Multi-Scale Densely Connected Dehazing Network

Single image dehazing is a challenging ill-posed problem. The traditional methods mainly focus on estimating the transmission of atmospheric-light medium with some priors or constraints. In this paper, we propose a novel end-to-end convolutional neural network (CNN) for image dehazing, called multi-scale densely connected dehazing network (MDCDN). The proposed network consists of a parallel multi-scale densely connected CNN network and an encoder-decoder U net. The parallel multi-scale dense-net can estimate transmission map accurately. The encoder-decoder U net is used to estimate the atmospheric light intensity. The all-in-one training can jointly learn the transmission map, atmospheric light, and dehazing images all together with jointly MSE error and a discriminator loss. We also create a dataset with indoor and outdoor data based on the LFSD, NLPR, and NYU2 depth datasets to train our network. Extensive experiments demonstrate that, in most cases, the proposed method achieves significant improvements over the state-of-the-art methods.