A Novel Color Image Inpainting Guided by Structural Similarity Index Measure and Improved Color Angular Radial Transform.
Sai Hareesh Anamandra, Venkatachalam Chandrasekaran · IPCV · 2010
ABSTRACTImageinpaintingisconsideredasapredictiveprocesstocom-pute the missing image data without introducing undesirableartifacts. Most of the existing methods in the literature workvery well for small regions but introduce blur for large holes.Inthispaper, weproposeanovelunifiedframeworkforaffineand flip invariant inpainting of color images. The proposedmethod combines structural similarity index measure, an im-proved version of color angular radial transform, frequencydomain-based image registration and Dr. Kekre’s LUV spacebased blending. It searches for best candidate regions that aresimilar to the neighbourhood of the inpainting domain eitherinthesameimageorinthelargedatabaseintermsofitsstruc-ture, color and texture simultaneously thereby improving theprediction accuracy. Experimental results indicate perceptu-ally satisfactory results.Keywords— affine and flip invariant inpainting, struc-turalsimilarityindexmeasure,colorangularradialtransform,frequency-based image registration, LUV space based imageblending1. INTRODUCTIONImage inpainting is the art of recovering the original imagefrom images which are generally incomplete due to variousfactors, including degradation due to ageing, damage due towear and tear, missing image details due to occlusion andloss of image data transmitted through a noisy communica-tion channel. In such situations, there is a need to predictthe missing image information without introducing undesir-able artifacts. To an observer, the inpainted image must lookauthentic without bearing any trace of being tampered with.A number of inpainting techniques have been proposed inthe literature. They can be divided into many classes namelymethods based on convolution using kernel [1], neighbour-hood diffusion using isophotes direction [2], Total Variation(TV) model [3], Curvature Driven Diffusion (CDD) model