Visual quality of printed surfaces: Study of homogeneity

Transcription

1 Visual qualiy of prined surfaces: Sudy of homogeneiy David Nébouy, Mahieu Héber, Thierry Fournel, Jean-Luc Lesur To cie his version: David Nébouy, Mahieu Héber, Thierry Fournel, Jean-Luc Lesur. Visual qualiy of prined surfaces: Sudy of homogeneiy. Proceedings of SPIE volume 906, Image Qualiy and Sysem Performance XI, 04, pp <hal > HAL Id: hal hps://hal.archives-ouveres.fr/hal Submied on 5 Mar 04 HAL is a muli-disciplinary open access archive for he deposi and disseminaion of scienific research documens, wheher hey are published or no. The documens may come from eaching and research insiuions in France or road, or from public or privae research ceners. L archive ouvere pluridisciplinaire HAL, es desinée au dépô e à la diffusion de documens scienifiques de niveau recherche, publiés ou non, émanan des élissemens d enseignemen e de recherche français ou érangers, des loraoires publics ou privés.

2 SPIE Elecronic Imaging 04 - Image Qualiy and Sysem Performance XI San Francisco - -6 February 04. Visual qualiy of prined surfaces: Sudy of homogeneiy D. Nébouy a,b, M. Héber a, T. Fournel a, J.-L. Lesur b a Universié de Lyon, Universié Jean Monne de Sain Eienne, CNRS UMR 556, Loraoire Huber Curien, F Sain Eienne, FRANCE; b Gemalo SA Avenue du pic de Beragne, ZA de Gémenos, BP.00 F-388 Gémenos Cedex, France. ABSTRACT This paper inroduces a homogeneiy assessmen mehod for he prined versions of uniform color images. This parameer has been specifically seleced as one of he relevan aribues of prining qualiy. The mehod relies on image processing algorihms from a scanned image of he prined surface, especially he compuaion of gray level cooccurrence marices and of objecive homogeneiy aribue inspired of Haralick's parameers. The viewing disance is also aken ino accoun when compuing he homogeneiy index. Resizing and filering of he scanned image are performed in order o keep he level of deails visible by a sandard human observer a shor and long disances. The combinaion of he obained homogeneiy scores on boh high and low resoluion images provides a homogeneiy index, which can be compued for any prined version of a uniform digial image. We esed he mehod on several hardcopies of a same image, and compared he scores o he empirical evaluaions carried ou by non-exper observers who were asked o sor he samples and o place hem on a meric scale. Our experimens show a good maching beween he soring by he observers and he score compued by our algorihm. Keywords: Prin qualiy assessmen, Halfoning, Qualiy index, Qualiy aribues, Uniformiy, Gray-level co-occurrence marix, Human Visual Sysem, Appearance, Scanner, Prining echnologies. INTRODUCTION Prining has evolved significanly he las few years, due o firs a diversificaion of processes, and also o he digial revoluion [] including an auomaion of he processes wihou he experise of he prinman a he prining ime. Wih his rapid evoluion, he qualiy of color images reproduced on hard suppors has considerly increased, and he amazing diversificaion of hese echniques permied by he new digial echnologies provides a large panel of soluions for image reproducion. This leads o an issue for comparing he differen echniques in erms of qualiy, an inuiive concep which is generally well apprecile visually bu difficul o assess wih objecive, measurle values. Scienific concern for prining qualiy is recen: In [], he prining qualiy is explored by defining various qualiy aribues: color, lighness, sharpness, conras, physical, arifacs. These aribues were esimaed from psycho-visual experimens carried ou by a panel of 5 observers wih various experise. In conras wih Pedersen e al. [3], [4], we believe ha an imporan disincion should be made beween prining qualiy (relaed o he degradaion of he original digial image when ransferred o he paper or plasic suppor) and image qualiy (including prining qualiy plus he perceived qualiy of he original image). Our opinion is ha, conras, lighness, and color raher qualify he image qualiy. The presen work focuses on he homogeneiy of ones i.e. he homogeneiy of any surfaces on which is prined an originally uniform digial color image. This aribue is considered as relevan in he sandard ISO 3660: 00, he only inernaional sandard describing a wide range of aribues of image qualiy dedicaed o qualiy assessmen for binary or monochrome prining sysems. [5]. In his sandard, moling is defined as a ype of homogeneiy defaul of paper-like surfaces and an index is suggesed in he spaial domain. A pass-band mehod was proposed [6] as an alernaive le o beer ake ino accoun he sensiiviy of he human visual sysem. A pass-band assessmen funded on waveles, followed by a second-order saisics calculaion on he gray-level co-occurrence marix was achieved in [7] o form an homogeneiy index. Our aim here is o inegrae human vision daa ino boh he filering sep and he second-order saisics calculaion. As in [6], our approach firs consiss in he digializaion of he prined surface using a high resoluion flabed office scanner, he scanned image being hen processed in order o compue an homogeneiy index. I hen consiss in comparing his score for many samples (several versions of a same digial image prined on differen suppors wih differen prining echniques) wih he empirical evaluaion by observers of hese samples on a scale.

3 In he nex secions, we firs explain how he surfaces are digialized and which ool of image processing we use before developing in deails our mehod and is unle parameers. We finally compare he score provided by he mehod wih empirical assessmen by observers on four ses of samples.. DIGITALIZATION OF THE PRINTS For he digializaion of he samples, a scanner is used. An iniial sudy on scanners have led o he conclusion ha i is a relile ool for image acquisiion due o is remarqule reproduciliy, mainly performed by a uniform illuminaion, and good geomery ha avoids specular reflecions when he surface is fla. However, spaial and color disorions induced by he acquisiion sysems should be aken ino accoun when assessing qualiy. The color gamu of our human visual sysem is much larger han ha of he scanner, which is unknown bu can be assumed o be quie close o srgb. Because of his, he color disorion problem is wofold: Firs, he so-called "seen by he scanner" colors are displaced wihin he very gamu of he scanner, bu also, i may appears ha differen prinle colors are ou of he gamu of he scanner and hus perceived by he scanner as having he same color. A color calibraion would seem necessary o solve a leas he firs issue. However, we noiced ha very ofen afer color correcion, images are noisier han before correcion. As par of our work (he assessmen of uniformiy), we chose no o apply color correcion o avoid he risk of disruping our measure. Regarding he spaial disorions, here is a significan horizonal drif. In he verical direcion, we do no noice geomeric drif, he moor is properly posiion conrolled. This drif can be correced by adjusing he posiion of he pixels by polynomial inerpolaion or spline on each line of he capured image. Bu such reamen would creae a change in he value of cerain pixels on each line, he mos affeced by his drif being on he edges of each line, and would lead o a loss of he naive resoluion of he image as well as local color unwaned disorions. I is preferle o analyze he scanner profile derives and locae he sample on he scanner glass in an area where he drifs are minimal. We recommend scanning samples a a high resoluion (e.g. 400 dpi), and hen reducing o 00 dpi, he iniial resoluion of our mehod, in order o perform conrolled down-sampling. 3. BACKGROUND: IMAGE HOMOGENEITY ANALYSIS The prined copy of a uniform color pach can be considered as a exure. In image processing, a classical ool in exure characerizaion relies on Haralick's parameers derived from he gray-level co-occurrence marix (GLCM) also called gray-one spaial-dependence marix, firs presened in [8]. In conras wih firs-order saisic ools (hisogram, mean, sandard deviaion, skewness, and kurosis), he GLCM is a saisic ool of nd order [9]. Firs order saisics are represenaive of gray-level disribuion of he pixels in he image regardless of heir spaial arrangemen, whereas second-order saisics involve wo pixels simulaneously. The consrucion of he GLCM wih H is defined as: h i j i 0.. b, j 0.. H (, ) n n () b h ( i, j) Card p, p I ; f( p) i and f( p ) j () n where I is a recangular image encoded on n b bis, f is an applicaion from I I 0.. b, Card means he cardinaliy of a se, is a given ranslaion vecor, and p, p is a pair of pixels in image I. The vecor may be specified by a uni vecor w and a disance D : Saring from an iniializaion o he null marix, p, p in image I, and incremening by one he enry D. w (3) H is incremenally buil by considering in urn every pair of pixels h f( p), f( p ) (see Figure ).

4 Afer normalizaion of his marix, each enry h i, j means he probiliy o ransi, along a vecor, from a pixel wih value i o a pixel wih value j : h i, j p i, j Card I T I (4) wih Tx means he ranslaion operaor of a vecor x. y yp+ yp x xp D j i Figure. Scanning of he image pixels for every poin x p, y p D D in order o build he gray level co-occurrence marix. D (a) (b) (c) 56 Figure. (a) Three original gray level images. (b) Gray level represenaion of he corresponding GLCM where he whie color means he max value. (c) Zoom on he cener of images (b). On such GLCM, Haralick defines several exure aribues among which an homogeneiy aribue comprised beween 0 and, defined as: Hom p i, j j 0 (i j ) nb nb i 0 (5)

5 If he GLCM H of image I is compued on a uniform image (op lef image on Figure ) or on a periodical image wih period (op middle image), hen for every pair of pixels f p, f p, we have f p f p. Therefore, only enries in he diagonal of H are incremened; and Hom. In he oher cases (for example for he op righ image of Figure ), H is no a diagonal marix, and 0 Hom. Formula (5) may be replaced wih oher ones, defined so as o give differen weigh o enries more disan from he diagonal. For example, we may generalize Eq. (5) by defining he following funcion weighing he ransiion probiliy p i, j : where a and b are unle real values (see Figure 3), and x i j. w, : x b a x (6) Using w, wih differen a and b values han he ones used in Eq. (5), we can une he response of he homogeneiy score o variaions of he pixel values in he image. w a,b (x) w funcion for hree differen ses of Hom: (a, b) = (, ) Hom: (a, b) = (0., 4) Hom: (a, b) = (0.6, 4) Hom3: (a, b) = (0.0, 4) x Figure 3.,, values. Original Haralick's aribue (denoed as Hom) corresponds o he w, funcion ploed in red line. w 0.,4 and w 0.6,4 funcions, respecively denoed as Hom and Hom, are ploed in blue, respecively green lines. 4. OVERVIEW OF THE METHOD The Homogeneiy assessmen mehod ha we propose relies on he following seps: Preprocessing, including image resizing and filering in order o simulae near vision a High Resoluion (), and far vision a Low Resoluion (), Applicaion of he Image Homogeneiy Analysis mehod described in Secion 3. Score compuaion from he GLCM on boh and images. 5. PREPROCESSING This secion inends o modify he raw scanned image ino a new one, closer o human percepion. A hree sep preprocessing of he scanned images is performed: Creaion of cusom color channels from he iniial RGB channels; filering in order o ake accoun of human visual resoluion; quanizaion (opional).

6 For he image, a 00 pixel per inch (ppi) image is recommended. A bicubic inerpolaion (scale of / if he inpu image is 400 ppi) wih anialiasing mehod is seleced o perform image resizing. The firs sep is o change he color space. Scanner yields R,G,B values for each pixel. We could process hese R,G, and B channels each ones as a grayscale image, bu hey do no conain he opimal informaion. In order o be more consisen wih human percepion, we prefer convering he RGB values ino CIELAB values by considering ha he inpu RGB values are represened in he Adobe RGB (998) color space, hen convering ino CIE93XYZ for a sandard observer, and finally convering in CIELAB by selecing he D65 illuminan. We analyze he L * (lighness) channel, and he h (hue) channel. CIELAB color space is chosen because i is a color appearance model [0] (pages 60-6). The defined as * a and * b Caresian coordinaes are ransformed ino cylindrical coordinaes * * * * C and h, respecively C a b (7) h * b * an a (8) Implemenaion remark: aan y x funcion yields an angle in, funcion is availle in mos compuaion sofwares, ofen denoed as aan, for every y and x. which is no defined if x 0. Alernaive y x, yielding an angle in,, defined Noe ha i is also possible o define he luminance image direcly from RGB values, using he NTSC luminance (Y) formula, firs presened in Ref. []: Y 0.99R0.587G0.4 B (9) The second sep corresponds o he simulaion of he blur produced by our human eyes. According o Hermann von Helmholz's observaions [], we consider he definiion of he "normal visual acuiy" as he iliy o resolve a spaial paern separaed by a visual angle α of one minue of arc. Thus, giving an observaion disance d (in cm) and a scan resoluion of he image r s (in ppi), we can define he convoluion kernel of he blurring filer (Figure 4). The size s of he kernel depends on he sandard deviaion used for he Gaussian profile: r s d an (0) s 6 () We choose o simulae he images as viewed a a shor disance ( d 5 cm), near he human puncum proximum, which corresponds o.06 pixels, i.e. s 3. d Figure 4. Definiion of knowing he visual angular resoluion ' degree a he viewing disance d. 60 Lasly, a quanizaion sep (e.g. on nb 8 bis for each channel), can evenually be performed, wih he consequence o decrease he size of he GLCM bu also o lose he leas significan bis (i.e. small grey-level variaions in he image). For he low resoluion images, we perform he same 3 seps wih a 40 ppi resoluion image (scale of /0 if he inpu is 400 ppi), wih.03 pixels and 7, which corresponds o he image viewed by an observer a ou 40 cm. The preprocessing sep leads o 4 images, denoed as Y, Y, h, h. α σ

7 6. APPLICATION OF THE IMAGE HOMOGENEITY ANALYSIS METHOD We apply he mehod presened in Secion 3 on he four images sep. Y, Y, h, h obained in he preprocessing For each of he wo images, we choose he following ranslaion vecors and o compue respecively GLCMs M and M : wih D u () D u (3) D 5 pixels (corresponding o a physical displacemen of ou 0.5 mm on 00 ppi image), and u y, u x he horizonal, respecively verical uni vecors (see Figure ). We hen compue a GLCM M as he average of M and M : M M M (4) We also compue a chroma value on he image, denoed as c and defined as he average of he chroma values compued from Eq. (7) on each pixel of he image. From marix M, we compue he homogeneiy aribue Hom according o Eq. (6) wih a value of a denoed a depending on c and given in Figure 5.a, and b 4. For each of he wo images, we perform similar processing as ove wih D 35 pixels (corresponding o a physical displacemen of ou 0.37 cm on 40 ppi images), and compue he Hom aribue using he w,4 funcion defined by Eq. (6), where a is given by Figure 5.b as a funcion of c compued as c on he chroma values of he image. These values for D, a, and b, have been chosen because hey seem he mos relevan in our sudy according o physical and empirical consideraions. a a () c () (a) a () c () (b) Figure 5. Variaion of he images. a and a values as a funcion of he chroma values c, respecively c, for and

8 7. CONSTRUCTION OF THE OBJECTIVE SCORE The compuaion of homogeneiy aribue on he GLCM of Y, Y, h, h yields four values o be combined in order o obain he final homogeneiy score. Le's denoe hese four values as H, H, H, H. Our experience shows ha relevan homogeneiy score in respec o human percepion is given by he following empirical formula: Y h Y h H c H H c H min, c c Y Y h h (5) Noice ha he chroma is aken ino accoun in his formula hrough he parameers c and c compued in Secion 6 because we noiced ha he influence of he homogeneiy parameer on he hue image is lower when he colors are no sauraed. The weigh aribued o he luminance channel is higher han he one aribued o he hue channel ( 0 c, and in pracice, on prins, i is raher 0c 0.6) as i is commonly done in video compression for digial elevision video encoding where he luminance channel has a double bandwidh han he chromaic ones (See for example [3]). The referenced parameers of our mehod are: 8. REFERENCED PARAMETERS OF THE METHOD Quanizaion (images encoded on n b number of bis): The less is quanizaion, he more careful deails are kep. I seems beer o keep all he deails. Insead of decreasing n b below he usual value of 8 bis, we prefer adjusing he a and b parameers in formula (6). Size of blurring kernel: i is se by he considered viewing disance d, which is ou 5 cm for near vision, and 40 cm for far vision. Scan resoluion: The size of one pixel in images is ou μm afer downsampling original 400 ppi scanned image o 00 ppi. This appears sufficienly accurae in comparison o he human vision and o he fines commercial prining sysems availle oday. D and D : They mus be represenaive of he characerisic size of he inhomogeneiies ha we wish o highligh, in respec o near sigh and far sigh observaions. and : horizonal and verical vecors used o compue he GLCM. These wo direcions have been seleced because hey coincide o he orienaion of classical defecs of mos prining sysems, because defauls ofen occur in he prining direcion (verical) and in he perpendicular direcion of i. 9. EXPERIMENTAL TESTING AND COMPARISON WITH PSYCHO-VISUAL ASSESSMENT In order o verify he relevance of he proposed score in respec o human observer evaluaion, we carried ou he following experimen: Uniform color paches have been prined wih various priners (laser, inkje, reransfer) and on various suppors (office paper, glossy phoograph paper, APCO paper, whie polymer). For he digializaion sep, all samples are scanned a 400 ppi in TIFF forma, 4 bis, by he EPSON Scan Ver. 3.8 FR, 00 driver, wih a EPSON Perfecion V700 PHOTO scanner. The No color correcion opion is seleced. An overview of some samples is presened below.

9 Figure 6. An overview of some scanned samples (in lower resoluion han he original ones). s_0 is prined on glossy phoograph paper, s_0 and s_08 on office paper, s_0 and s_ on APCO paper. The assessors were submied o a ranking es, using an ordinale scale, unsrucured, as i is presened by François Sauvageo in [4]. The assessors were presened randomly he samples o sor in respec o heir perceived homogeneiy, and were asked o place hem on a one meer scale. The posiions of he samples (in meer) are indicaed by he dos in Figure 7, where one do shape is aribued o each observer. Noe ha no rescaling of he posiions has been performed: he dispersion of he posiions for one sample parly comes from he differen scaling adoped by he observers in heir evaluaion. The coninuous line in Figure 7 represens he score compued wih our mehod, represened by formula (5). We see ha he scores follows fairly well he posiioning of he samples by he observers on he meric scale.,00 Homogeneiy ranking,00 Homogeneiy ranking 0,90 0,90 0,80 0,80 Grading 0,70 0,60 0,50 0,40 0,30 0,0 0,0 Rank Assessor Rank Assessor Rank Assessor 3 Rank Assessor 4 Rank Assessor 5 Rank Assessor 6 Rank Assessor 7 Our objecve mehod Grading 0,70 0,60 0,50 0,40 0,30 0,0 0,0 Rank Assessor Rank Assessor Rank Assessor 3 Rank Assessor 4 Rank Assessor 5 Rank Assessor 6 Rank Assessor 7 Rank Assessor 8 Our objecve mehod 0,00 s_0 s_06 s_9. s_07 s_0 s_5 s_ s_09 s_08 s_04 s_0 s_ Sample ID 0,00 s_06 s_0 s_07 s_9. s_09 s_5 s_0 s_08 s_ s_04 s_0 s_ Sample ID (a) (b),00 Homogeneiy ranking,00 Homogeneiy ranking 0,90 0,90 0,80 0,80 0,70 0,70 Grading 0,60 0,50 0,40 0,30 Rank Assessor Rank Assessor 4 Rank Assessor 3 Rank Assessor 6 Rank Assessor 7 Grading 0,60 0,50 0,40 0,30 Rank Assessor Rank Assessor 4 Rank Assessor 3 Rank Assessor 6 Rank Assessor 7 Rank Assessor 8 0,0 Our objecve mehod 0,0 Our objecve mehod 0,0 0,0 0,00 s_06 s_07 s_0 s_9. s_08 s_09 s_5 s_0 s_04 s_ s_0 s_ Sample ID 0,00 s_0 s_30 s_06 s_07 s_. s_0 s_09 s_08 s_ s_0 s_04 s_ Sample ID (c) (d) Figure 7. Comparison beween objecive and subjecive evaluaion for (a): a blue se of samples, (b): an orange se of samples, (c): a red se of samples, and (d): a gray se of samples.

10 In order o assess he performance of our mehod, six differen classes are defined by selecing six hreshold values of he score (0.8, 0.7, 0.6, 0.5, 0.3, 0). Assessors scorings are also quanified ino his six grades classificaion. The maching of he class given by observers and by our mehod is presened in Figure 8. Figure 8. Percenage of samples classified in he same of six caegories by assessors (subjecive evaluaion) and by our scoring mehod. 0. CONCLUSION The index compued according o he mehod ha we propose for he assessmen of he homogeneiy of prined colors is in accordance wih he empirical assessmen by a small panel of non-exper observers. In his sense, if he rend is confirmed for a larger se of observers, he algorihm ha we have developed will appear consisen wih he human visual percepion. This homogeneiy index is a firs aribue which should be combined wih oher aribues in order o ge a global qualiy score for prined naural images. I may also help o compare he performance of differen prining sysems in erms of visual qualiy. REFERENCES [] H. Kipphan, [Handbook of Prin Media] Springer Ed., (00). [] M. Pedersen, N. Bonnier, J. Y. Hardeberg e al., Aribues of a New Image Qualiy Model for Color Prins, Color and Imaging Conference, 009(), (009). [3] M. Pedersen, and S. A. Amirshahi, Framework for he Evaluaion of Color Prins Using Image Qualiy Merics, Conference on Colour in Graphics, Imaging, and Vision, 00(), 75-8 (00). [4] M. Gong, and M. Pedersen, Spaial pooling for measuring color prining qualiy aribues, Journal of Visual Communicaion and Image Represenaion, 3(5), (0). [5] Sandard, [IEC 3660 Informaion echnology Office equipmen Measuremen of image qualiy aribues for hardcopy oupu Binary monochrome ex and graphic images -ISO] IEC, Genevas, Swizerland: Inernaional Organizaion for Sandardizaion and Inernaional Elecroechnical Commission.(00). [6] A. Sadovnikov, P. Salmela, L. Lensu e al., [Moling Assessmen of Solid Prined Areas and Is Correlaion o Perceived Uniformiy] Springer Berlin Heidelberg, 4 (005). [7] M. Dube, F. Mairesse, J.-P. Boisver e al., Wavele Analysis of Prin Mole, IEEE Transacions on Image Processing, (005). [8] R. M. Haralick, K. Shanmugam, and I. H. Dinsein, Texural Feaures for Image Classificaion, Sysems, Man and Cyberneics, IEEE Transacions on, SMC-3(6), 60-6 (973). [9] G. Srinivasan, and G. Shobha, Saisical exure analysis, proceedings of world academy of science, engg & ech, 36, (008). [0] G. Sharma, [Digial Color Imaging Handbook] CRC Press, (00). [] D. H. Prichard, U.S. Color Television Fundamenals: A Review, SMPTE Journal, 86(), (977).

11 [] H. v. Helmholz, and J. P. C. Souhall, Helmholz s Treaise on Physiological Opics. Vol. II: The Sensaions of Vision, rans. JPC Souhall.(Translaed from he Third German Ediion), The Opical Sociey of America, (94), pp [3] ITU, [Sudio encoding parameers of digial elevision for sandard 4:3 and wide-screen 6:9 aspec raios], (0). [4] F. Depled, [Evaluaion sensorielle Manuel méhodologique] Tec E Doc, (009), pp. 30-3, in French.