Content-Based Image Retrieval -A New Emerging Trend For Image Description

The d i f f icu lt ies f aced by te xt bas ed r et r ieva l b eca me mo re and mo re s evere . The ef f ic ien t manage ment o f the rap id ly e xpand ing v is ua l in fo r mat ion beca me an u rg en t p rob le m. Th is need fo r med the d r iv ing fo rc e beh ind the e me rgen ce o f con ten t -bas ed image ret r ieva l te chn iques . Many d i f fe ren t app roaches fo r con ten t -b as ed imag e r et r ieva l hav e been p ropos ed in the l ite ratu re . Succes s fu l approa ches cons id e r no t on ly s imp le featu res l ike co lo r , bu t a ls o take the s t ructu r a l r e lat ions h ip bet ween ob je cts in to ac count . Th e ret r ieva l p ro ces s in o rde r to generat e perc ep tua l ly and s e man t ica l ly mor e me an ing fu l ret r iev a l res u lt s . W e pres en t a new app roa ch that s ign i f ican t ly au to mat es the e xa minat ion p roces s by re ly ing on image ana lys is techn iqu es . The gener a l appro ach is to us e p rev ious ly iden t i f ied con ten t (e .g . , con t raband images ) and to pe r fo r m fe atu r e e xt ract ion , wh ich cap tu res mathe mat ic a l ly the es s en t ia l p rope rt ies o f th e imag es . Bas ed on th is an a lys is , we bu i ld a featu re s et databas e that a l lo ws us to au to ma t ica l ly s can a target mach in e fo r images that a re s imi lar to the ones in the datab as e. 1 . INTRO DUCT IO N Con t en t -b as ed image ret r ieva l , a techn ique wh ich us es v is ua l con ten ts to s earch images f ro m la rge s ca le image databas es acco rd ing to us ers ' in te res ts , h as been an act iv e and f as t advan c ing res ea rch are a s ince the 1990s . Du r ing the pas t decade , re ma r kab le p rog res s has been made in bo th theo ret ic a l res ea rch and s ys tem deve lop ment e .g . D ig it a l fo rens ic inves t igat ions o ften requ i r e the e xa min at ion o f p ictu res found on the ta rget med ia . T wo typ ica l tas ks in that res pect are th e iden t i f ic at ion o f con t raband imag es and the iden t i f ic at ion o f cas e -s pec i f ic images , th e p r es ence o f wh ich can es tab l is h a fa ct o r a log ica l l in k re levan t to the inves t igat ion . Th e es s en t ia l p rob le m is that cu r ren t fo r ens ic too ls a re i l l -equ ipped to he lp the inv es t ig ato r g iven th e s ca le o f the tas k . To i l lus t ra te, we w i l l co l lect thous ands o f imag e f i les on a rando mly s elect ed mach ine in ou r co mput ing lab . Ev en i f the inves t igato r s pends on average a f ract ion o f a s econd on each imag e, it w i l l s t i l l requ i re s ever a l hours o f rou t ine , ted ious wo r k to b ro ws e th rough a l l o f the m. W e wi l l ma ke the e xa mine r’s t as k even mo re d i ff icu lt by re mov ing any inc en t ive fo r us ers to delete imag es . Thus , it is no t un re as onab le to e xp ect th at the hard d r ive o f the ave rage ho me us e r w i l l con ta in hund reds o f thous ands o f images . I f we cons ide r a ta rget s uch as a web hos t ing s erv ice th at can have t ens o f mi l l ions o f images , the p rob le m o f e xa min ing a l l imag es beco mes v i rtua l ly in t ract ab le and inves t igato rs w i l l need s ome means to nar ro w do wn the s earch s pace. S imi la r p rob le ms in t r ad it iona l fo rens ics (e .g . f inge rpr in t iden t i f icat ion ) have been tac k led by bu i ld ing la rge re fe renc e databas es that a l lo w ev id ence f ro m pr ev ious cas es to be au t o ma t ica l ly s earched . Clea r ly , a s ys te m cap ab le o f au to mat ica l ly iden t i fy ing con t raband images found on ta rget med ia by c ros s re fe renc ing a dat abas e o f kno wn imag es cou ld b e o f s ign i f ic an t he lp to inves t igato rs . Th e p rob le m, ho weve r , is that un l ike o the r fo rens ic a rt if acts , con t rab and images typ ica l ly canno t be s to red , even by la w enfo r ce ment agen c ies , fo r fu tu re re fe renc e. As id e f ro m the lega l bar r ie rs , bu i ld ing a s ize ab le r ef er ence databas e to be us ed on a rou t ine bas is by nu me rous agenc ies wou ld be a ch a l leng ing tas k. Fro m a techn ica l po in t o f v ie w , the s to rage and band wid th r equ i re ments wou ld be s tagger ing . Sca lab i l ity wou ld be d i f f icu lt to ach iev e as rep l ic at ion and d is t r ibu t ion o f s uch h igh ly s ens it ive mate r ia l wou ld have to be l imi ted . F ina l ly , a po ten t ia l s ecur ity b rea ch at s uch a s to rage fa c i l ity o r mis us e by au tho r ized pers onne l c an on ly b e co mpa red to a nuclea r acc id en t as fa r as the pub l ic ou tc ry is con ce rned . 2 . Con te n tB as e d Image Re trie va l 2 .1 O ve r vie w Depend ing on the que ry fo r mats , image ret r ieva l a lgor ith ms rough ly be long to t wo catego r ies : t e xt -bas ed app roaches and con ten t -b as ed methods (s e e F igu re 1 ) . The te xt -bas ed app roa ches as s ociat e key wo rds w ith ea ch s to red image . Thes e key wo rds are typ ica l ly gen er ated manua l ly . I mage ret r ieva l then beco mes a s tandard databas e manage ment p rob le m Ho weve r; manua l anno tat ion fo r a la rge co l le ct ion o f imag es is no t always avai lab le . Fu rthe r , it may be d i f f icu lt to des cr ib e image con ten t w ith a s ma l l s et o f key wo rds . Th is mot ivat es res ea rch on con t en t -b as ed image r et r ieva l ( CBI R) , wh er e ret r ieva l o f images is gu ided by p rov id ing a qu ery image o r a s ketch gene rat ed by a us er (e .g ., a s ketch o f a ho rs e ). Figu re 1 : Sche me d iag ra ms o f a t e xt -bas ed image ret r ieva l s ys te m (top) and a con ten t bas ed image r et r ieva l s ys te m (bo t to m) . In the pas t decade, many CBI R s ys tems have b een dev e loped . Exa mp les inc lud e the I BM QBI C Sys te m [FA LO94] , the MI T Photoboo k Sys te m [ P ENT 96] , the Be r ke ley Chabot [ OGLE95] and Blob wo r ld Sys te ms [ CA RS 02], the Vir age Sys te m [ GUP T 97] , Co lu mb ia ’s Vis ua lS EEK and W eb S EEK Sys te ms [ SM IT 96], the P ic Hunte r Sys te m [ COX00] , UCS B’s NeT ra Sys te m [MA 97] , UI UC’s MARS Sys te m [M EHR97] , the P ic To See k Sys te m [ GE VE00] , and Stan fo rd’s W BII S [W A NG98] and S IM P LIc ity Sys te ms [W A NG01] , to na me jus t a fe w . F ro m a co mputat ion a l pe rs pect ive , a typ ica l CBI R s ys tem v ie ws the que ry image and the images in the d atabas e as a co l lect ion o f featu res , and r an ks the re levance b et ween the query and any match ing image in p ropo rt ion to a s imi la r ity me as ure ca lcu lated f ro m th e featu res . Thes e featu res a re typ ica l ly e xt ra cted fro m s hape, te xtu re , in t ens ity , o r co lo r p rope rt ies o f the que ry image and the images in the dat abas e. Thes e featu res a re image s ignat u r es and ch ar acte r ize the con ten t o f images , w ith the s imi la r ity meas u re quan t i fy ing the res e mb lance in con ten t fe atu r es bet we en a pa i r o f imag es . S imi la r ity co mp ar is on is an impo rtan t is s ue in CBI R. In g ene ra l , the co mpa r is on is pe rfo r med e ith er g loba l ly , us ing techn iques s uch as h is togr a m ma tch ing and co lo r layou t inde xing , o r loca l ly , bas ed on deco mpos ed reg ions (ob je cts ) . As a re lat iv e ly matu re method , h is togra m match ing has been app l ied in many gener a l-pu rpos e image r et r ieva l s ys te ms s uch as I BM QBI C, M I T Photoboo k , Vir age Sys te m, and Co lu mb ia Vis ua lS EEK and W ebS EEK. A ma jo r d ra wbac k o f the g loba l h is togra m s ea rch l ies in it s s ens it iv i ty to in tens ity va r iat ions , co lo r d is to rt ions , and cropp ing . In a hu man v is ual s ys te m, a lthough co lo r and te xtu r e a re funda menta l as pects o f v is ua l pe rcep t ions , hu man d is ce rn ment o f cert a in v is ual con t en ts cou ld po t en t ia l ly be as s ociated w ith in te res t ing c las s es o f ob je cts , o r s e mant ic mean ings o f ob je cts in the image . A reg ion -bas ed ret r iev a l s ys te m s egments imag es in to reg ions (ob jects ) , and r et r ieves images bas ed [ 2 ] on the s imi lar ity bet we en reg ions . If image s egmentat ion is id ea l , it is re lat ive ly e as y fo r the s ys te m to iden t i fy ob jects in the image and to match s imi la r ob jects fro m d i f fe ren t images . 2.2 S a mple CB IR architecture 3. Methodo logy : Content-Based Image Retrieva l In con ten t -bas ed image r et r ieva l the us e o f s imp le featu res l ikes co lo r, s hape o r te xtu re is no t s u ff ic ien t . Ins tead , the u lt imate go a l is to cap tu re th e con ten t o f an image v ia e xt ra ct ing th e ob je cts o f the image . Us ua l ly images con ta in an inhe ren t s t ructu r e wh ich may be h ie ra rch ica l . An e xa mp le can be s een in fo l lo wing f igu re . In the fo l lo wing , we des cr ib e t wo mod e ls fo r imag e r epr es en tat ion and s imi la r ity meas u re ment [ 2 ] wh ich ta ke s t ru ctu r a l as we l l as con ten t fe atu r es l ike co lo r in to account .