Indirect Domain Shift for Single Image Dehazing

Despite their remarkable expressibility, convolution neural networks (CNNs) still fall short of delivering satisfactory results on single image dehazing, especially in terms faithful recovery fine texture details. In this paper, we argue that the inadequacy conventional CNN-based dehazing methods can be attributed to fact domain hazy images is too far away from clear images, rendering it difficult train a CNN for learning direct shift through an end-to-end manner and recovering details simultaneously. To address issue, propose add explicit constraints inside deep model guide restoration process. contrast learning, proposed mechanism shifts narrows candidate region estimation output via multiple confident neighborhoods. Therefore, capable consolidating expressibility different architectures, resulting more accurate indirect (IDS) images. We also two training schemes, including hard IDS soft IDS, which further reveal effectiveness method. Our extensive experimental indicate method based dramatically outperforms state-of-the-arts.