test haproxy public

Description

Image Cluster Compression using Partitioned Iterated Function Systems and efficient Inter Image Similarity Features Matthias Kramm Technical University of Munich Institute for Computer Science Boltzmannstr. 3 D 85748 Garching Email: [e-mail] Abstract—When dealing with large scale image archive sys compression rate of the individual images), which limits the tems, efficient data compression is crucial for the economic size of possible image groups. storage of data. Currently, most image compression algorithms Inthispaper,wepresentanovelalgorithmforimagegroups, only work on a per picture basis — however most image which is based on PIFS compression [6], and thus managesdatabases (both private and commercial) contain high redundan to exploit several high level redundancies, in particular scaledcies between images, especially when a lot of images of the same objects, persons, locations, or made with the same camera, exist. image parts. Inordertoexploitthosecorrelations,it’sdesirabletoapplyimage Compression of image sequences using PIFS was done compression not only to individual images, but also to groups of previously (in the context of video compression) in [7], [8]. images, in order to gain better compression rates by exploiting However, in these papers, both the frames/images contributinginter image redundancies.

Informations

• Publié par: kyann
• Publié le: 27 novembre 2012
• Nombre de lectures: 2
• Langue: English
• Poids de l'ouvrage: 1 Mo

Extrait

ImageClusterCompressionusingPartitioned IteratedFunctionSystemsandefficientInter-Image SimilarityFeatures MatthiasKramm TechnicalUniversityofMunich InstituteforComputerScience Boltzmannstr.3 D-85748Garching Email:[e-mail] Abstract—Whendealingwithlargescaleimagearchivesys- compressionrateoftheindividualimages),whichlimitsthe tems,efficientdatacompressioniscrucialfortheeconomic sizeofpossibleimagegroups. storageofdata.Currently,mostimagecompressionalgorithms Inthispaper,wepresentanovelalgorithmforimagegroups, onlyworkonaper-picturebasis—howevermostimage databases(bothprivateandcommercial)containhighredundan- whichisbasedonPIFScompression[6],andthusmanages ciesbetweenimages,especiallywhenalotofimagesofthesame toexploitseveralhigh-levelredundancies,inparticularscaled objects,persons,locations,ormadewiththesamecamera,exist. imageparts. Inordertoexploitthosecorrelations,it’sdesirabletoapplyimage CompressionofimagesequencesusingPIFSwasdone compressionnotonlytoindividualimages,butalsotogroupsof previously(inthecontextofvideocompression)in[7],[8]. images,inordertogainbettercompressionratesbyexploiting However,inthesepapers,boththeframes/imagescontributing inter-imageredundancies. Thispaperproposestoemployamulti-imagefractalPartitioned toonecompressiongroupaswellastheorderofthoseimages IteratedFunctionSystem(PIFS)forcompressingimagegroupsispredeterminedbythevideosequence.Furthermore,images andexploitingcorrelationsbetweenimages.Inordertopartition needtobeofthesamesize,whichcan’tbeassumedformost animagedatabaseintooptimalgroupstobecompressedwith thisalgorithm,anumberofmetricsarederivedbasedonthe real-worldimagedatabases.Here,wespecifyamulti-image normalizedcompressiondistance(NCD)ofthePIFSalgorithm. PIFSalgorithmwhichworksonimagesofarbitrarysizes,and Wecompareanumberofrelationalandhierarchicalclustering alsoallowstoclusterimagedatabasesintogroupssothat algorithmsbasedonthesaidmetric.Inparticular,weshowhowacompressionofeachgroupisoptimized. reasonablegoodapproximationofoptimalimageclusterscanbe Therestofthispaperisorganizedasfollows:Wefirstderive obtainedbyanapproximationoftheNCDandnCutclustering. themulti-imagePIFSalgorithmbygeneralizingthesingle- WhiletheresultsinthispaperareprimarilyderivedfromPIFS, theycanalsobeleveragedagainstothercompressionalgorithms imagePIFSalgorithm.Wealsodescribeawaytooptimizesaid forimagegroups. algorithmusingDFTlookuptables.Afterwards,wetakeon theproblemofcombiningthe“right”imagesintogroups,by I.I NTRODUCTION firstdescribingefficientwaystocomputeadistancefunction betweentwoimages,andthen,inthenextsession,comparing Extendingimagecompressiontomultipleimageshasnot anumberofclusteringalgorithmsworkingonsuchadistance. attractedmuchresearchsofar.Theonlyexceptionsarethe Thefinalalgorithmisevaluatedbycompressionrunsovera areasofhyperspectralcompression[1]–[3]and,ofcourse, photodatabaseconsistingof3928images. videocompression[4],whichbothhandlethespecialcase ofcompressinghighlycorrelatedimagesofexactlythesameII.T HECOMPRESSIONALGORITHM size. PIFSalgorithmsworkbyadaptivelysplittinganimageI Concerninggeneralizedimagegroupcompression,were- intoanumberofnon-overlappingrectangular“range”blocks centlyresearchedanalgorithmwhichworksbybuildinga R 1 ...R n (usingaquadtreealgorithmwithanerrorthreshold specialeigenimagelibraryforextractingprincipalcomponent  max ),andthenmappingeachrangeblock R ontoa“domain” basedsimilaritiesbetweenimages. block D (with D beingselectedfromanumberofrectangular Whilethealgorithmpresentedin[5]isquitefast,and overlappingdomainblocksD 1 ,...,D m fromthesameimage) managestomergelow-scaleredundancyfrommultipleimages, whichisscaledtothedimensionsofRbyanaffinetransform, itfailstodetectmoreglobalscaleredundancies(inparticular, resultinginablockD ˆ ,andishenceforthprocessedbya similarimagepartswhicharebothtranslatedandscaled), contrastscalingcandaluminanceshiftl: andalsohastheproblemofbecoming“saturated”quitefast (i.e.,themoreimagesinagroup,theworsetheadditionalR xy =cD ˆ xy +l(1)