Learning an Image Manifold for Retrieval

Size: px
Start display at page:

Download "Learning an Image Manifold for Retrieval"

Transcription

1 Learnng an Image Manfold for Retreval Xaofe He*, We-Yng Ma, and Hong-Jang Zhang Mcrosoft Research Asa Bejng, Chna, *Department of Computer Scence, The Unversty of Chcago ABSTRACT We consder the problem of learnng a mappng functon from low-level feature space to hgh-level semantc space. Under the assumpton that the data le on a submanfold embedded n a hgh dmensonal Eucldean space, we propose a relevance feedback scheme whch s naturally conducted only on the mage manfold n queston rather than the total ambent space. Whle mages are typcally represented by feature vectors n R n, the natural dstance s often dfferent from the dstance nduced by the ambent space R n. The geodesc dstances on manfold are used to measure the smlartes between mages. However, when the number of data ponts s small, t s hard to dscover the ntrnsc manfold structure. Based on user nteractons n a relevance feedback drven query-by-example system, the ntrnsc smlartes between mages can be accurately estmated. We then develop an algorthmc framework to approxmate the optmal mappng functon by a Radal Bass Functon (RBF neural network. The semantcs of a new mage can be nferred by the RBF neural network. Expermental results show that our approach s effectve n mprovng the performance of content-based mage retreval systems. Categores and Subject Descrptors H.3.1 [Informaton Storage and Retreval]: Content Analyss and Indexng Algorthms, Indexng methods. General Terms Algorthms, Management, Performance, Expermentaton. Keywords Image Retreval, Semantc Space, Manfold Learnng, Dmensonalty Reducton, Remannan Structure 1. INTRODUCTION Content-Based Image Retreval (CBIR [3][9][1][14][1] s a long standng research problem n computer vson and nformaton retreval. Most of prevous mage retreval technques buld on the assumpton that the mage space s Eucldean. However, n many cases, the mage space mght be a non-lnear sub-manfold whch s embedded n the ambent space. Intrnscally, there are two fundamental problems n mage retreval: 1 How do we represent an mage? How do we judge smlarty? Permsson to make dgtal or hard copes of all or part of ths work for personal or classroom use s granted wthout fee provded that copes are not made or dstrbuted for proft or commercal advantage and that copes bear ths notce and the full ctaton on the frst page. To copy otherwse, or republsh, to post on servers or to redstrbute to lsts, requres pror specfc permsson and/or a fee. MM 04, October 10 16, 004, New York, New York, USA. Copyrght 004 ACM /04/0010 $5.00. One possble soluton to these two problems s to learn a mappng functon from the low-level feature space to the hgh-level semantc space. The former s not always consstent wth human percepton whle the latter s what mage retreval system desres to have. Specfcally, f two mages are semantcally smlar, then they are close to each other n semantc space. In ths paper, our approach s to recover semantc structures hdden n the mage feature space such as color, texture, etc. In recent years, much has been wrtten about relevance feedback n content-based mage retreval from the perspectve of machne learnng [16][17][18][19][0], yet most learnng methods only take nto account current query sesson and the knowledge obtaned from the past user nteractons wth the system s forgotten. To compare the effects of dfferent learnng technques, a useful dstncton can be made between short-term learnng wthn a sngle query sesson and long-term learnng over the course of many query sessons [6]. Both short- and long-term learnng processes are necessary for an mage retreval system though the former has been the prmary focus of research so far. We present a long-term learnng method whch learns a radal bass functon neural net-work for mappng the low-level mage features to hghlevel semantc features, based on user nteractons n a relevance feedback drven query-by-example system. As we pont out, the choce of the smlarty measure s a deep queston that les at the core of mage retreval. In recent years, manfold learnng [1][4][11][13][15] has receved lots of attenton and been appled to face recognton [7], graphcs [10], document representaton [5], etc. These research efforts show that manfold structure s more powerful than Eucldean structure for data representaton, even though there s no convncng evdence that such manfold structure s accurately present. Based on the assumpton that the mages resde on a low-dmensonal submanfold, a geometrcally motvated relevance feedback scheme s proposed for mage rankng, whch s naturally conducted only on the mage manfold n queston rather than the total ambent space. It s worthwhle to hghlght several aspects of the framework of analyss presented here: (1 Throughout ths paper, we denote by mage space the set of all the mages. Dfferent from most of prevous geometrybased works whch assume that the mage space s a Eucldean space [8][1], n ths paper, we make a much weaker assumpton that the mage space s a Remannan manfold embedded n the feature space. Partcularly, we call t mage manfold. Generally, the mage manfold has a lower dmen- Ths work was done whle Xaofe He was a summer ntern at Mcrosoft Research Asa. 17

2 sonalty than the feature space. The metrc structure of the mage manfold s nduced but dfferent from the metrc structure of the feature space. Thus, a new algorthm for mage retreval whch takes nto account the ntrnsc metrc structure of the mage manfold s needed. ( Gven enough mages, t s possble to recover the mage manfold. However, f the number of mages s too small, then any algorthm can hardly dscover the ntrnsc metrc structure of the mage manfold. Fortunately, n mage retreval, we can make use of user provded nformaton to learn a semantc space that s locally sometrc to the mage manfold. Ths semantc space s Eucldean and hence the geodesc dstances on the mage manfold can be approxmated by the Eucldean dstances n ths semantc space. Ths ntuton wll be strengthened n our experments. (3 There are two key algorthms n ths framework. One s the retreval algorthm on mage manfold, and the other s an algorthm for learnng a mappng functon from feature space (color, texture, etc. to hgh-level semantc space. The learnng algorthm wll gradually flat the mage manfold, and make t better consstent wth human percepton. That s, f two mages are close (n the sense of Eucldean metrc to each other, they are semantcally smlar to each other. Here, by flat we mean that the mage manfold wll be ultmately equpped wth a flat Remannan metrc defned on t, at whch tme we call t semantc space. The rest of ths paper s organzed as follows: Secton descrbes the proposed retreval algorthm on mage manfold. Secton 3 descrbes the proposed framework for learnng a semantc space to represent the underlyng mage manfold. The expermental results are shown n Secton 4. We gve concludng remarks n Secton 5.. RELEVANCE FEEDBACK ON IMAGE MANIFOLD In many cases, mages may be vsualzed as ponts drawn on a low-dmensonal manfold embedded n a hgh-dmensonal Eucldean space. In ths paper, our objectve s to dscover the mage manfold by a localty-preservng mappng for mage retreval. We propose a geometrcally motvated relevance feedback scheme for mage rankng, whch s conducted on the mage manfold, rather than the total ambent space..1 The Algorthm Let Ω denote the mage database and R denote the set of query mages and relevant mages provded by the user. Our algorthm can be descrbed as follows: 1. Canddate generaton. For each mage x R, we fnd ts k- nearest neghbors C = {y 1, y,, y k }, y j Ω (those mages n R are excluded from selecton. Let C = C 1 C C R. We call C canddate mage set. Note that R C =.. Construct subgraph. Construct a graph G(V, where V=R C. The dstance between any two mages x, x j V s measured as follows: x x j f x x j < ε dst( x, x j = otherwse where ε s a sutable constant. The choce of ε reflects our defnton of localty. We put an edge between x and x j f dst(x, x j. Snce the mages n R are supposed to have some common semantcs, we set ther dstances to zero. That s, dst(x, x j = 0, x, x j R. The constructed graph models the local geometrcal structure of the mage manfold. 3. Dstance measure on mage manfold. To model the geodesc dstances between all pars of mage ponts on the mage manfold M, we fnd the shortest-path dstances n the graph G. The length of a path n G s defned to be the sum of lnk weghts along that path. We then compute the geodesc dstance dst G (x, x j (.e. the shortest path length between all pars of vertces of and j n G, usng Floyd s O( V 3 algorthm. 4. Retreval based on geodesc dstance. To retreve the mages most smlar to the query, we smply sort them accordng to ther geodesc dstances to the query. The top N mages are presented to the user. 5. Update query example set. Add the relevant mages provded by the user nto R. Go back to step 1 untl the user s satsfed.. Geometrcal Justfcaton Our algorthm deals wth fnte data sets of ponts n R n whch are assumed to le on a smooth submanfold M wth low dmensonalty. The algorthm attempts to recover M gven only the data ponts. A crucal stage n the algorthm nvolves estmatng the unknown geodesc dstance n M between data ponts n terms of the graph dstance wth respect to some graph G constructed on the data ponts. The natural Remannan structure on M (nduced from the Eucldean metrc on R n gves rse to a manfold metrc d M defned by: d M ( x, y = nf r { length( r} where r vares over the set of (pecewse smooth arcs connectng x to y n M. Note that d M (x, y s generally dfferent from the Eucldean dstance x-y. Our algorthm makes use of a graph G on the data ponts. Gven such a graph we can defne a metrc, just on the set of data ponts. Let x, y belong to the set {x }. We defne: d G (x, y = the length of the shortest path between x and y Gven the data ponts and graph G, one can compute d G wthout knowledge of the manfold M. Bernsten et al. [] show that the two dstance metrcs (d M and d G approxmate each other arbtrarly closely, as the densty of data ponts tends to nfnty. Here, we gve a smple example to show the advantage of geodesc dstances on manfold over Eucldean dstance, and the advantage of semantc space over low-level mage feature space. Fgure 1 shows a spral on a plane. Consder that the mages of our concern are sampled from the spral. Clearly, t s a onedmensonal manfold. Fgure 1(a shows the Eucldean dstance between data ponts A and B. Fgure 1(b shows the geodesc dstance along the spral. In ths example, the ntrnsc geometrcal structure can only be characterzed by the geodesc dstance. In many real world applcatons, one s often confronted wth the problem that the number of sample ponts s too small to descrbe the underlyng topology of the data. In ths case, the geodesc 18

3 (a (b mappng each mage nto a semantc space n whch the dstances between the mages are consstent wth human percepton. The problem we are gong to solve can be smply stated below: Let S denote the low-level feature space, and T denote the semantc space. Learn a nonlnear mappng functon from S to T, f : x z ( x S, z T whch preserves the local Remannan structure of the low-level feature space. (c Fgure 1. (a Eucldean dstance between data ponts A and B. (b Geodesc dstance between data ponts A and B. (c Seven data ponts sampled from the spral. The geodesc dstances between them can not be accurately estmated. (d 1-D representaton of the spral. dstance on the mage manfold can not be accurately estmated, as can be seen from Fgure 1(c. Fortunately, n mage retreval, the user provded nformaton can help us recover the underlyng structure of the mage manfold. In the next secton, we wll descrbe how to learn a contnuous functon whch maps the data ponts (mages nto a semantc space n whch the Eucldean dstances between mages are consstent wth human percepton, as llustrated n Fgure 1(d. The nonlnear Remannan structure of the manfold (Fgure 1(c can be nferred from the lnear Eucldean structure of the semantc space (Fgure 1(d. It mght be more nterestng to consder ths example n mage retreval doman. Suppose the pont A s the query mage and the other sx ponts denote the mages n database. If we conduct the retreval n the low-level feature space (Fgure 1(c, the pont B wll be selected as relevant mage, no matter what dstance metrc we use, Eucldean or geodesc. Ths s because the ntrnsc Remannan structure of the mage manfold can not be accurately detected due to the lack of suffcent sample ponts. However, f the retreval s conducted n the semantc space (Fgure 1(d, the pont B wll never be selected as relevant mage. Ths s because that, by ncorporatng user provded nformaton, the ntrnsc Remannan structure of the mage manfold can be accurately detected. Clearly, the retreval n semantc space s more consstent wth human percepton. 3. USING MANIFOLD STRUCTURE FOR IMAGE REPRESENTATION In the prevous secton, we have descrbed an algorthm to retreve the user desred mages by modelng the underlyng geometrcal structure of the mage manfold. One problem of ths algorthm s that, f the number of sample mages s very small, then t s dffcult to recover the mage manfold. In ths case, we propose a long-term learnng approach to dscover the true topology of the mage manfold usng user nteractons. To be specfc, we am at (d Our proposed soluton conssts of three steps: 1. Inferrng a semantc matrx B m m from user nteractons, whose entres are the dstances between pars of mages n semantc space T. m s the number of mages n database.. Fnd m ponts {z 1, z,, z m } R k whch preserve parwse dstances specfed n B m m. Laplacan egenmaps [1] s used to fnd such an embeddng. The space n whch the m ponts {z 1, z,, z m } are embedded s called LE semantc space n the rest of the paper. The user provded nformaton s ncorporated nto the LE semantc space. Note that, the LE semantc space s only defned on the mage database. In other words, for a new mage outsde the database, t s unclear how to evaluate ts coordnates n the LE semantc space. 3. Gven m par vectors, (x, z ( = 1,,,m, where x s the mage representaton n low-level feature space, and z s the mage representaton n LE semantc space, tran a radal bass functon (RBF neural network f that accurately predcts future z value gven x. Hence f(x s a semantc representaton of x. The space obtaned by f s called RBFNN semantc space. Note that, f(x z. That s, RBFNN semantc space s an approxmaton of the LE semantc space. However, RBFNN semantc space s defned everywhere. That s, for any mage (ether nsde or outsde the database, ts semantc representaton can be obtaned from the mappng functon. We descrbe the detal of these steps n the followng. 3.1 Inferrng a Dstance Matrx n Semantc Space from User Interactons In ths secton, we descrbe how to nfer a dstance matrx n semantc space from user nteractons. Some prevous work could be found n [6]. Here, we present a smple method to update the dstance matrx gradually. Let B denote the dstance matrx, B j = x x j. Intutvely, the mages marked by the user as postve examples n a query sesson share some common semantcs. Therefore, we can shorten the dstances between them. Let S denote the set of postve examples, S = {s 1, s,, s k }. We can adjust the dstance matrx as follows: B s s B / α ( s, s S s s where α s a sutable constant greater than 1. Smlarly, we can lengthen the dstances between the postve examples and negatve examples, as follows: R R β ( s S, t T s t s t j j 19

4 where T = {t 1, t,, t k } s the set of negatve examples, and β s a sutable constant greater than 1. As the user nteracts wth the retreval system, the dstance matrx wll gradually reflect the dstances between the mages n semantc space whch s consstent wth human percepton. 3. Usng Manfold Structure for Image Representaton In the above subsecton, we have obtaned a dstance matrx n semantc space. In ths subsecton, we dscuss how to fnd the semantc representaton for each mage n database, whle the dstances are preserved. Recently, there has been some renewed nterest [1][15][11] n the problem of developng low dmensonal representatons when data arses from samplng a probablty dstrbuton on a manfold. To choose a proper mappng algorthm, the followng two requrements should be satsfed: 1 Snce the mage dstrbuton n feature space s hghly rregular and nconsstent wth human percepton, the mappng algorthm must have the localty preservng property. The mappng algorthm should explctly take nto account the manfold structure. Based on these two consderatons, we use Laplacan Egenmaps [1] to fnd such a mappng. We frst compute the smlarty matrx as follows: exp( Bj /t f Bj < ε W j = 0 otherwse where t and ε s a sutable constant, and B s the dstance matrx obtaned n the prevous subsecton. Note that, the weght matrx has localty preservng property, whch s the key feature of Laplacan Egenmaps. Suppose y={y 1, y,, y m } s a one-dmensonal map of {x 1, x,, x m } n the LE semantc space. A reasonable crteron for choosng a good map s to mnmze the followng objectve functon under approprate constrants: mn ( y y j Wj y, j The objectve functon wth our choce of weghts W j ncurs a heavy penalty f neghborng ponts x and x j are mapped far apart. Therefore, mnmzng t s an attempt to ensure that f x and x j are close then y and y j are close as well. To mnmze ths objectve functon, t s equvalent to solve the followng egenvector problem: Ly = λdy where D s a dagonal matrx, whose entry s column sum (also row sum, snce W s symmetrc of matrx W, D = Σ j W j. L s called Laplacan matrx, L = D W. Let y (0, y (1,, y (n be the solutons of the above egenvector problem, ordered accordng to ther egenvalues, λ 0 λ 1 λ n. It s easy to show that λ 0 = 0, and y (0 = (1, 1. We leave out y 0 and use the next k egenvectors for embeddng n k-dmensonal Eucldean space. x z = ( y ( 1 ( ( k, y,, y ( j where y s the th entry of the egenvector y (j. z s a k- dmensonal map of mage x n the LE semantc space. In summary, our goal s to fnd a vector representaton (map n semantc space for each mage n database. Dmensonalty reducton tself s not our goal, though we can make the dmensonalty of the LE semantc space much lower than the feature space. 3.3 Learnng the Optmal Mappng Functon In the above secton, every mage n database s mapped nto the semantc space. Now, the problem s that, for a new mage outsde the mage database, t s unclear how to evaluate ts map n the LE semantc space, snce we don t have a mappng functon. Here we present an approach that apples neural network to approxmate the optmal mappng functon, whch ntrnscally dstngushes our framework from prevous work [6]. The optmal mappng functon f* s gven by mnmzng the followng cost functon: m f * = arg mn f ( x z f = 1 where m s the number of mages n database. Clearly, ths s a multvarate nonparametrc regresson problem, snce there s no a pror knowledge about the form of the true mappng functon whch s beng estmated. In ths work, we use radal bass functon (RBF networks, and the standard gradent descent s used as a search technque. The mappng functon learned by RBF networks can be represented by f ( x ω G ( x = h j j= 1 where h s the number of hdden layer neurons, ω j R are the weghts. G s the radal functon defned as follows: G x ( x = exp( σ c where c s the center for G, and σ s the bass functon wdth. The k-dmensonal mappng n semantc space can be represented as follows: x f x = ( f ( x, f ( x,, f ( ( 1 k x where f = [f 1, f,, f k ] s the mappng functon. Snce the mappng functon s approxmated by the RBFNN (radal bass functon neural network, we call ths semantc space RBFNN semantc space. In summary, the RBF neural network approxmates the optmal mappng functon from low-level feature space to semantc space. It s traned off-lne wth the tranng samples {x, z }. The computatonal complexty n retreval process wll be reduced as the dmensonalty of the semantc space s reduced. The mage representaton f(x n RBFNN semantc space s an approxmaton of mage representaton z n LE semantc space,.e., f(x z. For a new mage prevously unseen, t can be smply mapped nto the RBFNN semantc space by the mappng functon f. 4. EXPERIMENTAL RESULTS In ths paper, we focus on mage retreval based on user s relevance feedback to mprove the system s short-term and long-term performances. The user can submt a query mage ether nsde or outsde the database. The system frst computes low-level features of the query mage and then maps t nto semantc space usng the learned mappng functon. The system retreves and ranks the 0

5 retreval n top 0 results number of teratons retreval on mage manfold Ru's algorthm Fgure. Comparson of retreval on mage manfold wth Ru s algorthm. mages n the database. Then, the user provdes hs judgment of the relevance of retreval. Wth the user s relevance feedback, the system refnes the search result teratvely untl the user s satsfed. The accumulated relevance feedbacks are used to construct and update the semantc space, as descrbed n Secton 3. We performed several experments to evaluate the effectveness of our proposed approaches over a large mage dataset. The mage dataset we use conssts of 3,000 mages of 30 semantc categores from the Corel dataset. Each semantc category contans 100 mages. The 3,000 mages are dvded nto two subsets. The frst subset conssts of,700 mages, and each semantc category contans 90 mages. The second subset conssts of 300 mages, and each semantc category contans 10 mages. The frst subset s used as tranng set for learnng the optmal mappng functon. The second subset s for evaluatng the generalzaton capablty of our learnng framework. A retreved mage s consdered correct f t belongs to the same category of the query mage. Three types of color features (color hstogram, color moment, color coherence and three types of texture features (tamura coarseness hstogram, tamura drectonary, pyramd wavelet texture are used n our system. The combned feature vector s 435-dmensonal. We desgned an automatc feedback scheme to model the shortterm retreval process. We only requre the user to provde postve examples. At each teraton, the system selects at most 5 correct mages as postve examples (postve examples n the prevous teratons are excluded from the selecton. These automatc generated feedbacks are used as tranng data to perform shortterm learnng. To model the long-term learnng, we randomly select mages from each category as the queres. For each query, a short-term learnng process s performed and the feedbacks are used to construct the semantc space. The retreval s defned as follows: relevant mages retreved n top N returns Accuracy = N Four experments are conducted. The experment wth the new retreval algorthm on mage manfold s dscussed n Secton 4.1. In Secton 4., we show the mage retreval performance n the learned semantc spaces. The generalzaton capablty s also evaluated. In Secton 4.3 we further test the system s performance n semantc space wth dfferent dmensonaltes. We compare retreval n top 0 results retreval n LE semantc space retreval n RBFNN semantc space retreval n low-level feature space our new algorthm wth Ru s algorthm [1] n semantc space n Secton Retreval on Image Manfold We compare the performance of our proposed retreval algorthm on mage manfold wth the relevance feedback approach descrbed n Ru [1]. We ddn t compare t to other mage retreval methods because our prmary purpose s to analyze the geometrcal structure of the mage space. Specfcally, we am at comparng the Eucldean structure and manfold structure for data representaton n mage retreval. The comparson was made n the lowlevel feature space wth no semantc nformaton nvolved. Fgure shows the expermental result by lookng at the top 0 retrevals. As can be seen, our algorthm outperforms Ru s approach. One reason s that the mage manfold s possbly hghly nonlnear, whle Ru s approach can only dscover the lnear structure. 4. Retreval n Semantc Space number of teratons Fgure 3. Image retreval performance n lowlevel feature space, RBFNN semantc space and LE semantc space. The query mages are from the mage database (tranng set Query Image Insde the Database As we dscussed n Secton 3, there are two dfferent semantc spaces, LE semantc space and RBFNN semantc space. One lmtaton of the LE semantc space s that, t only contans those mages n database,.e., tranng set. It s unclear how to evaluate the map n the LE semantc space for new test data. To overcome ths lmtaton, a mappng functon f from low-level feature space to hgh-level semantc space (LE semantc space s learned by a RBF neural network. That s, the mage representaton n LE semantc space, z, s approxmated by f(x whch s the mage representaton n RBFNN semantc space. Intutvely, the retreval performance n RBFNN semantc space should not be better than that n LE semantc space, snce RBFNN semantc space s an approxmaton of LE semantc space. Fgure 3 shows the retreval performance n low-level feature space, LE semantc space and RBFNN semantc space. Our new retreval algorthm on mage manfold s used. We use the tranng set (700 mages as the mage database. We frst conduct the experment n low-level feature space. As the prevous experment n Secton 4.1, we randomly choose 0% of mages n each semantc class as queres to perform the retreval. The user s relevance feedbacks are used to learn a LE semantc space, a RBFNN 1

6 Retreval n top 0 results VS. Demensonalty of RBFNN semantc space dmensonalty Fgure 5. The ntal retreval (no feedback s provded n RBFNN semantc space wth dfferent dmensonaltes. Fgure 4. Image retreval performance n lowlevel feature space and RBFNN semantc space. The query mages are outsde the database. semantc space, as well as a mappng functon from 435- dmensonal low-level feature space to 30-dmensonal hgh-level semantc space. We wll dscuss how to determne the ntrnsc dmensonalty of the semantc space n the next secton. As can be seen, the retreval performance n semantc spaces s much better than that n low-level feature space. The performance dfference s especally sgnfcant at the 0 th retreval when no user s relevance feedback s provded, whch s our prmary goal. That s, the semantc representaton of the query mage can be learned by the RBF neural network f whch s traned by prevous users nteractons wth the system. In fact, n the real world, f the ntal retreval result s too bad, the user mght lose hs nterest to provde feedbacks. Another observaton s that, the retreval performances n LE semantc space and RBFNN semantc space are almost the same. Ths means that the optmal mappng functon f* can be accurately approxmated by the RBF neural network f. 4.. Query Image outsde the Database --- Generalzaton Capablty Evaluaton Whle usng RBF neural network to solve the regresson problem, a key ssue s ts generalzaton capablty. Generalzaton refers to the neural network producng reasonable outputs for nputs not encountered durng tranng. To evaluate the generalzaton capablty of our model, the 300 mages (testng set are used as queres outsde the mage database (tranng set for testng. These mages have no semantc representatons n LE semantc space, but we can obtan ther semantc representatons n RBFNN semantc space by the mappng functon f. Snce our ntenton s to evaluate the generalzaton capablty of our model, the ntal retreval result s especally mportant when no feedbacks are provded. The precson-scope curves are shown n Fgure 4. As can be seen, the retreval n RBFNN semantc space outperforms that n low-level feature space. Ths means that the semantc representaton of the prevously unseen mages can be accurately learned by the RBF neural network. 4.3 Retreval n Semantc Space wth Dfferent Dmensonaltes One ssue of learnng a semantc space s how to estmate ts ntrnsc dmensonalty. Even though the dmensonalty of lowlevel feature space s normally very hgh, the dmensonalty of semantc space s much lower. In ths secton, we evaluate the retreval performance n semantc spaces wth dfferent dmensonalty. As before, the 300 mages outsde the mage databases are used as the query mages. Both the mages n database and the query mages are mapped nto RBFNN semantc space by the mappng functon. Fgure 5 shows the results. As can be seen, the optmal dmensonalty s closely related to the number of semantc classes n the database. Ths observaton concdes wth that obtaned n [6]. If the mage database admnstrator has a pror knowledge about ths number, t can be used as a gudelne to control the dmensonalty of the semantc space. The system reaches the best performance (n terms of and effcency when the dmensonalty of the semantc space s close to the number of semantc classes. Further compresson of the semantc space wll start to cause nformaton loss and decrease the retreval. 4.4 Comparng Dfferent Retreval Algorthms n Semantc Space In prevous two subsectons, we have evaluated the retreval performance n semantc space usng our retreval algorthm. It s nterestng to see how Ru s algorthm [1] performs n semantc space. Fgure 6 shows the retreval results usng our retreval algorthm and Ru s algorthm. We use the same mage database and the same query mages as n Secton The retreval s conducted n RBFNN semantc space rather than the feature space. As can be seen, Ru s algorthm works almost the same as our algorthm. It s mportant to note that the baselne performance n semantc space s much hgher than that n low-level feature space. Ths observaton confrms our prevous ntuton that the semantc space gets more and more regular (flat and lnear as the user s relevance feedback s ncorporated. To be specfc, n the semantc space, the geodesc dstances are almost equal to the Eucldean dstances (see Fgure 1(d. Hence the Remannan structure of the mage manfold can be nferred from the Eucldean structure of the semantc space.

7 number of teratons retreval on mage manfold Ru's algorthm Fgure 6. The comparson of our retreval algorthm wth Ru s algorthm n semantc space. The performances of these two algorthms are very close. It shows that the semantc space s Eucldean (flat. 5. CONCLUSIONS In ths paper, under the assumpton that the data le on a submanfold hdden n a hgh dmensonal feature space, we developed an algorthmc framework to learn the mappng between low-level mage features and hgh-level semantcs. It utlzes relevance feedback to enhance the performance of mage retreval system from both short- and long-term perspectves. Ths framework gves a soluton to the two fundamental problems n mage retreval: how to judge smlarty and how to represent an mage. To solve the frst problem, the proposed retreval algorthm on mage manfold uses the geodesc dstance rather than Eucldean dstance as the smlarty measure between mages. It takes nto account the Remannan structure of the mage manfold on whch the data may possbly resde. To solve the second problem, two semantc spaces, LE semantc space and RBFNN semantc space, are learned from user s relevance feedback. A mappng functon s approxmated by a RBF neural network. The semantc space gves a Eucldean representaton of the Remannan mage manfold. Several questons reman unclear: 1. We do not know how often and n whch partcular emprcal contexts, the manfold propertes are crucal to account for the underlyng topology of mage data. Whle the results n ths paper provde some ndrect evdence for ths, there stll seems to be no convncng proof that such manfold structures are actually present.. Secondly, and most ntrgungly, whle the noton of semantc space s a very appealng one, the propertes of the true mappng from low-level feature space to hgh-level semantc space remans unclear. It s unclear whether the true mappng s one-to-one, or many-to-one, snce ntutvely two dfferent mages mght have totally the same semantcs. The mappng functon s learned n a statstcal sense. Though the experments show ts strong generalzaton capablty, t stll remans unclear n theory. REFERENCES retreval n top 0 results [1] M. Belkn and P. Nyog, Laplacan egenmaps for dmensonalty reducton and data representaton, Advances n Neural Informaton Processng Systesms, 001. [] B. Bernsten, V. de Slva, J. C. Langford, and J. B. Tenenbaum, Graph approxmatons to geodescs on embedded manfolds, Techncal report, Stanford Unversty, December 000 [3] E. Chang, K. Goh, G. Sychay, and G. Wu, CBSA: Content- Based Soft Annotaton for Multmodal Image Retreval Usng Bayes Pont Machne. IEEE Trans. on Crcuts and Systems for Vdeo Technology, vol. 13, No. 1, Jan [4] X. He and P. Nyog, Localty Preservng Projectons, n Advances n Neural Informaton Processng Systems 16, Vancouver, Canada, 003. [5] X. He, D. Ca, H. Lu and W.-Y. Ma, Localty Preservng Indexng for Document Representaton, n ACM SIGIR conference on Informaton Retreval, Sheffeld, 004. [6] X. He, O. Kng, W.-Y. Ma, M.-J. L, and H.-J. Zhang, Learnng a semantc space from user s relevance feedback for mage retreval, IEEE Trans. on Crcut and System for Vdeo Technology, Jan, 003. [7] X. He, S. Yan, Y. Hu and H.-J. Zhang, Learnng a localty preservng subspace for vsual recognton, n Proc. IEEE Conf. on Computer Vson, Nce, France, 003. [8] Y. Ishkawa, R. Subramanya and C. Faloutsos, MndReader: query databases through multple examples, 4 th Conf. on Very Large Databases, New York, [9] W.-Y. Ma and B. S. Manjunath, Netra: A toolbox for navgatng large mage databases, Multmeda System Journal, vol. 7, pp , [10] W. Matusk, H. Pfster. M. Brand, and L. McMllan, A data-drven reflectance model, n Proc. of SIGGRAPH, 003. [11] S.T. Rowes, and L.K. Saul, Nonlnear dmensonalty reducton by locally lnear embeddng, Scence, vol 90, December 000. [1] Y. Ru and T. S. Huang, Optmzng learnng n mage retreval, n Proc. of IEEE Conf. on Computer Vson and Pattern Recognton, Hlton Head, SC, June 000. [13] H.S. Seung and D. Lee, The manfold ways of percepton, Scence, vol 90, December 000. [14] J. Smth and S.F. Chang, VsualSEEk: A fully automatc conent-based mage query system, ACM Multmeda, [15] J.B. Tenenbaum, V.D. Slva, and J.C. Langford, A global geometrc framework for nonlnear dmensonalty reducton, Scence, Vol 90, December 000. [16] K. Teu and P. Vola, Boostng mage retreval, n Proc. IEEE Conf. on Computer Vson and Pattern Recogntno, Hlton head, SC, June 000. [17] S. Tong and E. Chang, Support vector machne actve learnng for mage retreval, n Proc. ACM Multmeda 001, Ottawa, Canada, 001. [18] N. Vasconcelos and A. Lppman, Learnng from user feedback n mage retreval systems, Advances n Neural Informaton Processng Systems, Denver, Colorado, [19] J. Wang and J. L, Learnng-based lngustc ndexng of pctures wth -D MHMMs, n Proc. ACM Multmeda, pp , Juan Les Pns, France, 00. [0] X. S. Zhou and T.S. Huang, Comparng Dscrmnatng Transformatons and SVM for Learnng durng Multmeda Retreval, n Proc. ACM Multmeda 001, Ottawa, 001. [1] L. Zhu, A. Rao and A. Zhang, A theory of keyblock-based mage retreval, ACM Trans. on Informaton Systems, vol. 0, No., pp. 4-57, 00. 3

Laplacian Eigenmap for Image Retrieval

Laplacian Eigenmap for Image Retrieval Laplacan Egenmap for Image Retreval Xaofe He Partha Nyog Department of Computer Scence The Unversty of Chcago, 1100 E 58 th Street, Chcago, IL 60637 ABSTRACT Dmensonalty reducton has been receved much

More information

LECTURE : MANIFOLD LEARNING

LECTURE : MANIFOLD LEARNING LECTURE : MANIFOLD LEARNING Rta Osadchy Some sldes are due to L.Saul, V. C. Raykar, N. Verma Topcs PCA MDS IsoMap LLE EgenMaps Done! Dmensonalty Reducton Data representaton Inputs are real-valued vectors

More information

Face Recognition University at Buffalo CSE666 Lecture Slides Resources:

Face Recognition University at Buffalo CSE666 Lecture Slides Resources: Face Recognton Unversty at Buffalo CSE666 Lecture Sldes Resources: http://www.face-rec.org/algorthms/ Overvew of face recognton algorthms Correlaton - Pxel based correspondence between two face mages Structural

More information

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points; Subspace clusterng Clusterng Fundamental to all clusterng technques s the choce of dstance measure between data ponts; D q ( ) ( ) 2 x x = x x, j k = 1 k jk Squared Eucldean dstance Assumpton: All features

More information

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Learning the Kernel Parameters in Kernel Minimum Distance Classifier Learnng the Kernel Parameters n Kernel Mnmum Dstance Classfer Daoqang Zhang 1,, Songcan Chen and Zh-Hua Zhou 1* 1 Natonal Laboratory for Novel Software Technology Nanjng Unversty, Nanjng 193, Chna Department

More information

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems A Unfed Framework for Semantcs and Feature Based Relevance Feedback n Image Retreval Systems Ye Lu *, Chunhu Hu 2, Xngquan Zhu 3*, HongJang Zhang 2, Qang Yang * School of Computng Scence Smon Fraser Unversty

More information

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data A Fast Content-Based Multmeda Retreval Technque Usng Compressed Data Borko Furht and Pornvt Saksobhavvat NSF Multmeda Laboratory Florda Atlantc Unversty, Boca Raton, Florda 3343 ABSTRACT In ths paper,

More information

Machine Learning 9. week

Machine Learning 9. week Machne Learnng 9. week Mappng Concept Radal Bass Functons (RBF) RBF Networks 1 Mappng It s probably the best scenaro for the classfcaton of two dataset s to separate them lnearly. As you see n the below

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1 4/14/011 Outlne Dscrmnatve classfers for mage recognton Wednesday, Aprl 13 Krsten Grauman UT-Austn Last tme: wndow-based generc obect detecton basc ppelne face detecton wth boostng as case study Today:

More information

CS 534: Computer Vision Model Fitting

CS 534: Computer Vision Model Fitting CS 534: Computer Vson Model Fttng Sprng 004 Ahmed Elgammal Dept of Computer Scence CS 534 Model Fttng - 1 Outlnes Model fttng s mportant Least-squares fttng Maxmum lkelhood estmaton MAP estmaton Robust

More information

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION SHI-LIANG SUN, HONG-LEI SHI Department of Computer Scence and Technology, East Chna Normal Unversty 500 Dongchuan Road, Shangha 200241, P. R. Chna E-MAIL: slsun@cs.ecnu.edu.cn,

More information

Cluster Analysis of Electrical Behavior

Cluster Analysis of Electrical Behavior Journal of Computer and Communcatons, 205, 3, 88-93 Publshed Onlne May 205 n ScRes. http://www.scrp.org/ournal/cc http://dx.do.org/0.4236/cc.205.350 Cluster Analyss of Electrcal Behavor Ln Lu Ln Lu, School

More information

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning Outlne Artfcal Intellgence and ts applcatons Lecture 8 Unsupervsed Learnng Professor Danel Yeung danyeung@eee.org Dr. Patrck Chan patrckchan@eee.org South Chna Unversty of Technology, Chna Introducton

More information

Lecture 5: Multilayer Perceptrons

Lecture 5: Multilayer Perceptrons Lecture 5: Multlayer Perceptrons Roger Grosse 1 Introducton So far, we ve only talked about lnear models: lnear regresson and lnear bnary classfers. We noted that there are functons that can t be represented

More information

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth

More information

Unsupervised Learning

Unsupervised Learning Pattern Recognton Lecture 8 Outlne Introducton Unsupervsed Learnng Parametrc VS Non-Parametrc Approach Mxture of Denstes Maxmum-Lkelhood Estmates Clusterng Prof. Danel Yeung School of Computer Scence and

More information

MULTI-VIEW ANCHOR GRAPH HASHING

MULTI-VIEW ANCHOR GRAPH HASHING MULTI-VIEW ANCHOR GRAPH HASHING Saehoon Km 1 and Seungjn Cho 1,2 1 Department of Computer Scence and Engneerng, POSTECH, Korea 2 Dvson of IT Convergence Engneerng, POSTECH, Korea {kshkawa, seungjn}@postech.ac.kr

More information

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION Paulo Quntlano 1 & Antono Santa-Rosa 1 Federal Polce Department, Brasla, Brazl. E-mals: quntlano.pqs@dpf.gov.br and

More information

Smoothing Spline ANOVA for variable screening

Smoothing Spline ANOVA for variable screening Smoothng Splne ANOVA for varable screenng a useful tool for metamodels tranng and mult-objectve optmzaton L. Rcco, E. Rgon, A. Turco Outlne RSM Introducton Possble couplng Test case MOO MOO wth Game Theory

More information

Optimizing Document Scoring for Query Retrieval

Optimizing Document Scoring for Query Retrieval Optmzng Document Scorng for Query Retreval Brent Ellwen baellwe@cs.stanford.edu Abstract The goal of ths project was to automate the process of tunng a document query engne. Specfcally, I used machne learnng

More information

S1 Note. Basis functions.

S1 Note. Basis functions. S1 Note. Bass functons. Contents Types of bass functons...1 The Fourer bass...2 B-splne bass...3 Power and type I error rates wth dfferent numbers of bass functons...4 Table S1. Smulaton results of type

More information

Learning a Locality Preserving Subspace for Visual Recognition

Learning a Locality Preserving Subspace for Visual Recognition Learnng a Localty Preservng Subspace for Vsual Recognton Xaofe He *, Shucheng Yan #, Yuxao Hu, and Hong-Jang Zhang Mcrosoft Research Asa, Bejng 100080, Chna * Department of Computer Scence, Unversty of

More information

Classifier Selection Based on Data Complexity Measures *

Classifier Selection Based on Data Complexity Measures * Classfer Selecton Based on Data Complexty Measures * Edth Hernández-Reyes, J.A. Carrasco-Ochoa, and J.Fco. Martínez-Trndad Natonal Insttute for Astrophyscs, Optcs and Electroncs, Lus Enrque Erro No.1 Sta.

More information

Data-dependent Hashing Based on p-stable Distribution

Data-dependent Hashing Based on p-stable Distribution Data-depent Hashng Based on p-stable Dstrbuton Author Ba, Xao, Yang, Hachuan, Zhou, Jun, Ren, Peng, Cheng, Jan Publshed 24 Journal Ttle IEEE Transactons on Image Processng DOI https://do.org/.9/tip.24.2352458

More information

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance Tsnghua Unversty at TAC 2009: Summarzng Mult-documents by Informaton Dstance Chong Long, Mnle Huang, Xaoyan Zhu State Key Laboratory of Intellgent Technology and Systems, Tsnghua Natonal Laboratory for

More information

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour 6.854 Advanced Algorthms Petar Maymounkov Problem Set 11 (November 23, 2005) Wth: Benjamn Rossman, Oren Wemann, and Pouya Kheradpour Problem 1. We reduce vertex cover to MAX-SAT wth weghts, such that the

More information

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification Introducton to Artfcal Intellgence V22.0472-001 Fall 2009 Lecture 24: Nearest-Neghbors & Support Vector Machnes Rob Fergus Dept of Computer Scence, Courant Insttute, NYU Sldes from Danel Yeung, John DeNero

More information

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur FEATURE EXTRACTION Dr. K.Vjayarekha Assocate Dean School of Electrcal and Electroncs Engneerng SASTRA Unversty, Thanjavur613 41 Jont Intatve of IITs and IISc Funded by MHRD Page 1 of 8 Table of Contents

More information

Manifold-Ranking Based Keyword Propagation for Image Retrieval *

Manifold-Ranking Based Keyword Propagation for Image Retrieval * Manfold-Rankng Based Keyword Propagaton for Image Retreval * Hanghang Tong,, Jngru He,, Mngjng L 2, We-Yng Ma 2, Hong-Jang Zhang 2 and Changshu Zhang 3,3 Department of Automaton, Tsnghua Unversty, Bejng

More information

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Parallelism for Nested Loops with Non-uniform and Flow Dependences Parallelsm for Nested Loops wth Non-unform and Flow Dependences Sam-Jn Jeong Dept. of Informaton & Communcaton Engneerng, Cheonan Unversty, 5, Anseo-dong, Cheonan, Chungnam, 330-80, Korea. seong@cheonan.ac.kr

More information

Fuzzy C-Means Initialized by Fixed Threshold Clustering for Improving Image Retrieval

Fuzzy C-Means Initialized by Fixed Threshold Clustering for Improving Image Retrieval Fuzzy -Means Intalzed by Fxed Threshold lusterng for Improvng Image Retreval NAWARA HANSIRI, SIRIPORN SUPRATID,HOM KIMPAN 3 Faculty of Informaton Technology Rangst Unversty Muang-Ake, Paholyotn Road, Patumtan,

More information

TN348: Openlab Module - Colocalization

TN348: Openlab Module - Colocalization TN348: Openlab Module - Colocalzaton Topc The Colocalzaton module provdes the faclty to vsualze and quantfy colocalzaton between pars of mages. The Colocalzaton wndow contans a prevew of the two mages

More information

GSLM Operations Research II Fall 13/14

GSLM Operations Research II Fall 13/14 GSLM 58 Operatons Research II Fall /4 6. Separable Programmng Consder a general NLP mn f(x) s.t. g j (x) b j j =. m. Defnton 6.. The NLP s a separable program f ts objectve functon and all constrants are

More information

Performance Evaluation of Information Retrieval Systems

Performance Evaluation of Information Retrieval Systems Why System Evaluaton? Performance Evaluaton of Informaton Retreval Systems Many sldes n ths secton are adapted from Prof. Joydeep Ghosh (UT ECE) who n turn adapted them from Prof. Dk Lee (Unv. of Scence

More information

CS47300: Web Information Search and Management

CS47300: Web Information Search and Management CS47300: Web Informaton Search and Management Prof. Chrs Clfton 15 September 2017 Materal adapted from course created by Dr. Luo S, now leadng Albaba research group Retreval Models Informaton Need Representaton

More information

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics Introducton G10 NAG Fortran Lbrary Chapter Introducton G10 Smoothng n Statstcs Contents 1 Scope of the Chapter... 2 2 Background to the Problems... 2 2.1 Smoothng Methods... 2 2.2 Smoothng Splnes and Regresson

More information

The Research of Support Vector Machine in Agricultural Data Classification

The Research of Support Vector Machine in Agricultural Data Classification The Research of Support Vector Machne n Agrcultural Data Classfcaton Le Sh, Qguo Duan, Xnmng Ma, Me Weng College of Informaton and Management Scence, HeNan Agrcultural Unversty, Zhengzhou 45000 Chna Zhengzhou

More information

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law)

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law) Machne Learnng Support Vector Machnes (contans materal adapted from talks by Constantn F. Alfers & Ioanns Tsamardnos, and Martn Law) Bryan Pardo, Machne Learnng: EECS 349 Fall 2014 Support Vector Machnes

More information

Unsupervised Learning and Clustering

Unsupervised Learning and Clustering Unsupervsed Learnng and Clusterng Why consder unlabeled samples?. Collectng and labelng large set of samples s costly Gettng recorded speech s free, labelng s tme consumng 2. Classfer could be desgned

More information

Support Vector Machines

Support Vector Machines /9/207 MIST.6060 Busness Intellgence and Data Mnng What are Support Vector Machnes? Support Vector Machnes Support Vector Machnes (SVMs) are supervsed learnng technques that analyze data and recognze patterns.

More information

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique //00 :0 AM Outlne and Readng The Greedy Method The Greedy Method Technque (secton.) Fractonal Knapsack Problem (secton..) Task Schedulng (secton..) Mnmum Spannng Trees (secton.) Change Money Problem Greedy

More information

y and the total sum of

y and the total sum of Lnear regresson Testng for non-lnearty In analytcal chemstry, lnear regresson s commonly used n the constructon of calbraton functons requred for analytcal technques such as gas chromatography, atomc absorpton

More information

SELECTION OF THE NUMBER OF NEIGHBOURS OF EACH DATA POINT FOR THE LOCALLY LINEAR EMBEDDING ALGORITHM

SELECTION OF THE NUMBER OF NEIGHBOURS OF EACH DATA POINT FOR THE LOCALLY LINEAR EMBEDDING ALGORITHM ISSN 392 24X INFORMATION TECHNOLOGY AND CONTROL, 2007, Vol.36, No.4 SELECTION OF THE NUMBER OF NEIGHBOURS OF EACH DATA POINT FOR THE LOCALLY LINEAR EMBEDDING ALGORITHM Rasa Karbauskatė,2, Olga Kurasova,2,

More information

Radial Basis Functions

Radial Basis Functions Radal Bass Functons Mesh Reconstructon Input: pont cloud Output: water-tght manfold mesh Explct Connectvty estmaton Implct Sgned dstance functon estmaton Image from: Reconstructon and Representaton of

More information

Support Vector Machines

Support Vector Machines Support Vector Machnes Decson surface s a hyperplane (lne n 2D) n feature space (smlar to the Perceptron) Arguably, the most mportant recent dscovery n machne learnng In a nutshell: map the data to a predetermned

More information

Determining the Optimal Bandwidth Based on Multi-criterion Fusion

Determining the Optimal Bandwidth Based on Multi-criterion Fusion Proceedngs of 01 4th Internatonal Conference on Machne Learnng and Computng IPCSIT vol. 5 (01) (01) IACSIT Press, Sngapore Determnng the Optmal Bandwdth Based on Mult-crteron Fuson Ha-L Lang 1+, Xan-Mn

More information

Discriminative Dictionary Learning with Pairwise Constraints

Discriminative Dictionary Learning with Pairwise Constraints Dscrmnatve Dctonary Learnng wth Parwse Constrants Humn Guo Zhuoln Jang LARRY S. DAVIS UNIVERSITY OF MARYLAND Nov. 6 th, Outlne Introducton/motvaton Dctonary Learnng Dscrmnatve Dctonary Learnng wth Parwse

More information

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization Problem efntons and Evaluaton Crtera for Computatonal Expensve Optmzaton B. Lu 1, Q. Chen and Q. Zhang 3, J. J. Lang 4, P. N. Suganthan, B. Y. Qu 6 1 epartment of Computng, Glyndwr Unversty, UK Faclty

More information

Out-of-Sample Extensions for LLE, Isomap, MDS, Eigenmaps, and Spectral Clustering

Out-of-Sample Extensions for LLE, Isomap, MDS, Eigenmaps, and Spectral Clustering Out-of-Sample Extensons for LLE, Isomap, MDS, Egenmaps, and Spectral Clusterng Yoshua Bengo, Jean-Franços Paement, Pascal Vncent Olver Delalleau, Ncolas Le Roux and Mare Oumet Département d Informatque

More information

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task Proceedngs of NTCIR-6 Workshop Meetng, May 15-18, 2007, Tokyo, Japan Term Weghtng Classfcaton System Usng the Ch-square Statstc for the Classfcaton Subtask at NTCIR-6 Patent Retreval Task Kotaro Hashmoto

More information

SVM-based Learning for Multiple Model Estimation

SVM-based Learning for Multiple Model Estimation SVM-based Learnng for Multple Model Estmaton Vladmr Cherkassky and Yunqan Ma Department of Electrcal and Computer Engneerng Unversty of Mnnesota Mnneapols, MN 55455 {cherkass,myq}@ece.umn.edu Abstract:

More information

A Binarization Algorithm specialized on Document Images and Photos

A Binarization Algorithm specialized on Document Images and Photos A Bnarzaton Algorthm specalzed on Document mages and Photos Ergna Kavalleratou Dept. of nformaton and Communcaton Systems Engneerng Unversty of the Aegean kavalleratou@aegean.gr Abstract n ths paper, a

More information

A Robust Method for Estimating the Fundamental Matrix

A Robust Method for Estimating the Fundamental Matrix Proc. VIIth Dgtal Image Computng: Technques and Applcatons, Sun C., Talbot H., Ourseln S. and Adraansen T. (Eds.), 0- Dec. 003, Sydney A Robust Method for Estmatng the Fundamental Matrx C.L. Feng and Y.S.

More information

Learning-Based Top-N Selection Query Evaluation over Relational Databases

Learning-Based Top-N Selection Query Evaluation over Relational Databases Learnng-Based Top-N Selecton Query Evaluaton over Relatonal Databases Lang Zhu *, Wey Meng ** * School of Mathematcs and Computer Scence, Hebe Unversty, Baodng, Hebe 071002, Chna, zhu@mal.hbu.edu.cn **

More information

UB at GeoCLEF Department of Geography Abstract

UB at GeoCLEF Department of Geography   Abstract UB at GeoCLEF 2006 Mguel E. Ruz (1), Stuart Shapro (2), June Abbas (1), Slva B. Southwck (1) and Davd Mark (3) State Unversty of New York at Buffalo (1) Department of Lbrary and Informaton Studes (2) Department

More information

Graph-based Clustering

Graph-based Clustering Graphbased Clusterng Transform the data nto a graph representaton ertces are the data ponts to be clustered Edges are eghted based on smlarty beteen data ponts Graph parttonng Þ Each connected component

More information

Lecture #15 Lecture Notes

Lecture #15 Lecture Notes Lecture #15 Lecture Notes The ocean water column s very much a 3-D spatal entt and we need to represent that structure n an economcal way to deal wth t n calculatons. We wll dscuss one way to do so, emprcal

More information

Hermite Splines in Lie Groups as Products of Geodesics

Hermite Splines in Lie Groups as Products of Geodesics Hermte Splnes n Le Groups as Products of Geodescs Ethan Eade Updated May 28, 2017 1 Introducton 1.1 Goal Ths document defnes a curve n the Le group G parametrzed by tme and by structural parameters n the

More information

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision SLAM Summer School 2006 Practcal 2: SLAM usng Monocular Vson Javer Cvera, Unversty of Zaragoza Andrew J. Davson, Imperal College London J.M.M Montel, Unversty of Zaragoza. josemar@unzar.es, jcvera@unzar.es,

More information

Private Information Retrieval (PIR)

Private Information Retrieval (PIR) 2 Levente Buttyán Problem formulaton Alce wants to obtan nformaton from a database, but she does not want the database to learn whch nformaton she wanted e.g., Alce s an nvestor queryng a stock-market

More information

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces Range mages For many structured lght scanners, the range data forms a hghly regular pattern known as a range mage. he samplng pattern s determned by the specfc scanner. Range mage regstraton 1 Examples

More information

An Optimal Algorithm for Prufer Codes *

An Optimal Algorithm for Prufer Codes * J. Software Engneerng & Applcatons, 2009, 2: 111-115 do:10.4236/jsea.2009.22016 Publshed Onlne July 2009 (www.scrp.org/journal/jsea) An Optmal Algorthm for Prufer Codes * Xaodong Wang 1, 2, Le Wang 3,

More information

Wishing you all a Total Quality New Year!

Wishing you all a Total Quality New Year! Total Qualty Management and Sx Sgma Post Graduate Program 214-15 Sesson 4 Vnay Kumar Kalakband Assstant Professor Operatons & Systems Area 1 Wshng you all a Total Qualty New Year! Hope you acheve Sx sgma

More information

Proper Choice of Data Used for the Estimation of Datum Transformation Parameters

Proper Choice of Data Used for the Estimation of Datum Transformation Parameters Proper Choce of Data Used for the Estmaton of Datum Transformaton Parameters Hakan S. KUTOGLU, Turkey Key words: Coordnate systems; transformaton; estmaton, relablty. SUMMARY Advances n technologes and

More information

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS ARPN Journal of Engneerng and Appled Scences 006-017 Asan Research Publshng Network (ARPN). All rghts reserved. NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS Igor Grgoryev, Svetlana

More information

Lecture 4: Principal components

Lecture 4: Principal components /3/6 Lecture 4: Prncpal components 3..6 Multvarate lnear regresson MLR s optmal for the estmaton data...but poor for handlng collnear data Covarance matrx s not nvertble (large condton number) Robustness

More information

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 1. SSDH: Semi-supervised Deep Hashing for Large Scale Image Retrieval

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 1. SSDH: Semi-supervised Deep Hashing for Large Scale Image Retrieval IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY SSDH: Sem-supervsed Deep Hashng for Large Scale Image Retreval Jan Zhang, and Yuxn Peng arxv:607.08477v2 [cs.cv] 8 Jun 207 Abstract Hashng

More information

Geometry-aware Metric Learning

Geometry-aware Metric Learning Zhengdong Lu LUZ@CS.UTEXAS.EDU Prateek Jan PJAIN@CS.UTEXAS.EDU Inderjt S. Dhllon INDERJIT@CS.UTEXAS.EDU Dept. of Computer Scence, Unversty of Texas at Austn, Unversty Staton C5, Austn, TX 78712 Abstract

More information

Semi-Supervised Discriminant Analysis Based On Data Structure

Semi-Supervised Discriminant Analysis Based On Data Structure IOSR Journal of Computer Engneerng (IOSR-JCE) e-issn: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 3, Ver. VII (May Jun. 2015), PP 39-46 www.osrournals.org Sem-Supervsed Dscrmnant Analyss Based On Data

More information

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like:

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like: Self-Organzng Maps (SOM) Turgay İBRİKÇİ, PhD. Outlne Introducton Structures of SOM SOM Archtecture Neghborhoods SOM Algorthm Examples Summary 1 2 Unsupervsed Hebban Learnng US Hebban Learnng, Cntd 3 A

More information

Announcements. Supervised Learning

Announcements. Supervised Learning Announcements See Chapter 5 of Duda, Hart, and Stork. Tutoral by Burge lnked to on web page. Supervsed Learnng Classfcaton wth labeled eamples. Images vectors n hgh-d space. Supervsed Learnng Labeled eamples

More information

Mathematics 256 a course in differential equations for engineering students

Mathematics 256 a course in differential equations for engineering students Mathematcs 56 a course n dfferental equatons for engneerng students Chapter 5. More effcent methods of numercal soluton Euler s method s qute neffcent. Because the error s essentally proportonal to the

More information

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz Compler Desgn Sprng 2014 Regster Allocaton Sample Exercses and Solutons Prof. Pedro C. Dnz USC / Informaton Scences Insttute 4676 Admralty Way, Sute 1001 Marna del Rey, Calforna 90292 pedro@s.edu Regster

More information

A Gradient Difference based Technique for Video Text Detection

A Gradient Difference based Technique for Video Text Detection A Gradent Dfference based Technque for Vdeo Text Detecton Palaahnakote Shvakumara, Trung Quy Phan and Chew Lm Tan School of Computng, Natonal Unversty of Sngapore {shva, phanquyt, tancl }@comp.nus.edu.sg

More information

PRÉSENTATIONS DE PROJETS

PRÉSENTATIONS DE PROJETS PRÉSENTATIONS DE PROJETS Rex Onlne (V. Atanasu) What s Rex? Rex s an onlne browser for collectons of wrtten documents [1]. Asde ths core functon t has however many other applcatons that make t nterestng

More information

A Gradient Difference based Technique for Video Text Detection

A Gradient Difference based Technique for Video Text Detection 2009 10th Internatonal Conference on Document Analyss and Recognton A Gradent Dfference based Technque for Vdeo Text Detecton Palaahnakote Shvakumara, Trung Quy Phan and Chew Lm Tan School of Computng,

More information

IMAGE matting is an important but still challenging problem

IMAGE matting is an important but still challenging problem IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 1 Patch Algnment Manfold Mattng Xuelong L, Fellow, IEEE, Kang Lu, Member, IEEE, Yongsheng Dong, Member, IEEE, and Dacheng Tao, Fellow, IEEE arxv:1904.07588v1

More information

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints Australan Journal of Basc and Appled Scences, 2(4): 1204-1208, 2008 ISSN 1991-8178 Sum of Lnear and Fractonal Multobjectve Programmng Problem under Fuzzy Rules Constrants 1 2 Sanjay Jan and Kalash Lachhwan

More information

FINDING IMPORTANT NODES IN SOCIAL NETWORKS BASED ON MODIFIED PAGERANK

FINDING IMPORTANT NODES IN SOCIAL NETWORKS BASED ON MODIFIED PAGERANK FINDING IMPORTANT NODES IN SOCIAL NETWORKS BASED ON MODIFIED PAGERANK L-qng Qu, Yong-quan Lang 2, Jng-Chen 3, 2 College of Informaton Scence and Technology, Shandong Unversty of Scence and Technology,

More information

Unsupervised Co-segmentation of 3D Shapes via Functional Maps

Unsupervised Co-segmentation of 3D Shapes via Functional Maps Unsupervsed Co-segmentaton of 3D Shapes va Functonal aps Jun Yang School of Electronc and Informaton Engneerng, Lanzhou Jaotong Unversty, Lanzhou 730070, P. R. Chna yangj@mal.lzjtu.cn Zhenhua Tan School

More information

Three supervised learning methods on pen digits character recognition dataset

Three supervised learning methods on pen digits character recognition dataset Three supervsed learnng methods on pen dgts character recognton dataset Chrs Flezach Department of Computer Scence and Engneerng Unversty of Calforna, San Dego San Dego, CA 92093 cflezac@cs.ucsd.edu Satoru

More information

Machine Learning: Algorithms and Applications

Machine Learning: Algorithms and Applications 14/05/1 Machne Learnng: Algorthms and Applcatons Florano Zn Free Unversty of Bozen-Bolzano Faculty of Computer Scence Academc Year 011-01 Lecture 10: 14 May 01 Unsupervsed Learnng cont Sldes courtesy of

More information

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 15

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 15 CS434a/541a: Pattern Recognton Prof. Olga Veksler Lecture 15 Today New Topc: Unsupervsed Learnng Supervsed vs. unsupervsed learnng Unsupervsed learnng Net Tme: parametrc unsupervsed learnng Today: nonparametrc

More information

High-Boost Mesh Filtering for 3-D Shape Enhancement

High-Boost Mesh Filtering for 3-D Shape Enhancement Hgh-Boost Mesh Flterng for 3-D Shape Enhancement Hrokazu Yagou Λ Alexander Belyaev y Damng We z Λ y z ; ; Shape Modelng Laboratory, Unversty of Azu, Azu-Wakamatsu 965-8580 Japan y Computer Graphcs Group,

More information

Semi-Supervised Kernel Mean Shift Clustering

Semi-Supervised Kernel Mean Shift Clustering IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. XX, NO. XX, JANUARY XXXX 1 Sem-Supervsed Kernel Mean Shft Clusterng Saket Anand, Student Member, IEEE, Sushl Mttal, Member, IEEE, Oncel

More information

Programming in Fortran 90 : 2017/2018

Programming in Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Exercse 1 : Evaluaton of functon dependng on nput Wrte a program who evaluate the functon f (x,y) for any two user specfed values

More information

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET 1 BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET TZU-CHENG CHUANG School of Electrcal and Computer Engneerng, Purdue Unversty, West Lafayette, Indana 47907 SAUL B. GELFAND School

More information

Solving two-person zero-sum game by Matlab

Solving two-person zero-sum game by Matlab Appled Mechancs and Materals Onlne: 2011-02-02 ISSN: 1662-7482, Vols. 50-51, pp 262-265 do:10.4028/www.scentfc.net/amm.50-51.262 2011 Trans Tech Publcatons, Swtzerland Solvng two-person zero-sum game by

More information

Fitting & Matching. Lecture 4 Prof. Bregler. Slides from: S. Lazebnik, S. Seitz, M. Pollefeys, A. Effros.

Fitting & Matching. Lecture 4 Prof. Bregler. Slides from: S. Lazebnik, S. Seitz, M. Pollefeys, A. Effros. Fttng & Matchng Lecture 4 Prof. Bregler Sldes from: S. Lazebnk, S. Setz, M. Pollefeys, A. Effros. How do we buld panorama? We need to match (algn) mages Matchng wth Features Detect feature ponts n both

More information

Corner-Based Image Alignment using Pyramid Structure with Gradient Vector Similarity

Corner-Based Image Alignment using Pyramid Structure with Gradient Vector Similarity Journal of Sgnal and Informaton Processng, 013, 4, 114-119 do:10.436/jsp.013.43b00 Publshed Onlne August 013 (http://www.scrp.org/journal/jsp) Corner-Based Image Algnment usng Pyramd Structure wth Gradent

More information

Edge Detection in Noisy Images Using the Support Vector Machines

Edge Detection in Noisy Images Using the Support Vector Machines Edge Detecton n Nosy Images Usng the Support Vector Machnes Hlaro Gómez-Moreno, Saturnno Maldonado-Bascón, Francsco López-Ferreras Sgnal Theory and Communcatons Department. Unversty of Alcalá Crta. Madrd-Barcelona

More information

Face Recognition Based on SVM and 2DPCA

Face Recognition Based on SVM and 2DPCA Vol. 4, o. 3, September, 2011 Face Recognton Based on SVM and 2DPCA Tha Hoang Le, Len Bu Faculty of Informaton Technology, HCMC Unversty of Scence Faculty of Informaton Scences and Engneerng, Unversty

More information

Improving Web Image Search using Meta Re-rankers

Improving Web Image Search using Meta Re-rankers VOLUME-1, ISSUE-V (Aug-Sep 2013) IS NOW AVAILABLE AT: www.dcst.com Improvng Web Image Search usng Meta Re-rankers B.Kavtha 1, N. Suata 2 1 Department of Computer Scence and Engneerng, Chtanya Bharath Insttute

More information

Quality Improvement Algorithm for Tetrahedral Mesh Based on Optimal Delaunay Triangulation

Quality Improvement Algorithm for Tetrahedral Mesh Based on Optimal Delaunay Triangulation Intellgent Informaton Management, 013, 5, 191-195 Publshed Onlne November 013 (http://www.scrp.org/journal/m) http://dx.do.org/10.36/m.013.5601 Qualty Improvement Algorthm for Tetrahedral Mesh Based on

More information

Unsupervised Learning and Clustering

Unsupervised Learning and Clustering Unsupervsed Learnng and Clusterng Supervsed vs. Unsupervsed Learnng Up to now we consdered supervsed learnng scenaro, where we are gven 1. samples 1,, n 2. class labels for all samples 1,, n Ths s also

More information

KIDS Lab at ImageCLEF 2012 Personal Photo Retrieval

KIDS Lab at ImageCLEF 2012 Personal Photo Retrieval KD Lab at mageclef 2012 Personal Photo Retreval Cha-We Ku, Been-Chan Chen, Guan-Bn Chen, L-J Gaou, Rong-ng Huang, and ao-en Wang Knowledge, nformaton, and Database ystem Laboratory Department of Computer

More information

Query Clustering Using a Hybrid Query Similarity Measure

Query Clustering Using a Hybrid Query Similarity Measure Query clusterng usng a hybrd query smlarty measure Fu. L., Goh, D.H., & Foo, S. (2004). WSEAS Transacton on Computers, 3(3), 700-705. Query Clusterng Usng a Hybrd Query Smlarty Measure Ln Fu, Don Hoe-Lan

More information

User Authentication Based On Behavioral Mouse Dynamics Biometrics

User Authentication Based On Behavioral Mouse Dynamics Biometrics User Authentcaton Based On Behavoral Mouse Dynamcs Bometrcs Chee-Hyung Yoon Danel Donghyun Km Department of Computer Scence Department of Computer Scence Stanford Unversty Stanford Unversty Stanford, CA

More information

Visual Thesaurus for Color Image Retrieval using Self-Organizing Maps

Visual Thesaurus for Color Image Retrieval using Self-Organizing Maps Vsual Thesaurus for Color Image Retreval usng Self-Organzng Maps Chrstopher C. Yang and Mlo K. Yp Department of System Engneerng and Engneerng Management The Chnese Unversty of Hong Kong, Hong Kong ABSTRACT

More information