Abstract: To address the issues of halo artifacts, edge blurring, color distortion, and contrast imbalance commonly encountered in traditional physical model-based dehazing algorithms, this paper proposes a single-image dehazing algorithm integrating multi-scale processing and lightweight networks. Initially, this paper generates images with varying exposure levels to capture texture features across different illumination regions, enabling comprehensive scene characterization through fusion of dark channel information under multi-exposure conditions. Subsequently, wavelet transform is employed to integrate multi-scale dark channel prior transmission estimates with multi-scale gamma-corrected image transmission estimates, which are further optimized using unsharp mask-guided filtering to enhance edge preservation capabilities. An adaptive regularization constraint mechanism is introduced to dynamically adjust smoothing intensity, effectively balancing detail retention and noise suppression. Finally, multi-scale luminance guidance combined with a lightweight network is adopted for atmospheric light estimation, enabling haze-free image reconstruction through inverse atmospheric scattering model computation. Extensive quantitative and qualitative experiments conducted on SOTS, RTTS, and HSTS datasets demonstrate that compared with conventional dehazing algorithms, the proposed method exhibits superior performance in both objective metrics and subjective visual quality, particularly showing enhanced restoration effectiveness in regions with abrupt depth variations.

Key words: image dehazing, image fusion, gamma correction, unsharp mask, dark channel prior

摘要： 针对传统物理模型去雾算法处理后易产生光晕效应、边缘模糊、颜色失真与对比度失衡等问题，提出一种结合多尺度处理和轻量级网络的单幅图像去雾算法。通过生成不同曝光水平的图像，捕获不同照度区域的纹理特征，融合不同曝光下的暗通道信息，实现了更全面的场景描述。利用小波变换对多尺度暗通道先验估计透射率和多尺度伽马校正图像估计透射率进行融合，并通过非锐化掩膜引导滤波进行优化，增强其边缘保持能力；同时引入自适应正则化约束机制调整平滑强度，有效平衡细节保留与噪声抑制。采用多尺度亮度引导结合轻量化网络估计大气光值，最终依托大气散射模型逆运算复原无雾图像。所提算法在SOTS、RTTS、HSTS三个数据集上进行了大量定量和定性实验对比，实验结果表明，相对于传统去雾算法，所提算法在图像去雾的定量指标与主观视觉效果上均表现出一定优势，尤其是在景深突变区域图像复原效果较为突出。

关键词: 图像去雾, 图像融合, 伽玛校正, 非锐化掩膜, 暗通道先验