CAA-Net: End-to-End Two-Branch Feature Attention Network for Single Image Dehazing

Gang Jin, Jingsheng Zhai, Jianguo WeiIEICE Trans. Fundam. Electron. Commun. Comput. Sci 2023 · OpenReview

Abstract

Single image dehazing is a critical task in computer vision, often hindered by the non-uniform distribution of haze and the computational complexity of high-performing models. This paper introduces CAA-Net, a novel dual-branch architecture that combines hierarchical feature extraction with an efficient attention mechanism. By leveraging Residual Dense Blocks (RDB) for structural reconstruction and a Feature Fusion branch with a lightweight Feature Attention (FPA) module, CAA-Net achieves superior image quality and faster processing speeds compared to existing state-of-the-art methods like FFA-Net.

1. Introduction

Single image dehazing is an ill-posed problem where light scattering by atmospheric particles reduces contrast and color fidelity. Traditional methods rely on the Atmospheric Scattering Model to estimate transmission maps and atmospheric light. While deep learning has transitioned the field toward end-to-end mapping, challenges remain in effectively fusing multi-scale features and addressing the spatial non-uniformity of haze.

2. CAA-Net Architecture

The proposed CAA-Net (Channel and Pixel Attention Network) utilizes a two-branch design to balance reconstruction and feature selection.

  • U-Net Branch: This branch employs an encoder-decoder structure enhanced with Residual Dense Blocks (RDB). RDBs facilitate continuous memory by reading states from previous layers, ensuring that fine-grained textures are preserved throughout the network.
  • Feature Fusion Branch: Composed of Feature Fusion based on Attention (FEPA) structures, this branch adaptively learns the importance of different features for the final output.
The overall architecture of CAA-Net, featuring the dual-branch design with U-Net and FEPA components.

3. The Feature Attention (FPA) Module

The core of the network's efficiency is the FPA module. It integrates three key components to handle varying haze densities and optimize feature weights:

  • Local Residual Learning: Stabilizes training and facilitates gradient flow.
  • Efficient Channel Attention (ECA): A lightweight mechanism that captures cross-channel interactions without the high parameter count of standard attention blocks.
  • Pixel Attention (PA): Focuses computational resources on hazy regions while ignoring clear areas, effectively handling non-uniform haze.
The structure of the Feature Attention (FPA) module, integrating ECA and PA for adaptive feature weighting.

4. Experimental Evaluation

The model was trained on the RESIDE Outdoor Training Set (OTS) and evaluated on the Synthetic Object Test Set (SOTS). Additionally, a custom dataset of 200 maritime images was used to verify real-world performance in challenging environments.

Quantitative Performance

CAA-Net achieved a PSNR of 32.96 dB and an SSIM of 0.9835 on the SOTS dataset, slightly outperforming the previous state-of-the-art, FFA-Net.

Efficiency and Speed

CAA-Net demonstrated a significant speed advantage, with an average runtime of 0.24 seconds per image, nearly twice as fast as FFA-Net's 0.42 seconds. This efficiency is largely attributed to the use of the lightweight ECA module instead of more complex attention mechanisms.

5. Conclusion

CAA-Net provides a robust and efficient solution for single image dehazing. By utilizing a dual-branch architecture and lightweight attention modules, it achieves a superior balance between accuracy and speed. Future work will explore the integration of perceptual loss and GAN-based frameworks to further enhance the visual sharpness and aesthetic quality of the dehazed results, moving beyond the limitations of standard L1 loss functions.

Summary updated

BibTeX
@article{jin2023caa,
  author = {Gang Jin and Jingsheng Zhai and Jianguo Wei},
  title = {{CAA-Net: End-to-End Two-Branch Feature Attention Network for Single Image Dehazing}},
  journal = {IEICE Trans. Fundam. Electron. Commun. Comput. Sci.},
  year = {2023},
  url = {https://openreview.net/forum?id=so2MUsYVgng},
}
Figure