Illumination Compensation for 2-D Barcode Recognition Basing Morphologic

Transcription

1 Research Joural of Applied Scieces, Egieerig ad Techology 5(12): , 2013 ISSN: ; e-issn: Maxwell Scietific Orgaizatio, 2013 Submitted: March, 09, 2012 Accepted: August 15, 2012 Published: April 10, 2013 Illumiatio Compesatio for 2-D Barcode Recogitio Basig Morphologic 1, 2 Jia-Hua Li, 1 Yi-We Wag, 1 Yi Che ad 1 Meg Zhag 1 Guo-Cheg Wag ad 1 Pig Li 1 Key Laboratory of Electroic Thi Films ad Itegrated Devices Uiversity of Electroic Sciece ad Techology of Chia, Chegdu Chia Iformatio Egieer School of Nachag Uiversity Abstract: Improvemet of image quality has bee highly demaded i digital imagig systems. This study presets a ovel illumiatio ormalizatio approach for 2-D barcode recogitio uder varyig lightig coditios. MMs (Morphological trasformatios) are employed to origial images usig big scale multiple SEs (structurig elemets). The we make use of etropy to fuse images. The performace of proposed methodology is illustrated through the processig of images with differet kids of 2-D barcodes uder differet backgrouds. The experimetal results show that this approach ca process differet kids of 2-D barcodes uder varyig lightig coditios adaptively. Compared with other covetioal methods, our proposed approach does a better job i processig 2-D barcode uder o-uiform illumiatio. Keywords: Mathematical morphology, morphological cotrast, structure elemet, 2-D barcode INTRODUCTION With the popularity of had held camera-eabled digital devices, amog the wide rage of automatic idetificatios such as RFID, magetic card, barcode is becomig icreasigly popular i recet years. Barcode techique was first itroduced by Westighouse Lab i 1920s. It is a symbol composed by bars ad blaks accordig to a certai ecodig rule ad it ca store umbers ad letters i a small space (Hu et al., 2009). The key to 2-D barcode idetificatio lies i the processig of acquired image. Images acquired by image acquisitio equipmet like PDA (Persoal Digital Assistat) will usually degeerate because of the complex ad chageable eviromet. Amog may factors causig image degeeratio, illumiatio variatio is a importat oe ad brigs great challege to 2-D barcode idetificatio. Thus, i order to get a high barcode recogitio rate, we must use light equalizatio to process acquired images. Nouiform illumiatio is still a challege i 2-D barcode recogitio research area. A variety of approaches have bee proposed to solve this problem, such as histogram equalizatio, logarithm trasform, gamma correctio ad mathematical morphology. Histogram equalizatio techique ca balace gray distributio of the whole image. The basic idea is to combie gray levels with fewer frequecies ito eighborig gray levels i order to decrease the umber of gray levels ad icrease global cotrast ad clarify image details. But this global processig techique is still difficult to solve the ueve illumiatio because the global properties of the image caot be properly applied i the local (Sherrier ad Johso, 1987). To solve this problem, regio-based histogram equalizatio, adaptive histogram equalizatio ad block-based histogram equalizatio have bee proposed (Sha et al., 2003; Park et al., 2008). Sha et al. (2003) proposed a ormalizatio method called quotiet illumiatio relightig based o the assumptio that lightig modes of the image are kow or ca be estimated. Che et al. (2006) proposed a logarithmic DCT method. This method takes the average of the max ad miimal values of the whole image as image s backgroud. As a result, local image teds to be degeerated ad the cotrast might ot be correctly ehaced i images with poor lightig whe the image backgroud has damaged chages i lumiace. Jimeez-Sachez et al. (2009) proposed a mathematical morphology techique based o Che et al. (2006), which divides image ito several areas, the uses average of the max ad miimal gray values of every area to replace the gray value of that area. By this way, Jimeez-Sachez et al. (2009) ca get a relatively good result. But images processed by this method ted to have blockig effect especially whe the divided areas are too large. Xu et al. (2008) proposed a algorithm which uses big-scale elemets to implemet Top-hat Correspodig Author: Jia-Hua Li, Key Laboratory of Electroic Thi Films ad Itegrated Devices Uiversity of Electroic Sciece ad Techology of Chia, Chegdu Chia

2 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 trasformatio o images uder varyig lightig coditios. This algorithm is simple ad beig widely used. However, it caot acquire good processig results uder complex lightig coditios due to usig sigle structurig elemet. The mai purpose of morphological operatio is to use SE with certai sizes ad shapes to detect images ad the obtai the relative structural features. So the sizes ad shapes of SEs play crucial roles i such type of processig. However, we ca ever fid a SE which is suitable for all kids of images, so i recet years, researchers begi to search a ew mathematical morphological method which is ot sesitive to SE. Multi-scale morphological trasformatio like geodesic erosio, geodesic dilatio, opeig recostructio ad closig recostructio are all good algorithms (Mukhopadhyay ad Chada, 2001, 2003). Morphological trasformatio usig multiple SE ca ot oly filter peak ad low oise but also maitai the image details. Operators cosist of multiple SE like ROBERTS, Cay ad Laplacia have their ow advatages i image filterig, edge detectio ad so o Acto ad Mukherjee (2000) ad Beediktsso et al. (2003). This study improves Xu et al. (2008) ad Park et al. (2008) algorithm. Firstly, we use big scale multiple SEs to implemet Top-hat trasformatio o origial images. The calculate the iformatio etropy of these processed images. Fially, fusig these images that implemeted by Top-hat trasformatio to get illumiatio ormalize image based o the ratio of iformatio etropy. Basis of algorithm: A morphological trasform: MM which is a extesio of Murkowski s set theory were first systematically examied by Gozalez ad Woods (1987). It has bee a well-kow techique used i image processig ad computer visio. Morphological operators cosist of erosio, dilatio, opeig, closig ad other derived trasformatio. It is well kow that the erosio ad dilatio is a pair of dual operators. I gray scale, erosio of a image ƒ by a SE B is defied as Gozalez ad Woods (1987): [ ε ] ( ) ( ) mi ( ) B f x = f x+ b b B (1) [ f ] γ ( f) = δ ε ( ) B B B [ f ] φ ( f) = ε δ ( ) B B B (3) (4) Usually, opeig ad closig operatios are used i morphological filters. They smooth the image with differet behaviors. Opeig a image with a disk SE will smooth the cotour ad elimiates small islads ad sharp peaks or capes. Closig a image with a disk structurig elemet smooth the cotours, elimiates small holes ad fills gaps o the cotour. Choosig proper morphological filter depeds o target's sharp, size ad directio. Usig opeig ad closig operatios of a SE which is ot uiform with relative structure i image ca elimiate these relative structures; these structures ca be restored through the differece betwee origial image ad its opeig or closig results. These morphological trasformatios are called WTH (white top-hat) ad BTH (black tophat). The WTH trasformatio obtais all bright features smaller tha SE: white g = f γ ( f) B (5) Similarly a BTH sieves out the dark features smaller tha SE: black g = f φ ( f) (6) Itroductio to iformatio etropy: Iformatio etropy is a measure of the disorder degree of system. It is calculated by the weighted average of each gray level probability's logarithm ad represets how much iformatio the image cotais: B H = k pilg 2 pi m (7) I Eq. (7), i represets the iformatio type, p i represets the possibility of correspodig iformatio type, k is the weight factor. For a M N gray image, iformatio type i [0,255] k = 1 meas etropy of gray image is: Dilatio is deoted ad defied as: [ δ ] ( ) ( ) max ( ) B f x = f x+ b b B (2) The opeig operatio performs erosio first, followed by dilatio, while the closig operatio performs dilatio first, followed by erosio (Haralick et al., 1987): H p lg p = 0 i 2 i (8) Approach to ormalize o-uiform illumiatio image with multiple SE: As 2-D barcodes are used uder complex eviromets ad images are acquired uder varyig lightig coditios, images acquired by acquisitio equipmet usually have bright gradiets

3 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 Fig. 1: PDF417 ad QR code with ueve lightig backgroud which might cause threshold techique uavailable i followig process, therefore lead to a false aalysis. I order to improve barcode idetificatio rate, we must implemet illumiatio ormalizatio o acquired images. I a gray image, the variatio of image iformatio is more severe tha the lightig, which meas coected areas of lightig are Bigger tha coected areas of image iformatio. Gray image ca be treated as a set due to its graphs ad sub graphs. We kow from Eq. (3) that SE moves aroud the set from the iside. Smaller set structures which stretch ito the backgroud were elimiated ad bigger size backgroud structures do ot chage. Barcode images are usually take uder dark lightig, so we ca choose relatively big-scale SEs ad use BTH to filter all image iformatio except o-uiform illumiatio backgroud of a image. The traditioal MM operatio employs sigle SE to do backgroud estimatio upo images, the most used SEs iclude square ad disc. As show i Fig. 1, whe ueve lightig degree is large, image backgroud has strog fluctuatio. As sigle SE is oly sesitive to edges with the same directio, thus edges with differet directios from the SE will be smoothed. Therefore, takig use of sigle SE to do backgroud estimatio caot keep good image details, causig a big deal of backgroud leakage. Whe sigal oise ratio is low, the gray value of backgroud leakage would possibly be higher tha the gray value, makig severe degeeratio after image processig, while employig multiple SEs ca keep details of image udamaged. From the aalysis above, it is evidet that multistructure morphological trasformatio has more advatages tha sigle structure morphological trasformatio, so we used big scale multi-structure elemet to do WTH trasformatio o image ƒ, the took use of etropy to fuse processed images, which is show as follow: B( f) = ωisi 1 (9) I Eq. (9) S i is a image upo which WTH trasformatio has bee doe usig differet SEs. ω i is image's weight factor. From Eq. (7), we ca get: ω li, = 255 p l= 1 0 lg li, 2 li, p lg li, 2 li, (10) I Eq. (10), p (l, i) represets the probability of gray value i i image category l. I this study, Fig. 2 illustrates the structurig diagram of the ew algorithm based o multiple structures, the complex algorithm ca be briefly described as follows: Establish the shape ad size of multi-structure SEs: I MM, multi-structure SEs ca be used to keep image details i differet directios. I this study we choose liear SEs with differet directios as is show i Fig. 3. p p Fig. 2: The structurig diagram of the ew algorithm 3275

4 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 w = 4 H = 1 H (13) Fig. 3: Differet SEs used i proposed algorithm Employ multi-structure SE based WTH trasformatio o origial image: I WTH trasformatio, the purpose of opeig operatio is to elimiate foregroud to get backgroud, so the scale of SE should be bigger tha the scale of foregroud structure, to be specific, bigger tha the biggest coected area i 2-D barcode. So we eed to estimate the scale of whole image's foregroud (structure of 2-D barcode). I a lower visio 2-D barcode, the biggest coected area accouts a higher proportio i the whole image. I the lowest visio 2-D barcode, idetificatio accouts almost 1/4 proportio of the whole image. Cosiderig that acquired image ot oly iclude 2-D barcode but also iclude some backgroud area, so idetificatio accouts less tha 1/4 proportio of the whole image. Thus we take 1/4 of the smaller scale of image lie ad colum as the scale of SE. Through WTH trasformatio we ca get images cotaiig iformatio i differet directios. From Eq. (5) we ca get: gg ii = ƒ γγ BBii i (1, 2, 3, 4) (11) I Eq. (11), ƒ is the origial image, B i are SEs as show i Fig. 2 whose scales are 1/4 of the smallest scale of image lie ad colum, g i ad are the WTH trasformatio results of correspodig SEs. Fuse images: Based o the cocepts of idex fuzzy etropy ad weighted etropy (Xu et al. 2008), we calculate the membership fuctio s correspodig fuzzy weighted etropy ad take each pixel s correspodig membership fuctio as the weight of weighted etropy. I fusio processig, the greater value of idex fuzzy etropy, the more iformatio it cotais. From Eq. (9) we ca obtai image s iformatio etropy processed by big scale WTH trasformatio: 255 H p lg p = i 2 i 0 (1, 2, 3, 4) (12) From Eq. (11) we ca obtai image s weight factor as follow: 3276 From Eq. (9) we ca obtai the fial fused image: f = w * g TEST RESULTS AND DISCUSSION (14) There are various kids of 2-D barcodes today. Accordig to the shape of image, 2-D barcodes ca be classified as stack code ad matrix code. The most widely used 2-D barcodes iclude PDF417 ad QR code, Data Matrix, Maxi code, code49 ad code 16K. I this study the PDF417, QR code which are the most popular are used to evaluate the proposed approach. Experimets maily focus o verifyig proposed approach s adaptive capacity o differet barcode s scales, tile agles ad textures. Verifyig proposed approach s adaptive capacity o differet barcodes scales: I MM, SE plays a key role. The processig result would be degeerated with the decrease of match degree betwee SE ad cocered features. I order to obtai a good processig result, SE is usually chose accordig to the priority kowledge of cocered features scale i image. I autoidetificatio aalysis, oe caot kow the scale of cocered features i advace, so fidig a scale-adaptive MM is of great sigificace. Relative scales iclude the size of cocered features ad the proportio of cocered features accoutig the whole image. Figure 4a got a 2-D barcode accouted about 4/5 proportio of the whole image. Figure 4b got a 2-D barcode accouted about 1/10 proportio of the whole image. Figure 4c is the processig result of Fig. 4a usig our algorithm. Figure 4d is the processig result of Fig. 4b usig our algorithm. From the experimetal results we ca see that proposed algorithm ca adaptively process images with differet resolutios ad differet proportios of 2-D barcodes as well. It idicates that proposed algorithm has a good adaptive capacity o differet barcode s scales. Verifyig proposed approach s adaptive capacity o differet barcodes tile agles: There are two kids of 2-D barcode idetificatio devices: fixed ad hadheld. Fixed 2-D barcode idetificatio device placed i fixed termial; hadheld device iclude 2-D barcode scaer gu ad portable PDA device. Whe users capture image with a PDA camera, it is difficult

5 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 (c) (d) Fig. 4: Processig results of barcode images with differet scales uder ueve lightig (c) (d) Fig. 5: Processig results of barcode images with differet tile agles uder ueve lightig 3277

6 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 (c) (d) Fig. 6: Processig results of differet kids of barcode images uder ueve lightig (c) (d) 3278

7 Res. J. Appl. Sci. Eg. Techol., 5(12): , 2013 Fig. 7: Images uder ueve lightig processed by differet algorithms for them to hold devices at a perfectly right agel. I MM, SEs i differet directios ca get image s iformatio o correspodig directios. As a result, image s iformatio ca be kept complete eve though image has bee iclied. I this study, we used four SEs with differet directios to process images. Figure 5a ad 5b are o-uiform illumiatio images with differet tile agles. Figure 5c ad 5d are images processed by proposed algorithm. From the experimetal results we ca see that proposed algorithm ca adaptively process o-uiform illumiatio images uder differet tile agles. This idicates that proposed algorithm has a good adaptive capacity o differet barcodes tile agles ad ca resolve iclied image s o-uiform illumiatio problem as well. Verifyig proposed approach s adaptive capacity o differet types of 2-D barcodes: Accordig to shape, 2-D barcodes ca be classified ito stacked code ad matrix code. Differet types of 2-D barcodes have differet textures. I MM, processig image eeds to choose differet SEs accordig to differet textures. Textures i 2-D barcode iclude image sigature, size of coected areas of data code ad differece betwee shapes ad scales. Although differet 2-D barcodes have differet textures, these textures have relatively small coected areas tha backgroud ad ca be extracted by TOP- HAT trasformatio usig big scale SE. Figure 6a is stacked PDF417 2-D barcode, Fig. 6b is matrix QR code; they were take uder the same o-uiform illumiatio coditio. Figure 6c ad 6d are the processig results of origial images. From the results we ca see that proposed algorithm ca adaptively process differet kids of 2-D barcodes. This idicates that proposed algorithm has a good adaptive capacity o differet kids of 2-D barcodes. (e) Compariso with other algorithms: Jimeez- Sachez et al. (2009) ad Mukhopadhyay ad Chada (2003) ormalized images usig the same filter model ad differet illumiatio ormalizatio approaches. The mai differece betwee the two images lies o the process of backgroud. Mukhopadhyay ad Chada (2003) put the average of the max ad miimal values as the whole image's backgroud; Jimeez-Sachez et al. (2009) divided image ito several parts with the same size ad put average of the max ad miimal values of each area as the backgroud of correspodig area. Soille (1988) used sigle bigscale SE to do Top-hat trasformatio ad got uiform illumiatio image. Figure 7a is the origial matrix QR code. Figure 7b, 7c, 7d ad 7e are images processed by algorithms proposed by Mukhopadhyay ad Chada (2003), Jimeez-Sachez et al. (2009) ad Soille (1988) ad ours respectively. I Fig. 7b, correctio might be deleterious to some iformatio of image, image appears to be either too dark or too bright, idicatig that algorithm proposed by Mukhopadhyay ad Chada (2003) caot properly apply global properties to the local; compared with (7-a), algorithm i Fig. 7c improved local properties, but obvious blockig effects ad profiles arose because iappropriate values may be obtaied, idicatig that algorithm proposed by Soille (1988) uable to process good global ormalizatio o o-uiform illumiatio images. Algorithm i Fig. 7d greatly improved illumiatio ormalizatio, but ueve illumiatio still exits. I Fig. 7e, illumiatio is ormalized idicatig that our proposed algorithm is more advatageous tha others. CONCLUSION A ovel illumiatio ormalizatio approach which is used to 2-D barcodes is proposed i this study. Proposed approach is divided ito three stages. I the first stage, we proposed a set of liear structurig elemets i four differet directios, which were 0, 3279

8 Res. J. Appl. Sci. Eg. Techol., 5(12): , , 90, 135, respectively. The scale of SE is the 1/4 of the smaller scale of image lie ad colum. I the secod stage, we took use of SEs acquired i the first stage to perform WTH trasformatio o origial image, obtaiig images cotaiig differet iformatio about differet directios. Fially, we perform weighted fusio o images acquired i the secod stage to obtai the fial image. Weight factor was determied by iformatio etropies. I this study we verified proposed algorithm o several aspects icludig barcodes scales, tile agles ad types of barcodes. Meawhile, we compared our proposed algorithm with several commoly used illumiatio ormalizatio algorithm, results show that our algorithm is more advatageous tha others. However, our proposed algorithm requires large amout of computatio ad storage space. The further studies will be carried out to optimize this algorithm to reduce its computatio work ad storage space ad implemet this algorithm through ASIC (applicatio specificatio itegrate circuit). REFERENCES Acto, S.T. ad D.P. Mukherjee, Scale space classificatio usig area morphology. IEEE T. Image Proc., 9(4): Beediktsso, J.A., M. Pesaresi ad K. Amaso, Classificatio ad feature extractio for remote sesig images from urba areas based o morphological trasformatios. IEEE T. Geosci. Rem. Ses., 41(9): Che, W.l., J.E. Meg ad S.Q. Wu, Illumiatio compesatio ad ormalizatio for robust face recogitio usig discrete cosie trasform i logarithm domai. IEEE T. Syst. Ma Cy. B, 36(2): Gozalez, R.C. ad R.E. Woods, Digital Image Processig. Addiso-Wesley Publishig Co., Readig, MA. Haralick, R.M., S.R. Sterberg ad X.H. Zhuag, Image aalysis usig mathematical morphology. IEEE T. Patt. Aal., 9(14): Hu, H.Q., W.H. Xu ad Q. Huag, A 2D barcode extractio method based o texture directio aalysis. Proceedig of the 5th Iteratioal Coferece o Image ad Graphics, pp: Jimeez-Sachez, A.R., J.D. Mediola-Satibaez, I.R. Terol-Villalobos, G. Herrera-Ruiz, J.J. Garciaescalate ad A. Lara-Guevara, Morphological backgroud detectio ad ehacemet of images with poor lightig. IEEE T. Image Proc., 18(3): Mukhopadhyay, S. ad B. Chada, Fusio of 2D gray-scale images usig multiscale morphology. IEEE T. Patter Recog., 34(10): Mukhopadhyay, S. ad B. Chada, Multiscale morphological segmetatio of gray-scale images. IEEE T. Image Proc., 5(12): Park, G.H., H.H. Cho ad M.R. Choi, A cotrast ehacemet method usig dyamic rage separate histogram equalizatio. IEEE T. Cosum. Elec., 54(4): Sherrier, R.H. ad G.A. Johso, Regioally adaptive histogram equalizatio of the chest. IEEE T. Med. Imag, 1(6): 1-7. Soille, P., Morphological Image Aalysis: Priciples ad Applicatios. 2d Ed., Spriger- Verlag, Berli. Sha, S.G., G. We, B. Cao ad D.B. Zhao, Illumiatio ormalizatio for robust face recogitio agaist varyig lightig coditios. IEEE Iteratioal Workshop o Aalysis ad Modelig of Face ad Gestures, pp: Xu, G.B., Z.B. Su, J. Wag, Y.X. Yi, Y. L. She, A adaptive morphological filter based o multiple structure ad multi-scale elemets. Proceedig Coferece o Itelliget Iformatio Techology Applicatio, pp: