为了解决单幅图像的去雾问题,提出一种新型端到端的网络,该网络利用改进的多尺度特征循环生成对抗网络。不同于以往的模型,所提的网络不依赖于传统的大气散射模型,并且在训练的过程中不需要对应匹配图像,大大简化了训练过程。接着设计一种新型的多尺度生成器,采用双通道融合的特征金字塔结构来最大程度地提取图像中的特征,同时引入多个全局和局部的鉴别器来改善网络性能与图像质量。实验结果表明,所提的模型在不同的数据集上都可以取得很好的结果。

In order to solve the problem of dehazing a single image, a new end-to-end network is proposed, which uses an improved multi-scale feature loop to generate a confrontation network. Unlike previous models, the proposed network does not rely on traditional atmospheric scattering models, and does not need to correspond to matching images during the training process, which greatly simplifies the training process. Next, a new type of multi-scale generator is designed, which uses a dual-channel fusion feature pyramid structure to extract the features in the image to the greatest extent, and introduces multiple global and local discriminators to improve network performance and image quality. Experimental results show that the proposed model can achieve good results on different datasets.