Font Size: a A A

A Video Inpainting Method Based On Structural Feature Constraint

Posted on:2020-01-12Degree:MasterType:Thesis
Country:ChinaCandidate:R YangFull Text:PDF
GTID:2428330605450780Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Image and video inpainting technology is a rapidly evolving digital technology in recent years.The purpose of the digital video inpainting is to inpaint the missing or damaged part of the video body by using the intact unbroken area information in the video body,so that the restored video picture has a reasonable visual effect.According to the basic principle of video inpainting technology,based on the analysis and simulation of existing video inpainting algorithms,this thesis studies digital video inpainting technology from the perspective of temporal and spatial correlation of video and structural constraint of video content to improve the visual effect of inpainted video.The main research contents and innovation of the thesis are as follows:1.The research significance of digital video inpainting technology and the research status at home and abroad are systematically reviewed.The key technologies and main difficulties in digital video inpainting are analyzed in depth.2.For the debounced video,it is prone to flicker and produce edge breakage,this thesis proposes a video debounce algorithm based on the Criminisi inpainting algorithm in order to better maintain the spatiotemporal continuity of the debounced video sequence.Firstly,the proposed algorithm introduces the average motion vector of the video sequence in the motion estimation stage,and uses the angle difference between the motion vector of each frame and the average motion vector to find the video frames with abnormal motion;then uses the geometric transformation(translation,rotation)of the image for abnormal video frames to conform them to the video global motion mode by correcting their positions,according to the relational information of the current frame,the adjacent frames and the motion state matching frame,the priority and the search method of matching block in the Criminisi inpainting algorithm are improved.Finally,the improved Criminisi algorithm is used to inpaint the edge vacancies caused by the geometric transformation of the abnormal video frames and complete the debounce of the video.The experimental results show that the proposed video debounce algorithm possesses better spatiotemporal continuity,no flickering,and the peak signal-to-noise ratio(PSNR)of the video sequence after debounce is increased by 3?5 d B on average,the structural similarity(SSIM)is also improved.3.For the consistent video inpainting algorithm,only the color and motion characteristics of the video are used to optimize the objective function,which leads to the inability to accurately predict the pixel values of the unknown region.This thesis introduces the structural informationof the video into the objective function and constrains it by increasing the structural feature of the video.To optimize the global objective function,the paper uses the average edge vector field model to locate the edge,and proposes a structurally constrained video inpainting method.Firstly,the proposed algorithm makes a spatiotemporal pyramid transform for the broken video,forms the multilayer pyramid decomposition layers,and uses the improved edge tracking algorithm to find the initial boundary of the broken part of each layer.Then,uses the weighted average method to inpaint the pixels on the boundary of broken part in the first pyramid from top to bottom.When all the pixels on the boundary have been inpainted,the boundary is updated,and the pixels on the new boundary are inpainted,the first layer of the pyramid is inpainted.The layers in the pyramid are inpainted in the same way until all the broken pixels are inpainted.Finally,the video is reconstructed by using the inpainted video pyramid layers to complete the inpainting of the entire broken video.The experimental results show that the algorithm obtains better visual effect and maintains better structure and edge continuity.The peak signal-to-noise ratio(PSNR)increases by an average of 2?5d B compared to similar related algorithms,structural similarity(SSIM)has also improved.
Keywords/Search Tags:video inpainting, spatiotemporal pyramid, global objective function, structural constraint factor, approximate nearest neighbour search, boundary tracking, edge vector field, texture synthesis, video debounce
PDF Full Text Request
Related items