Progressive deep video dehazing without explicit alignment estimation

To solve the issue of video dehazing, there are two main tasks to attain: how align adjacent frames reference frame; restore frame. Some papers adopt explicit approaches (e.g., Markov random field, optical flow, deformable convolution, 3D convolution) neighboring with frame in feature space or image space, then various restoration methods used achieve final dehazing results. In this paper, we propose a progressive alignment and method for which fuses process simplifies complexity network models. The aligns consecutive stage by without using flow estimation convolution. is not only implemented under but also uses refinement improve performance whole network. proposed networks include four fusion one decrease parameters networks, three share same first stage. Extensive experiments demonstrate that achieves outstanding against the-state-of-art methods.