Video Denoising Based On MeshFlow

This thesis addresses the problem of video denoising.As a kind of recordable information medium,digital images and videos have always played a very important role in human civilization.With the rapid development of electronic and multimedia technology,a series of new occupations based on media communication have been developed for years.However,digital images and videos often introduce noise in a series of processes such as acquisition,transmission and storage,which brings a bad visual experience to the receiver.How to remove noise so that it can present information to the greatest extent is particularly important.Some traditional image and video denoising algorithms introduce artificial noise or greatly blur the image while filtering noise,so there are great limitations;some have a better denoising effects,but edge information lost,or computational complexity,resulting in inefficiencies.Based on it,we propose an efficient video denoising approach that produces clean videos by utilizing the meshflow motion model for the camera motion compensation.The meshflow is a spatially-smooth sparse motion field with motion vectors located at the mesh vertexes.The model is not only efficient but also effective for the purpose of the multi-frames denoising due to its internal characteristics such as lightweight,nonparametric representation,and spatially-variant motion compensation.We extract pixel profiles instead of adopting real motion trajectories for the efficient motion accumulation.A pixel profile is a collection of motion vectors along the time at a pixel's spatial location.In meshflow,pixel profile is referred to as vertex profile,because only sparse motions at the mesh vertex positions,instead of every pixel location,are of the interest.A denoised frame is generated by fusing of several registered frames in a spatial and temporal manner with consistency verification for outlier rejections.When a noisy video is inputted,we estimate the meshflow between adjacent frames,then All the neighboring frames within a local temporal window are warped towards the central frame according to the meshflow.The denoised frame is generated by fusing of warped frames with consistent pixel identification.the window is moved forward by one frame at each time and the video is denoised frame by frame.Moreover,we optimize the algorithm framework in the researching process.We show that it is possible to align frames to common “key” frames to save the computations without sacrificing the denoising quality.Combined with several efficient solutions,our method can denoise videos(720P)with the speed of 40 fps.Moreover,our method can not only denoise offline video,but also can be adapted for online processing.Experiments on stabilizing challenging video scenes demonstrate the effectiveness and robustness of our technique.