基于Retinex的多尺度单幅图像去雾网络

Multi-scale Single Image Dehazing Network Based on Retinex

针对现有去雾算法应用于真实场景时容易产生颜色失真、雾霾残留严重等问题,提出一种多尺度单幅图像去雾网络。首先基于Retinex理论的去雾模型,设计了多尺度残差光照图估计模块用于生成初步去雾后的图像,其次设计了精细化去雾模块来优化粗糙的去雾图像,从而获得去雾更加彻底、细节更加丰富的清晰图像。多尺度残差光照图估计模块利用不同尺寸的感受野捕获全局背景特征和局部细节特征,精细化去雾模块采用U-net结构,通过多个跳跃连接融合不同阶段的特征。实验结果表明,本文网络在定量和定性方面均有所提升,对于真实场景下的雾霾图像有较好的去雾能力。

The existing dehazing algorithms have defects in real scenes,which are prone to color distortion and serious haze residue.In order to solve these problems,a multi-scale single image dehazing network was proposed.Firstly,based on the dehazing model of Retinex theory,a multi-scale residual illumination estimation module was designed to obtain the preliminary dehazing image after extracting the features of the hazy image.Then the refined dehazing module was used to optimize the rough dehazing image,and finally a clearer image with more thorough dehazing and richer details was obtained.The multi-scale residual illumination estimation module uses receptive fields of different sizes to capture global background features and local detail features.The refined dehazing module adopts U-net structure and fuses the features of different stages through multiple skip connections.Experiments show that the performance of the proposed network is better than other dehazing algorithms in quantitative and qualitative aspects.It has better dehazing ability for the haze image in the real scene.