FFA-Net: Feature Fusion Attention Network for Single Image Dehazing

Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, Huizhu Jia

Abstract

In this paper, we propose an end-to-end feature fusion attention network (FFA-Net) to directly restore haze-free images. The FFA-Net architecture consists of three key components: a novel Feature Attention (FA) module combining Channel Attention with Pixel Attention; a basic block structure utilizing Local Residual Learning and Feature Attention; and an Attention-based Feature Fusion (FFA) structure. By treating different features and pixels unequally, FFA-Net provides flexibility in handling non-uniform haze distributions. Experimental results demonstrate that FFA-Net surpasses previous state-of-the-art methods by a large margin, boosting the best published PSNR from 30.23 dB to 36.39 dB on the SOTS indoor test dataset.

Introduction

Haze significantly degrades image quality by causing color distortion, blurring, and loss of contrast, which negatively impacts downstream computer vision tasks like object detection and autonomous navigation. Traditional dehazing methods often rely on the Atmospheric Scattering Model, which requires estimating intermediate components like global atmospheric light and transmission maps. However, these estimations are frequently inaccurate.

While deep learning has advanced the field, many existing networks treat all pixels and feature channels equally. This is suboptimal because haze is rarely uniform across an image, and different feature channels contain varying levels of useful information. FFA-Net addresses these limitations by employing a hierarchical attention mechanism to focus on the most relevant features and spatial regions.

Methodology

The FFA-Net architecture is designed to bypass the limitations of physical model estimation by directly restoring haze-free images in an end-to-end manner. Its core consists of three hierarchical levels:

Feature Attention (FA) Module

The FA module combines Channel Attention (CA) and Pixel Attention (PA). CA weights different feature maps based on their importance, while PA allows the network to focus on specific spatial regions, such as areas with thicker haze. This unequal treatment enables the network to adaptively process different types of information.

Basic Block and Group Architecture

The network is constructed from Basic Blocks that integrate Local Residual Learning with the FA module. These blocks are further organized into Group Structures. Local residual connections allow the network to bypass "easy" information—such as thin haze or low-frequency regions—enabling the main architecture to focus on restoring complex details.

The overall FFA-Net architecture showing the group structures and fusion mechanism.

Feature Fusion Attention (FFA)

Instead of static stacking or simple concatenation, FFA-Net uses an attention-based fusion structure. It adaptively learns weights to combine features from different depths, ensuring that critical information from shallow layers (like edges and textures) is preserved and effectively integrated into deeper layers.

Experimental Results

The authors evaluated FFA-Net using the RESIDE (REalistic Single Image DEhazing) benchmark, including the Indoor Training Set (ITS) and the Synthetic Objective Testing Set (SOTS).

Quantitative Performance

FFA-Net achieved a breakthrough in performance on the SOTS indoor dataset, reaching a PSNR of 36.39 dB and an SSIM of 0.9886. This represents a massive improvement over previous state-of-the-art models like GCANet (30.23 dB).

Qualitative Assessment

Visual comparisons demonstrate that FFA-Net produces images with higher color fidelity and sharper details. Unlike earlier methods that often suffer from over-saturation or halo artifacts, FFA-Net restores natural-looking scenes even in the presence of heavy haze.

Comparison showing FFA-Net's superior detail and color restoration compared to other methods.

Conclusion

FFA-Net represents a significant advancement in single image dehazing. By moving away from traditional physical models and implementing a sophisticated feature fusion attention mechanism, it achieves unprecedented results on standard benchmarks. The flexibility of the FFA-Net framework suggests it could be effectively adapted for other image restoration tasks, such as deraining and denoising.

Summary updated

Abstract

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.
BibTeX
@inproceedings{qin2020ffa,
  author = {Xu Qin and Zhilin Wang and Yuanchao Bai and Xiaodong Xie and Huizhu Jia},
  title = {{FFA-Net: Feature Fusion Attention Network for Single Image Dehazing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=LuDWLXesy2},
}
Figure