Source-linked AI summary
FFA-Net: Feature Fusion Attention Network for Single Image Dehazing
Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, Huizhu Jia
TL;DR
Single-image dehazing must recover clean images from haze that introduces visible degradation and makes restoration underdetermined when atmospheric variables are unknown. FFA-Net directly restores haze-free images using feature attention, local residual learning, and adaptive feature fusion, and it reportedly surpasses previous state-of-the-art methods by a very large margin.
Problem
Haze causes visible image degradation, while recovering a haze-free image is underdetermined without the global atmosphere light and transmission map.
Method
FFA-Net directly restores haze-free images using channel-and-pixel feature attention, local residual learning, and attention-based fusion of different-level features.
Results
FFA-Net surpasses previous state-of-the-art single-image dehazing methods quantitatively and qualitatively by a very large margin.
Takeaways & Limitations
FFA-Net is reported to improve restoration of image detail and color fidelity, particularly in thick-haze and rich-texture regions.
Abstract
from arXiv · showhide
In this paper, we propose an end-to-end feature fusion at-tention network (FFA-Net) to directly restore the haze-free image. The FFA-Net architecture consists of three key components: 1) A novel Feature Attention (FA) module combines Channel Attention with Pixel Attention mechanism, considering that different channel-wise features contain totally different weighted information and haze distribution is uneven on the different image pixels. FA treats different features and pixels unequally, which provides additional flexibility in dealing with different types of information, expanding the representational ability of CNNs. 2) A basic block structure consists of Local Residual Learning and Feature Attention, Local Residual Learning allowing the less important information such as thin haze region or low-frequency to be bypassed through multiple local residual connections, let main network architecture focus on more effective information. 3) An Attention-based different levels Feature Fusion (FFA) structure, the feature weights are adaptively learned from the Feature Attention (FA) module, giving more weight to important features. This structure can also retain the information of shallow layers and pass it into deep layers. The experimental results demonstrate that our proposed FFA-Net surpasses previous state-of-the-art single image dehazing methods by a very large margin both quantitatively and qualitatively, boosting the best published PSNR metric from 30.23db to 36.39db on the SOTS indoor test dataset. Code has been made available at GitHub.
Introduction
Single-image dehazing seeks to recover clean images degraded by haze, but the task is underdetermined and existing methods face uneven, channel-dependent information. FFA-Net addresses this with feature attention, local residual learning, and attention-based fusion, achieving strong qualitative and quantitative results.
- Haze causes color distortion, blurring, and low contrast, complicating both image restoration and downstream visual tasks.
- The atmosphere scattering model makes dehazing underdetermined when global atmosphere light and transmission are unknown.
- Prior-based methods estimate haze using image statistics, while deep networks regress transmission maps or haze-free images directly.
- FFA-Net combines channel and pixel attention to weight feature information unequally because haze distribution and channel importance vary across images.
- Its local residual learning bypasses thin-haze and low-frequency information, while attention-based fusion preserves shallow features and passes them to deeper layers.
- FFA-Net surpasses previous state-of-the-art dehazing methods quantitatively and qualitatively, with reported advantages in thick haze, detail, and color fidelity.
Related Work
Earlier dehazing methods relied on physical scattering models and handcrafted priors, whereas later approaches used deep networks to estimate haze effects or directly reconstruct clean images.
- Traditional methods depend on the ill-posed scattering model and use priors to compensate for information lost during image corruption.
- The dark channel prior estimates transmission but can be unreliable when scene objects resemble atmospheric light.
- End-to-end models such as DehazeNet and AOD-Net learn haze-related representations or directly generate clean images from hazy inputs.
Fusion Feature Attention Network (FFA-Net)
FFA-Net combines feature attention, local residual learning, and adaptive multi-level feature fusion to restore haze-free images while emphasizing informative features and preserving shallow information.
- FFA-Net processes a hazy image through shallow feature extraction, grouped architectures, feature fusion, reconstruction, and global residual learning.
- Feature Attention: Feature Attention combines channel-wise and pixel-wise attention to weight different features and pixels unequally.Channel attention derives channel weights, while pixel attention produces a 1 × H × W attention map emphasizing informative regions such as thick haze and high-frequency areas.
- Basic Block Structure: The basic block uses local residual learning to bypass thin-haze or low-frequency information, allowing the main network to focus on more effective information.Experiments report that this structure improves network performance and training stabilization.
- Feature Fusion: Adaptive feature fusion weights outputs from different group architectures, retaining low-level information while passing it into deeper layers.The weighting mechanism emphasizes thick haze regions, high-frequency texture, and color fidelity.
- Implementation Details: The implementation uses three group structures, 19 basic blocks per group, 3 × 3 convolutions except for 1 × 1 channel attention, and 64 output filters per group.The network is optimized with a simple L1 loss using ground-truth and hazy input images.
Experiments
FFA-Net is evaluated against established dehazing methods on synthetic SOTS and realistic hazy images, using quantitative metrics and qualitative comparisons. The authors report large PSNR and SSIM advantages alongside improved detail and color fidelity.
- Quantitative and qualitative evaluation: The evaluation compares FFA-Net with DCP, AOD-Net, DehazeNet, and GCANet on the RESIDE-based SOTS benchmark.The study uses PSNR and SSIM for quantitative comparison and includes visual comparisons on SOTS and realistic hazy images.
- Quantitative and qualitative evaluation: FFA-Net outperforms the four compared methods by a very large margin in PSNR and SSIM.The reported metrics are cited from prior studies, including Li et al. (2018) and Qu et al. (2019).
- Qualitative comparisons: DCP shows color distortion, AOD-Net leaves haze and produces low-brightness images, while DehazeNet produces excessively bright reconstructions.The authors attribute DCP’s distortion to its underlying prior assumptions and describe these visual differences relative to ground truth.
- Qualitative comparisons: FFA-Net better preserves realistic details and colors, including distant towers, wet-road texture, and raindrops in real hazy images.The comparison also reports nonexistent spots on a building in GCANet’s result.
Ablation Analysis
The ablation study evaluates Feature Attention, Local Residual Learning, and Feature Fusion Attention separately and in combination. The results indicate that each factor contributes to performance, with FFA especially important.
- Configurations: The ablation study varies the FA module, the combination of LRL and FA, and the FFA structure.Configurations are evaluated on the SOTS indoor test set using 48 × 48 crops and 3 × 10^5 training steps.
- Results: 35.77db PSNR is achieved when the paper’s full implementation details are used.This value is reported in the ablation discussion.
- Results: Every considered factor contributes to network performance, with the FFA structure identified as especially important.The authors also report that FA alone remains competitive with previous state-of-the-art methods.
- Results: Local Residual Learning stabilizes training while improving performance, and combining FA with feature fusion produces a high-performing configuration.The stated ablation interpretation links these effects to the roles of LRL, FA, and FFA.
Conclusion
The paper concludes that FFA-Net is a simple end-to-end network with strong single-image dehazing performance. The authors emphasize its restoration of image detail and color fidelity and suggest applicability to other low-level vision tasks.
- Conclusion: FFA-Net demonstrates strong performance for single-image dehazing as an end-to-end Feature Fusion Attention Network.The conclusion presents this as the paper’s central contribution.
- Conclusion: The authors report that FFA-Net surpasses previous state-of-the-art methods by a very large margin.The conclusion frames the architecture as simple despite this reported performance advantage.
- Conclusion: FFA-Net is reported to have advantages in restoring image detail and color fidelity.The authors connect these strengths to its potential relevance for other image-restoration tasks.
- Conclusion: The authors suggest that FFA-Net may address other low-level vision tasks, including deraining, super-resolution, and denoising.This is presented as an expected extension rather than an evaluated result in the supplied conclusion passage.