convolution: Dilated convolution expands the receptive field without reducing resolution, but separate input dependencies can produce gridding artifacts.For dilation rate r and kernel size k, the receptive field increases from k to r*(k−1)+1.
- Smoothed Dilated Convolution: Smoothed dilated convolution adds an extra separable, shared convolution before dilation to introduce interactions among input units and alleviate gridding artifacts.The added convolution uses kernel size 2r−1 and modifies input dependencies before dilated convolution.
- Gated Fusion: The gated fusion sub-network combines feature maps from different levels before decoding, using a 3x3 convolution whose input concatenates Fl, Fm, and Fh.Its output channel number is 3.
- Network Structure: The network uses a simple auto-encoder with three convolutional encoder layers, seven residual blocks, one deconvolutional decoder layer, and two final convolutional layers.The seven residual blocks use dilation rates (2, 2, 2, 4, 4, 4, 1), while intermediate convolutional layers use 64 channels.
- Loss Function: GCANet trains with mean square error on the haze residue rather than directly on the final haze-free image.The predicted residue is added to the input during runtime; the authors state that loss design is not the paper’s focus.
4. Experiments
Experiments evaluate GCANet on benchmark dehazing and deraining datasets using quantitative, qualitative, ablation, and visual comparisons. GCANet outperforms prior methods, while smoothed dilation and the full component combination improve restoration quality.
- Dehazing evaluation: GCANet outperforms six state-of-the-art dehazing methods on RESIDE SOTS indoor evaluation by a large margin.Compared methods include DCP, CAP, GRM, AOD-Net, DehazeNet, and GFN, using PSNR and SSIM.
- Dehazing evaluation: Qualitatively, GCANet removes haze while preserving original brightness better than competing dehazing methods in indoor and outdoor examples.DCP and CAP produce relatively dark results, while AOD-Net may leave haze incompletely removed.
- Ablation analysis: The complete combination of smoothed dilation, gated fusion, and instance normalization achieves the best ablation performance.Four configurations are evaluated incrementally, and final performance keeps increasing as components are added.
- Ablation analysis: Smoothed dilated resblocks reduce gridding artifacts and preserve color fidelity near object boundaries and texture regions.Regular exponentially dilated resblocks are associated with gridding artifacts and color shift in the compared examples.
- Deraining evaluation: GCANet generalizes to image deraining and exceeds the previous best method by more than 3 dB in PSNR on the DID-MDN test dataset.The comparison includes seven prior deraining methods, and GCANet also produces clearer visual results than competing methods.
5. Conclusion
GCANet uses smoothed dilation and gated feature fusion for end-to-end image restoration, outperforming previous state-of-the-art dehazing methods and achieving state-of-the-art deraining performance.
- Smoothed dilated convolution eliminates gridding artifacts in the dilated-convolution-based network.The technique is used to address artifacts caused by conventional dilated convolution.
- A gated subnetwork fuses features from different levels.
- GCANet surpasses previous state-of-the-art image dehazing methods by a large margin.
- Applying the proposed network to image deraining also achieves state-of-the-art performance.The deraining application is presented as evidence of the method's generality.