WDMamba: When Wavelet Degradation Prior Meets Vision Mamba for Image Dehazing

TL;DR

WDMamba tackles image dehazing by exploiting a haze-specific wavelet degradation prior that concentrates in low-frequency components. It decomposes the task into low-frequency restoration with a Mamba-based network and high-frequency detail enhancement via a CNN, guided by a self-guided contrastive regularization that uses the coarse restoration as a hard negative. The approach delivers state-of-the-art quantitative results on synthetic datasets and strong qualitative performance on real hazy scenes, while maintaining efficiency through linear-complexity components. This framework offers a principled, two-stage strategy that balances global structure restoration with fine-detail recovery, making it practical for real-world dehazing with improved naturalness and color fidelity.

Abstract

In this paper, we reveal a novel haze-specific wavelet degradation prior observed through wavelet transform analysis, which shows that haze-related information predominantly resides in low-frequency components. Exploiting this insight, we propose a novel dehazing framework, WDMamba, which decomposes the image dehazing task into two sequential stages: low-frequency restoration followed by detail enhancement. This coarse-to-fine strategy enables WDMamba to effectively capture features specific to each stage of the dehazing process, resulting in high-quality restored images. Specifically, in the low-frequency restoration stage, we integrate Mamba blocks to reconstruct global structures with linear complexity, efficiently removing overall haze and producing a coarse restored image. Thereafter, the detail enhancement stage reinstates fine-grained information that may have been overlooked during the previous phase, culminating in the final dehazed output. Furthermore, to enhance detail retention and achieve more natural dehazing, we introduce a self-guided contrastive regularization during network training. By utilizing the coarse restored output as a hard negative example, our model learns more discriminative representations, substantially boosting the overall dehazing performance. Extensive evaluations on public dehazing benchmarks demonstrate that our method surpasses state-of-the-art approaches both qualitatively and quantitatively. Code is available at https://github.com/SunJ000/WDMamba.

Figures (10)

• Figure 1: Comparison of model performance and complexity on the Haze4K dataset liu2021synthetic across various state-of-the-art methods. FLOPs are calculated at a resolution of 256$\times$256.
• Figure 2: Wavelet degradation prior of haze. By exchanging the low-frequency subbands between paired hazy and haze-free images during wavelet decomposition, it can be observed that haze-related degradation is predominantly concentrated in the low-frequency components. DWT and IWT denote the Discrete Wavelet Transform and the Inverse Wavelet Transform, respectively, while Low-freq refers to the low-frequency subband of the decomposed image.
• Figure 3: The overall framework of the proposed WDMamba. It consists of three main components: the Low-Frequency Restoration Network (LFRN), the Detail Enhancement Network (DEN), and the Self-Guided Contrastive Regularization (SGCR). LFRN operates on the degraded low-frequency components, employing linear complexity to restore global structures and produce a coarse restored image. Subsequently, DEN enhances local details, generating the final output. SGCR treats the coarse restored image as a hard negative sample, encouraging the final dehazed image to better approximate the clean image.
• Figure 4: The detailed architecture of the low-frequency restoration network. It follows a U-Net design, incorporating Mamba blocks to model long-range feature dependencies for effective global information restoration. To further preserve structural integrity, Haar wavelet downsampling is employed in place of conventional downsampling operations, thereby alleviating information loss and enhancing feature fidelity throughout the restoration process.
• Figure 5: Illustration of 2D selective scanning mechanism.