小波域条件扩散与跨频率协同建模的水下图像增强方法

Wavelet-Domain Conditional Diffusion with Cross-Frequency Collaborative Modeling for Underwater Image Enhancement

摘要: 水下图像增强在海洋探测与水下视觉感知中具有重要应用价值，但水体的光吸收与散射效应会导致严重的颜色失真与细节退化。现有方法多依赖空间域建模，缺乏对结构与纹理的解耦建模；扩散模型虽具备强生成能力，但计算开销大且难以兼顾多频信息恢复。为此，本文提出一种融合小波域条件扩散与跨频率协同建模的水下图像增强方法，实现结构与细节的协同恢复。具体而言，通过两级小波分解将图像解耦为低频结构与高频细节，并在低频域引入条件扩散模型以恢复全局结构；设计方向感知高频增强模块(HFEM)，通过方向建模与子带交互强化纹理表达；提出跨频率校正模块(CFC)，实现高低频信息的双向融合与一致性约束，从而提升整体重建质量。在UIEB数据集上的实验结果表明，所提方法的PSNR达23.75 dB，且在SSIM、LPIPS等指标上均优于现有方法；同时在真实场景中表现出更自然的颜色恢复与更清晰的细节重建，展现出良好的泛化能力。

Abstract: Underwater image enhancement is of significant importance for marine exploration and underwater visual perception. However, complex light absorption and scattering effects in water often lead to severe color distortion and detail degradation. Existing methods predominantly rely on spatial-domain modeling and lack explicit decoupling of structural and textural information. Although diffusion models have demonstrated strong generative capability, they typically incur high computational cost and struggle to jointly recover multi-frequency information. To address these challenges, we propose a wavelet-domain conditional diffusion framework with cross-frequency collaborative modeling for underwater image enhancement, enabling effective joint restoration of structure and fine details. Specifically, a two-level wavelet decomposition is first employed to separate the input image into low-frequency structural components and high-frequency detail components. A conditional diffusion model is then introduced in the low-frequency domain to recover global structures. Meanwhile, a direction-aware high-frequency enhancement module (HFEM) is designed to refine texture details via directional modeling and sub-band interactions. Furthermore, a cross-frequency correction module (CFC) is proposed to facilitate bidirectional fusion and consistency between low- and high-frequency representations, thereby improving overall reconstruction quality. Experimental results on the UIEB dataset demonstrate that the proposed method achieves a PSNR of 23.75 dB and consistently outperforms existing approaches in terms of SSIM and LPIPS. In addition, it produces more natural color restoration and clearer details in real-world scenarios, indicating strong generalization capability.