Today s Image Capturing Needs: Going beyond Color Management

Size: px

Start display at page:

Download "Today s Image Capturing Needs: Going beyond Color Management"

Chastity Fletcher
5 years ago
Views:

1 Today s Image Capturig Needs: Goig beyod Color Maagemet Chris Tuij ad Wim Cliquet Agfa-Gevaert N.V., GS/EPS/R&D B-264 Mortsel, Belgium Abstract Oe of the mai cocers i both desktop ad pre-press eviromets is reliable color reproductio. This problem is addressed by the color maagemet systems which are aimig at the productio of so-called facsimile color. I order to use color maagemet systems, oe should kow very well what the color space of the digital represetatio of the source image is. If this kowledge is ot available, the CMS work-flows caot be followed ad more itelliget adaptive color correctio techiques are required. Eve if the source of the images ad the scaig equipmet is well-kow, people ofte wat to reproduce their origials better. I order to produce more appealig images, so-called color editig is required. This kid of editig icludes rage adjustmets, toal adjustmets, saturatio ehacemet, global ad selective color trasformatios etc. I order to icrease productivity, these color correctios should be carried out automatically. The mai goal of automatic color correctio techiques thus cosists of brigig the origial images (the source of which might ot be kow ito a well-kow calibrated RGB space such that the reproductio of the images is appealig to the viewer. I order to achieve these goals, the images have to be aalyzed ad referece poits have to be detected. This paper is orgaized as follows. I the first sectio, we will itroduce a geeral purpose model for automatic image correctio. The geeral techiques exposed i that sectio will be illustrated by several case studies i the followig sectios. I the first case study, we will itroduce a automatic toal correctio which has bee used i the ewspaper busiess for black ad white images. The secod case study will briefly describe adaptive techiques which have bee used i order to covert egatives to a well-calibrated positive RGB space. The complexity of this techique is relatively low sice it oly ivolves a global color correctio through the idicatio of a eutral poit (which is equivalet to the specificatio of a global cast. I the third case study, we will deal with the geeral problem of automatic image correctio of color images from ukow sources. I the last sectio, we will summarize the obtaied results ad idicate topics for future research. A Geeric Approach to Automatic Image Correctio First, we will itroduce a formal model that defies a geeral framework for automatic image correctio. Both spatial image correctios (such as usharp maskig, de-screeig, oise removal etc. as well as color correctios (such as toal correctios, selective color correctios or a combiatio (such as the correctio of colored patters ca be cosidered. Basically, ay automatic image correctio scheme ca be split up i 4 steps. The first step is the most difficult oe ad deals with the geeral problem aalysis. Followig steps deal with image correctio as such. Step : Problem aalysis by studyig a test set of images First, the problem eeds to be defied. Ofte, this is doe based o a umber of images (I :..K to be corrected ad the maually corrected images (I :..K. A classificatio of the test set ca be realized by studyig both the origial set ad the corrected set. This classificatio ca be formalized by a set of parameters (α :..L which ca be calculated for each source image. The parameter set ca be used to derive a set of M correctios (Γ :..M such that for each image I i the test set : M ' o Γi (I I i= The idea, of course, is that the compositio of the trasformatios should produce good results o arbitrary images as well. The problem is to fid a umber of characterizig parameters which cotai eough iformatio to classify the images ad to geerate good trasformatios. There is o geeral method to describe how to defie those parameters. There is a dager i both usig too may parameters as well as i usig ot eough of them. The extreme cases are either usig the empty set of parameters or takig all iput pixels ito cosideratio. Clearly oe of these approaches makes much sese. If oly color (o-spatial correctios are to be applied, usig a dowsampled versio of a iput image might tur out to be useful. Other characterizig parameters are f.i.: oe-dimesioal histograms multi-dimesioal histograms frequecy aalysis (usig either classical FFT techiques, widowed FFT techiques or multi-resolutio aalysis filtered versios of a complete image or smaller regios etc. Also parameters that are special to the problems to be addressed ca be used. We hereby thik, e.g., of parameters such as the hue of the lightest ad darkest poit, average colors of regios of iterest showig ski, sow, grass, sky, etc. Chapter I Color Maagemet 9

2 The type of the correctios to be applied i order to geerate the output images ca be determied either automatically from the corrected test set or by discussig the applied correctios with the skilled scaer operator. I the past, we have bee doig experimets to determie the parameters of well-defied parameterized trasformatios automatically usig eural etworks. The learig set cosists of K elemets (oe for each image mappig the aalysis parameters ( :..L to the parameterized represetatio of the trasforms (Γ :..M. α I I... α ( I ( L Γ ( I I '... Γ M ( I Although reasoable results were obtaied at that time, ofte better results ca be obtaied usig classical methods from umerical aalysis based o liear ad o-liear regressio techiques. 4,5 The followig steps (2, 3 ad 4 describe how a particular image will be trasformed. Step 2: detectio of well-kow parameters Now, we eed to calculate the characterizig set of parameters as described above. I this step, we basically determie which class(es the give image belogs to. Ofte, heuristic techiques (based o fuzzy logic are applied; rulebased systems or Prolog-based search egies ca be used i this cotext as well. Step 3: geeratig color trasforms based o these parameters This step calculates the image trasforms based o the characterizig set of parameters as retured i Step 2. Now, the image correctio ca be applied to the source image. Step 4: learig mode The user is able to request small adjustmets to the (automatically proposed image trasforms. The modified parameters are the fed back to the algorithm ad used later o. We advise a additioal optio to eable/disable the learig mode. We ow preset a few case studies that illustrate our approach. Case Study : Toal Correctio of Black/White Images for the Newspaper Busiess I cotrast with the traditioal offset presses, the ewspaper pritig has specific costraits ad, therefore, special gamut mappig techiques are required i order to produce good reproductios. I particular, the total ik limit costraits are very importat ad should be used as a drivig force behid the gamut mappig. Although the solutios to this problem are far from trivial, they already have bee studied i the past. Ofte, skilled scaer operators apply subtle toal chages o top of the color maagemet trasforms described above. These toal chages appear to deped heavily o the type of the origial. I order to get a better feelig of what kid of correctios usually are applied, we asked a experieced scaer operator to collect a umber of represetative grey-scale images ad to correct those images. This experimet resulted i a database of images ad their corrected versios. It tured out that the correctios could be reduced to a global remappig of the itesity values. Most ofte, a small cotrast ehacemet or reductio was applied together with some correctios i the highlights or shadows. More sophisticated spatial filters to create local cotrast chages (such as dodgig were ot cosidered. The test set that we got back from our skilled scaer operator cosists of 3 images some of which: are toally good; are overexposed; are uderexposed; have flat shadow parts; have a high cotrast i the shadows; have a low cotrast i the highlights; have a too high cotrast i the highlights; have flat mid-toes; have too much cotrast i the mid-toes; etc. All of the images i the test set have oe or more of the characterics listed above. Most of the items i the list ca be derived by lookig at the histogram or cumulative distributio fuctio (CDF. By plottig all CDF s i a diagram, we get a fairly good idea of what types of origials we ca expect. Most of the CDF s are positioed i a kid of hysteresis shape. I Figure, we show a few CDF curves ad how they ca be iterpreted. The horizotal axis represets the dot percetages; the vertical axis represets the accumulated frequecy percetages. The curves ca be iterpreted as follows: curve is the CDF of a image which will be perceived as overall too dark; curve 2 is the CDF of a image with o mid-toes; curve 3 is the CDF of a image with a uiform histogram; curve 4 is the CDF of a image which has ehaced mid-toes at the cost of little highlight ad shadow areas; ad, curve 5 is the CDF of a image which will be perceived as overall too light. Cumulated distributio Dot percetage Figure. CDF sample curves Curve Curve 2 Curve 3 Curve 4 Curve 5 2 Recet Progress i Color Maagemet ad Commuicatios

3 A straightforward method to improve the CDF would be to apply a so-called histogram equalizatio (as poited out i []. It turs out that, although the toally compressed areas come out much better after histogram equalizatio, the result ofte looks artificial ad is uacceptable. Therefore, we propose to use the test set created by the skilled operator. After careful ivestigatio of the test set, the test images ca be divided i 5 differet families: Family CO, cotaiig 5 images : uderexposed images without cotrast problems; Family CH, cotaiig images : overexposed images without cotrast problems; Family CN, cotaiig 5 images : images with ormal exposure; Family CL, cotaiig 9 images : images with cotrast problems (usually located ear the mid-toes; ad, Family CU, cotaiig image : images with extremely high cotrast ad lackig mid-toes. Case Study 2: Adaptive Color Correctios for Negative to Positive Coversio The problem of scaig egative film ad covertig the egative sigal ito a well-kow positive RGB space has bee addressed extesively i the past. 8,,,2 The mai problem cosists of calculatig appropriate iversio tables i order to covert from egative to positive; these iversio curves are based o the characteristic film curves of the egative film as perceived by the scaer as show i Figure 3 (see also [9]. It turs out, however, that the characteristic film curves do ot oly deped o the type of film, but ca vary from batch to batch ad are also heavily iflueced by the developmet of the egative film. O top of these problems, the circumstaces uder which the picture was take also ifluece the fial result (shutter speed, opeig, type of camera, light coditios, etc The families were derived by aalyzig the problem ad lookig at the proposed correctio. After averagig the curves proposed by the expert user, followig correctios could be derived for each family: 2.5 Red Gree Corrected dot percetage Dot percetage Figure 2. Corrrectio for the differet families FmCO FmCH FmCN FmCL FmCU We ow are ready to start with the actual correctio of ay give image. First of all, we determie the family to which the image belogs. This is doe by meas of a distace fuctio. Good results ca be obtaied by usig a sum of squares of differeces i a limited umber of poits. If F deotes the set of families f, CDF(f the cumulative distributio fuctio of f, Fm(f the proposed correctio as i Figure 2, ad a distace fuctio i a fuctio space as described previously, the the toal correctio TF to be applied o a image g ca be calculated as follows: TF(g = f F f F Fm( f (CDF( f,cdf(g (CDF( f,cdf(g Additioal spatial correctios ca be carried out to sharpe the images; the parameters for these USM filters ca also be determied automatically but are ot covered i this article Figure 3. Characteristic curves of Agfa HDC film as see by Agfa s DuoSca The above argumets suggest that it is almost impossible to take all these parameters ito accout. We ca, however, get a reasoable idea of the characteristic curves of a particular egative film by cosiderig a test wedge (made o the same type of film, captured by the same scaer ad use these curves as a basis to determie future correctios. I Figure 3, the characteristic film curves of a Agfa HDC film are show as measured by the Agfa DuoSca. 2 The X-axis cotais the exposig itesities o a logarithmic scale; the Y-axis cotais the desities of the developed egative wedge i DuoSca scaer space. I order to get rid of measurig errors, we use a parameterized curve (as described i [6]; the parameters of these curves are determied i a least square sese. 4,5 The adaptive correctios to these curves eed to be obtaied by aalyzig the actual egative film strip ad the specific picture o this strip which we wat to sca ad covert. The correctios we have i mid here are defied as correctios o the characteristic film curves ad are thus iheretly global. Local or so-called selective color correctios will ot be cosidered but better results ca be obtaied by applyig additioal multi-dimesioal color correctios o the iverted image to produce more reliable color (see [,2]. I the traditioal photo-fiishig labs, however, the oly parameters that ca be iflueced to geerate the prit o photo paper are the exposure parameters Blue Chapter I Color Maagemet 2

4 of the light-source i red, gree ad blue (which are, of course, global correctios. These global correctio parameters ca be specified as red, gree ad blue desities i the scaer iput space. The meaig of this desity triplet is that a desity poit of the egative film should be mapped to a eutral poit of a give itesity i the iverted (positive space. As such, the desity triplet iflueces both toal ad color balace. The color balace correctio is based o a statistical aalysis of our film strip ad ivolves the calculatio of a histogram of the ear-eutrals. This iformatio is combied with statistical iformatio o the average characteristic film curves of a represetative set of various films (of differet vedors. All these data are weighted ad result i the so-called TFS curves 3 which defie the adapted characteristic film curves for this specific film. Now the frame to be coverted is aalyzed further i order to get rid of local color casts due to, e.g., the illumiatio or other evirometal parameters (such as flashlight. The fial correctio is obtaied as a weighted average of the geeral mea desities, the desities of the eareutrals ad the TFS curve. The last parameter to be determied is related to the dyamic rage ad idicates a mid-toe eutral. Parameters which are take ito accout to determie this midpoit are: the miimum desity of the frame; the maximum desities of cocetric zoes to fid out where the mai object of the frame resides; the desities of particular areas of iterest showig ski, white areas (skiig, etc.; correctios based o the classificatio of scees (portrait, ladscape,... the orietatio of the frame;... The algorithm above thus results i three desities which are used to calculate the iversio curves. Basically, these desities will traslate the characteristic curves horizotally such that said desities (o Y axis origiate from a triplet with equal values, i.e., eutral light (o the X axis. The dyamic rage of the selected frame (exposure latitude is also take ito accout whe calculatig the iversio curves. For further detail, we refer to [8]. The user has the optio to specify small chages to the proposed parameters. I Agfa s FotoLook 3. 4 scaig software, this ca be doe by pickig a color cast i a hue/ saturatio color-wheel. The proposed chages ca be stored i a database ad be used later o to ifluece the system s behavior. Summarizig, we coclude that we basically follow the same work-flow as described i the first sectio: Step : geeral problem aalysis; Step 2: aalysis of the origial (the film strip ad oe specific frame o this strip, resultig i a histogram of ear eutrals ad may other parameters resultig i three characterizig parameters; Step 3: calculatio of the color correctio based o the three parameters; ad, Step 4: small correctios to characterizig parameters suggested by user ad learig facility. Case Study 3: Automatic Correctio of Color Images from a Ukow Source The geeral problem of automatic image correctio of images from ukow sources is very complex. The complexity is ot oly caused by the iheret, techical problems related to the recogitio of objects, patters, etc., but also stems from the fact that the proposed correctios will always be very subjective. I this sectio, we will ot cover our solutio to automatic image correctio i detail but we will rather outlie some ideas we have bee pursuig ad which seem of fudametal importace. I particular, we will summarize the correctios we have bee usig for image improvemet i geeral ad some of the idicators that allowed the determiatio of the parameters for the trasformatios. Global color correctios: Global correctios ca be defied as color correctios which are applied to oe or more color compoets idepedetly of each other. They ca be implemeted usig oe-dimesioal toal LUT s (look-up tables. The most importat global color correctios are: dyamic rage adjustmets: ofte, the digital image is ot usig its full rage of values. This suggests that the dyamic rage should be stretched to the allowed values. This should be carried out carefully, however, based o a classificatio the origial. Otherwise, very saturated colors or pastel tits might be ruied. toal correctios: a more geeral toal correctio is eeded if the origial appears to be overall too light or too dark. Oe way to determie this is to study the CDF as discussed i Case. This method ca be improved by also takig ito cosideratio the spatial activity i certai areas, the mai idea beig that a spatially active area ofte eeds toal ehacemet. color cast removal: for color cast removal, aalogous techiques to Case 2 ca be used. Local Color Correctios: Sometimes, particular areas i the color space require further ehacemet. I order to determie such color correctios (which are also kow as selective color correctios, we eed to detect referece objects i the origial. This detectio will be based o both spatial ad color cotet. Areas which are cosidered to be very importat are, e.g., areas cotaiig ski, sky, meadows etc. The correctio the cosists of mappig the color of the detected area to a cofigurable color usig a selective color trasformatio. The learig facilities (cf. Step 4 above cosist of both ifluecig the detectio of the special categories as well as the determiatio of the target colors. Spatial Correctios: O top of color correctios, spatial correctios might be eeded for further ehacemet. We hereby thik of: sharpeig; oise removal ad removal of other artifacts (such as artifacts from lossy compressio schemes; scratch removal; dodgig; Recet Progress i Color Maagemet ad Commuicatios

5 This type of correctio requires a study of the spatial characteristics of the image by meas of Fourier or wavelet aalysis. 3 To coclude, I would like to emphasize that the obtaied results must be iterpreted i a stadard color space that has bee determied up frot. The classifyig parameters as well as the proposed correctios deped o the choice of this exchage space. Cadidates for such a space are moitor spaces havig a specific gamma value. The chromaticities of of this space should be wide eough to spa a reasoable gamut. Preferably, it should be specified as a ICC profile (cf. [2]. Oce the images are i this exchage space, stadard CMS techiques ca be used to trasform the images from this space to ay other space. Coclusios I this documet, we described a global approach to automatic image correctio. Although substatial results have bee reached so far, it is obvious that the geeral problem of correctig colored images automatically will ever have a completely satisfyig solutio ad therefore will eed ogoig attetio ad improvemets. The mai techique which has bee itroduced here is essetially based o a statistical aalysis of a test set of images usig a umber of parameters which describe importat features allowig to make some kid of classificatio. After the aalysis of the test set, a relatioship is established betwee the classifyig parameters ad the correctios proposed by a experieced user. I operatioal mode, the algorithm will calculate the image specific parameters ad come up with a correctio as proposed by the expert. I additio, the ed-user will be able to apply mior chages to the proposed correctios; the system will lear about the subjective chages proposed by a particular user ad will correct its future behavior accordigly. The iferece egie which establishes a relatioship betwee the iput parameters ad image correctio is essetially a rule-based system. It might be worthwhile to study the usability of other formalisms to establish this relatioship such as, e.g, eural etworks or Prolog-based egies. The quality of the obtaied results is heavily depedet o the classifyig parameters. Much of the future work will be cocetrated o tryig to come up with ew classifyig parameters ad more sophisticated correctios. For spatial correctios, we ca make use of wavelets ad the associated multi-resolutio aalysis; for color correctios, further aalysis of 3-D histograms ca give better isight i what type of origial scee we are dealig with. A eve more advaced solutio might be based o a 5-dimesioal spatial/color aalysis; this is, e.g., ecessary to deal with colored patters of cloth. Refereces. Gozalez R.C. ad Witz P., Digital Image Processig, 2d Editio, Addiso-Wesley Publishig Compay, Readig, MA, Iteratioal Color Cosortium, ICC Profile Format Specificatio, Versio 3.3, November Kaiser, G., A Friedly Guide to Wavelets, Birkhäuser, Bosto, Lawso C.L. ad Haso R., Solvig Least Squares Problems, Pretice Hall, Eglewood Cliffs, NJ, Marquardt, D.W., Joural of the Society for Idustrial ad Applied Mathematics, vol., pp , Pytela, O., ad Majer, J., J. of Imagig Sciece, vol. 35.6, pp , Quatitative Iterpretatio of Sesitometric Curves of Photographic Materials, Stroebel, L., Compto, J., Curret, I. ad Zakia, R., Photographic Materials ad Processes, Focal Press, Bosto, Tuij C., Iput Calibratio for Negative Origials, Proceedigs of IS&T/SPIE s Symposium o Electroic Imagig: Sciece ad Techology, Device-Idepedet Color Imagig II, Volume 244, Sa Jose, CA, 995, pp Tuij, C., Hardware ad Firmware Requiremets for the Ideal Negative Scaer, Agfa-Gevaert N.V., Iteral Memo, Mortsel, 996. Tuij, C., The Ifluece of Precompesatio Curves o Multidimesioal Color Modelig, Proceedigs of IS&T/ SPIE s Symposium o Electroic Imagig: Sciece ad Techology, Color Imagig: Device-Idepedet Color, Color Hard Copy, ad Graphic Arts, Volume 2658, Sa Jose, CA, 996, pp Tuij, C., A Adaptive Approach to Negative Scaig, Proceedigs of SPIE/Europto s Coferece o Imagig Scieces ad Display Techologies, Berli, FRG, 996, pp Tuij, C., Scaig Color Negatives, Proceedigs of IS&T/ SID Fourth Color Imagig Coferece: Color Sciece, Systems ad Applicatios, Scottsdale, AZ, 996, pp Agfa HDC is a registered trademark of Agfa-Gevaert N.V. 2 Agfa DuoSca is a registered trademark of Agfa-Gevaert N.V. 3 TFS stads for Total Film Scaig ad refers to the fact that, i order to determie the correctio parameters for a particular frame of the egative film roll, the etire strip (i.e., the total film is scaed first. TFS is a registered trademark of Agfa-Gevaert N.V. 4 Agfa FotoLook 3. is a registered trademark of Agfa-Gevaert N.V. This paper was previously published i IS&T/SID 5th Color Imagig Coferece Proc., p. 23 (997. Chapter I Color Maagemet 23

APPLICATION NOTE UNDERSTANDING EFFECTIVE BITS

APPLICATION NOTE UNDERSTANDING EFFECTIVE BITS APPLICATION NOTE AN95091 INTRODUCTION UNDERSTANDING EFFECTIVE BITS Toy Girard, Sigatec, Desig ad Applicatios Egieer Oe criteria ofte used to evaluate a Aalog to Digital Coverter (ADC) or data acquisitio