Mutual Information-driven Triple Interaction Network for Efficient Image Dehazing

Hao Shen, Zhong-Qiu Zhao, Yulun Zhang, Zhao Zhang

Introduction

Image dehazing is a fundamental "ill-posed" problem in computer vision, where the goal is to recover a clear scene from a single hazy input. While deep learning has significantly advanced the state-of-the-art (SOTA), many existing models suffer from three persistent limitations: the neglect of frequency-domain information, weak interaction between network stages, and high feature redundancy.

The Mutual Information-driven Triple Interaction Network (MITNet) addresses these challenges by splitting the dehazing process into amplitude recovery and phase refinement, utilizing a novel interaction module and a mutual information minimization constraint to achieve high performance with low computational overhead.

Methodology

Two-Stage Spatial-Frequency Processing

MITNet operates on the physical insight that haze primarily degrades the amplitude of an image's Fourier transform, while the phase—which encodes structural information—remains relatively intact. The architecture is divided into two specialized stages:

  1. Stage 1 (Amplitude-guided): Focuses on restoring the amplitude spectrum to effectively strip away haze.
  2. Stage 2 (Phase-guided): Focuses on the phase spectrum to refine fine-grained background structures.

Both stages utilize custom Residual Amplitude Blocks (RAB) and Residual Phase Blocks (RPB) that process spatial and frequency information in parallel, ensuring the network captures both global context and local details.

Adaptive Triple Interaction Module (ATIM)

To prevent the loss of rich intermediate features, the authors introduced the ATIM. This module acts as the network's connective tissue, aggregating features across three critical dimensions:

  • Cross-domain: Merging spatial and frequency features.
  • Cross-scale: Integrating low-resolution and high-resolution data.
  • Cross-stage: Linking features between the first and second stages.

The ATIM employs an Adaptive Dynamic Filter Block (ADFB) to generate content-specific filters on the fly, allowing the model to adapt dynamically to the unique haze patterns present in each image.

Mutual Information Constraint (MIC)

In many multi-stage networks, different stages often end up learning nearly identical, redundant features. MITNet solves this by applying a Mutual Information Minimization loss. By forcing the mutual information between Stage 1 and Stage 2 features to be low, the network is incentivized to learn complementary information, ensuring that the second stage adds genuine value to the restoration process rather than simply repeating the work of the first.

Experimental Results

Performance Benchmarks

MITNet was evaluated on several standard benchmarks, including RESIDE (ITS, OTS, SOTS), Dense-Haze, and NH-Haze. The results demonstrate significant improvements over previous methods:

  • SOTS-Indoor: MITNet achieved a PSNR of 40.23 dB, outperforming the previous best (FSDGN) by 1.6 dB and the Transformer-based Dehamer by 3.6 dB.
  • NH-Haze: On this challenging real-world dataset featuring non-homogeneous haze, MITNet achieved a PSNR of 21.26 dB, proving its robustness beyond synthetic data.

Efficiency and Complexity

Despite its high accuracy, MITNet is remarkably lightweight. It utilizes approximately 2.7M parameters, making it more efficient than many heavy-weight Transformer-based models. By leveraging smart architectural priors—specifically Fourier domain processing—the model achieves a superior balance between FLOPs and restoration quality.

Conclusion

MITNet introduces a principled approach to image dehazing by combining frequency-domain insights with a novel interaction framework. The use of mutual information minimization as a regularizer ensures feature diversity, while the two-stage spatial-frequency design allows for precise restoration of both color and structure. While future work may address the computational overhead of Fast Fourier Transforms (FFT) on specific hardware and extend the model to video consistency, MITNet currently stands as a highly efficient and effective solution for single-image dehazing.

Summary updated

BibTeX
@inproceedings{shen2023mutual,
  author = {Hao Shen and Zhong-Qiu Zhao and Yulun Zhang and Zhao Zhang},
  title = {{Mutual Information-driven Triple Interaction Network for Efficient Image Dehazing}},
  booktitle = {ACM Multimedia},
  year = {2023},
  doi = {10.1145/3581783.3612299},
}
Figure