Research On An Adaptive Image Inpainting Algorithm

Image inpainting，which is a research hotspot of computer vision and computergraphics，is to complete a painting by filling the useful information into the damageddomains of a given image. Digital inpainting techniques have found broad applicationsin digital restoration of ancient paintings for conservation purposes, vision analysis,restoration of old photographs or films with scratches or missing patches, text removaland objects removal in images for special effects, and so on.The main ideal of models of image inpainting is based on partial differentialequation (PDE). The most commonly used models are BSCB model, TV model andCDD model. They propagate only the local useful information around the inpaintingdomain automatically, so they are applied only for structure images. On the basis ofthese local PDE-based inpainting models, a nonlocal CDD model for image inpaintingis introduced by incorporating the nonlocal differential operators into the CDD model inthis paper. This method utilize the pixels of similar structures to estimate the lost pixels，which makes this new model perform very efficiently for inpainting images, especiallyfine texture images.Therefore, for the advantages and disadvantages of each model, in this paper, anadaptive image inpainting algorithm based on non-local Curvature Driven Diffusion isproposed. First, the damaged area of the image can be classified into structure andtexture part automatically. Then the improved non-local CDD model is used to repairthe fine texture part and an adaptive image inpainting algorithm is used to inpainting theremaining part. The improved non-local CDD model can take full advantage of theglobal information of the image and increase the rate of diffusion. According to theimage edge details, adaptive parameters q can be chosen for the suitable inpaintingmodel, which makes the algorithm choose C&E model when damaged area containsmany edges, and otherwise choose the QCDD model to obtain a better inpainting effectand reduce the inpainting time.The performances of the different models are analyzed by comparing with theexperimental results of TV model, CDD model, improved CDD model and the proposedadaptive model. The experimental results show the superiority of the proposed adaptivemodel.