An Evolutionary Stochastic Approach for Efficient Image Retrieval using Modified Particle Swarm Optimization

Transcription

1 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, A Evolutoary Stochastc Approach for Effcet Image Retreval usg Modfed Partcle Swarm Optmzato Hads Hedar Departmet of Computer Egeerg Raz Uversty Ira, Kermashah Abdolah Chalechale Departmet of Computer Egeerg Raz Uversty Ira, Kermashah Abstract Image retreval system as a relable tool ca help people reachg effcet use of dgtal mage accumulato; also fdg effcet methods for the retreval of mages s mportat. Color ad texture descrptors are two basc features mage retreval. I ths paper, a approach s employed whch represets a composto of color momets ad texture features to extract low-level feature of a mage. By assgg equal weghts for dfferet types of features, we ca t obta good results, but by applyg dfferet weghts to each feature, ths problem s solved. I ths wor, the weghts are mproved usg a modfed Partcle Swarm Optmzato (PSO) method for creasg average Precso of system. I fact, a ovel method based o a evolutoary approach s preseted ad the motvato of ths wor s to ehace Precso of the retreval system wth a mproved PSO algorthm. The average Precso of preseted method usg equally weghted features ad optmal weghted features s 49.85% ad 54.6%, respectvely. 4.3% crease the average Precso acheved by proposed techque ca acheve hgher recogto accuracy, ad the search result s better after usg PSO. Keywords color momets; cotet based mage retreval; partcle swarm optmzato (PSO); texture feature I. ITRODUCTIO The developmet of dfferet mages oblgates the use of effcet techques of maagg the vsual formato by ts cotet []. A mage retreval system s used the color, shape, ad texture features to exact retreve mages from datasets [2]. Cotet based mage retreval (CBIR) s a ope area research for retreval of formato usg ts cotets [3]. From past decade, studes o CBIR have bee a actve research because may large mage databases, tradtoal techques of mage retreval have prove to be suffcet. CBIR system extracts vsual formato of each mage the dataset ad stores features form ad the system extract the related mages that are smlar to the query mage. Color, shape ad texture features are used CBIR systems ad practcal applcatos [4]. Oe of the famous mage retreval systems s QBIC [5]. Shape feature represets the geometrcal formato [6] ad s dvded to boudary based shape ad rego based shape Dr. Abdolah Chalechale, Departmet of Computer Egeerg, Faculty of Egeerg, Uversty of Raz, Kermashah, Ira. descrptors. The shape of the outer boudary s cosdered by boudary based shape descrptors. Zere momets descrptors s a rego based shape descrptors that descrbe the etre rego of a shape [7]. Texture formato ca be used for recogzg a object [8] ad structural methods are used [9] to descrbe t. The fe feature descrptor s appled to reach the truly matched mages [0]. Color hstogram s a color feature [] that captures the umber of pxels havg proper propertes [2]. A combed use of color ad texture would provde better performace tha that of color or texture aloe [3] ad the feature vector cossts of the color ad texture features [4]. Most of the mage retreval methods are ot stochastc; cosequetly, searchg dfferet soluto space s ot possble [5]. By usg equal weghts for the features we ca t have approprate average Precso, ad Recall but applyg dfferet weghts to each feature s a proper soluto. For example, Partcle Swarm Optmzato (PSO) s a approprate approach. Applyg dfferet weghts to each feature ad optmzg PSO algorthm s a method to crease the average precso mage retreval system. Dscrete wavelet trasform ad partcle swarm optmzato was proposed by Qurash et al. for optmzg mage retreval system [6]. The multlevel thresholds mage segmetato approach ad mproved partcle swarm optmzato was proposed by Hogme et al. [7]. A multlevel threshold-based mage segmetato method ad ew partcle swarm optmzato was proposed by Jag et al. [8]. A color mage ehacemet method was preseted by Gora et al. [9] mage retreval system. Also, a hstogram equalzato approach ad PSO algorthm was preseted by Masra et al. [20]. Luo et al. troduced a wavelet-hstogram mage retreval techque ad PSO CBIR systems [2]. A ovel method based o PSO algorthm was proposed by Ye et al. for mage retreval [22]. Also, a ew approach based o PSO ad wavelet was proposed by We et al. that PSO was employed to optmze the weghts [23]. Most of the CBIR systems may ot perform robustly o mage retreval usg the dfferet features. Cosequetly, the ey motvato ths wor has bee to develop a more robust ad accurate mage retreval method whch ca be effectve. I ths paper modfed PSO s used to effectve retreval CBIR 05 P a g e

2 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, systems. I fact, ths paper, a ovel method based o color ad texture mage retreval techque ad modfed PSO s preseted to retreval of mages huge databases to do color textured mage retreval. It s mportat to meto that oe of the approprate methods for extracto of the color features s color momets. I ths wor, color momets cludg mea, stadard devato, ad sewess are used ad etropy, stadard devato, local rage ad cotrast s appled to extracto of the texture features I fact, the ey cotrbuto of ths paper s gve the followg: A proposed of a mage retreval system usg the color ad texture features. A proposed of a optmzato algorthm for creasg average precso of CBIR system. The remder of ths wor s orgazed as follows. Secto II dscusses algorthms covetoal PSO ad the modfed PSO algorthm. Secto III explas a ovel method for the mage retreval systems. Secto IV llustrates the expermetal results ad fally, Secto V provdes cocluso ad future wor. II. THE PARTICLE SWARM OPTIMIZATIO ALGORITHM I ths secto, some formato about algorthms covetoal PSO ad the modfed PSO algorthm used ths study s provded. A. The Stadard Partcle Swarm Optmzato Algorthm PSO s a heurstc techque ad a evolutoary computato model developed by Keedy ad Eberhart 995 [24] that s related to geetc algorthms ad evolutoary programmg. PSO s cosdered robust solvg problems featurg o-dfferetablty, o-learty, ad hgh dmesoalty [25] ad s used eural etwors [26]. If X be the decso vector a cost fucto f (X) the t must be mmze the optmzato problem. I the PSO algorthm, all partcles have radom coordates - dmesoal space. For each partcle, pbest ad gbest are the best coordates each partcle ad the best coordates amog overall partcles, respectvely that the partcles move based o pbest ad gbest. X ad V are curret posto vector ad velocty vector for each partcle, respectvely. At the th tme step (terato), the velocty vector s updated as follows: V d W V d c rad ( pbest c2 rad 2 ( gbest X d ) Also, the posto vector of the th partcle s chaged as follows: X d X d Xd Vd d,2,..., p d,2,...,max terato ) () (2) That p s the umber of partcles. c ad c 2 are the relatve attracto toward pbest ad gbest, respectvely ad rad ad rad 2 are radom umbers uformly dstrbuted betwee [0,]. Also, W s erta weght parameter. The PSO algorthm ca be expressed as Fg.. Fg.. Stadard PSO algorthm Each dmeso of X ad V must be lmted betwee lower boud ad upper boud that are determed based o the parameter of the problem. These parameters must be optmzed optmzato algorthms. B. The Improved Partcle Swarm Optmzato Algorthm A mportat problem PSO method s to determe lmts search space [27]. I stadard PSO algorthm, executo tme creases wth larger search space. Cosequetly, the doma of each dmeso of vector X s lmted ad the stadard PSO route s called. Improved PSO algorthm s gve Fg. 2 that d s appled to lmt search space. Fg. 2. Improved PSO algorthm III. THE PROPOSED IMAGE RETRIEVAL SYSTEM I ths secto, proposed mage retreval procedure s provded. I addto, because of color ad texture are two of the most wdely used features. 06 P a g e

3 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, A. The Basc Cocepts of CBIR Systems Oe of the approprate ways of accessg vsual data s mage retreval that use to color, shape, ad texture [28]. Feature extracto s oe of mportat steps CBIR systems. Extracto of features of the mages s stored feature vectors form. The put mage s called the query mage. The query mage feature vector s compared wth all feature vectors the dataset. Cosequetly, the approprate mages retreve usg dstace measuremet techque. Fg. 3 llustrates the archtecture of CBIR systems. The user terface s cossts of a query formulato part ad a vsualzato part, s the frot page of most systems dealg wth put ad output. The matchg process does smlarty measurg ad the ecessary comparsos. The dexes of those mages whch are selected to retreve are passed to the mage poters process. It obtas mage poters (mage d s), ad the fetchg process physcally retreves the mages from the dataset. B. CBIR Systems usg the Fuso of Texture Features ad Color Momets I ths secto, texture features ad color momets are vestgated. Etropy, local rage, stadard devato ad cotrast measures are used to extract the texture features. Texture = (Etropy + Local Rage + Stadard devato + Cotrast) Etropy ca be used to descrbe the texture of the put mage that ca be calculated as: ET M P log P Where, ET, M, ad P are etropy, total umber of samples, ad probablty of occurreces, respectvely. Maxmum value of chose pxel-mmum value of chose pxel s called local rage. Stadard devato ca be calculated as follows. S X ( X X ) That, s umber of pxels the mage. Cotrast represets the qualty of pcture a mage ad s calculated by (5) F co 4 S 2 4 X 2 (3) (4) (5) 4 4 ( X (, j) X ) m Where, s the 4 th momet of the mea X, S 2 s the 4 varace of the gray values mage. I ths wor, mea, stadard devato, ad sewess to extract color features s used. Mea, stadard devato, ad the sewess are effectve represetg color dstrbutos of mages. Color momets are descrbe as follows. Momet : Mea Pj j That P j s the value of the -th color chael at the j-th mage pxel Momet 2: Stadard devato Momet 3: Sewess S m j j 3 j ( P ) j ( P j ) I ths wor, the combato of texture features ad color momets s used. Etropy, local rage, stadard devato ad cotrast measures are used to extract the texture features ad 3 features are appled. Features = (Texture Features + Color Features) C. Stochastc CBIR usg Improved Partcle Swarm Optmzato Cost fucto must be mmzed optmzato algorthm. A mmzato tool s the stochastc PSO method. The solutos space s made of the features f=,,f, that are calculated o every dataset mage. I the algorthm, the value of F s adjusted to 3. These 3 features are etropy, stadard devato, local rage, cotrast, mea of red compoet, stadard devato of red compoet, sewess of red compoet, mea of gree compoet, stadard devato of gree compoet, sewess of gree compoet, mea of blue compoet, stadard devato of blue compoet, ad sewess of blue compoet, respectvely. 3 2 (6) (7) (8) 07 P a g e

4 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, User Query Features Query feature extracto (ole) Query Image User terface cludg: - Query formulato 2- Vsualzato Resultg mages Matchg process Image Poters Fetchg process Idexes Idexg Algorthms Features Feature extracto (offle) Image Database Fg. 3. Archtecture of CBIR systems The mages of the database x j, j=,, DB represet a dscrete set of pots ad the partcles ca move wth the features space. To assocate every partcle wth the earest mage, a weghed cty bloc dstace (WCBD) or the Mahatta dstace s used whch s expressed as follows. F f f f WCBD ( x q, x j ) xq x j w, f By usg equal weghts for each feature we ca t have good average Precso, ad Recall. Dfferet weghts to each feature are a good soluto that s optmzed usg PSO algorthm. Weghg vector w calculate aga at each terato. The proposed algorthm shows Fg. 4. I the frst terato (=), a query mage wth a feature vector x q =[,,,, ] s selected. The, the dstaces from all the dataset mages x j ; j=,, DB are computed as WCBD(x q,x j ). The speed vector each partcle s set by radomly selectg a value over the features space, ad the the stochastc optmzato s doe. The related ad the rrelevat mages are updated each terato ad the ew features weghts are computed. After classfy the swarm based o the ftess of each partcle, the th terato s completed. Fally, each partcle to the earest mage the dataset s assocated ad the best FB s show to the user. Whle a predefed umbers of teratos (9) are reached, the optmzato process eds. The, the relevat solutos are show. IV. EXPERIMETAL RESULTS The performace of a mage retreval system s computed usg the Recall ad Precso values. The Recall s defed as the rato of the umber of relevat mages retreved ad the umber of relevat mages class. The Precso s defed as the rato betwee the umber of relevat mages retreved ad the total umber of mages retreved [29]. Precso ad Recall s computed as: umber of relevat mages retreved Recall Total umber of relevat mages umber of relevat mages retreved Precso () Totalumber of mages retreved Also, the average Precso s computed by: Average_ Precso Aq p( ) A That tem belogs to the qth category (A q ). The fuso of color ad texture features wth optmal weghts to a subset of MPEG-7 dataset s used. q (0) (2) 08 P a g e

5 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, Query Image WCBD Frst Iterato? yes Swarm Italzato o yes Covergece? Ed o Features Re-weghtg Ftess Evaluato Swarm Update Fg. 4. Flowchart of the proposed method The mages the dataset are categorzed 23 classes ad each class cotas 0 pctures JPEG format that samples of MPEG-7 mage dataset are show Fg. 5. method. These results show that the performace of the proposed method s better tha the other methods. I expermets, Precso versus Recall curves to evaluate retreval effcecy s adopted. (a) Fg. 5. Samples of MPEG-7 mage database Four texture features clude etropy, stadard devato, local rage, ad cotrast ad e colour features (mea, stadard devato, ad sewess, for R, G, ad B compoets RGB space). Precso ad Recall are evaluato parameters our expermets ad mplemetatos s doe usg a PC wth Itel Petum 2.5 GHz ad 4 GB RAM. Fg. 6 llustrates the results geerated from proposed system usg optmal weghted features that show the effcecy of proposed (b) Fg. 6. Cotet based mage retreval results. (a) put mage for retreval. (b) usg the mproved PSO Fg. 7 shows average Precso for whe texture features (TF), the color momets (CM), the combato of the texture features ad color momet usg equally weghted features (TCEW) ad optmal weghted features (TCOW), respectvely extracted from mages. 09 P a g e

6 Precso Precso Average Precso Average Precso (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, The average Precso these four methods s 4.87, 45.64, 49.85, ad 54.6 percet, respectvely. features ad usg the mprove PSO, respectvely that the results of optmal weghted features show better average Precso ad Recall. The average Precso of proposed approach usg mproved PSO are 54.6%. Also, for the proposed method, the maxmum average Precso of 00% at Recall value s 0%, ad the Precso value decreases to 25.92% at 00% of Recall. Table I shows the quattatve results obtaed by the optmal weghted features to the dataset that a total average of 54.6% retreved mages s achevable usg mproved PSO algorthm. Fg. 9 llustrates the comparso of average precso the proposed method wth the other methods [8], [9], ad [0]. TF CF TCEW TCOW Fg. 7. The Average precso chart for CBIR system usg the texture features, the color momets, equally weghted features ad system wth usg mproved PSO method Recall Recall (a) Fg. 8. The Average precso/recall chart for (a) CBIR system usg equally weghted (b) CBIR system wth usg mproved PSO method Fg. 8 (a) ad (b) show the Precso-Recall graph for the proposed mage retreval system usg equally weghted (b) Fg. 9. The Average precso for the proposed method ad the other methods Fg. 0 shows the Precso versus Recall result of 30 query mages that the proposed method s better tha the fxed weghtg. TABLE I. Recall (%) PRECISIO AD RECALL OF THE PROPOSED METHOD Precso (%) for the equally weghted features Precso (%) for the preseted method usg mproved PSO AR = 55% AP = 49.85% AP = 54.6% 0 P a g e

7 Average Precso Precso Optmal weghted features Fx weghed features (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, I ths wor, to show the effectvty of proposed PSO algorthm, oly color ad texture features are selected, selectg more features wll acheve better retreval effect, whch s our further wor. Furthermore, the approach ca be exteded for traslato ad rotato propertes, so that the retreval effcecy ca be creased Fg. 0. Average precso vs. recall I our expermets, sze of swarm s 260 partcles ad sum of c ad c 2 varables s smaller 3. Also, the average precso dfferet teratos s ot same. Fg. shows the Precso chart dfferet teratos Recall Iterato Fg.. Precso chart dfferet teratos V. COCLUSIO AD FUTURE WORK I ths paper, a method PSO to mprove the accuracy ad ablty for mage retreval s preseted ad the mproved PSO algorthm for mage retreval to get hgher accuracy s employed. I fact, the use of a stochastc optmzato algorthm to acheve a proper CBIR system was vestgated. The expermetal results showed that the proposed method was effectve to the smlarty search mages dataset after usg PSO. To ehace the retreval performace, cost fucto was mmzg. Proposed method was evaluated usg Precso, Recall, ad average Precso that the average Precso ad the average Recall of proposed method are 54.6% ad 55.00%, respectvely. REFERECES [] S. Agarwal, A. Verma, ad P. Sgh, Cotet Based Image Retreval usg Descrete Wavelet Trasform ad Edge Hstogram Descrptor, Iteratoal Coferece o Iformato Systems ad Computer etwors, 203, pp [2] D. Pedroette ad R. Torres, Usupervsed measures for estmatg the effectveess of mage retreval systems, Iteratoal Coferece o Graphcs, Patters ad Images, 203, pp [3] A. Ahmad, A. Chalechale, ad H. Hedar, Parallelzed Computato for Edge Hstogram Descrptor Usg CUDA o the Graphcs Processg Uts (GPU), The 7 th CSI Iterattoal Symposum o Computer Archtecture & Dgtal Systems, 203, pp. -6. [4] S. Youssef, ICTEDCT-CBIR: Itegratg curvelet trasform wth ehaced domat colors extracto ad texture aalyss for effcet cotet-based mage retreval, Computers ad Electrcal Egeerg 38, 202, pp [5] A. Bhagat ad M. Atque, Desg ad Developmet of Systems for Image Segmetato ad Cotet Based Image Retreval, IEEE, 202, pp. -5. [6] C. Sgh ad Pooja, A effectve mage retreval usg the fuso of global ad local trasforms based features, Optcs & Laser Techology 44, 202, pp [7] P. Mapoochelv ad K. Mueeswara, Sgfcat Rego Based Image Retreval Usg Curvelet Trasform, Iteratoal Coferece o Recet Advacemets Electrcal, Electrocs ad Cotrol Egeerg, 20, pp [8] K. Km ad S. Kwo, Image Retreval Scheme Based o Adaptve Feature Weghtg, IEEE, 202, pp [9] W. Yua, C. Feg, ad Y. Jao, A effectve method for color mage retreval based o texture, Computer Stadard & Iterfaces 34, 202, pp [0] C. L, D. Huag, Y. Cha, K. Che, ad Y. Chag, Fast color-spatal feature based mage retreval methods, Expert Systems wth Applcatos 38, 20, pp [] W. Che, W. Lu, ad M. Che, Adaptve Color Feature Extracto Based o Image Color Dstrbutos, IEEE Trasacto o Image Processg, 200, pp [2] B. Syam, S. Vctor, ad Y. Rao, Effcet Smlarty Measure va Geetc Algorthm for Cotet Based Medcal Image Retreval wth Extesve Features, IEEE, 203, pp [3] H. Hedar, A. Chalechale, ad A. Ahmad, Acceleratg of Color Momets ad Texture Features Extracto Usg GPU Based Parallel Computg, 8 th Iteratoal Coferece o Mache Vso ad Image Processg, 203, pp [4] A. Salahudd, A. aqv, K. Mujtaba, ad J. Ahtar, Cotet based Vdeo Retreval usg Partcle Swarm Optmzato, 0 th Iteratoal Coferece o Froters of Iformato Techology, 202, pp [5] M. Brolo, P. Rocca, ad F. atale, Cotet-Based Image Retreval by a Sem-Suppervsed Partcle Swarm Optmzato, IEEE, 2008, pp [6] M. Qurash, K. Dhal, J. Paul, ad M. De, A ovel Hybrd Approach to Ehace Low Resoluto Images Usg Partcle Swarm Optmzato, 2 d IEEE Iteratoal Coferece o Parallel, Dstrbuted ad Grd Computg, 202, pp [7] T. Hogme, W. Cuxa, H. Lyg, ad W. Xa, Image Segmetato Based o Improved PSO, IEEE Iteratoal Coferece o Computer ad Commucato Techologes Agrculture Egeerg, 200, pp [8] F. Jag, M. Frater, ad M. Pcerg, Threshold-based Image Segmetato Through a Improved Partcle Swarm Optmzato, IEEE, 202, pp. -5. P a g e

8 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, [9] A. Gora ad A. Ghosh, Hue-Preservg Color Image Ehacemet Usg Partcle Swarm Optmzato, IEEE, 20, pp [20] S. Masra, P. Pag, M. Muhammad, ad K. Kpl, Applcato of Partcle Swarm Optmzato Hstogram Equalzato for Image Ehacemet, IEEE Colloquum o Humates, Scece & Egeerg Research, 202, pp [2] T. Luo, B. Yua, ad L. Ta, Blocg Wavelet-hstogram Image Retreval by Adaptve Partcle Swarm Optmzato, The st Iteratoal Coferece o Iformato Scece ad Egeerg, IEEE, 2009, pp [22] Z. Ye, B. Xa, D. Wag, ad X. Zhou, Weght Optmzato of Image Retreval Based o Partcle Swarm Obtmzato Algorthm, IEEE, 2009, pp. -3. [23] K. We, T. Lu, W. B, ad H. Sheg, A Kd of Feedbac Image Retreval Algorthm Based o PSO, Wavelet ad Sub-bloc sortg thought, 2 d Iteratoal Coferece o Future Computer ad Commucato, 200, pp. -6. [24] J. Keedy ad R. Eberhart, Partcle swarm optmzato, IEEE Iteratoal Coferece eural etwors, 995, pp [25] A. Taher, A. Karma, ad M. Hasa, A ew method for optmal locato ad szg of capactors dstorted dstrbuto etwors usg PSO algorthm, Smulato Modelg ad Theory 9, 20, pp [26] M. Brolo ad F. atale, A Stochastc Approach to Image Retreval Usg Relevace Feedbac ad Partcle Swarm Optmzato, IEEE Trasacto o Multmeda, 200, pp [27] R.A. Vural, O. Der, ad T. Yldrm, Ivestgato of Partcle Swarm Optmzato for Swtchg Characterzato of Iverter Desg, Expert Systems wth Applcatos, Vol. 38, o. 5, 20, pp [28] F. Mal ad B. Baharud, Aalyss of dstace metrcs cotetbased mage retreval usg statstcal quatzed hstogram texture features the DCT doma, Computer ad Iformato Scece 38, 202, pp. -2. [29] C. Rubert, L. Cque, Decomposto of two-dmesoal shapes for effcet retreval, Image ad Vso Computg 27, 2009, pp P a g e