基于不变学习的真实雾霾去除方法

Real-world dehazing method with invariant learning

针对有监督去雾方法面临的质量干扰问题,提出基于不变学习的真实雾霾图像去除方法.该方法使用傅里叶特征变换将网络提取的特征线性化表示,针对线性化特征进行全局加权并求解协方差,去除特征之间的相关性.不变学习使网络更加关注特征与去雾图像之间的本质关系,可以使网络获得稳定的跨域特征.分析并解释了数据和所提出方法的不同作用.该方法既实现了对大气光散射模型的改进,又构建了适用于真实雾霾场景的新数据集.通过大量实验,验证了该方法在真实世界的良好去雾效果.

A real hazy image removal method based on invariant learning was proposed to solve the problem of quality interference in image dehazing.The Fourier feature transform was used to linearize the features extracted by the network.Then global weighting was performed and covariance was solved for the linearized features.The correlation between the features was removed.Invariant learning makes the network more concerned about the essential relationship between features and the dehazing image,which can enable the network to obtain stable cross-domain features.The different roles of the data and the proposed method were analyzed and explained.The improvement of the atmospheric scattering model was realized,and a new dataset suitable for real haze scenarios was constructed.The real-world dehazing effect of the method was verified through extensive experiments.