A Multi-scale Dehazing Network with Transmission Range Stretching

Image dehazing has become a significant research area in recent years. However, the traditional dehazing algorithms based on statistics priors cannot adaptive to various conditions of natural hazy images. And those algorithms based on Data-driven learning such as some dehazing networks for estimating transmission almost have the problem that the range of the estimated transmission is too narrow for those haze images where hazy density changes largely. So in this paper, we present a novel Dehazing Network to learn the relationship between the hazy image and its corresponding transmission map. It uses jump connection and the layer of Multi-scale features fusion to obtain more feature related to haze density and use both max pooling and average pooling which in turn remove some details of the transmission map and make the gained transmission map more accurate. Moreover, we also propose a linear stretching algorithm based on dark channel prior to extent the transmission range. The experimental result demonstrate that proposed algorithm achieves favorable result against existing dehazing algorithms on both synthetic images and natural images.