Research And Implement On The Improved Algorithm Of Image Inpainting

The research of digital image processing has been paid more and more attention with thedevelopment of the digital age. As a new subject in the digital image processing, image inpaintingtechniques are widely used in other fields of digital image processing, such as image compression,image enhancement and image restoration. The study of image inpainting is of great significance,because it can contribute significantly to the research and development in the other fields. Moreover,the image inpainting technology already becomes a hotspot of research, because it can be widelyused in real life.Image inpainting techniques can be divided in two basic groups. One group of imageinpainting techniques is based on partial differential equations (PDE). This group of PDE-basedtechniques can achieve good results, but they often consume a lot of time to inpaint image each time,which is unacceptable. They are only suitable to inpaint small damaged regions. Another group ofimage inpainting techniques is based on texture synthesis. They not only can inpaint small damagedregions, but also can achieve good results for the big damaged regions or removing big objects. It’stime complexity is lower than the image inpainting techniques based on PDE. This thesis mainlyresearches on the Criminisi algorithm, and presents some meaningful improving methods.The first chapter introduces the background, purpose and significance of this thesis’s research,and describes the digital image inpainting problem. The second chapter describes two classicmodels based on PDE and three classic algorithms based on texture synthesis, and describes theCriminisi algorithm in detail, including algorithm steps and implementation details. In chapter 3, animproved algorithm with variable-size-exemplar is proposed. This chapter first analyzes thelimitations of fixed size exemplar in Criminisi algorithm which can’t properly achieve good results.For that the method of variable size exemplar is proposed. The defects with no considering thepatch matching precision are analyzed, and an improving strategy of updating confidence isproposed. Subsequently this chapter analyzes the global searching strategy in Criminisi algorithmthat leads to a high time complexity, then presents a local searching strategy which can effectivelyimprove the algorithm efficiency. At last the experimental results given in this chapter confirm theeffectiveness of the improved algorithm. Some meaningful improving thoughts are proposed inchapter 4. Using the advantages of multiresolution framework, this chapter introduces ahigh-order-based generalized dissimilarity measure in order to improve the precision of the texturepatch matching. Then the local searching strategy proposed in chapter 3 is further improved. Thisimproved algorithm in this chapter can well inpaint image structure information, but it can not achieve very good results of texture synthesis. Finally, through the analysis of experimental results,the advantages, disadvantages of the improving algorithm in this chapter are pointed out. In chapter5 the conclusion is draw.