结合光吸收和大气散射模型的可解释图像去雾网络

An Interpretable Network for Image Dehazing via A Light Absorption-Based Atmospheric Scattering Model

在雾霾天气下拍摄的图像存在对比度低和颜色偏移的问题,这会进一步影响后续高级计算机视觉应用的准确性。大气散射模型(Atmospheric Scattering Model,ASM)作为图像去雾中的热门技术之一,越来越受到研究者的关注。然而,现有的基于ASM的方法由于忽略了光陷现象而会引起暗淡效应。为了解决这个问题,构建了一种基于光吸收的可解释图像去雾网络。具体地,首先构建了一种基于光吸收的大气散射模型(Light Absorption-based ASM,LA-ASM),在ASM中引入了由纹理密度和场景深度定义的光吸收系数。此外,构建了一种可解释网络LA-ASMNet,该网络通过学习LA-ASM的雾气成分来辅助图像去雾。基于RESIDE数据集的仿真结果证明了LA-ASMNet在缓解图像去雾暗淡效应方面的有效性。

Images taken under haze weather suffer from low contrast and color shift,which can further negatively affect the accuracy of the following high-level computer vision applications.As one popular technology in image dehazing,the atmospheric scattering model(ASM)is gaining attention from researchers.However,the existing ASM-based methods suffer from the dim effect because they ignore the light-trapping phenomenon.To address this problem,a light absorption-based interpretable network for image dehazing is proposed.Specifically,a light absorption-based ASM(LA-ASM)is first constructed,in which a light absorption coefficient defined by the texture density and the scene depth is introduced.Moreover,an interpretable network,dubbed LA-ASMNet,in which haze patterns of LA-ASM are learned to assist image dehazing,is built.Simulation results based on RESIDE dataset demonstrate the effectiveness of LA-ASMNet in alleviating the dim effect for image dehazing.