Introduction
Image dehazing remains a critical challenge in computer vision, primarily due to its ill-posed nature. Traditional approaches often struggle to balance the need for global context—essential for understanding large-scale haze patterns—with the computational efficiency required for high-resolution image processing. While Convolutional Neural Networks (CNNs) excel at local feature extraction, they lack a global receptive field. Conversely, Vision Transformers (ViTs) provide global modeling but suffer from quadratic computational complexity.
WDMamba introduces a novel paradigm by integrating State Space Models (SSMs), specifically Vision Mamba, with a frequency-based physical prior. By recognizing that haze predominantly degrades low-frequency image components, the framework employs a coarse-to-fine restoration strategy that is both effective and computationally efficient.
The Wavelet Degradation Prior
A core contribution of this work is the empirical validation of the wavelet degradation prior. Through wavelet transform analysis, the authors demonstrate that haze-related information is concentrated in low-frequency subbands. In experiments where high-frequency components were swapped between hazy and clean images, the perceived haze remained largely unchanged. However, swapping low-frequency components significantly altered the haze density. This insight justifies a frequency-split architecture: treating global structure (low frequency) and fine detail (high frequency) as distinct restoration tasks.
Methodology: The WDMamba Framework
WDMamba decomposes the dehazing process into two specialized stages:
1. Low-Frequency Restoration Network (LFRN)
The LFRN utilizes the Discrete Wavelet Transform (DWT) to isolate the image's low-frequency "base." This stage leverages Mamba blocks, which utilize State Space Models to capture long-range dependencies with linear complexity. This allows the model to analyze the entire scene's haze distribution without the memory overhead of self-attention, effectively restoring the global structure and removing the bulk of the haze.
2. Detail Enhancement Network (DEN)
Following the global restoration, the DEN employs a CNN-based architecture to refine the output. Since CNNs are inherently suited for local texture extraction, this stage reinstates fine-grained edges and details that may have been smoothed during the initial stage. A Frequency Domain Enhancement (FEM) module is integrated here to further sharpen high-frequency features.
Self-Guided Contrastive Regularization (SGCR)
To optimize training, the authors introduce a Self-Guided Contrastive Regularization strategy. Unlike standard contrastive learning that uses the hazy image as a negative sample, SGCR utilizes the intermediate output from the LFRN as a "hard negative." This forces the DEN to specifically learn the nuances required to bridge the gap between a "mostly dehazed" image and the ground truth, leading to superior texture recovery.
Experimental Results
WDMamba was evaluated against several state-of-the-art (SOTA) models across four major benchmarks: Haze4K, RESIDE-6K, HSTS, and O-HAZE.
- Performance: On the Haze4K dataset, WDMamba achieved a PSNR of 35.88 dB, surpassing previous leaders like ConvIR-L by 1.38 dB. On the real-world O-HAZE benchmark, it reached 27.22 dB, demonstrating robust handling of complex physical haze.
- Efficiency: Despite its performance, WDMamba is remarkably lightweight. It utilizes only 11.25M parameters and 38.84G FLOPs. In comparison, models like DehazeFormer-L require over 25M parameters and 280G FLOPs to achieve lower accuracy scores.
Conclusion and Future Work
WDMamba demonstrates that combining the global modeling capabilities of Mamba with frequency-domain priors results in a highly efficient and effective dehazing solution. By treating haze as a predominantly low-frequency degradation, the model achieves SOTA results with a fraction of the computational cost of Transformer-based alternatives. Future research may extend this framework to non-homogeneous weather conditions, such as sandstorms or patchy fog, and explore its application in real-time video dehazing.
Summary updated