Depth-Guided Video Dehazing With Multi-Stage Fusion

Dehazing is an important problem in computer vision,researches on video dehazing based on deep learning have made much progress.However,there are still some shortcomings in these jobs.For example,they did not extract the information of reference frame thoroughly,they did not consider the concentration of fog and the indoor dataset they used differs from the real existence environment of fog.In this paper,we propose a video dehazing model guided by depth information,which includes dehazing sub-network and depth prediction sub-network.There are three innovations in the network.First,as for the use of reference frames,we insert reference frames into the network stage by stage in the dehazing sub-network,and we process each frame in different stages,so as to fully extract the information of reference frames.Second,in order to distinguish the fog concentration,we estimate the depth map of current frame,and we fuse the depth map and the feature of dehazing sub-network with a non-local structure.Third,as for the depth prediction,we use the sequence information in our depth prediction sub-network.In order to guide the depth prediction of current frame,we get a prediction result at every stage of the depth prediction sub-network and fuse them with the attention mechanism.Besides,we consider the uncertainty of multiple depth maps and we optimize the final depth map with an uncertainty map.In addition,we select 50 videos from the automated driving dataset KITTI,and add two types of fog with different concentrations on every video.By this way,we propose an outdoor video dehazing dataset composed of 100 sequences.We compare our method with several dehazing algorithms with our dataset,the quantitative analysis and visual comparison demonstrate that the proposed method is well behaved in video dehazing.Besides,we perform several ablation experiments,which verifies the necessity of each module in DGVDN.