In view of the fact that the existing defog methods often ignore the key atmospheric light estimation, a method based on quadtree decomposition is proposed, which avoids the influence of bright white area on atmospheric light estimation and accurately estimates atmospheric light in the sky region. In order to avoid the limitation of manual feature extraction, three convolution scales are used to check the original fog image for convolution operation, and the propagation map to be refined is obtained after a series of feature learning of the network, and then the image fusion method is used to refine it. Finally, the estimated parameters are brought into the atmospheric scattering model to deduce a clear image. The quantitative and qualitative experimental results of synthetic and real-world grain depot fog and dust images show that the algorithm has a good effect on image texture details and sky region processing, and has high robustness and universality.

鉴于现有的除雾方法经常忽略关键的大气光估计，提出了一种基于四叉树分解的方法，该方法避免了明亮的白色区域对大气光估计的影响，并准确地估计了天空区域的大气光。为了避免人工特征提取的局限性，采用三种卷积比例尺对原始雾图像进行卷积运算，经过一系列的网络特征学习后，得到待细化的传播图，然后进行图像融合方法用于优化它。最后，将估计的参数引入大气散射模型，以得出清晰的图像。