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: 1) A novel Feature Attention (FA) module that combines Channel Attention with a Pixel Attention mechanism, accounting for the uneven distribution of haze and the varying importance of channel-wise features. 2) A basic block structure incorporating Local Residual Learning and Feature Attention, allowing the network to bypass less important information and focus on effective features. 3) An Attention-based Feature Fusion (FFA) structure where feature weights are adaptively learned, preserving shallow layer information for deeper layers. Experimental results demonstrate that FFA-Net surpasses previous state-of-the-art methods by a significant margin, improving the PSNR metric from 30.23 dB to 36.39 dB on the SOTS indoor test dataset.
Introduction
Single image dehazing is a challenging ill-posed problem. Traditional methods often rely on the atmospheric scattering model, which requires estimating global atmospheric light and transmission maps. While deep learning has advanced this field, early CNN-based models often treated every pixel and channel with equal importance.
In reality, haze distribution is rarely uniform; it varies in thickness across the background and foreground and affects color channels differently. Standard networks struggle with this unevenness, often leaving artifacts in thick haze regions or losing fine details in clearer ones. FFA-Net addresses this by implementing an attention mechanism that allows the model to focus computational power on the most obscured parts of the image.
Architecture and Methodology
FFA-Net moves away from simple pixel-wise regression toward a sophisticated attention-based architecture. It is built upon three main pillars:
- Feature Attention (FA) Module: This module combines Channel Attention, which weights feature maps based on importance, and Pixel Attention, which creates a spatial weight map to focus on regions with thicker haze.
- Basic Block with Local Residual Learning: The network utilizes blocks that combine the FA module with skip connections. This allows "easy" information (like thin haze or low-frequency areas) to bypass main processing, focusing the network's capacity on more complex image regions.
- Attention-based Feature Fusion (FFA): Instead of simple sequential processing, FFA-Net concatenates features from different depths and uses an attention mechanism to adaptively weight them. This ensures critical structural information from shallow layers is retained throughout the deep network.
The final architecture employs three "Group Architectures," each containing 19 "Basic Blocks," creating a deep and highly expressive model for feature extraction and restoration.
Experimental Evaluation
The authors evaluated FFA-Net using the RESIDE (REalistic Single Image DEhazing) benchmark, specifically the Indoor Training Set (ITS) and the Synthetic Objective Testing Set (SOTS).
Quantitative Results
FFA-Net achieved a massive performance leap over previous benchmarks:
- PSNR: 36.39 dB on the SOTS indoor dataset (compared to the previous best of 30.23 dB).
- SSIM: 0.9886, indicating near-perfect structural similarity to ground truth images.
Qualitative Results
Visually, the model excels at recovering color fidelity and sharp textures. Unlike earlier models like DCP or AOD-Net, which often produce "ghosting" artifacts or dull images, FFA-Net maintains clarity even in areas previously obscured by thick haze.
Conclusion
FFA-Net demonstrates that treating pixels and channels unequally is essential for effective single image dehazing. By combining pixel-wise and channel-wise attention with multi-level feature fusion, the network achieves state-of-the-art performance. While the model's depth introduces computational complexity, its success suggests that attention-based fusion is a versatile framework applicable to other low-level vision tasks such as deraining and super-resolution.
Summary updated