Profile Based Sub-Image Search in Image Databases

Transcription

1 Profile Bsed Su-Imge Serh in Imge Dtses Vishwkrm Singh 1, Amuj K. Singh 2 Deprtment of Computer Siene, University of Cliforni, Snt Brr, USA 1 vsingh@s.us.edu, 2 muj@s.us.edu Astrt Su-imge serh with high ury in nturl imges still remins hllenging prolem. This pper proposes new feture vetor lled profile for keypoint in g of visul words model of n imge. The profile of keypoint ptures the sptil geometry of ll the other keypoints in n imge with respet to itself, nd is very effetive in disriminting true mthes from flse mthes. Su-imge serh using profiles is single-phse proess requiring no geometri vlidtion, yields high preision on nturl imges, nd works well on smll visul odeook. The proposed serh tehnique differs from trditionl methods tht first generte set of ndidtes disregrding sptil informtion nd then verify them geometrilly. Conventionl methods lso use lrge odeooks. We hieve preision of 81% on omined dt set of syntheti nd rel nturl imges using odeook size of 500 for top-10 queries; tht is 31% higher thn the onventionl ndidte genertion pproh. I. INTRODUCTION AND MOTIVATION Community ontriuted imge sites (e.g. Flikr.om) nd stok photo sites (e.g. Gettyimges.om) re witnessing n unpreedented growth in the reent dede. Serh of imges y exmple is one of the most ommon tsks performed on these dt sets. A relted tsk is su-imge retrievl [16], [19], [11]. It extends trditionl full-imge serh y llowing users to selet region in n imge nd then, serh for similr regions in repository. Su-imge serh is very importnt tool for hrnessing the potentil of ever growing imge repositories. Su-imge serh s tool is lso eing inorported in online shopping (e.g. like.om) to help ustomers serh produts using imges. It is lso n importnt tool for nlysis nd study of iologil nd medil imge dtsets. Community ontriuted repositories ontin imges of senes or ojets tken under vrying imging onditions. These imges hve ffine, viewpoint, nd photometri differenes nd lso vrying degrees of olusion. Lol desriptors like SIFT [10], [15], [14], [13] re used in literture [16], [22], [7], [20] with fir suess to mesure similrity etween nturl imges. Imges re snned to detet keypoints [5], ovrint regions re extrted round eh point, nd finlly high dimensionl lol feture vetor [10] representtion of eh region is otined [15], [13]. We know the sptil lotion, geometry of the ovrint region, nd high dimensionl feture vetor representtion of eh keypoint deteted in n imge. Thus, eh imge is trnsformed into g of feture vetors. Reserhers hve pursued two-phse tehniques to retrieve similr imges using lol desriptors. The first phse onsists of ndidte genertion disregrding sptil reltionships etween keypoints nd the seond phse onsists of geometri () Query () Result Fig. 1. A pthologil se for su-imge serh using g of words model nd L p norm. Flse mth () is otined in the top-5 results for query imge () euse sptil reltionship is not onsidered. verifition. Cndidtes re generted using two ommon pprohes. One pproh [16], [20] trnsforms eh imge into n orderless g of visul words. Feture vetors of ll the imges re lustered nd eh feture vetor is ssigned the symol of its luster lled visul word. Thus, ll the feture vetors of n imge re trnsformed into visul words nd n imge is finlly represented s histogrm of these visul words. This enles leverging nd extending existing text-retrievl tehniques to imge serh. Distne etween imge histogrms is omputed using L p norm. A given query imge is lso trnsformed into histogrm of visul words nd its top-k ndidtes re retrieved. Another pproh [7], [10] finds top-k nerest neighors of eh keypoint feture vetor of the query imge in the dtset of lol feture vetors of ll the imges. It rnks imges sed on the totl numer of nerest neighors deteted in them nd retins top-k imges s ndidtes. Both pprohes re mde effiient y using indexing tehniques [1], [4]. We present two pthologil results otined y su-imge serh using the first-phse methods s desried previously. Figure 1() is query su-imge nd Figure 1() is n imge in the top-5 results retrieved from the dtset. We n see tht the omponents of the query pttern re sttered rndomly in Figure 1() nd it is not meningful mth. A lol desriptor like SIFT is omputed just using neighorhood pixels round keypoint. Similr keypoints will e omputed in oth the query nd the given result imge euse result imge ontins ll the omponents of the query. The only wy to distinguish etween these two imges is to onsider the sptil lyout of the keypoints. These methods generte this flse mth euse they do not tke sptil reltionships etween keypoints into ount. Figure 2() is nother query su-imge nd Figure 2() is true mth not returned in the top-5 results. An explntion of this nomly is tht the query region is very smll frtion of the whole imge nd therefore, rndom outliers sore etter thn the true mth when similrity is omputed just using orderless g of words representtion of imges. Both these exmples motivte the

2 () Query () Result Fig. 2. A pthologil se for su-imge serh using g of words model nd L p norm. True mth () is not retrieved euse su-imge query () is very smll frtion of the tul imge (). neessity of using sptil reltionships etween mthed points in su-imge serh to get high ury. Existing literture uses geometril verifition (Hough [10], Rnsk [16], [7], [17]) on the ndidte imges to find the est mth or rernk the top-k ndidtes. Generting smll set of high qulity ndidtes is very importnt for ll the onventionl pprohes to redue the high ost of geometri verifition. Geometri verifition lgorithms re itertive in nture nd ostly. Su-imge serh with high ury is very hllenging nd still remins n open prolem. There is need of further development of etter tehniques to retrieve high qulity imges. Suess of su-imge serh tools in optimlly utilizing imge repositories for vrious pplitions is diretly onstrined y the ury of their serh results. The fous of our work is to develop feture vetor nd su-imge similrity mesure using this feture vetor tht yields high ury for suimge serh. In this pper, we present new feture vetor lled profile. We onstrut onentri profile for eh keypoint in g of visul words representtion of n imge (Setion III). A keypoint s profile pproximtely ptures the sptil reltionships of other keypoints in the imge with respet to itself. Suimge retrievl for given query is single-phse serh of the est mthing profile in the dtse (Setion IV). Our feture produes high qulity results for su-imge serh without geometri verifition using single-step serh nd smll visul odeook. We perform experiments to empirilly vlidte our method in Setion V. We show tht our profile hieves omprtively higher ury on the lndmrk dtset provided y Philin et l. [16]. We lso prepred dtset whih is olletion of nturl nd syntheti imges. We used smll odeook of size 500 for g of words representtion of imges ompred to thousnds or even million [16] used in the literture. A onventionl user mostly views top-k results of query [21] returned y serh engine. Therefore, high ury in top-k results is relly importnt. We speifilly fous on empirilly nlyzing ury of top-k mthing imges retrieved using our feture vetor. Our pproh hieves 81% preision on our dtset for top-10 results using ode ook size of 500; this is 31% higher thn the preision of onventionl methods. The min ontriution of our pper is the development of roust feture vetor for eh keypoint tht ptures the sptil reltionship of keypoints in n imge, nd mesure of similrity etween the feture vetors for retrieving similr imges. II. RELATED WORK Svetln et l. [8] proposed to onstrut single feture vetor for the whole imge y ontenting histogrms of lol fetures of su-regions otined y repetedly splitting the imge t vrious sles in prinipled wy (Sptil Pyrmid). They used it for full-imge tegoriztion. This feture is glol feture vetor nd is not diretly designed for su-imge serh. Feture vetor my fil to find those mthing su-regions of dtse imges for given query whih get split into multiple regions during feture vetor onstrution. The similrity sore of su-region of n imge with its full-imge using this feture vetor my e smll, depending on the similrity mesure used, mking it diffiult for the true mth to e distinguished from flse mthes. Further, this feture is neither rottion nor trnsltion invrint euse of the use of orthogonl split lines. Our profile is omputed for eh keypoint of n imge nd mthes re retrieved sed on keypoint profile similrity rther thn fullimge similrity. Therefore, our profile is nturlly suited to su-imge serh. We disuss sle, rottion, nd trnsltion invrine of our profile in Setion III. Weijun et l. [22] proposed to sptilly luster lol desriptors per imge with ound on the numer of points in the luster nd its rdius. They represented eh luster s g of visul words nd eh imge s olletion of these lusters. Imges were rnked sed on the ount of the lusters tht were top-k nerest neighors of the lusters of given query imge. They hieved preision of 65%. Philin et l. [16] lustered the lol desriptors to uild visul odeook of size 1 million nd represented eh imge s g of visul words. They used the L 2 mesure to find ndidtes nd then vlidted those using LO-RANSACK for restrited sets of trnsformtions. They reported men verge preision of 66% on speilized dtset of lndmrk imges using 1 million visul words nd geometri verifition. Yn et l. [7] proposed to generte ndidtes y nerest neighor serh using Lolity Sensitive Hshing [4] nd vlidted using RANSACK for su-imge retrievl for ner duplite imge detetion nd opyright protetion. They experimented only on syntheti dtset. Lowe et l. [10] performed nerest neighor serh using BBF lgorithm [2] nd geometri vlidtion y Hough trnsform to reognize ojets in imges. III. PROFILE CREATION In this setion, we design our new feture vetor lled profile whih is reted for eh keypoint in n imge. A

3 n 1 = 2 n 2 =4 n 3 =8 n 4 =16 n 5 =1 Fig. 3. An imge with its keypoints feture vetors represented s visul words. Conentri irles re drwn round keypoint nd keypoint. The profile of keypoint is defined y set of histogrms. Eh histogrm summrizes the keypoints in ring. The numer of keypoints in ring doules s we move outwrds from the enter. profile of keypoint is struturl representtion of the sptil lyout of ll other keypoints round it. We ssume tht n imge I hs een preproessed nd trnsformed into n orderless g of visul words. We lso know the oordinte of eh keypoint deteted in n imge. To form the profile, we drw onentri irles round eh keypoint p of n imge. Figure 3 shows onentri irles for keypoint nd keypoint. Eh ring is represented s histogrm h of visul words of keypoints lying in it. Profile H = (h 1, h 2,, h m ) of keypoint p, where m is the numer of rings, is ontented list of ring histogrms ordered from the enter towrds the outer rings. The dimension of profile is diretly proportionl to the odeook size nd the numer of rings round keypoint. The numer of points in ring (lled size) inreses s we move wy from the enter, nd is defined y Size(h i ) = 2 Size(h i 1 ). The numer of points in the first ring is n 0 whih is user defined nd is sme for every keypoint s profile ross ll the imges. The ith ring ontins 2 i 1 n 0 points exept in the se of the lst ring where the required numer of points my not exist. It s worth noting tht we do not fix the rdii of the rings ut fix the numer of points in ring, whih is funtion of n 0. As result, numer of rings in the profiles of ll the keypoints in given imge will e the sme ut the rdii of the rings my vry. Rdii of the rings re sed on the sptil density of other keypoints round given keypoint. An imge hving more keypoints will hve more rings thn n imge with less keypoints irrespetive of the size of the imges. Rings of n imge with high sptil density of keypoints will hve nrrower rings ompred to n imge hving low sptil density of keypoints. The profile of eh keypoint of n imge will differ depending on its position in the imge euse keypoints ptured in the rings of different profiles will e different. Similrity etween profiles: The similrity etween two profiles H i nd H j, where the numer of rings in H i is m nd in H j is n respetively, is given y Sim(H i, H j ) = min(m,n) k=0 e λk S k S k = Sim(h k (H i ), h k (H j )) Here, λ is deying prmeter lerned empirilly nd is lwys positive. The similrity etween orresponding ring histogrms (S k ) n e omputed using Jrd s Coeffiient or Cosine mesure. The distne etween two profile is omputed s omplement of their similrity. For the purpose of explntion, ssume Jrd s Coeffiient s the mesure of similrity. For this mesure, the mximum vlue of similrity etween two orresponding histogrms is 1. Therefore, the est vlue of similrity etween two profiles is Sim mx (H i, H j ) = min(m,n) k=0 The distne etween two profiles is given y e λk 1 D(H i, H j ) = Sim mx (H i, H j ) Sim(H i, H j ) = min(m,n) k=0 e λk (1 S k )

4 Fig. 4. The profile of keypoint in query imge Q is sme s its orresponding profile in rotted imge I. This n e represented s reurrene min(m,n) D(H i, H j ) = D(H i, H j ) l + e λk (1 S k ) k=(l+1) where D(H i, H j ) l is the distne omputed for the first l orresponding ring histogrms. Other similrity mesures nd their omplements n e used to find similrity nd distne etween two profiles respetively. We found Jrd s Coeffiient s mesure of similrity etween histogrms to perform etter thn other methods empirilly. The similrity in the proximity of keypoint should e weighed higher ut should e weighed less s we move wy to deimte the effet of noise nd overfitting in mthing. This is the reson for the exponentilly deying ggregtion of similrity etween orresponding rings of profiles nd keeping inresingly more keypoints in the rings wy from the enter. A keypoint is lolized in n imge nd its lol desriptor ptures the property of its smll neighorhood in the imge. Lolity of keypoint mkes it nturlly fit for suimge serh. Our onentri profile ptures sptil lyout of keypoints of the whole imge with respet to given keypoint. The profile of keypoint gives it the strutured glol view of the whole imge mking it semntilly riher thn the keypoint itself. Therefore, the profile of keypoint n distinguish etween true mth nd flse mth effetively. A su-imge serh sed on profiles would hve higher ury thn g of visul words sed serh. Profile Vetor Roustness: Visul words whih form our profile s histogrm re otined from lol desriptors. It should e noted tht lthough lol desriptors like SIFT provide invrine to ffine hnges for similrity serh, feture vetors reted y sptil division of imge round keypoint my ompromise invrine to ffine hnges. Our profile vetor retins invrine to rottion, sling, trnsltion, nd olusion. We perform imge division only in rdil diretion. Therefore, our method yields sme onentri profile for given keypoint irrespetive of the rottion of n imge mking it rottion invrint. We see in Figure 4 tht the profile of keypoint in query imge Q remins sme in rotted imge I. The sling of n imge lters the reltive distne etween points, s seen in imge I of Figure 5. Therefore, fixed rdius Fig. 5. The profile of keypoint in imge Q nd imge I remin the sme in spite of sling. The ring rdius hs inresed in imge I with n inrese in sle. onentri division will fil to provide sle invrine even though the lol desriptors re sle invrint. We keep equl numer of keypoints in rings while onstruting the profile of given keypoint irrespetive of the size nd sle of n imge. This tehnique genertes the sme profile for given keypoint in two imges irrespetive of the sle. The profile of keypoint of imge Q in Figure 5 is the sme s the profile of keypoint in the sled imge I. Therefore, our profile feture remins invrint to sling. Our serh methodology utomtilly preserves trnsltion invrine nd provides roustness to olusion s disussed in Setion IV. IV. SUB-IMAGE SEARCH In this setion, we disuss the method used to ompute the similrity etween pir of profiles. This similrity is used for su-imge serh. Similr keypoints will e extrted from query imge Q nd mthing su-region of dtse imge I y lol desriptor methods. Therefore, the keypoints ptured in the more weighted histogrms of the profile of keypoint q in query imge Q would e sme to the profile of keypoint p in the mthing su-region of dtse imge I. This will give high similrity sore etween the profiles. We exploit this property to retrieve est mthing su-regions from the dtse. All the imges, represented s gs of words, re onverted into gs of profiles. Query Q is lso proessed to generte g of profiles. We serh the est mthing profile p i in dtse for eh query profile q j nd finlly, hoose the highest soring pir (q j, p i ) mong these. The su-region round the keypoint p i of the highest soring pir is the est mthing imge. If Q hs m profiles nd N is the totl numer of profiles in the dtset then the est mthing su-imge is the region round ith profile where i is otined y rg i mx j m { mx i N Sim(H q j, H qi ) Our serh lgorithm inherently provides trnsltion invrine euse it serhes for the est mthing profile. We rete profile for eh keypoint of n imge. Sine, lol desriptors re trnsltion invrint, reltively similr keypoints will e deteted in mthing su-region of trnslted imge nd the given query. Therefore, profiles of the query will hve }

5 Fig. 6. Keypoint is trnslted in imge I. The profile of keypoint in imge I hs high similrity to its profile in imge Q. Fig. 7. The profile of keypoint in imge I hs high similrity with profile of in imge Q even fter olusion. high similrity with the profiles of the keypoints deteted in the mthing su-region ompred to the profiles of rndom keypoints nd our lgorithm will suessfully find these trnslted mthes. Imge I in Figure 6 hs trnsltion of few ojets in imge Q s well s dditionl new ojets. We see tht the onentri profile of keypoint in imge I is very similr to the orresponding profile of in imge Q. Our lgorithm will lso find prtilly oluded ut true mthing su-region of dtse. Profile of keypoint deteted in the preserved prt of the mthing su-region will hve reltively high similrity sore with some keypoint profile of the given query. Profile of keypoint in imge I of Figure 7 will hve high similrity with the profile of keypoint in imge Q. Therefore, the true mthing su-region will rnk higher on similrity sore ompred to rndom profiles. V. EXPERIMENTAL EVALUATION In this setion, we present omprtive study with stte of rt tehnique using the uthor s dtset. Next, we desrie our own dtset nd perform experiments to show tht our profile sed pproh yields high preision for su-imge serh with smll visul odeook. A. Comprtive Evlution We pplied our profile feture sed serh on the Oxford dtset provided y Philin et l. 1 [16]. This dtset hs g of visul words representtion of 5, 062 high resolution (1, vgg/dt/oxuildings/index.html Fig su-imge queries mrked with yellow ox from Oxford Dtset. 768) imges of 11 different Oxford uildings olleted from Flikr. It hs 55 queries with ground truth for omprtive vlidtion. Philin et l. [16] used the L 2 distne on g of words model to rnk ll the imges for eh query nd then omputed the verge preision. They reported men verge preision of 64.5% without geometri vlidtion nd men verge preision of 66% with geometri vlidtion. We hose the top-36 most hllenging queries whih were etter representtive of su-imge serh s seen in Figure 8, nd speifilly left out the queries whih were more of full imge serh. For given k, we retrieved top-k similr imges for eh of the 36 queries from this dtset. We omputed the rtio of true results retrieved to the totl imges retrieved (preision). We verged the preision over ll the 36 queries. We lso retrieved top-k results for eh query y just omputing the L 2 distne on g of words model s proposed y Philin et l. nd omputed men preision. We present the omprtive perentge preision in Tle I for vrying k from 1 to 10. For profile sed serh, we n see tht the top-5 results hve men preision s high s 94% nd it drops to only 86% for top-10 results. For g of words sed serh, men preision drops to 77% for top-10 results. A typil user generlly looks t top few results of query returned y serh engine. We see tht our profile sed feture vetor returns highly urte results for hllenging queries on the lndmrk imges in top-10 ompred to the stte of the rt method.

6 Rnk k % preision for profile sed serh TABLE I % preision for g of words using L 2 sed serh COMPARATIVE PRECISION FOR 36 OXFORD BUILDING QUERIES FOR VARYING k. B. Dtset Preprtion We downloded nturl imges from Flikr. We lso mnully photogrphed lrge numer of senes under vrying onditions. We hose 35 rndom queries from this rel dtset of nturl imges for our experiments. We synthetilly reted some test imges to put our method to even tougher hllenges. We rndomly hose 17 su-imges from the nturl imges nd emedded them into other lrge nturl imges. We lso dded vrying noise levels to these imges s disussed in [12]. Some of the opertions to dd noise were rottion, sling, shering, gussin lur, nd verging noise. We lso split su-imge nd sttered its frgments into other imges. Finlly, we otined dtset of 1, 000 imges (800-nturl nd 200-syntheti). A snpshot of the dtset is shown in Figure 9. We used totl of 52 queries (35-nturl nd 17- syntheti). Next, we reted g of words representtion of eh imge. We extrted ovrint regions [15] from eh imge nd the orresponding 128 dimensionl SIFT vetor. We lustered the feture vetors y piking 500 rndom enters using k- mens [9] lustering. We hose k-mens lustering euse it is not only simple ut emrrssingly prllelizle. A nive omputtion of k-mens hs omplexity of O(mnt) where m is the numer of luster enters, n is numer of dt points, nd t is the numer of itertion for k-mens onvergene. This n e redued y using metri property [3]. We repeted this proess 10 times to hoose the luster set whih hs the minimum verge sum of the squred error or stter. We ssigned the symol to eh luster. We mpped eh SIFT feture vetor to the luster symol to whih it elonged. This gve us g of words representtion of n imge. For reting profiles, we hose n 0 =50 points in the first ring nd used λ=1/3 s deying prmeter for ggregtion. We found the Jrd s Coeffiient to yield etter results thn other distne mesures nd used it to ompute similrities etween ring histogrms of our profiles. C. Preision Test We omputed preision of top-k results otined using our profile sed serh nd ompred it ginst the preision of onventionl methods whih ompute similrity only on the Fig. 9. A snpshot of our experimentl dtset whih hs mny vritions nd is lose representtive of nturl imge dtset. g of words without tking sptil reltionships into ount. We used Cosine, L 1, nd Jrd s Coeffiient s the mesure of similrity for onventionl methods. We onsidered oth the shemes of stndrd tf-idf weighting [18] nd without tfidf weighting of visul words for the ndidte genertion pproh. In tf-idf pproh, ommonly ourring visul words re weighed less s they re less disrimintive. We did not onsider the tf-idf weighting of visul words for our profile sed pproh. Figure 10 nd Figure 11 show the omprtive preision of vrious methods for the nonweighted nd the weighted ses, respetively. We find tht our pproh yields 81% preision rte for top-10 results. Without tf-idf weighting, onventionl method with osine mesure gives the est preision of 50% whih is 31% less thn profile sed method. With tf-idf weighting, onventionl method with Jrd s oeffiient gives the est preision of 39% for top-10 results. Serh with tf-idf weighting hieves less preision ompred to the serh without tf-idf weighting for onventionl methods. We lso experimented y vrying λ for our profile similrity nd hieved more thn 20% higher preision for every λ 1 ompred to onventionl methods. We lso experimented y weighting the symols with the re of ovrint regions ut hieved less preision. We see tht our method hieves higher preision thn onventionl methods used for ndidte genertion on smll odeook size nd without ny geometri verifition. D. Visul Results We present the top-5 visul results for 4 rel queries from our serh in Figure 12. Our profile sed pproh retrieves high qulity results irrespetive of the kind of noise present in the dtset. We outline the mthing su-region in result imge with red ox. We got ll true mthes in the top-5

7 S. No. Query 1st 2nd 3rd 4th 5th Fig. 12. Top-5 results for 4 rel queries over generl imge dtset using profile sed serh Profile Cosine Jrd L Preision Preision 100 Profile Cosine Jrd L No. of Results (k) 8 10 Fig. 10. Comprison of onventionl methods, using different distne mesures without tf-idf weighting, with profile sed pproh. results for the 3rd query whih got pthologil result using onventionl methods shown in Figure 1. We lso verified for the 2nd pthologil se shown in Figure 2 nd found tht the result ws returned in top-10 results nd ll the etter soring imges were true mthes. VI. S CALABILITY TO LARGE DATASETS We my detet thousnds of keypoints in n imge depending on its omplexity nd size. Therefore, we my hve huge set of profiles to serh for given point profile in query imge. This poses hllenge to rel time slility. Though the fous of this pper is to propose new feture vetor tht ptures sptil lyout, here we lso disuss steps to sle our profile sed serh. We hoose disrimintive profiles from eh imge sed on redundny hek using greedy lgorithm. Profiles in given No. of Results (k) 8 10 Fig. 11. Comprison of onventionl methods, using different distne mesures nd tf-idf weighting, with profile sed pproh. imge my overlp to lrge extent depending on the sptil proximity of the orresponding points. A profile is onsidered redundnt if its similrity with ny of the lredy seleted profiles is more thn user given prmeter 1. We luster ll the seleted profiles in the dtset using slle lustering tehnique with progressive refinement. First, we only use the first ring of ll the profiles to generte initil lusters using the tehnique proposed y Hveliwl et l. [6]. We use the remining rings to refine eh luster. It n e seen from Eqution 1 tht the distne does not derese with inlusion of more rings. Refinement is done using the nerest neighor join pproh. Two lusters re merged only if the distne etween ny two profiles in the merged luster is less thn 2. We ssign symol lled visul onept to eh luster.

8 A profile in n imge is represented y the symol of the luster to whih it elongs. We get g of profile symols or visul onepts representtion of eh imge. A visul word just represents n interest point hving very little informtion of its neighorhood wheres visul onept ptures the sptil lyout of keypoints in its neighorhood nd hene is more meningful. Next, we disuss the serh proess for given query. A profile of query imge is ssigned the symol of the luster whih hs minimum inrese in the dimeter. The dimeter of luster is defined s the mximum distne etween ny two profiles in it. This wy we onvert query imge into g of onepts. Similrity etween two imges is omputed using Jrd s oeffiient. Su-linerity in the serh n e hieved y using the LSH [4] indexing tehnique. VII. CONCLUSION In this pper, we developed simple ut effetive profile sed feture vetor tht ptures the sptil reltionships etween keypoints of n imge. Inlusion of sptil lyout in feture vetor improves the serh result drmtilly without the need of geometri verifition. We proposed method to mesure the similrity etween profiles. Our tehnique produed higher ury on lndmrk imges ompred to the stte of the rt method. We evluted our tehnique on mixture of syntheti nd rel nturl imges dtset over 52 queries nd otined 81% preision for top-10 mthes using smll ode ook size of 500. This ws 31% higher thn the onventionl methods. [14] K. Mikoljzyk nd C. Shmid. A performne evlution of lol desriptors. IEEE Trns. PAMI, 27(10), [15] K. Mikoljzyk, T. Tuytelrs, C. Shmid, A. Zissermn, J. Mts, F. Shfflitzky, T. Kdir, nd L. V. Gool. A omprison of ffine region detetors. IJCV, 65. [16] J. Philin, O. Chum, M. Isrd, J. Sivi, nd A. Zissermn. Ojet retrievl with lrge voulries nd fst sptil mthing. In CVPR, pges 1 8, [17] T. Quk, B. Leie, nd L. Vn Gool. World-sle mining of ojets nd events from ommunity photo olletions. In CIVR, [18] G. Slton nd C. Bukley. Term-weighting pprohes in utomti text retrievl. In Informtion Proessing nd Mngement, pges , [19] N. See, M. S. Lew, nd D. P. Huijsmns. Multi-sle su-imge serh. In MM 99: Proeedings of the seventh ACM interntionl onferene on Multimedi, [20] J. Sivi nd A. Zissermn. Video google: A text retrievl pproh to ojet mthing in videos. In ICCV, [21] Z. un nd E. G. Cutrell. An eye trking study of the effet of trget rnk on we serh. In Proeedings of the SIGCHI onferene on Humn ftors in omputing systems, [22] W. Wng, Y. Luo, nd G. Tng. Ojet retrievl using onfigurtions of slient regions. In Proeedings of the 2008 interntionl onferene on Content-sed imge nd video retrievl, REFERENCES [1] S. Ary, D. M. Mount, N. S. Netnyhu, R. Silvermn, nd A. Y. Wu. An optiml lgorithm for pproximte nerest neighor serhing fixed dimensions. J. ACM, 45(6): , [2] J. S. Beis nd D. G. Lowe. Shpe indexing using pproximte nerestneighour serh in high-dimensionl spes. In CVPR, pges , [3] C. Elkn. Using the tringle inequlity to elerte k-mens. In ICML, [4] A. Gionis, P. Indyk, nd R. Motwni. Similrity serh in high dimensions vi hshing. In VLDB, [5] C. Hrris nd M. Stephens. A omined orner nd edge detetion. In Proeedings of The Fourth Alvey Vision Conferene, pges , [6] T. H. Hveliwl, A. Gionis, nd P. Indyk. Slle tehniques for lustering the we (extended strt). In WeDB 2000,In onjuntion with ACM SIGMOD, [7] Y. Ke, R. Sukthnkr, nd L. Huston. An effiient prts-sed nerduplite nd su-imge retrievl system. In MM 04: Proeedings of the 12th nnul ACM interntionl onferene on Multimedi, [8] S. Lzenik, C. Shmid, nd J. Pone. Beyond gs of fetures: Sptil pyrmid mthing for reognizing nturl sene tegories. In CVPR, [9] S. P. Lloyd. Lest squres quntiztion in pm. IEEE Trnstions on Informtion Theory, 28(2): , [10] D. G. Lowe. Distintive imge fetures from sle-invrint keypoints. IJCV, 60:91 110, [11] J. Luo, Nsimento, nd M. A. Content sed su-imge retrievl vi hierrhil tree mthing. In MMDB 03: Proeedings of the 1st ACM interntionl workshop on Multimedi dtses, [12] Y. Meng, E. Chng, nd B. Li. Enhning dpf for ner-repli imge reognition. CVPR, pge 416, [13] K. Mikoljzyk nd C. Shmid. An ffine invrint interest point detetor. In Pro. Europen Conf. Computer Vision, pges Springer Verlg, 2002.