Australian Journal of Basic and Applied Sciences. Fastest Color Model for Image Processing Using Embedded Systems

Transcription

1 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: AENS Journals Ausralian Journal of Basic and Applied Sciences Journal home page: Fases Color Model for mage Processing Using Embedded Ssems Dmiri Skopin, Jamil Al-Azzeh, Jihad Nader, Ashraf Abu-Ein and Dr. Mazin Al Hadidi Compuer Engineering, Al Balqa Applied Universi Amman, 11134, Jordan. ARTCLE NFO Aricle hisor: Received 13 November 013 Received in revised form 0 December 013 Acceped 3 Januar 014 Available online 1 Februar 014 Kewords: Digial image processing, color models, embedded ssems. ABSTRACT This research inroduces and proposes new color model for image processing. The model consiss of hree channels; wo channels ransmi richromaic coefficiens of he red and green colors, and he hird channel ses he inensi componen. n purpose of applicaion of he new color model for digial image processing, an equaion for convering colors from RGB o new model and is inverse has been creaed wih respec o he archiecure of modern embedded ssems. Since he proposed color model is no expanded o a clindrical form, he risk o specif he coordinaes which lie ouside he color gamu has been evaluaed, and some boundar condiions are defined. To prove he usefulness and he validi of new color model, several experimens and analsis of he resuls are presened and demonsraed. Evaluaion of he performance shows ha proposed model around en imes faser compared o sandard HS color model for image processing. 013 AENS Publisher All righs reserved. To Cie This Aricle: Dmiri Skopin, Jamil Al-Azzeh, Jihad Nader, Ashraf Abu-Ein and Dr. Mazin Al Hadidi., Fases Color Model For mage Processing Using Embedded Ssems. Aus. J. Basic & Appl. Sci., 7(14): 83-89, 013 NTRODUCTON is well known ha man algorihms of image processing in spaial, frequenc and ransform domains designed for gra scale images (Keih Jack, 008; János Schanda, 007). To be applicable o color images such gra-scale mehods require reformulaion using color models ha decouples he color and gra-scale informaion in an image, making i suiable for man of he gra-scale image processing algorihms. The color model is an absrac mahemaical model describing he wa colors can be represened as upelos of numbers, picall as hree or four values or color componens. Tpical compuer graphics, digial signal processing and analsis are represened in eiher RGB color model (used in video ssems and graphical files) (Keih Jack, 008); YQ, YUV or YCbCr (used in video ssems); or CMYK used in color priners (János Schanda, 007). The CE 1931 color space (Wrigh, William David, 007) was derived from a series of experimens done b W. David Wrigh (Wrigh 198) and John Guild (Guild 1931). n he CE color model each color is represened as hree componens, bu he sum of is reurn is a consan one. The RGB color model iself has limied applicaions since he R, G, and B componens of an objec s color in a digial image, or digial signal processing and analsis are correlaed wih he amoun of ligh hiing he objec. As a resul, he colors are correlaed wih each oher, and image descripions in erms of hose componens are difficul o discriminae he objec. Descripions in erms of hue, sauraion, value (HSV) or hue, sauraion, inensi (HS) are ofen more relevan. The HSV and HS color spaces were developed for color manipulaing and allow independed represenaion of color informaion and brighness, which is imporan for image processing algorihms (Salvador, E., 004). The difference beween HS and HSV is in compuaion of brighness componen ha can be V or according o he model. HLS color model (hue, lighness, sauraion) is similar o HS bu he erm lighness is used raher han inensi. Tradiionall he HS color model is bes for digial image processing echniques in spaial (such as convoluion, equalizaion, hisograms and so on) and in frequenc (digial image filraion and specrum analsis) domains. These image processing algorihms manipulae he brighness componen when color componens sae unchanged during image processing. The HSL, HSV, HS, or relaed models are ofen used in compuer vision and image analsis for feaure deecion or image segmenaion. The applicaions of such ools include objec deecion, for insance in robo vision; objec recogniion, for insance of faces, ex, or license plaes; conen-based image rerieval; and analsis of medical images. Some new color models were inroduced in (Salvador, E., 004) and (Gevers, T. and A.W.M. Smeulders, 000) which are relaivel robus agains Corresponding Auhor: Dmiri Skopin, Compuer Engineering, Al Balqa Applied Universi Amman, 11134, Jordan. d_skopin@mail.ru

2 84 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: illuminaion changes. These color models are based on he dichromaic reflecion model and show good characerisics from viewpoin of he performance. However, hese color models are no coninuous and irreversible. n his research, we have inroduced a new color model named R G where wo componens sore informaion abou decorrelaed colors, and he hird one abou inensi. As i will be shown, he proposed model is around en imes more efficien in erms of performance compare o sandard color models for digial image processing. This model is especiall acual for such embedded ssems as modern smar phones, console devices and pla saions equipped wih low cos processors. The remainder of he paper is organized as follows: Secion represens descripion of proposed color model, secion 3 describes resuls of performance and quali analsis, secion 4 shows resuls of gami evaluaion of he color model, and secion 5 provides he conclusion.. Developmen and evaluaion of new color model: Sandard HS color model for image processing and analsis is no opimal in erms of performance. According o he equaions proposed b Gonzalez and Woods (009) conversion from RGB o HS color space (direc ransform) mees a se of complex (in erms of compuaional complexi) mahemaical operaions such as exracion of arccosine funcion, square roo operaor and searching for he minimum elemen of he se. This echnique requires compuaion of cosine funcion and floaing-poin division when convering from HS o RGB. All operaions in direc and inverse ransforms based on floaing poin require arihmeic ha definiel causes oal degradaion of image processing performance. Perhaps, his quesion is non-acual for modern deskops ssem equipped wih las generaion of muli-core processors wih high performance of floaing poin uni, bu for low cos segmen of embedded ssems ha emulaes floaing-poin arihmeic sandard color models is no opimal in poin of view of performance. Nowadas, among all porable devises mos popular are operaing under Android operaing ssem (Goesman, B., 01). According o performance ips of Android developers [8], he programmer should avoid using floaing-poin compuaions since floaing-poin few imes slower han ineger on Android-powered devices. n speed erms, here is no difference beween floa and double and on be and ineger on he more modern hardware. Also, even for ineger daa pe, some processors have hardware mulipl bu lack of hardware divide. n such cases, ineger division and modulus operaions are performed in sofware. Furhermore, programmer of Android applicaions should avoid mehod calls for simple mehods ha called ver ofen in a loop, consider o inline he code (Reo Meier, 01). Thereb, calls of cosine, arccosine or square roo mehods in sandard color models are quie compuaionall expensive. n our poin of view, he fases color model can be elaboraed using CE color space where each richomaic coefficien defined as a raio beween appropriae RGB componen and sum of all RGB componens. Each richomaic coefficien in CE space is represened as a fracional number in he range from 0 o 1. n our color model, we have used wo richomaic coefficiens for red and blue colors denoed as R, G respecivel. Third componen of our model is inensi, and i is designaed as. A logical shif biwise operaion (Morris Mano, M., 01) has been used o avoid floaing poin arihmeic and o adjus he range for desired be represenaion from 0 o 55. Convering from RGB color space o R G color space is represened b equaions , if R + G + B 0 S R + G + B,oherwise R << 8 R S G << 8 G S (1) () (3) S >> 4 (4) where "<<" and ">>" denoes lef and righ logical shif respecivel. is obvious o see, ha convering from RGB o R G color space includes onl simple compuaional operaions, and he inensi componen is defined b shifing 4 bis o he righ of he sum of RGB componens. Definiion he inensi componen according o he equaion 4 allows simplificaion of inverse ransformaions for convering he R G o RGB color space ha can be done according o he equaions 5-7. R R >> 6 (5)

3 85 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: G G B >> 6 ( 55 R G ) >> 6 (6) (7) is obvious o see ha equaions 5-7 are ideal in erms of microprocessor performance because he conain onl fas compuaional operaions such as ineger muliplicaion, subracion and logical shif. Since inensi componen of R G has reduced resoluion, 19 grades compare o 56 grades of sandard HS model for mos popular 3 bis image formas, he proposed R G is loss daa color model, meaning ha some original image informaion is los and canno be resored, possibl affecing image quali. Resuls of appling he R G model for digial image processing are represened in figure 1. The figure shows ha original color image is rerieved even i wen hrough convering and inverse ransforms process in he proposed color space model. Fig. 1: Appling sandard and proposed color models for digial image processing. (a) Original image. (b) The same image afer direc and inverse ransforms using proposed color model. (c) Conras sreching of original image using sandard HS color model. (d) Conras sreching of original image using proposed color model. also shows ha inensi modificaion during conras sreching process did no aler he values of color componens, and i did no impac he overall color percepion. Visual analsis of he images represened in figure 1 b naked ees as well as digial zooming allows o come o conclusion ha proposed daa loss color model does no change image quali significanl. To sud degradaion of image quali caused b R G loss daa model, a pilo sud is conduced for sandard and elaboraed color models. Around 100 digial images wih differen resoluions have been analzed hrough compuaion for parameers such as imum difference beween pixel brighness of original and ransformed image (Max. Dif.). The also have been analzed hrough normalized cross-correlaion (NCC) (Rawashdeh, N.A., 007), mean squared error (Keelan, B., 00), peak signal-o-noise raio (PSNR) (Simoncelli, E.P., 005) and srucural similari (Wang, Z., 004; Loza, e al., 006). The PSNR is mos commonl used as a measure of quali of reconsrucion of loss compression codecs. We have used he PSNR coefficien o evaluae new color model in daa loss erms. When original image was compared wih he image reconsruced from R G color space, he PSNR was used as an approximaion o human percepion of reconsrucion quali (Welsead, Sephen T., 1999). Higher values of PSNR coefficiens indicae ha he reconsrucion is of higher quali. As for he range of validi of his meric, i is onl conclusivel valid when i is used o compare resuls from he same codec (or model) and same conen (Hunh-Thu, Q., 008). Tpical values for he PSNR in loss image formas and video compression are beween 30 and 50 db, where higher is beer (Welsead, Sephen T., 1999). Accepable values for wireless ransmission quali loss are considered o be abou 0 db o 5 db. n curren paper, he PSNR coefficien is defined according o he equaion 8: PSNR 0log 10 ( 1 m n m 1 n 1 i 0 j 0 L ) [ X ( i, Y ( i, ] (8)

4 86 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: where L is he dnamic range of pixel values, X ( i, - original image of he m x n pixels resoluion, Y ( i, is nois image o be evaluaed. The srucural similari (SSM) index is a mehod for measuring he similari beween wo images (Wang, Z., 004). Like PSNR coefficien he SSM index is a full reference meric, he measuring of image quali based on an iniial uncompressed or disorion-free image as reference. The difference wih respec o oher echniques menioned previousl such as PSNR, is ha hese approaches esimae perceived errors on he oher hand SSM considers image degradaion as perceived change in srucural informaion. Compuaion of SSM index provided in curren work according o he equaion 9: (µ xµ + [ k1l] )(σ x + [ kl] ) SSM ( µ + µ + [ k L] )( σ + σ + [ k L] ) x 1 x (9) where µ x, µ are average values of pixels in images X ( i, σ x is covariance of images, k1 and k are correcion coefficiens assigned in curren work as 0.01, Y ( i, correspondingl, σ, x σ are variances, and 0.03 accordingl. Table 1 summarizes he resuls of appling elaboraed color model o 100 digial images acquired wih differen resoluions from 5 o 1 Megapixels. The daa, represened in he able, shows he average values of each parameer. MSE is mean square error, Max. Dif denoes he imum difference of pixels brighness beween original and ransformed images, NCC is normalized cross-correlaion, PSNR - peak signal o noise raio, when SSM denoes he srucural similari index, ± denoes sandard deviaion. Table 1: Accurac of proposed color model compare o sandard HS color space and JPEG codec. Name of model MSE Max NCC PSNR SSM (codec) Dif. HS R G.73 ± ± ± ± ±0.001 JPEG (QF95) 3.67 ± ± ± ± ±0.001 JPEG (QF75) ± ± ± ± ±0.00 is possible o see ha unlike sandard lossless HS model, our proposed color model cause sligh, bu no an annoing disorion of he image quali ha can no be deeced b human persepion. The degradaion of he image quali consiues less disorion compared o he quali of JPEG codec applied wih highes quali facor (QF95) and much beer quali compared o JPEG quali facor 75 ha is characerized as high quali of JPEG compression in modern phoo indusr. Resuls reveal ha our proposed color model is beer in erms of quali compared o JPEG codec ha is widel used as file forma in digial cameras. Compared o sandard HS color model, our model has worse resul since sandard model is lossless.. Performance esing of new color model: To es he performance of he new model, a porable device (smarphone) has been chosen as an embedded ssem. The smarphone is equipped wih ARM 11 processor, and operaes wih Android operaing ssem. is well known ha Android operaing is he mos common OS (Goesman, B., 01) as well as i is an open source. The ARM 11 processor famil provides he engine ha powers man smar phones in producion oda and is also widel used in consumer, home, and embedded applicaions. For example such popular models as Nokia 5800, HTC Hero, iphone 3G, Samsung Galax, LG P500 are based on ARM 11 processor. The archiecure of his processor includes ineger core uni, vecor floaing poin coprocessor, and shifer ha allow fas (one machine ccle) hardware based logical shif (shifer) wih regiser conrolled shif disance. Presence of hardware shifer in archiecure of modern embedded ssems allows o perform fases (single clock) shif operaion independenl o shif disance. Since proposed color model based on logical shif operaions (see equaions -7) i is obvious o see ha he model adaped wih archiecure of embedded ssems. To address he performance of he proposed color model, a pilo sud was conduced for randoml seleced 100 images acquired wih differen resoluions up o 5 Megapixels. All experimens were provided wih a device ha includes following parameers: CPU MHz ARM 11, 51 MB RAM, Android OS, v. (Froo). To provide numerical parameers of performance of he new proposed color model, an Android program was creaed using Java language and Android Developmen Ki. The program execues basic image-processing algorihms in spaial domain using sandard HS and proposed R G color models wih measuremen ime

5 87 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: laenc required for algorihms. Figure 3 illusraes he acual measuremen resuls of program laenc according o he image resoluion for sandard (.b) and proposed (.a) color models when able shows numerical daa of provided ess Direc RG nverse RG (a) Direc HS nverse HS (b) Fig. : Relaion beween ime laenc and image resoluion during image processing using (a) proposed model, (b) - sandard HS color model. The value of speedup is defined as he raio beween HS and R G color model laencies for direc (from RGB), inverse (o RGB) conversions and basic image processing algorihms using sandard and proposed color models. Table : Speedup of proposed color model. d Applied algorihm Speedup, S. Dev. mean value 1 Direc ransform nverse ransform mage brighness correcion Conras sreching Hisogram equalizaion The pracical resuls show ha proposed model is around 13 imes faser in direc and inverse ransforms compared o sandard model. This achievemen can be explained b he fac ha onl fas processor operaions involved o compuaions. On he oher hand floaing poin arihmeic of HS color model, using rigonomeric funcions and compuaion of square roos according o (Gonzales, R., R. Woods, 009) give large amoun of ime laenc and as consequence an excellen value of speedup of proposed model. During rise he compuaional complexi of image processing algorihms speedup is decreasing up o he value 8.4 obained for hisogram equalizaion algorihm. This fac can be explained b definiion of speed given b equaion 10: S 1 c1 c + + p p (10) Speed up is defined as he raio of image processing ime for sandard (HS) model and proposed R G color models. mage processing ime is sum of conversion ime c (direc and inverse) and ime of image processing algorihm p. is obvious o see ha incremen of image processing ime ha is common for wo models cause decremen of speedup value. 4. Evaluaion of new color model for is gamu: As i was menioned above, color image processing based on ransforming from RGB color space o R G color space, modificaion in his space he inensi componen hen ransforming back o RGB color space. These ransformaions poeniall ma generae a gamu problem, when he values of he variables ma no be in heir respecive inervals. The spaial allocaion of proposed R G color model is illusraed in he figure 3. (a) (b)

6 88 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: Fig. 3: Spaial configuraion in fron (a) and in back (b) of he R G color model. Figure 3 is rendered using one poin perspecive projecion creaed b OpenGL AP (Dave Shreiner, 013). Since R G model is based on ineger arihmeic some poins of R G virual spaces are unfilled, conain brakes or jes ha can be seen in he origin of coordinae ssem (see he figure 3.b). is obvious o see ha unlike o HS ssem, proposed color model is no expanded o a clindrical form ha causes he risk o specif he coordinaes, which lie ouside he color gamu. n order o preven his risk, some boundar elaboraed and described below. is well known ha geomericall, he RGB ssem can be represened as a cube wih he red, green and blue axes defining he x,, and z vecors respecivel. The lengh of he cube edge for mos popular 4 bis image formas equal o 55 unis because each color componen encoded using 8 bis. According o equaions - 4 he range of richomaic red and green componens of R G model also limied b 55 bu imum value of inensi (see equaion 4) does no exceed 191. During image processing original image is convered o R G color model, and hen inensi componen is modified according o he image processing algorihm. Afer inverse conversion o RGB color space (equaion 5-8) resul should be limied b he value 55 for each color componen of RGB color model. To describe he boundar condiion in general case we have designaed as a imum value of inensi componen of R G color model, R - he imum size of RGB cube edge. According o equaions 5-7 he value of inensi componen should be limied b he value compued according o he equaion 11: R << 6 min( R, G,55 R G ) (11) n paricular case for processing 4 bis images equaion 11 can be simplified o he equaion 1: 1630 min( R, G,55 R G ) Using he value of inensi limied b he value componens will be in heir respecive inervals. (1) is a guaranee ha resuling values of R, G, B 5. Conclusion: Processing of color images based on sandard color models such as HSV, HS, HSL where wo componens conain color informaion and hird one inensi value is ver popular in digial image processing echniques. According o algorihms he inensi componen is modified during image processing hen inverse ransformaion applied o reurn resuls o original RGB forma. This approach allows processing of color images using algorihms originall designed for gra scale images. However, sandard models are no opimal in erms of speed since heir equaions include complex mahemaical operaions for microprocessor ssems such as floaing poin arihmeic, compuaion of square roo and riginomeric funcions. Especiall his problem is acual for embedded ssems ha are ofen based on low cos processors wih poor performance compare o performance of modern deskop ssems. n curren paper new color model has been suggesed for image processing using low cos embedded ssems. The model designed wih respecing he archiecure of low cos mobile processors and performance ips of Android operaing ssem. Conversion from original RGB color space o proposed R G ssem and is inverse according o equaions 1-7 based on fas processor operaions such as logical shif and basic ineger arihmeic such as addiion, subracion, muliplicaion and division. Performance esing provided using ARM 11 processor shows more han en imes speed improvemen in digial image processing compare o sandard HS model defined b Gonzalez and Woods (009). Since proposed color model is no expanded o a clindrical form, boundar condiions for general case (equaion 11) and paricular case of mos popular 4 bis images (equaion 1) are defined. According o equaions 1-7 proposed color model can be characerized as daa loss forma, bu our invesigaions show ha degradaion of quali is small o be deeced visuall even using zooming ool (see he figure ). Saisical invesigaions of original and processed images using sandard and proposed color model confirm minimum change of image parameers even when compared wih JPEG codec applied wih imum quali of image. Therefore, proposed color model can be recommended o be included o sofware of modern embedded ssems for fas digial image processing of color images.

7 89 Dmiri Skopin e al, 013 Ausralian Journal of Basic and Applied Sciences, 7(14) December 013, Pages: REFERENCES ARM 1176JZF-S Technical Reference Manual. Revision: r0p7. ARM Limied 010. Bernd Jähne, 01. Digial mage Processing, Springer; 7h ed. ediion, 01, SBN-10: , SBN-13: Dave Shreiner, Graham Sellers, John Kessenich, Bill Licea-Kane, 013. OpenGL Programming Guide: The Official Guide o Learning OpenGL, Version 4.3, 8h Ediion, 013, SBN-10: , SBN-13: Gevers, T. and A.W.M. Smeulders, 000. PicToSeek: combining color and shape invarian feaures for image rerieval, EEE Trans. on mage Processing, 9: CrossRef. Gonzales, R., R. Woods, 009. "Digial image processing", hird ediion, Pearson Prenice Hall, SBN Goesman, B., 01. "Mobile Operaing Ssems - Readers' Choice Awards 01: Smarphones and Mobile Carriers", PCMag.com, Rerieved Hunh-Thu, Q., M. Ghanbari, 008. "Scope of validi of PSNR in image/video quali assessmen". Elecronics Leers, 44(13): doi: /el: chiro Masuda, Yukio Nomoo, Kei Wakabaashi and Susumu oh, 008. Lossless Re-encoding of JPEG images using block-adapive inra predicion. Proceedings of he 16h European Signal Processing Conference (EUSPCO 008). János Schanda, 007. Colorimer. Wile-nerscience. SBN Keelan, B., 00. Handbook of mage Quali Characerizaion and Predicion, Marcel Dekker, nc., New York. Keih Jack, 008. Digial Video and DSP, Elseiver nc, SBN: Loza, e al., 006. "Srucural Similari-Based Objec Tracking in Video Sequences", Proc. of he 9h nernaional Conf. on nformaion Fusion. Morris Mano, M., Michael D. Cilei, 01. "Digial Design: Wih an nroducion o he Verilog HDL", Prenice Hall; 5 ediion, SBN Rawashdeh, N.A., 007. Characerizaion of Defecs in mages wih Complex Conen, Ph.D. Disseraion, Deparmen of Elecrical and Compuer Engineering, Universi of Kenuck, Lexingon, KY. Reo Meier, 01. "Professional Android 4 applicaion design" John Wile &Songs, nc, ndianpolis, ndiana, SBN: Salvador, E., A. Cavallaro and T. Ebrahimi, 004. Cas shadow segmenaion using invarian color feaures, Proc. of Compuer Vision and mage Undersanding, 95(): CrossRef. Simoncelli, E.P., 005. Saisical modeling of phoographic images. n Handbook of mage and Video Processing. Sirner, M. and G. Seelmann, 007. mproved Redundanc Reducion for JPEG Files. Proc. of Picure Coding Smposium (PCS 007), Lisbon, Porugal, 7-9. Wang, Z., A.C. Bovik, H.R. Sheikh and E.P. Simoncelli, 004. "mage quali assessmen: From error visibili o srucural similari," EEE Transacions on mage Processing, 13(4): Welsead, Sephen T., Fracal and wavele image compression echniques. SPE Publicaion, pp: SBN Wrigh, William David, 007. "Golden Jubilee of Colour in he CE The Hisorical and Experimenal Background o he 1931 CE Ssem of Colorimer". n János Schanda. Colorimer. Wile nerscience, pp: 9-4. doi:10.100/ ch. SBN