Content Based Color Image Retrieval via Wavelet Transforms

Transcription

1 8 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 Content Bsed Color Imge Retrevl v Wvelet Trnsforms Mrs.Y. M. Lth Dr.B.C.Jng V.S.K.Reddy, GNITS,JNTU,Ind Rector,JNTU,Ind IIT,Khrgpur.Ind Summry Content-Bsed Imge Retrevl (CBIR) llows to utomtclly extrctng trget mges ccordng to objectve vsul contents of the mge tself. Representton of vsul fetures nd smlrty mtch re mportnt ssues n CBIR. In ths pper novel CBIR method s proposed by explot the wvelets whch represent the vsul feture. We use Hr nd D wvelet to decompose color mges nto multlevel scle nd wvelet coeffcents, wth whch we perform mge feture extrcton nd smlrty mtch by mens of F-norm theory. Furthermore, we lso provde progressve mge retrevl strtegy to cheve flexble CBIR. We tested fve ctegores of color mges n the experments. The retrevl performnce of D nd Hr wvelet s compred wth wvelet hstogrms n terms of recll rte nd retrevl speed. Experment results reflect the mportnce of wvelets n CBIR nd F-norm theory long wth progressve retrevl strtegy cheves effcent retrevl. Key words: F-norm theory, progressve mge retrevl strtegy.. Introducton Content-Bsed Imge Retrevl (CBIR) s consdered s the process of retrevng desred mges from huge dtbses bsed on extrcted fetures from the mge themselves (wthout resortng to key word). Fetures re derved drectly from the mges nd they re extrcted nd nlyzed by mens of computer processng[].cbir s bottleneck of the ccess of multmed dtbses tht del wth text, udo, vdeo nd mge dt whch could provde us wth enormous mount of nformton. Mny commercl nd reserch CBIR systems hve been bult nd developed (e.g.: QBIC, Vrge, Pchunter, vsul SEEK, Chbot, Exclbur, photobook, Jcob) []. Content bsed mge retrevl []-[], llowng to utomtclly extrct trgets ccordng to objectve vsul contents of mge tself(e.g. color, texture nd shpe) hs become ncresngly ttrctve, n Multmed Informton Servce System (MISS). Wth ppelng tme frequency loclzton nd mult-scle propertes, wvelet trnsform proved to be effectve n vsul feture extrcton nd representton. It cn be used to chrcterze textures usng sttstcl propertes of the gry levels of the ponts/pxels comprsng surfce mge. In CBIR, wvelet pproches mnly nclude wvelet hstogrm nd wvelet moment of mge, etc. []. Wvelet trnsform cn be used to chrcterze textures usng sttstcl propertes of the gry levels of the pxels comprsng surfce mge []. The wvelet trnsform s tool tht cuts up dt or functons or opertors nto dfferent frequency components nd then studes ech component wth resoluton mtched to ts scle. In ths pper, we used D nd Hr wvelet trnsforms to decompose color mges nto multlevel scle nd wvelet coeffcents, wth whch we perform mge feture extrcton nd smlrty mtch by mens of F-norm theory. We lso present progressve retrevl strtegy, whch contrbutes to flexble compromse between the retrevl speed nd the recll rte. The retrevl performnces re compred wth the exstng wvelet hstogrm technque. The effcency n terms Recll rte nd retrevl speed s tested wth fve types of mges nd the results reflect the mportnce of wvelets n CBIR.Fnorm theory long wth progressve retrevl strtegy mproves retrevl performnce. The rest of ths pper s orgnzed s follows. In secton, we ntroduced the generl structure of the proposed CBIR system. Secton provdes the mge decomposton usng wvelets. Secton descrbes Feture Extrcton nd Smlrty crter. Secton 5 presents progressve CBIR strtegy. Secton 6 descrbes the mplementton nd expermentl results. Fnlly conclusons re offered n secton 7.. Generl structure of Proposed CBIR system Our proposed CBIR lgorthm s bsed on decomposton of the dtbse mges usng Hr nd D wvelets n the offlne s well s n onlne for query mge. Wth resultng coeffcents usng F-norm theory we extrct the fetures nd perform hghly effcent mge mtchng. In the dtbse we used four typcl groups of color mges nmely Sunflowers,,,Lles Mnuscrpt receved December 5, 7 Mnuscrpt revsed December, 7

2 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 9 nd Fshes s shown n the Fg.. On lne processng Query Imge Imge decomposton (Wvelets) Feture extrcton (F-norm Theory) Off lne processng Dt bse Imges. Imge decomposton wth wvelets We used two wvelet pproches for color mge decomposton, nmely Hr, D wvelets. These resultng decomposton coeffcents re employed to perform mge feture extrcton nd smlrty mtch by vrtue of F-norm theory. Hr Wvelet Trnsform Inputs Sun Flowers Fg..Structure of proposed CBIR system Hr trnsform Progressve Retrevl Strtegy Feture extrcton (F-norm Theory) Imge decomposton (Wvelets) Lles Fg.. Hr wvelet forwrd trnsform Retreved Imges Fshes Averges Wvelet coeffcents If dt set S, S.S N- contns N elements, there wll be N/ verges nd N/ wvelet coeffcent vlues. The verges re stored n the upper hlf of the N element rry nd the coeffcents re stored n the lower hlf s shown n the Fg.. The verges become the nput for the next step n the wvelet clculton, where for terton +, N + N /. The recursve tertons contnue untl sngle verge nd sngle coeffcent re clculted. Ths replces the orgnl dt set of N elements wth n verge, followed by set of coeffcents whose sze s n ncresng power of two (Ex:,,... N/). The Hr equtons to clculte n verge nd wvelet coeffcent c from n odd nd even element n the dt set re: s + s+ s - s+ c () Forwrd Hr trnsform for n eght element sgnl s shown n the Fg. Here sgnl s multpled by the forwrd trnsform mtrx. é ù é ù é ù és ù c - s s Ü c c s - s c c s5 c s 6 - ë c û ë c û ë s7 û - ë û Fg..Hr Forwrd wrd trnsform for 8 element sgnl The rrow represents splt operton tht reorders the result so tht the verge vlues re n the frst hlf of the vector nd the coeffcents re n the second hlf. To complete the forwrd Hr trnsform there re two more steps. The next step would multple the verge vlues by x trnsform mtrx, genertng two new verges nd two new coeffcents whch would replce the verges n the frst step. The lst step would multply these new verges by x mtrx genertng the fnl verge nd the fnl coeffcent..the Dubeches D Wvelet Trnsform D wvelet Trnsform N/ sclng coeffcents N/ wvelet coeffcents Fg.. D wvelet forwrd trnsform

3 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 The D trnsform hs four wvelet nd sclng functon coeffcents s shown n the Fg.. The sclng functon coeffcents re: h h + - h + - h () Ech step of the wvelet trnsform pples the sclng functon to the dt nput. If the orgnl dt set hs N vlues, the sclng functon wll be ppled n the wvelet trnsform step to clculte N/ smoothed vlues. In the ordered wvelet trnsform the smoothed vlues re stored n the lower hlf of the N element nput vector. The wvelet functon coeffcent vlues re: g h ; g -h ; g h ; g -h. () Ech step of the wvelet trnsform pples the wvelet functon to the nput dt. If the orgnl dt set hs N vlues, the wvelet functon wll be ppled to clculte N/ dfferences (reflectng chnge n the dt). In the ordered wvelet trnsform the wvelet vlues re stored n the upper hlf of the N element nput vector. The sclng nd wvelet functons re clculted by tkng the nner product of the coeffcents nd four dt vlues. The equtons re shown below. Dubeches D sclng functon h s + h s+ + hs+ + hs + [ ] h s[ ] + h s[ + ] + h s[ + ] + h s[ ] () + Dubeches D wvelet functon c c g s + gs+ + g s+ + g s+ [ ] g s[ ] + g s[ + ] + g s[ + ] + g s[ ] (5) + Ech terton n the wvelet trnsform step clcultes sclng functon vlue nd wvelet functon vlue. The ndex s ncremented by two wth ech terton, nd new sclng nd wvelet functon vlues re clculted. T D D forwrd trnsform mtrx for 8 element sgnl s shown n the Fg.5. éh h h h ù és ù g g g g s h h h h s g g g g s h h h h s g g g g s5 h h h h s 6 ë g g g gû ë s7 û Fg.5. D forwrd trnsform mtrx for 8 element sgnl. Feture Extrcton nd Smlrty crter Our CBIR lgorthm s bsed on drect wvelet decomposton of mge n RGB color spce nd utlzes the query by exmple method. Wth pproches mentoned bove, dtbse mges re decomposed offlne nto mult-level coeffcents from - to -J levels, wth whch, we cn generte color feture dtbse nd perform smlrty mtch between mges. After decomposton, ech resultng sub mge s n fct coeffcent mtrx, where, by specl processng, lrge coeffcents wth more energy cn be dstrbuted n the upleft re, therefore, wth F-norm theory[7], we cn well decrese the dmenson of mge feture nd perform hghly effcent mge mtchng.. Feture Vector Suppose A s squre mtrx nd A s ts th order sub mtrx where é... n ù é... ù A , A ( ~ n) ë n... nn û ë... û The F-norm of A s gven s: æ ö A F ç åå kl (6) è k l ø Let Δ A A F - A - F nd A F, we cn defne the feture vector of A s: V AF {ΔA, ΔA ΔA n } (7). Smlrty crter Defne the Smlrty α of ΔA nd ΔB s:

4 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 ìmn( DA, DB ) / mx( DA, DB ) DA ¹ ordb ¹ (8) í î DA ordb nd we cn thus gve the smlrty α of the mtrces A nd B s: n å c (9) - n Where, C (,,..., n) nd n å C. Obvously, the smlrty α. 5. Progressve Retrevl Strtegy We use progressve retrevl strtegy n order to blnce between computtonl complexty nd retrevl ccurcy.. Rough flterng: Strtng from the mxml decomposton level J, Wth the resultng LL coeffcents, we clculte stndrd vrnces vectors of the query mge nd the dtbse mge s (σ r q, σ g q,, σ b q ) nd (σ r d, σ g d, σ b d )respectvely, fter whch,we cn roughly flter dtbse mges s follows F (βσ q r < σ d r < σ q r /β) && (βσ q g < σ d g < σ q g /β) && (βσ q b < σ d b < σ q b /β) Where, the flterng constnt β Є (, ) s used to djust the sfted dtbse mges. If F s flse, then dtbse mge cn be dentfed s fr prt from Query mge nd therefore s dscrded; else, dtbse mge be kept for further mtch.. Progressve rough flterng: Consderng the effect of hgh frequency components wth LH nd HL wvelet coeffcents n step.. More precse flterng: Wth the obtned LL coeffcents, whch best reflect the generl feture of mge; we use the smlrty crter (9) to determne more precse mges. If α exceeds gven threshold, t mens tht msmtch occurs nd t should be dscrded; else, t wll be kept for further mtch.. Iterton: JJ-, nd terte step ~ tll J. Fnlly, retrevl results re fed bck to users n the order of ther smlrty vlues. 6. Expermentl Results The generl flow of the experments strts wth the decomposton of dt bse mge usng Hr nd D wvelet n offlne. The mxml decomposton level J. We repeted the sme decomposton n onlne for query mge. Wth F-norm theory we extrcted the mge feture vector nd performed hghly effcent mge mtchng. We used progressve retrevl strtegy to blnce between computtonl complexty nd retrevl ccurcy. We focus on the comprson of two mportnt retrevl ndces, nmely retrevl ccurcy nd the speed. The test mge dtbse contns over mges of bts true color. They re dvded nto groups ech contnng 6 mges. The four groups re, Fshes, Sunflowers, nd. For smplcty, ll mges re pre processed to be 56x56 szes before decomposton. Smple dt bse mges for ech ctegory re shown n the Fg.6. The Recll rte s defned s the rto of the number of relevnt (sme ctegory) retreved mges to the number of relevnt tems n collecton [8]. Number of relevnt tems retreved Recll rte Totl number of relevnt tems n collecton SAMPLE DATA BASE IMAGES Fshes Sunflowers

5 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December Lles Fg. 6. Smple dt bse mges of four ctegores Smple retreved mges for the gven query mge usng Hr nd D wvelets re shown n the Fg.6 nd 7. The number t the foot of ech mge ndctes ts smlrty (α) to the exmple mge.α vlues ndcte the smlrty between the Query mge nd the mges n the dtbse.for smlr mges, α vlue s.greter the α vlue greter s the smlrty between the Query mge nd the dtbse mge. Retreved mges re sorted ccordng to ther lph vlues (degree of smlrty). RETRIEVED IMAGES BASED ON HAAR WAVELET Retreved Imges Retreved Imges: Retreved Imges: Retreved Imges Fg.7.Retreved Imges bsed on Hr wvelet RETRIEVED IMAGES BASED ON D WAVELET Retreved Imges Retreved Imges

6 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December Retreved Imges:

7 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 The progressve retrevl strtegy contrbutes to flexble compromse between the retrevl speed nd the ccurcy. Retreved Imges Comprson of recll rte Retreved Imges Recll rte Hr wvelet D wvelet wvelet hstogrm Fshes Sunflowers Comprson of recll rte Recll rte Fshes Sunflowers Retreved Imges: Hr wvelet D wvelet wvelet hstogrm Fg.8.Retreved Imges bsed on D wvelet scheme The followng grphs nd Tble. shows the Recll rte cheved by the two schemes nmely D,Hr Wvelet Trnsform nd exstng wvelet hstogrm for ech ctegory of color mges. D wvelet trnsform cheves the best Recll rte. Tble llustrte the retrevl speed. Dfferent verge retrevl tmes of the two pproches n Tble., mnly cused by decomposton pproches, could roughly reflect ther quntttve computtonl complexty. Accordng to the expermentl results we obtn: In CBIR, representng mge usng Hr nd D wvelets cheves effcent recll rte. Ths s lso the most pprent dvntge of the wvelets n rel tme pplctons. Tble..Comprson of Recll rte D wvelet D lftng Fshes Sunflowers Tble..Comprson of Retrevl speed(s) Hr wvelet D wvelet.5.55 Fshes.. Sunflowers

8 IJCSNS Interntonl Journl of Computer Scence nd Network Securty, VOL.7 No., December 7 5 Retrevl speedn seconds 5 Comprson of retrevl speed Hr wvelet 7. Conclusons D wvelet Wve let hstogrm Fshes Sunflowers In ths pper we ntroduced novel CBIR pproch v wvelet decomposton of mges, followed by feture extrcton nd smlrty mtch under F-norm theory. We compred the retrevl performnce of D wvelet, Hr wvelet schemes wth the exstng technque wvelet hstogrms. It turns out tht D wvelet hs gretly speeded retrevl s well s ensured enough recll rte comprble wth ts Hr wvelet nd wvelet hstogrm. In ddton, the progressve retrevl strtegy helps to cheve flexble compromse mong retrevl ndces. Fnlly we conclude from the results tht wvelets cheve hgh retrevl performnce n rel tme CBIR systems. References [] Chrstopher C. Yng, Content-bsed mge retrevl: comprson between query by exmple nd mge browsng mp pproches, Journl of Informton Scence, pp. 5-67,. [] Ru Y. & Hung T. S., Chng S. F. Imge retrevl: current technques, promsng drectons, nd open ssues. Journl of Vsul Communcton nd Imge Representton,, 9-6, 999. [] Vshl Chtkr, Color-Bsed mge Retrevl Usng Bnry Sgntures. Techncl Report TR -8, Unversty of Ulbert, Cnd, My. [] Qsm Iqbl nd J. K. Aggrwl, Combnng Structure, Color, nd Texture for Imge Retrevl: A performnce Evluton.6 th Interntonl Conference on Pttern Recognton (ICPR), Quebec Cty, Cnd, August -5., vol., pp. 8- [5] R. Clderbnk, I. Dubeches, W. Sweldens nd Y.Boon-Lock, Wvelet trnsform tht mp ntegers to ntegers,appl. Comut. Hrmon. Anl., 5(), -69, 998 [6] [] M.G. Mostf, M.F. Tolb, T.F. Ghrb, M.A. Megeed, Medcl ImgeSegmentton Usng Wvelet bsed Multresoluton EM Algorthm. IEEE Interntonl Conference on Industrl Electroncs, Technology, &Automton., Cro IETA. [7] Huhu Hung; We Hung; Zhgng Lu; Werong Chen; Qngqun Qn; Content-bsed color mge retrevl v lftng scheme Autonomous Decentrlzed Systems, 5. ISADS Proceedngs Aprl 5 Pge(s): [8] Hennng Muller, Wolfgng Muller, Dvd McG. Squre nd Therry Pun, Performnce Evluton n Content-Bsed Imge Retrevl: Overvew nd Proposls. Computng Scence Center, Unversty of Genev, Swtzerlnd,. Y.Mdhveelth, born on Jnury, 97, completed her grduton nd post grduton from JNT Unversty, Hyderbd. She hs been workng t G.Nrynmm Insttute of Scence nd Technology, JNT Unversty,Hyderbd. Her reserch nterest ncludes Imge nd Vdeo processng,multmed Systems Desgn. She s Lfe Member of Insttute of Electroncs nd Telecommuncton Engneers nd Lfe Member of ISTE. Dr.B.C. Jng, born on June, 95, grduton from Krntk Unversty, Dhrwd, Post Grduton from Regonl Engneerng, Wrngl nd Ph. D. from Indn Insttute of Technology, Delh. He hs been wth JNTUnversty snce yers. Prof. Jng ws occuped vrous key postons t JNT Unversty nd presently Rector t JNT Unversty. Hs reserch nterest ncludes sgnl Processng nd Codng Technques. He receved Best Techer Awrd for the yer - Awrded by Government of Andhr Prdesh. He s fellow of Insttute of Engneers(Ind)nd Fellow of Insttute of Electroncs nd Telecommuncton Engneers V.S.K.Reddy, born on June 967, grduted from S.V. Unversty, Trupt, Ind, nd Post Grduton from JNT Unversty, Hyderbd. Hs Doctorl Reserch from IIT Khrgpur n the re of VLSI Sgnl Processng. He hs been workng t Srndh Insttute of Scence nd Technology snce 998. Hs speclzed felds re Imge nd Vdeo Processng, TCP/IP Networks nd Protocols, VLSI Archtectures Sgnl Processng nd DSP Processors. He s fellow of Insttute of Electroncs nd Telecommuncton Engneers nd Member of IEEE.