Improved Adaptive Brovey as a New Method for Image Fusion

Transcription

1 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < Improved Adaptve Brovey as a ew Metod for Image Fuson Hamd Reza Sadoost Abstract An deal fuson metod preserves te Spectral nformaton n fused mage and adds spatal nformaton to t wt no spectral dstorton. Among te exstng fuson algortms, te contourlet-based fuson metod s te most frequently dscussed one n recent publcatons, because te contourlet as te ablty to capture and lnk te pont of dscontnutes to form a lnear structure. Te Brovey s a popular pan-sarpenng metod owng to ts effcency and g spatal resoluton. Ts metod can be explaned by matematcal model of optcal remote sensng sensors. Ts study presents a new fuson approac tat ntegrates te advantages of bot te Brovey and te cotourlet tecnques to reduce te color dstorton of fuson results. Vsual and statstcal analyzes sow tat te proposed algortm clearly mproves te mergng qualty n terms of: correlaton coeffcent, ERGAS, UIQI, and Q4; compared to fuson metods ncludng IHS, PCA, Adaptve IHS, and Improved Adaptve PCA. Index Terms Image fuson, Pan-sarpenng, Brovey Transform, Contourlet. I I. ITRODUCTIO n remote sensng systems, scenes are observed n dfferent portons of electromagnetc spectrum; terefore, te remote sensng mages vary n spectral and spatal resoluton. To collect more potons and mantan sgnal to nose rato smultaneously, te multspectral sensors, wt g spectral resoluton, ave a lower spatal resoluton compared wt pancromatc mage wt a ger spatal resoluton, and wde spectral bandwdt. Wt proper algortms, t s possble to fuse tese mages and produce magery wt te best nformaton of bot, namely g spatal and g spectral resoluton. Te most commonly used mage-fuson metods are tose based on te IHS (also known as HSI) [], and prncpal component analyss (PCA) [2], [3]. However, tese metods can cause spectral dstorton n te results [2]. Cavez proposed HPF(Hg Pass Flterng) fuson algortm wc as sown better performance n terms of g-qualty syntess of spectral nformaton [4]. Prncple of ts metod s to extract te g-frequency nformaton from te mage and nject t nto te low resoluton multspectral () mage, formerly resampled to matc te pxel sze. Box car flters are used as a low pass flter n ts metod. However, te rpple n te frequency response of box car flters as some negatve mpact on ts metod. Te well-known Mallat s fuson algortm [5] uses an ortonormal bass and can effectvely preserve spectral nformaton, but because of down-samplng operators, wc are used to mplement ordnary dscrete wavelet transforms, tese transforms are not sft-nvarant and can lead to problem n data fuson [6]. To avod above-mentoned problem, te dscrete wavelet transform known as à trous algortm [7] was proposed by elmnatng te decmaton operators n wavelet structure. It s a sft-nvarant and redundant wavelet transform algortm based on a multresoluton dyadc sceme. However, wavelets cannot capture te smootness along te contours [8]. An effcent decomposton of te mage sould deally possess propertes of multresoluton, localzaton, drectonalty, and ansotropy. An alternatve multresoluton approac, called contourlet transform, provdes an effcent drectonal decomposton and s also as ablty of capturng spatal structures of te natural mage along te smoot contours [9]. Te Brovey, based on te cromatcty transform [], [], [2], s a smple metod for combnng data from dfferent sensors. It also preserves a g degree of spatal detal of te mage; owever, dstorton of te spectral nformaton s not acceptable at all. To overcome te defcency of Brovey metod, a new fuson approac s proposed n ts letter. We call ts algortm Adaptve-Brovey because t vares te amount of spatal njecton dependng on te ntal and pancromatc () mages. To mnmze spectral dstorton, we use SCT (onsubsampled Contourlet Transform) to add ger frequency components of Adaptve Brovey result to te ntal mages. As we know, all of te lower frequency components of H (Hg Resoluton ) are avalable n te mage and we can use ts spectral nformaton to mprove Adaptve Brovey. Ts mproved adaptve Brovey metod ntegrates te advantages of bot te Brovey and te SCT tecnques to preserve bot spatal and spectral nformaton of and mages. To verfy te effcency of te proposed metod, vsual and statstcal assessments are carred out on and data. II. ADAPTIVE BROVEY AGORITHM Image formaton model sows tat te rato between Te autors s wt te Department of Electrcal and Computer Engneerng, at Unversty of Tarbyat Modares at Tearn, Tel ; fax: , IRA (e-mal: Hamdrezasadoost@gmal.com).

2 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < 2 and b s ndependent of te sensor band [3]. If b represents a pancromatc band and a ow resoluton multspectral band, ten we wll ave: were and = () are pxel values of te and te correspondng pxel values of te RM, respectvely. Dfferent metods are used to compute wll ave: were s estmated from = = n [3-5]. Te n te Brovey metod. So we (2) s te t fused multspectral band at te resoluton of Pan mage. Wen Brovey s used, njecton of spatal nformaton nto te mages s more tan desrable, leadng to te spectral dstorton n te results [3]. l Te rato between and n equaton (2) = s te gan of sarpenng. To avod over-sarpenng of te Brovey metod, we propose te Adaptve Brovey to control te spatal nformaton of te fused mage were between = ( ) a b = (3) b values are determned by mnmzng te RE b and mage, usng metod = ntroduced n [6], and a s te parameter wc can control te njecton of spatal nformaton. If a equals zero, no fuson s performed and f a equals one, equaton (2) and (3) are te same. To fnd a sutable a parameter, we vary te value of a from zero to one by an arbtrary step lke.5 and fuse te and mages by equaton (3). Te best a value s determned by mnmzng te QR ndex. Recently proposed by Alparone et al. [7], te QR ndex evaluates te qualty of te fused mage wtout requrng te HRM mage and combnes te two dstorton ndexes of te radometrc and geometrc dstorton ndexes. In ts manner, fused mage wt smultaneously g spectral and spatal resoluton can be aceved. Te man advatntage of Brovey on wc we empasze s tat te spatal nformaton of eac fused pxel s relatve to ts value. Ts means tat Brovey uses local gan n ts fuson procedure. For example, vegetaton as almost no response n te blue range, wc exbts dark dgts n (multspectral) mages [8]. Te fused result of equaton (4) or (5) remans zero n ts case (vegetaton area n te blue range), but oter metods nject spatal nformaton of oter bands nto te blue band n ts example. Te context-baseddrven algortms (CBD) [9] try to add spatal nformaton relatve to ts local energy, wereas te appled wndow wc s used n suc algortms takes nto account negbor pxels and may lead to spectral and spatal dstorton n some cases suc as vegetaton area and around t. III. SCT TRASFORM Ts secton revews te SCT as presented n [9] and [2]. Te nonsubsampled contourlet transform (SCT) s used to decompose nput mages nto low-frequency and gfrequency parts. Te contourlet as te ablty to capture and lnk te pont of dscontnutes to form a lnear structure or contours. SCT can be dvded nto two sft-nvarant parts: ) nonsubsampled pyramd structure (SP) and 2) a nonsubsampled drectonal flter bank structure (SDFB). ) SP: SP s aceved by usng two-cannel nonsubsampled 2-D flter banks. Ts expanson s conceptually smlar to te -D SWT computed wt te ATWT (à trous wavelet Transform) algortm and as n+ redundancy, were n s te number of decomposton stages. Flters of te subsequent stages are upsampled verson of tose of te frst stage. Te desred passband support of te low-pass π π 2 flter at te nt stage s te nterval [ ( ),( )]. n n 2 2 Accordngly, te deal regon of te equvalent g-pass flter s te complement of te low-pass one,.e., te 2 2 nterval[ π π ( ),( )] /[ ( ),( )] n n n n 2 2 π π 2 2. Wtout usng any addtonal flter desgn, ts gves a multscale property. At eac stage one bandpass mage s produced. Tus, certan parts of te nose spectrum are fltered n te processed pyramd coeffcents. 2) SDFB: Te SDFB s constructed by elmnatng te downsamplers and upsamplers operator n te drectonal flter bank (DFB). By swtcng off te downsamplers/upsamplers n eac two-cannel flter bank n te DFB tree structure and upsamplng te flters accordngly, a sft-nvarant drectonal expanson s obtaned. Tese results n a tree are composed of two-cannel SFBs. By combnng te SP and te SDFB, te SCT s constructed. IV. IMPROVED ADAPTIVE BROVEY In te past few years, several researcers proposed dfferent and mage-fuson metods based on ybrd concept

3 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < 3 to ntegrate te advantages of two fuson metods. Undecmated Wavelet PCA (udwpc) [2], Undecmated Wavelet IHS (udwi) [2], IHS-wavelet [2], IHS- retnanspred [2], FFT-enanced IHS [22], and Improved adaptve PCA [8] are te most commonly used ybrd metods. However, Adaptve Brovey tres to preserve spectral and spatal nformaton smultaneously and t s faster tan oter smlar algortms suc as [2]. In order to mnmze spectral dstorton n te Adaptve Brovey metod, we use SCT to preserve te lower frequency of mages. As we know, te mssed g frequency components of H mages are only needed to mprove te spatal resoluton of mages, so we use SCT to decompose mages and fused result of Adaptve Brovey nto low-frequency and g-frequency parts. Improved Adaptve Brovey replaces te detal contourlet coeffcents of te result by tose of Adaptve Brovey mages. Because ger frequency components of mages ave zero mean, usng te lower frequency component of (wc s equal n and reference mage) n any fuson metod can at least remove bas between Fused result and te reference mage. One can fnd a sutable a parameter by varyng te value of a and fuse te and mages by Improved Adaptve Brovey. Te best a value s determned by mnmzng te QR ndex between Improved Adaptve Brovey result and ntal and mages. V. EXPERIMETA RESUTS Four-band multspectral QuckBrd data was used for our experments. Tese mages are acqured by a commercal satellte, QuckBrd. Te QuckBrd data set was taken over te Pyramd area of Egypt n 22. Te subscenes of te raw mages are used as and. QR ndex data X:.94 Y: Fg.. QR ndex Curve versus a Two recent adaptve metods consdered for comparson are Adaptve IHS and Improved Adaptve PCA. To mplement Adaptve IHS metod, we use software wc s ntroduced n [6]. A MATAB toolbox tat mplements te SCT s used for mplementaton of Improved Adaptve PCA and Improved Adaptve Brovey. Orgnal and mages are spatally degraded down to a lower resoluton n order to compare fused product to te genune references resoluton. Fg. sows te QR ndex versus a n our experment. As can be seen from Fg., te value of a wc maxmze te QR equals.94. a b c

4 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < 4 d e f Fg.2. (a). (b). (c) Reference mage. (d) Fused mage by Improved Adaptve PCA. (e) Fused mage by Adaptve IHS. (f) Fused mage by Improved Adaptve Brovey. TABE I : Spectral Qualty Metrcs for Fg. 2 PCA Improved Adaptve PCA IHS Adaptve HIS Improved Adaptve Brovey CC ERGAS UIQI Q Vsual comparson of te fused mages s te frst step of qualty assessment. To do ts, several aspects of te mage qualty suc as surfaces, lnear features, edges, color, blurrng, and bloomng can be taken nto account. Te vsual performances of te fused mages are sown n Fg. 2. Fg. 2 llustrates fuson results at nferor level. Accordng to Fg. 2, te spatal detals pretty well enanced n te merged mage usng Improved Adaptve PCA metod, but te colors ave been canged durng te fuson process wc presents te spectral dstorton. As llustrated n Fg 2, te sarpenng n te Adaptve IHS and Improved Adaptve Brovey metod s sutable; owever, some false edges can be seen n Adaptve IHS at pyramd area. Tese artfacts probably come from unsutable parameters of te edge detector (see [6]). Over-enanced edges are evdent n vegetaton area of bot Adaptve IHS and Improved Adaptve PCA (upper rgt corner of te mage). To sow smlarty between te reference mage and fused mages, stogram of te reference mage and fused mages for green band are represented n Fg. 3. Te stogram of proposed metod s more smlar to tat of reference mage n ts fgure. In addton to te vsual nspecton, te performance of eac metod sould be analyzed quanttatvely. In order to assess te qualty of te merged mages at te nferor level, four objectve ndcators were used [23-29] Ref Image 2 Adaptve IHS Improved Adaptve Brovey 2 Improved Adaptve PCA 2 Fg. 3 Hstogram of te green band of Fg. 2. ) Correlaton coeffcent Ts metrc ndcates te degree of lnear dependence between te orgnal reference and fused mage. If two mages are dentcal, te correlaton coeffcent wll be maxmal and equals. It s defned as follows:

5 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < 5 Corr( A, B) = npx j= npx ( A - m )( B - m ) j A j B npx ( A - m ) ( B - m ) j A j B j= j= 2) ERGAS ERGAS (or relatve global dmensonal syntess error) s as follows: = 2. (4) 2 RE ( B) ERGAS = (5) l M were s te resoluton of te g spatal resoluton mage, l s te resoluton of te low spatal resoluton mage, and M s te mean radance of a specfc band nvolved n te fuson. RE s te root mean square error and can be computed by te followng expresson: RE ( B) = Bas ( B) + SD ( B) (6) metods ave te least color dstortons and desrable spatal nformaton. I. COCUSIO A new pan-sarpenng metod based on ybrd concept s proposed n ts paper. To avod te over-enanement of te Brovey metod, we proposed te Adaptve Brovey to control te spatal nformaton of te fused mage. In order to mnmze spectral dstorton n te Adaptve Brovey metod, te Improved Adaptve Brovey uses SCT to preserve te lower frequency of mages. Proposed Improved Adaptve Brovey metod mantans spectral and spatal nformaton smultaneously. Fnally, te vsual results sow tat Improved Adaptve Brovey can aceve better performance n comparson wt some new metods. In addton to te vsual nspecton, te performance of proposed metod and some former well-known metods were analyzed quanttatvely. We appled correlaton coeffcent, ERGAS, UIQI and Q4 metrcs to obtan te qualty values. Te performance evaluaton metrcs confrmed te superorty of te Improved Adaptve Brovey metod. Te lower te value of te ERGAS s, te ger te qualty of te merged mages s. 3) UIQI It s defned as follows: σ 2mm 2σσ Q=.. σσ σ σ AB A B A B A B ma + mb A + B were A and B are fused and reference mages respectvely. Te unversal mage qualty ndex (UIQI) models any dstorton as a combnaton of tree dfferent factors: loss of lnear correlaton, contrast dstorton and lumnance dstorton. 4) Q 4 Q s composed of tree dfferent factors: Te frst s te 4 modulus of te yper-complex CC between te two spectral pxel vectors wc s senstve to bot te loss of correlaton and to spectral dstorton between te two data sets. Contrast canges and mean bas on all bands are measured by te second and trd factors, respectvely. Te Table I sows te values of CC, ERGAS, UIQI and Q 4 for proposed metod and oter mentoned metods. As can be seen from Table I, te qualty ndexes obtaned by applyng proposed metod are pretty good compared wt tose obtaned by applyng oter metods. Tese statstcal assessment fndngs agree wt tose of te vsual analyss. Vsual and statstcal assessments sow tat te proposed (7) REFERECES [] T. M. Tu, S. C. Su, H. C. Syu, and P. S. Huang, A new look at IHSlke mage fuson metods, Inf. Fuson, vol. 2, no. 3, pp , Sep. 2. [2] Sadoost, H.R. and Gasseman, H., 22, May. Spatal PCA as a new metod for mage fuson. In Artfcal Intellgence and Sgnal Processng (AISP), 22 6t CSI Internatonal Symposum on (pp. 9-94). IEEE. [3] Sadoost, H.R. and Gasseman, H., 26. Combnng te spectral PCA and spatal PCA fuson metods by an optmal flter. Informaton Fuson, 27, pp.5-6. [4] P. S. Cavez, Dgtal mergng of andsat TM and dgtzed HAP data for : 24. scale mage mappng, Potogramm. Eng. Remote Sens., vol. 52, pp , Oct [5] S. Mallat, A teory for mult-resoluton sgnal: Te wavelet representaton, IEEE Trans. Pattern Anal. Mac. Intell., vol., no. 7, pp , Jul [6] D. A. Yocky, Artfacts n wavelet mage mergng, Opt. Eng., vol. 53, pp , July 995. [7] J. uñez, X. Otazu, O. Fors, A. Prades, V. Palà, and Romàn Arbol, Multresoluton-based mage fuson wt addtve wavelet Decomposton, IEEE Trans. Geosc. Remote Sensng, vol. 37, pp. 24 2, May 999. [8] V. P. Sa,. H. Younan, and R.. Kng, An effcent pan-sarpenng metod va a combned adaptve PCA approac and contourlets, IEEE Trans. Geosc. Remote Sens., vol. 46, no. 5, pp , May 28. [9] A.. da Cuna, J. Zou, and M.. Do, Te nonsubsampled contourlet transform: Teory, desgn, and applcatons, IEEE Trans. Image Process., vol. 5, no., pp , Oct. 26. [] J. G. u, Smootng flter-based ntensty modulaton: A spectral preserve mage fuson tecnque for mprovng spatal detals, Int. J. Remote Sens., vol. 2, no. 8, pp , 2. [] A. R. Gllespe, A. B. Kale, and R. E. Walker, Color enancement of gly correlated mages II. Cannel rato and cromatcty transformaton tecnques, Remote Sens. Envron., vol. 22, pp , 987. [2] M. González-Audícana, J.. Saleta, R. G. Catalán, and R. García, Fuson of multspectral and pancromatc mages usng mproved IHS

6 > REPACE THIS IE WITH YOUR PAPER IDETIFICATIO UMBER (DOUBE-CICK HERE TO EDIT) < 6 and PCA mergers based on wavelet decomposton, IEEE Trans. Geosc. Remote Sens., vol. 42, no. 6, pp , Jun. 24. [3] Z. Wang, D. Zou, C. Armenaks, D., and Q., A Comparatve Analyss of Image Fuson Metods, IEEE Trans. Geosc. Remote Sens., vol. 43, no. 6, Jun. 25. [4] J. G. u and J. M. Moore, Pxel block ntensty modulaton: Addng spatal detal to TM band 6 termal magery, Int. J. Remote Sens., vol. 9, no. 3, pp , 998. [5] Y. Zang, A new mergng metod and ts spectral and spatal effects, Int. J. Remote Sens., vol. 2, no., pp , 999. [6] S. Raman, M. Strat, D. Merkurjev, M. Moeller, and T. Wttman, An adaptve IHS pan-sarpenng metod, IEEE Geosc. Remote Sens. ett., vol. 7, no. 4, pp , Oct. 2. [7]. Alparone, B. Aazz, S. Baront, A. Garzell, F. encn, and M. Selva, Mutltspectral and pancromatc data fuson assessment wtout refrence, Potogramm. Eng. Remote Sens., vol. 74, no. 2, pp. 93 2, Feb. 28. [8] C. Tomas, T. Rancn,. Wald, and J. Canussot, Syntess of multspectral mages to g spatal resoluton: A crtcal revew of fuson metods based on remote sensng pyscs, IEEE Trans. Geosc. Remote Sens., vol. 46, no. 5, pp. 3 32, May 28. [9] B. Aazz,. Alparone, S. Baront, and A. Garzell, Context-drven Context fuson of g spatal and spectral resoluton mages based on oversampled multresoluton analyss, IEEE Trans. Geosc. Remote Sens., vol. 4, no., pp , Oct. 22. [2] A. G. Mayar, and M. Yazd, Pancromatc and Multspectral Image Fuson Based on Maxmzaton of Bot Spectral and Spatal Smlartes, IEEE Trans. Geosc. Remote Sens., vol. 49, no. 6, pp , Jun. 2. [2] S. Danesvar, H. Gasseman, MRI and PET mage fuson by combnng IHS and retna-nspred models, Inf. Fuson, vol., no. 2, pp. 4 23, Apr. 2. [22] Y. ng, M. Elers, E.. Usery and M. Madden, FFT-enanced IHS transform metod for fusng g-resoluton satellte mages, Potogramm. Eng. Remote Sens., vol. 6, no. 6, pp , Feb. 27. [23] Sadoost, H.R. and Gasseman, H., 25. Fuson of and mages preservng spectral qualty. IEEE Geoscence and Remote Sensng etters, 2(3), pp [24] Sadoost, H.R., 27. and mage fuson by combnng Brovey and wavelet metods. arxv preprnt arxv: [25] Mogadam, F.V. and Sadoost, H.R., 27. A ew Multfocus Image Fuson Metod Usng Contourlet Transform. arxv preprnt arxv: [26] Sadoost, H.R. and Javaer,., 27. Pansarpenng of clustered and Pan mages consderng mxed pxels. IEEE Geoscence and Remote Sensng etters, 4(6), pp [27] Sadoost, H.R. and Gasseman, H., 25. Multspectral and Pancromatc Image Fuson by Combnng Spectral PCA and Spatal PCA Metods. Te Modares Journal of Electrcal Engneerng, (3), pp [28] Sadoost, H.R. and Merab, A., 27. MRI and PET mage fuson usng structure tensor and dual rpplet-ii transform. Multmeda Tools and Applcatons, DOI: ttps://do.org/.7/s , pp [29] Sadoost, H.R. and Merab, A., 28. Multmodal mage fuson usng sparse representaton classfcaton n tetrolet doman. Dgtal Sgnal Processng, Vol. 79, PP. 9-22, 28.