Comparative Studyon Super-Resolution ofImages
I. I. Ibrahim, Mohamed Ahmed, Z. B. Nossairl, Ali Mohamed Nabil Allam · 2006
Super-resolution ofimages hasbecomeaveryimportant research topic nowadays. Therearemanyalgorithms thathave beendeveloped toenhance theresolution ofimages. Inthis paper, weundertake astudy forevaluating andcomparing three ofthese algorithms. Thesethreealgorithms are: neuralnetwork algorithm, waveletextremaextrapolation algorithm, and hallucinating facesalgorithm. Ourstudyindicated that:the better performance comesattheexpense ofhigher complexity, large database, andmorecomputational time. Thehallucinating facesalgorithm gives thelargest PeakSignal toNoiseRatio (PSNR)whenmagnifying lowdimensional faces andgives better output whenthedatabase contains larger numberofimages. The neuralnetworkalgorithm givesbetterresults forhigh dimensional faces, butitneedslongtimefortraining. Thewavelet extremaextrapolation algorithm gives better results forhigh dimensional facesthanforlowdimensional faces.The performance ofthesethreealgorithms getsbetteras the dimension ofinput faces getshigher andonlythehallucinating faces cangivegoodresults forlowerdimensional faces suchas 64x48pixels. I.INTRODUCTION Super-resolution istheprocess ofobtaining animage ata resolution higher thanthat afforded bythephysical sensor. In manyapplications suchascomputer vision, remote sensing, industrial inspection, medical imaging, video enhancement, enlarging consumer photo andautomatic target recognition [12],super-resolution isverydesirable andtherefore itbecomes a veryimportant research topic innowadays. Inthisstudy wearespecifically addressing theissue of image magnification. Commonly,magnification is accomplished through convolution ofimagesamples witha single kernel suchasthebilinear, bi-cubic [3], orcubic Bspline [4]. These kernels arevery commonimage interpolation functions buttheysuffer fromblurring ofedges andimage details, todeblur these images recent attempts toimprove on cubic spline interpolation [5-7] havemetwithlimited success. Schreiber andcollaborators [5]proposed asharpen Gaussian interpolator function tominimize information spillover between pixels andoptimize flatness insmooth areas. Schultz andStevenson [8]haveusedaBayesian methodforsuperresolution. Another approach tointerpolation istolearn howto interpolate froma setofhigh-resolution training samples, together withcorresponding lowresolution versions ofthem. InFreeman andPaztor [9-10], thehigh-resolution image is modeled asaMarkov network, whereeachpixel isattached to its neighbors andthecorresponding pixel inthelow-resolution image.FrankCandocia [11] proposed aprocedure, whichis equivalent toconvolution ofimagewithafamily ofkernels developed fromatraining image. SimonBakerandTakeo Kanade [12] usedapyramid-based algorithm tolearn aprior on thederivatives ofthehigh-resolution image asafunction ofthe spatial location intheimage.ZoranCvetkovic andMartin Vetterli [13]proposed a wavelet-based method, which estimates thehigher resolution information needed tosharpen theimage. Inthis study, weevaluate andcompare three ofthese algorithms, whichare:wavelettransform extrema extrapolation, neural network implementation ofinterpolation withfamily ofkernels, andhallucinating faces. Theevaluation ofeachoneofthese algorithms andthecomparison among themarebased onthequality ofoutput images, processing time, sizeofdatabase, PSNRandthedimension ofinput images that canbemagnified andresolution enhanced. We areprimarily interested inhumanfaces inour comparison togaugehowdifficult resolution enhancement is forfaces.