基于小波双分支建模的频率感知单幅图像去雾

Frequency-Aware Single Image Dehazing via Wavelet-Based Dual-Branch Modeling

摘要: 在单幅图像去雾任务中，实现有效去雾的同时保持颜色保真度和边缘结构细节仍然是一个具有挑战性的问题。现有的基于深度学习的去雾方法在复杂有雾条件下往往容易出现整体颜色失真、视觉伪影以及边缘模糊等问题。为克服上述局限性，本文基于频率感知图像去雾框架提出了TripleD-Net图像去雾模型，该模型利用离散小波变换对图像的低频分量与高频分量进行分离，低频分量主要表征图像的整体结构和颜色一致性，因此通过全局上下文建模机制对其进行建模，以增强对全局语义信息的感知能力；相比之下，高频分量包含丰富的边缘结构与纹理细节，使用多尺度卷积框架对其进行处理，从而有效保留精细的结构信息。通过对低频分支与高频分支的共同优化，所提出的方法在颜色一致性、边缘清晰度以及伪影抑制等方面均表现出更优的恢复效果。在RESIDE数据集上的大量实验结果表明，TripleD-Net在定量指标和定性视觉对比方面均优于近期具有代表性的去雾方法。

Abstract: In single-image dehazing tasks, achieving effective haze removal while preserving color fidelity and edge structural details remains a challenging problem. Existing deep learning-based dehazing methods often suffer from overall color distortion, visual artifacts, and edge blurring under complex hazy conditions. To overcome these limitations, this paper proposes TripleD-Net, an image dehazing model based on a frequency-aware framework. The model employs the Discrete Wavelet Transform (DWT) to separate image components into low-frequency and high-frequency bands. Since low-frequency components primarily represent the overall structure and color consistency of the image, they are modeled via a global context modeling mechanism to enhance the perception of global semantic information. In contrast, high-frequency components contain rich edge structures and texture details; thus, a multi-scale convolutional framework is utilized to process them, effectively preserving fine structural information. Through the joint optimization of the low-frequency and high-frequency branches, the proposed method demonstrates superior restoration performance in terms of color consistency, edge sharpness, and artifact suppression. Extensive experiments on the RESIDE dataset show that TripleD-Net outperforms recent representative dehazing methods in both quantitative metrics and qualitative visual comparisons.