针对传统单幅图像去雾算法容易受到雾图先验信息制约而导致颜色失真,以及现有深度学习去雾算法受网络模型限制而存在去雾残留等问题,提出了一种基于双域分解的多尺度深度学习单幅图像去雾方法,设计了一个包含低频去雾子网和高频去雾子网的多尺度深度学习网络模型。首先采用双边滤波器对有雾图像进行分解,得到雾图的高、低频子图,然后通过设计的网络模型分别学习雾图高、低频子图与高、低频透射率之间的映射关系,再将模型学习得到的高、低频透射率进行融合,得到原始雾图对应的场景透射率图,最后根据大气散射模型实现有雾图像到无雾图像的恢复,采用雾图数据集对该模型进行训练测试。结果表明,所提方法在合成有雾图像和真实自然雾图的实验中均能取得良好的去雾效果,在主观评价和客观评价方面均优于其他对比算法。

The traditional single image dehazing algorithms are susceptible to the prior information of hazy images, resulting in color distortion. Furthermore, the deep-learning dehazing algorithms are limited by the network model, leading to residual haze. To overcome these problems, this study proposes a single image dehazing method of multiscale deep-learning based on dual-domain decomposition. This method develops a multiscale deep-learning network model that includes low- and high-frequency dehazing subnets. Firstly, the hazy image is decomposed using bilateral filters to obtain high- and low-frequency sub-images of the hazy image. Subsequently, the mapping relations between the high- and low-frequency sub-images as well as the high- and low-frequency transmissivity of the hazy image are learned using the developed network model. The high- and low-frequency transmissivity obtained by model learning is fused to obtain the scene transmissivity of the original hazy image. Finally, the hazy image is restored to the dehazed image based on the atmospheric scattering model, which is trained and tested using the hazy image dataset. The experimental results denote that the proposed method can achieve a good dehazing effect for the synthetic hazy images and real natural hazy images and that it is superior to other contrast algorithms in subjective and objective evaluations.