Research Of Image Inpainting Based On Improved Criminisi Algorithm

Image Inpainting is an important part of the digital image processing field.It is the process of filling information in the missing part of the image with the effective area of the image.It is the process of filling information in the missing part of the image by using the known information.Its purpose is to restore the damaged image as much as possible so that the observer cannot perceive the image has been rehabilitated.It has a high application value in terms of cultural relics protection,old photo repair,specific target removal,etc.It has received extensive attention and in-depth research by scholars.At present,digital image Inpainting technology is mainly divided into traditional methods and methods based on deep learning.The traditional methods can be divided into:Image Inpainting based on PDE,and Image Inpainting based on texture synthesis.We will introduce the research background and significance of digital image inpainting,discusses the Image Inpainting based PDE,and analyzes and compares the three classic models:BSCB model,TV model,and CDD model.Then the current popular method based on deep learning is elaborated,and a more classic image inpainting method based on deep learning is given.Highlights of this thesis is Criminisi image inpainting algorithm which is based on texture synthesis.This kind of algorithm is effective in repairing large area of damaged area.It is one of the most popular methods in current Image Inpainting technology.The thesis proposes some improvements to the deficiency of the original algorithm.After improving the algorithm,it proposes a more intelligent specific target removal method,as follows:(1)The priority of the original algorithm will quickly approach zero due to the repair process.The priority function is hierarchically defined,and the edges,textures,and smooth regions that can better distinguish the image damage area are introduced.The structural features of the image block and the variance between the blocks constrain.(2)The original algorithm needs to manually determine the size of the sample block,and the size of the sample block does not change with the repair process.For this defect,a repair algorithm is proposed based on the matching error to make the sample block size adaptive.(3)At present,in the removal of a specific target,the target to be removed needs to be manually annotated.For this disadvantage,our proposes an automatic segmentation based on deep learning and then uses the improved algorithm to remove the target.