From Coarse to Fine: A Two-Stage Network for Dense Haze removal

Dense haze degrades details and obscures color information, which increases the difficulty of extracting global features at deep layers of CNN. In this paper, we propose a two-stage progressive network for single image dehazing. Unlike previous typical learning-based approaches that use hazy and clean image pairs to train a mapping between the two directly, our method learns different levels of haze with different supervision in each stage. Hence the network is divided into rough and refinement as two steps and each step is modeled based on encoder-decoder architecture. To handle different size objects and uneven haze distribution, we utilize the special designed Pyramid Dilated Convolution Block (PDCB) and the latest Fusion Attention Module (FAM) which are embedded in the refinement network to form the final dehazed image with R, G and B color channels. Based on this model, we formulate a loss function that adopts contrast enhancement and color consistency constraint to recover distinct contrast and natural color. We perform extensive experiments on our new dense-haze dataset synthesized with benchmark RESIDE and real-world images, and show our network is effective to deal with previously challenging hazy images.