Research On Exemplar-based Image Inpainting Algorithms

Digital image restoration technique is to repair the damaged region or blank regions after removing redundant objects,and then maintain the integrity of image and the overall visual effect.It is one of the important contents in the field of computer graphics and computer vision.Digital image inpainting algorithm can be divided into two categories: one is based on partial differential equation,which is suitable to restore small damaged region,such as scratches and spots.The other is based on texture synthesis,which is suitable for the restoration of large missing region.It is the research center of image inpainting field and represents the study direction in the future.This paper mainly focuses on the exemplar-based texture synthesis algorithm,in which the damaged region is inpainted by iterative matching and copying patches from known region to data missing region,and then the image restoration is achieved.Firstly,this thesis presents the background and research status,as well as the significance,of digital image inpainting.We focus on analyzing the exemplar-based texture synthesis algorithm,and clarify the various implementation stages,as well as their effects on final effect.It is pointed that the image restoration effect is mainly determined by the calculation of priority,confidence update and similarity measure function.Secondly,this thesis summarizes the advantages and disadvantages of the traditional exemplar-based texture synthesis algorithm.We point out that the priority value easily tends to zero quickly due to the rapid falling of the traditional confidence value.It will result in a wrong guided direction and a not-satisfactory inpainting effect.In addition,the matching criterion of Sum of Squared Differences(SSD),which is widely used in traditional algorithm to search for the optimal matching block,is too single for only considering the difference of color information,thus the matching precision is greatly limited due to the lacking of robustness.To solve the problems above,we put forward a new confidence update function to suppress the fast attenuation of confidence and then improve the accuracy of guided repairing direction.At the same time,the Census transform matching criterion is introduced and combined with the traditional SSD criterion to improve the matching precision.Experimental simulation results show that satisfactory results can be received even in complex structural images by the proposed algorithm.Finally,this thesis proposes an image restoration algorithm based on multi-scale decomposition.In order to overcome the problem of losing edge and detail information in the process of down-sampling,Weighted Guided Image Filtering is introduced due to its perfect effect in edge preserving.Moreover,the curvature and structure characteristic information are introduced to calculate the priority in each scale,thus the filling and repairing order become more reasonable.Experimental simulation results prove the effectiveness of the proposed algorithm.