Deep Flow-Guided Video Dehazing Based on Multi-scale Recurrent Network

Abstract

We present a deep flow-guided video dehazing method based on multi-scale recurrent network dehazing. First, optical flow estimation is used for extracting inter-frame information between adjacent frames. Next, the adjacent frames and the target frames are aligned after deformation operation. Finally, the aligned frames are fed into a multi-scale recurrent network for dehazing, which employs an encoder-decoder structure that incorporates residual blocks. Moreover, this method introduces a hybrid loss method including hard flow example mining loss, contrastive regularization loss and bright channel loss to improve the training accuracy. To further improve the generalization ability, we added real datasets to train the network, which is different from the previous training method that used only synthetic datasets. Extensive experimental results demonstrate that the proposed video defogging algorithm outperforms other mainstream algorithms.