Home

Colour management and colour correction in

1. T T T T shadows X highlights f midtones shadows highlights midtones 0 54 128 158 255 nival intancity 0 54 128 156 255 pixel intensity nival intancity PASA nivriony Figure 5 5 Lightness weighting curves Bernstein polynomials left and Gaussian right equal to the lightness of the old colour and the result is converted back to RGB If the HLS model is used calculations are again based on Foley van Dam et al s algorithm If the L a b is used conversions are carried out through littleCMS and the colour management system previously implemented in this project If the user has specified a profile for the image this will be taken into consideration to achieve colorimetrically optimal results If no profile has been specified a standard RGB profile will be assumed I will elaborate further on the differences resulting from the use of HLS and L a b respectively in the evaluation in chapter 6 3 This only leaves the interpretation of the brightness ranges of shadows midtones and highlights to be specified They are implemented as simple weighting functions applied to the manipulation The weight ranges between 0 0 and 1 0 and is calculated in function of the original pixel intensity For shadows the weighting function returns a high weight for pixels of low brightness For midtones the weight peaks in the middle For highlights it increases with increasing brightness Two alternative weighting
2. header with header with meta data meta data ges T look up table LUT look up table LUT XYZ E nl L i em image profile display profile image data in RGB stored in file device link Figure 3 1 Colour management conversions using profiles Figure 3 1 illustrates the conversion from an image s colour space to a monitor s colourspace and the possibility of making it more efficient using a device link profile 24 CMS architecture Justification of the decision 3 1 2 A colour management module littleCMS Having chosen to incorporate the ICC profile standard the next step was to find a library to read profiles and to apply them to image data In the standard architecture of a colour management system such a library is commonly referred to as a colour management module The architecture of a colour management system commonly consists of three components a colour management module a set of profiles and the colour management system CMS itself The profiles represent the environment depending configuration of the system They customise the system for the devices in use such as printers or monitors The CMM is the engine that drives the conversions It is an arithmetic unit that reads the profiles and uses their look up tables and matrices to convert the input data The colour management system is the enclosing unit that ties everything together and makes it accessible
3. On top of the profile cache transformations are cached Again the implementation provides a function to retrieve a transformation and a function to return it The function to request a transformation is essentially a wrapper around littleCMS s transformation generation function It compiles the parameters such as the profile handles into a hash key and checks the cache Transformations are somewhat different from profiles Opening a profile takes no noticeable time it is equivalent to opening a file Profiles are usually kept by the requester for a long time An image profile for example is kept during the whole lifetime of the GImage structure This is not a problem because profile handles are merely sophisticated file handles The files themselves are only loaded into memory as needed Transformations on the other hand are often bigger and lie entirely in memory To allow efficient cache handling they are usually requested for one transformation and then returned immediately Pre computing transformations into a device link is time intense On the PII 300 that development was done on it took very noticeable time about 2 seconds Because of these differences transformations need a slightly more sophisticated caching system than profiles It should not only avoid keeping the same transformation open several times but also unnecessary recalculation of device links As a first measure a copy of every calculated transformation is stored
4. This is where the problem with colour reproduction in computer graphics starts Usually colours are stored in files as proportions of red green and blue light which are applied to the three light sources of a monitor to reproduce a colour according to the stored specification However there is no guarantee that the red green and blue on one monitor is the same as on another Admittedly the differences between different CRT monitors are subtle but noticeable 15 The differences become more obvious if you compare CRT monitors to TFT displays You may compare it to cooking If you follow the exact same recipe i e if use the exact same colour specification but in one case you use fresh vegetables in the other tinned vegetables i e if the primaries change the outcome will be different Colours will not necessarily appear the same on each monitor if they are simply specified as proportions of its red green and blue light sources When you go to a department store compare the TVs set up on its shelves They all show the same program they all receive the exact same input signal but you will easily observe that the colours can appear quite differently from set to set This states the first problem Storing colours as red green and blue values is not an accurate colour description The same colour can appear different on different monitors The fact that no set of three primaries can represent all colours the HVS is capable of percei
5. Point 3 can obviously be satisfied by look profiles It somewhat overthrew the initial design decisions Look profiles were initially intended to be stored in external files for a maximum of flexibility Look profile files could be stored in any directory and applying a profile would correspond to selecting the file through a file selection dialog Such a system however would not have been suitable for the quick and repetitive tests and checks that the editor was carrying out as he was applying previously used effects Still it seemed important to make the file nature obvious to the user and to give him the freedom to store profiles in any directory and to move them between different workstations The final design tries to combine the flexibility of files with a seamless integration with the interface This was achieved by deciding that all profiles used at a given point have to be stored in a single directory Profiles can then be presented in a simple list This list can be kept on the screen at all time and allows to easily switch looks on and off Predefining a directory as in the case of colour management profiles seemed more restricted than necessary so the list dialog allows changing the current directory which makes the file nature apparent to the user When the user saves a new profile the directory that is currently used in the look profile dialog will be offered as the default storage directory and the look profile dialog will be updated
6. Bernstein polynomials stretch the three ranges further but the results are smoother See figure 5 in the appendix for a comparison More explanations on the weighting curves are given in the next section of this chapter The button Save Profile will allow you to store your manipulation in a look profile You will be prompted to enter a file name and a number of sample points The number of sample points is the number of colour samples taken per colour channel to fill the multidimensional look up table of the profile The default of 13 will result in 13x13x13 samples to be taken As you may assume a higher number of samples leads to higher precision but more expensive computation A look profile allows you to preview other images or whole sequences of images with a manipulation applied without actually changing the image data You can create several profiles for a scene watch the scene through each of them using the flipbook and finally decide which one you like best You are also able to commit the manipulations to the image data meaning permanently changing the data in accordance with the profile Look profiles are used through the look profile dialog figure 5 3 in the menu View gt Look Profiles Each display owns it own dialog The dialog shows you all profiles that were found in a given storage directory You select the storage directory by clicking the open icon The dialog will remember the storage directory and the profile save dialog
7. and saturation and they need ways of inspecting and manipulating them effectively and reliably Common classes of tools can be found in many professional software packages such as Adobe Photoshop Discreet Combustion Apple Final Cut Pro or Apple Shake Two of the most popular ones are colour curves that describe colour changes by visually manipulating a mathematical curve relating input colour to output colours and colour wheels that allow to change the colour balance of an image shifting a handle over a colour wheel This project will 9 Colour Correction focus on colour wheels With an increasing number of films being edited digitally the possibilities of artistic expression are increasing too An increasingly used concept is look management which allows storing colour corrections in data structures called look profiles which can then be used to preview or to apply the same corrections on other images This allows the colourist to create a collection of looks to experiment with Studios currently use their own tools to achieve this There are no publicly available implementations yet as far as I am aware Kodak was the first to present a Look Management System at the American National Association of Broadcaster s NAB Spring Convention in April this year It is however focused on simulating the look of their different film materials rather than on artistic looks Advances in colour correction tools are unfortunately often made in the
8. but allow the reapplication of profiles without any loss of quality caused by interpolation Other areas of improvement will likely involve the handling of profiles such as the possibility to enter meta information when the profile is created or to arrange profiles into categories in the look profile dialog It can only be hoped that CinePaint will carry on in its promising development and eventually become the professional and robust tool it is aspiring to be It is a very ambitious project with a very friendly group of developers that to work in has been an interesting experience and a great pleasure 55 look profiles 56 References Colour 1 2 3 4 5 6 7 8 9 Color Models amp Color Wheels from Watercolors Bruce MacEvoy 2002 http handprint com HP WCL color6 html Colormaking Attributes from Watercolors Bruce MacEvoy 2002 http handprint com HP WCL color2 html The watercolors website contains an enormous wealth of information about colours from an artistic point of view The two pages quoted contain a detailed overview of colour models and their origin and a good description of the colour making attributes Computers amp Colors Bruce Fraser chapter 2 in Real World Color Management Peachpit Press 2003 also at http www creativepro com story feature 19500 html a highly readable introduction to colour encoding its origins and development and the problems of device depe
9. If we then use the flipbook to move to the next frame we can keep the look profile switched on and see the same effect on other frames 49 weight functions colour models The fact that each display owns its own look profile dialog provides a maximum of flexibility It allows using different displays for seeing an same image with and without a profile at the same time 6 2 2 Testing the Implementation A suitable test image for the colour corrector is a simple grey gradient from black to white It consists of all possible lightnesses and therefore clearly shows the different ranges of shadows midtones and highlights Besides grey has a saturation of zero leaving a lot of room for colour shifts First the weight functions were tested Figure 5 of the appendix shows the results It depicts a correction of 50 saturation towards red Lightness is preserved in all cases the used model is L a b You can clearly distinguish between the application to shadows midtones and highlights The figure also compares the two different weighting functions Gaussian and Bernstein polynomials You can see how the ranges are spread out more widely for Bernstein polynomials but the results are smoother Especially when shadows midtones and highlights are applied together the Gaussian functions add up too high in some parts Next the correctness of the implementations of the colour models HLS and L a b was considered The correctness of the implemen
10. a given hue manipulating the proportions of the image s red green and blue channels This means either adding the given hue or subtracting the complementary hue the hue that cancels the selected hue to white Adding or subtracting intensity on the image s channels may change the image s overall lightness so the algorithm needs to provide means to maintain the lightness constant There is no standard way of preserving lightness Different approaches sometimes differ in the quality of the result sometimes in the desired effect It was explained before that it proved difficult to specify the exact effect required Most likely there is no singular definition and no best approach and having different approaches to attempt for a given task may be advantageous Therefore the implementation provides two alternatives and leaves their use at the discretion of the user One is based on the simple but flawed HLS model to maintain lightness the other one is based on the more complex L a b model The algorithm is divided into two steps Changing the colour balance and readjusting the channels to maintain lightness Changing the colour balance of an image in the physical sense means changing the spectrum it radiates adding or subtracting intensities of certain wavelengths To model this colour model which is linear with intensity is needed a model that allows simple addition of intensities This is the case for RGB and XYZ RGB will be less expensive i
11. are still widely used in graphics applications but are not recommended for reliable colour calculations see for example question 36 in 4 They are still useful however as a basis for interface elements and as means of simple visualisation It would be more correct to call them colour ordering systems instead of colour models They provide an easy and accessible way to find colours and to move between them but do not offer the means of precise calculation and prediction a physical model does Calculations based on these models will mostly be rough approximations only However other intuitive models and these are models in the physical sense were developed 18 by the CIE Instead of abstract lightsources X Y and Z they use accessible concepts like L a b lightness L a b and L u v are based on XYZ and allow direct manipulation of concepts amp L u v such as lightness L or hue for example via a b where a representstransition from red to green and and b from yellow to blue L a b and L u v are useful models for colour calculations a sound foundation to base an algorithm on that will generally give better results than HSV or HLS in chapter 5 It should have become clear from the above that there is a wealth of models and ordering systems that all have different advantages and drawbacks In the end they are all only handles onto the mystery that colour still is The difficulty is to know which one to use to achieve the des
12. defining an overwhelming and confusing amount of options 4 2 The User s View 4 2 1 Configuration To be able to use colour management you have to configure it for your environment The most important step is providing and selecting the necessary ICC profiles Profiles are selected on the colour management tab of CinePaint s preference dialog which is located in the menu File gt Preferences see figure 4 1 By default profiles are searched for in the directories usr share colorAcc and color icc The idea is that the first path is used for system wide profiles and the second for user specific ones You can change these settings on the directories tab of the preference dialog On the colour management tab there are three profile selections to be made a default image profile a workspace profile and a display profile This structure has become a de facto standard or at least good practice in standard colour management workflows and related applications such as Adobe Photoshop or Apple ColorSync The selections for default image and workspace profile will contain all profiles that were found in the profile directories The selection for display profile will only show profiles of class display profile To correctly display an image CinePaint essentially needs to know two profiles the profile that 31 Display Interface Environment Directories Global Settings Color Management Color Management Profile Settings Default image p
13. film editing Secondly I phoned the BBC to find out about the possibility of witnessing a colour grading session Eventually I was able to sit through a day s worth of colour correction using yet another package Avid Symphony which is the industry standard in video editing from which I took away a couple of valuable points The user interface is one of the most crucial parts of designing colour correction tools As the former Director of Software Development for Leitch Video www leitch com pointed out to me The back end processing algorithms were never the issue For the product to succeed it had to intuitive to operate and fit in the job flow model used in video and film post Building an interface blind it is hard to get the UI designed and laid out just right so it will work in a production environment Defining the underlying algorithm proved equally difficult Few guidelines were available about what exactly is desired There is specialised literature on colours some of which might have helped to some extent but it is hard to access unless one is willing to buy it Neither the university library nor the city library nor the art college had anything available The central question to answer was which colour models to base the implementation on 5 2 The user s view The colour correction tool is accessible via the menu Image gt Colors gt Color Correction see figure 5 1 Its main feature are three colour wheels that allow intuit
14. functions were implemented and left for the user to choose Bernstein polynomials and Gaussian bell curves A Gaussian function gives more confined ranges but does not necessarily add up well when manipulations of different ranges are combined A Gaussian function can be computed effectively because of its symmetry It is sufficient to compute half the function and then reflect it Bernstein polynomials have the nice property of adding up to one at any point which results in very smooth transitions between the different lightness ranges Figure 5 5 depicts the two different functions The weight functions are precalculated for all 256 possible lightness values and stored in an array To extend colour wheels to higher precisions it will be necessary to interpolate or to experiment with direct calculation 5 3 3 Look Profiles Chapter 3 1 4 explained that look profiles can be implemented as abstract ICC profiles The previous chapter on the implementation of colourmanagement explained that the image data is transformed through the image s profile and the display s profile before it is displayed These two ideas can readily be combined In the creation of transformations in littleCMS the number of profiles is irrelevant littlkeCMS receives an array of profile pointers and merges them into a single device link transformation Look profiles can therefore be simply combined with the remaining profiles The transformation then starts with the image profil
15. in a temporary file on disc If ever it is evicted from memory it can be reloaded quickly and without the need to recompute it In addition because transformations are usually requested for short times only but often for several subsequent times the last used profiles are kept in memory even when the reference counter has reached zero How many profiles are kept is a compile time option If all places are taken transformations are evicted on a least recently used basis This is simply implemented as a ordered list of transformation handles for which the reference count is zero When the last reference to a transformation is returned the transformation 1s added to the end of the list If a transformation is requested again it is removed from the list If at any point the list is full the transformation that comes first in it is evicted to make room 34 Transformation cache The standard usecase in movie editing is keeping many open images simultaneously Q9 them using the same image profile since they are frames of the same film and the same di profile since they are 7o viewed on the same screen In this situation the caching mechanism is particularly effective Without it each image would open its own copy of the profile and generate its own transformation keeping all transformations in memory at the same time or heavily swapping between memory and disc The caching mechanism ensures that the image profile is opened exactly
16. is often subtle The human visual system HVS is very good at evening out errors in images A white shirt in an image may be slightly pink due to a wrong camera setting still at first sight the brain will suggest otherwise because it knows that the shirt is white This phenomena is called colour constancy the ability of the HVS to make colours appear the way they should This often makes it very difficult to judge an image Take a look at the tow images in figure 1 of the appendix and only this figure They look presentable Then take a look at figure 2 on the following page You will suddenly notice that the previous images were not very good at all that they had fairly strong yellow and red casts on them Still without experience it takes most people the comparison between the two pictures to notice it clearly Colours are not easy to get firm control over Changing colours is much harder than changing the formatting in a word processor Sometimes a little tweaking makes a big difference sometimes a huge shift does not get you anywhere This stresses the importance of appropriate tools They need to afford intuitive manipulations They need to contain in their design some measure of what can be done and what might have to be done to correct a given image Suitable tools provide handles onto colour in abstract ways Artists do usually not think in terms of the red green and blue that a monitor works with They work with abstractions such as hue
17. management conversions ICC profiles 22 3 1 2 A colour management module little CMS eene 24 3 1 3 Handling many profiles Storage and Caching esee 25 3 1 4 A format for look profiles ICC abstract profiles eee 25 3 2 Aside familiarisation with the code base binary fixed point sues 26 3 3 Cine Paint tendida qi ERE eC EYES 27 4 Implementation I Colour Management 4 eee eee ee ee eee een ee ones to seo en seen sten e tasso e eaaea 29 LAKESIDE E 29 4 2 The Users Vie Was ee dd pe ele er Detto eo Debe RO t epe HARDER CEU 30 4 2 T Configuratiof x oe teen e teet te ale cite e Dade ie Ee Pee nds 30 A OEC E P oes ceeded 32 45 Behind the SCENES cinco tail tat ted hee UE deta RR e ARS 32 4 3 1 Efficiency I The Profile cache seen nennen 32 4 3 2 Efficiency II Display buffer eene trennen renes 33 4 3 3 Fault Tolerancia tl e e a e EH e ient 34 5 Implementation II Colour Correction e eese eee eere eee ee eee ee en een stans ss 35 Sl Challenges eere teet eet ete t let ld 36 522 PNG USERS Vi Wat eret DATA te de to ec RE Pe Rer Ie e PEN EE Tete 36 5 3 Behind th Scenes ed o Tee RR e eue 39 5 3 1 Drawing the colour Wheel noir ds 39 5 32 C olo rcortection ud eee er esser ehe RU eee ec tee e RE ena 40 5 3 3 TOOK Profiles sisi eei E Ne Tete vod
18. once that the transformation is calculated only once and that this transformation can be kept permanently in main memory for fast access 4 3 2 Efficiency II Display buffer To make the actual transformation of image data more efficient parts of the CinePaint rendering engine had to be restructured The rendering engine was inherited from the GIMP and had grown in rather unfortunate and inconcise ways over the years Since this project is not mainly concerned with this part of CinePaint refactoring was limited to the necessary changes to allow the introduction of a display buffer for every display of an image that holds its final screen version Before the rendering engine only used one buffer for each image not one for each display of it Besides the use of the buffer was inconcise For example layer composition was done within the buffer whilst channel decomposition which happens if for example you switch off the red channel in an image was done on the fly as the image buffer was copied onto the display The new display buffer however contains the final screen version of the image A screen refresh thus only involves copying from the buffer to the screen without any further transformations apart from scaling to the zoom level It is necessary to have a buffer for each display and not only for each image because with the introduction of look profiles in the next chapter the data shown in each display can differ The new buffer struc
19. only starting to be explored by developers of professional applications Acknowledgments The author would like to thank the following Robin Rowe at CinePaint for inviting me to work on the CinePaint project for invaluable advice in particular during the early stages of the project and for many interesting discussions Julian Bradfield for making this project possible in the first place by supervising me Kai Uwe Behrmann at CinePaint for useful feedback and cooperation John Matthews at the BBC for taking the time to show me his work in spite of a very busy schedule Contents 1 IntEOOUC OD as 7 E2 MottyatiOns i aue tet Pet e be e bet he toii iet Due b deb e cute tete 7 1 3 Cin Pa iht oit e Rr Eee de et REN DA 10 1 4 Goals 5e nee e nee te elt eee deret ee cei e RA ee e 10 1 5 Result eet eite citer Rien P ERE UE HE RE nud 11 1 6 Remaining Chapters iun aie ad un Neel de cU ea ANS 11 A R P 13 2 1 Deseribing Colours ore ee d e het ER FEL ges tee t edge reni sheet 13 2 2 Matching mixing colours from colours eee 14 2 3 Physical limits problems with matching eese eterne 15 2 4 Colour Management 3 2 o iii 16 2 5 Visual Colour Models and colour wheels eee 17 2 6 Coloutr Correction recti A A bu PDA Ce A 18 STI P e EE 21 3 1 Fundamental design decisions eese enne enne 21 3 1 1 A format for colour
20. requirements was to obtain enough insight into the area of application Unfortunately there is little information available on the art of colour correction in film Asked by Robin Rowe the CinePaint project leader about where to learn more about film colour correction a colourist at Flash Filmworks the studio that used CinePaint for The Last Samurai answered Only in a training program or on the job experience at a film lab or color suite as far as I am aware Colour correction is a skill that is based on a lot of experience and the advice one receives depends on the colourist asked and the kind of tools he used in the past f you get this kind of a strange green cast in the blacks it usually helps to tweak here a little And adjusting this is always a good idea too Asking three colourists will likely result in five answers And after all there was no colourist designated to this project to answer questions For obvious reasons a trip to Hollywood to visit one of the studios CinePaint is used at was out of question I tried to close the gap in two ways Firstly by exploring packages available to me at the university These were Adobe Photoshop Adobe Premiere and Discreet Combustion Adobe Photoshop is a still image retouching program and currently the standard that everyone has to bear comparison with Adobe Premiere is a hobbyistic video editing tool and Discreet Combustion is already a considerable way up the professional scale towards
21. slider The first switch in the option dialog 39 options saving look profiles using look profiles FT preserve lightness when shifting across wheel rcalour model for lightness preservation v HLS fast but often less accurate Lab slower but usually better rweighting curves for shadows mids and highs A Bernstein palynomials smoother but wider y Gaussian more confined OK Cancel Figure 5 2The Colour Correction Options Dialog allows you to switch this correction off resulting in a simple manipulation of the colour components irrespective of the caused lightness changes See figure 6 in the appendix for an example If you choose to preserve lightness the dialog allows you to choose the colour model that the calculations to preserve lightness are based on HLS is simpler and faster but often less accurate You will notice that in many cases the colour cast tends to be too light For an example see figure 7 in the appendix L a b is slower but the colorimetrically correct model It even takes into consideration your particular image s colourspace if you have assigned an image profile for colour management As a last option you can choose the weighting curves that are used to distinguish between highlights midtones and shadows A Gaussian will give you more confined ranges but especially when adjusting more than one range at a time with the adjustments adding up the result sometimes turns out to light
22. such as a flipbook that allows to easily load store and move through sequences of images 1 3 Goals The aim of this project is to equip CinePaint with full display colour management on the basis of the ICC standard The purpose of display colour management is to maintain colour appearance the same between different display devices such as a CRT and a TFT This project is not concerned with colour management in printing but the results will be readily extensible The reason printing was disregarded is that it is probably almost harder to set up a good demonstration than to do the actual implementation Within the display CMS it will be possible to specify image and monitor characteristics in the form of ICC profiles which will be explained in 3 1 1 and to correctly reproduce colours on any display It will be possible to convert between colour representations Considering that the conversions needed to correct the colours will cause an extra time strain on the work of the rendering engine special attention will be paid to efficiency The aim is further to provide a new colour correction tool to CinePaint CinePaint already has a 10 curves tool which was inherited from the GIMP but does not have a colour corrector based on colour wheels The goal is to provide such a colour corrector This involves designing an accessible interface and working out an underlying algorithm The corrector will be designed with a future integration with the cur
23. such colour spaces are called device dependent What is needed is a device independent space And it has to be a gamut that can represent all colours the HVS is capable of perceiving The colour matching experiments showed that it is impossible to match any visible colour by a weighted sum of red green and blue because it would require negative light However the fact that negative luminance cannot exist in reality does not prevent assuming its existence just as a fourth and fifth dimension can be assumed The colour matching functions could then be used to describe any colour even if it cannot physically be generated on the three sources the colour matching functions are extracted from because it would require negative luminance 16 It seemed awkward though to assume the existence of negative luminance Besides the XYZ a new problem that there is no clear definition of the primaries persists The exact red green and blue colour model used to reproduce a colour depend on the monitor that is used Therefore in 1931 the Commision International de L Eclairage International Commission on Illumination CIE www cie co at cie an international body devoted to cooperation and exchange in the science and art of lighting defined three standard primaries called X Y and Z to replace red green and blue in the matching process They are positive over the whole visible spectrum and they can be used to describe all visible colours Obviously
24. to the user It manages the profiles requests the CMM to perform transformations as necessary and lets the user set default options Figure 3 2 shows these relations The idea behind separating the CMM is to make it plug replaceable For Apple s ColorSync CMS for instance there is a variety of CMMs from different manufacturers such as Kodak or Apple itself For Linux there are three CMMs available or under development little CMS www littlecms com ArgyllCMS web access net au argyll argyllcms html and gCMS gcms sourceforge net In spite of their names they all are CMMs in the architecture outlined above LittleCMS was chosen for this project for several reasons its developer Marti Marias provides good and on going support It has a well defined and stable interface which would in principle allow to later replace littleCMS with another CMM implementing the same interface It is being used by an increasing number of prominent software packages on Linux such as Scribus a desktop publishing package www scribus net The GIMP www gimp org ImageMagick www imagemagick com and Qt the KDE interface toolkit www trolltech com product qt It has the best prospects of developing into a standard for Linux In contrast to ArgyllCMS and gCMS it is also available on Windows and MacOS which Profiles LI user options QN CMM i y littleCMS corrected images images Figure 3 2 Architect
25. will use the same directory 40 e eveninglighticc thenewlook icc Figure 5 3 The Look Profiles dialog The handling of profiles is similar to layers To watch an image through a look profile simply click onto the white space on the left of its name The profile will be switched on symbolised by a little pair of sunglasses You can switch on several profiles at the same time To commit the current look profile settings to an image click onto the second icon in the look profile dialog The last icon lets you commit the profiles to all frames currently loaded into the flipbook In either case the image data will be manipulated and the profiles will be switched off since after committing them they are already part of the image data 5 3 Behind the Scenes 5 3 1 Drawing the colour wheel In contrast to all other interface elements like buttons and inputboxes that are already provided as elements of CinePaints interface toolkit gtk the colour wheels had to be implemented manually The underlying gtk interface element is a drawing area providing simple drawing primitives There is one drawing area for each wheel and each lightness slider The slider is drawn by iterating over the pixel positions and translating them into RGB gray values between 0 0 0 and 255 255 255 equal proportions of R G and B give grey The position on the slider is scaled to a range of 0 100 and determines the lightness The wheel is drawn by iterating ove
26. 41 6 UI rii Mese tai 43 6 1 Colour Manage ccoononccnocnnnnonnnncononcnnnoncnncnnonncn nono enne nennen nenne trennen treni nennen e nete trenes 44 6 T I Interface JUST COM erre entree erre tren avia 44 6 1 2 Testing the Implementation esee enne enne enne enne 44 6 2 Colour Correction ined E 45 6 2 2 Testing the Implementation eee eese enne 45 Rodi M 49 741 Achievements nene ee tpe pelo boa dee deste nont 49 7 2 Scope for future AA de RR ee tee gebe cde Cope dete ue cet tua 50 References Appendix letroduction The goal of this project is to provide means of reliable colour reproduction and colour manipulation for the open source image retouching program CinePaint 1 1 Motivation Colours are an important means of communication and expression Colours convey structure in diagrams they create emotions and atmosphere The French artist Yves Klein 1928 1962 developed a major part of his works on one single blue that he created specifically and which he patented as International Klein Blue IKB He claimed it to be the physical representation of cosmic energy that otherwise invisible floats in the air freely People have favourite colours companies have corporate colours Colour photography the first colour television and more recently the breakthrough of colour in newspaper caused excitement and were usually seen as a Colour Management big improvement over their black and white predecessors Colour p
27. Adobe Photoshop Claims that even colour blind people could colour correct which may be a little far fetched but it still gives a very thorough and understandable basis 20 Photoshop Restoration amp Retouching chapter 4 Working with Colour Katrin Eismann New Rider http www digitalretouch org 2edition downloads extras PSRR 2ed ch4 colorcorrect pdf A broad and thorough overview of all the tools Photoshop offers for colour correction and which to use for what 21 Avid Color Correction User s Guide 2003 Avid Technology Inc available from http support01 avid com support ProdRef nsf DocsR3 ReadForm amp sStart 1 amp Count 3 O amp Expand 4 6 How to perform colour correction using the tools provided in Avid Symphony the industry standard video editing suite 59
28. University of Edinburgh Division of Informatics Colour Management and Colour Correction in CinePaint 4 Year Project Report Computer Science Stefan Klein May 26 2004 Abstract Before the advent of affordable desktop computers colour reproduction used to be the task of trained specialists As it moved to the desktop and into the reach of everyone systems failed to provide the means to handle it well Users are still surprised when they find that colours do not appear the same on all screens or printers To provide means of reliable colour reproduction is the goal of colour management Colour correction or colour grading refers to the manipulation of colours In film grading gives a film a consistent look It can set an atmosphere for a scene for example by bathing it in warm evening light It can also be a means of artistic expression such as symbolising the loss of joy of a film character through the use of bleached colours This project is set within the open source project CinePaint an image retouching application that has been used for films such as Harry Potter and The Last Samurai The goal is to equip CinePaint with a display colour management system This will make CinePaint the second open source application to implement such a system The goal is further to provide tools for colour correction namely a colour corrector based on colour wheels and look profiles Whilst colour wheels are a widely used concept look profiles are
29. and Engineering Dawn Waller 04 2000 International Color Consortium http www color org iccprofiles html An introduction to reading and writing ICC profiles Has code samples and is generally more readable and informative than the specification itself Contains a good introduction to the different processing models ICC profiles offer 16 Colour Calibration for Colour Reproduction P Emmel R D Hersch Ecole Polytechnique F d rale de Lausanne Switzerland at IEEE International Symposium on Circuits and Systems 05 2001 Geneva Switzerland http diwww epfl ch w3lsp publications colour ccfcr pdf contains a good summarized description of ICC profiles in the first half of the article Colour Correction 17 Color Correction in Final Cut Pro Steve Martin Ripple Training twosense net macworld sf pdfs C65 pdf slide show of the colour correction tools available in final cut pro Since it is slides it has little depth but gives a reasonable first overview of the scope of colour correction in film 58 18 Color Correction Finishing in FCP Andrew Balis Kenneth Stone Photography 09 09 2002 www kenstone net fcp homepage cc legal fcp3 html step by step explanations of colour correction in Apple Final Cut Pro 3 a film editing package Even though product specific it is a very good overview of colour correction explaining many general concepts 19 Color Correction By the Number An introduction to the principles of colour correction using
30. ck and 255 255 255 representing white 16 bit unsigned integer between 0 0 0 and 65535 65535 65535 and 16 bit and 32 bit float between 0 0 0 0 0 0 and 1 0 1 0 1 0 Today s displays cannot diplay colour of more than 8 bit depth 256 different settings for red green and blue but film scanners can often read data in higher resolution For many image manipulations the result is better if the data is stored and manipulated in a higher bit depth with more detail and only downcast when it has to be displayed Along the same lines it can be advantageous to be able to store values that are whiter than white White is always relative It is whatever the display device or the paper define it to be Paper often appears white until it is compared to another sheet that is even whiter White on a monitor is 255 255 255 Values higher than this would simply appear white too However there are reasons to store these values anyway An example may be a picture that contains a lot of bright values Fires and explosions are common examples in film Increasing the brightness of such a picture will result in some values going beyond the white limit If this limit is enforced values beyond it are clipped and simply set to the maximum value If there is a possibility to store over white such values will store be stored as they are for example at 1 2 even though on the screen they will appear white all the same The advantage becomes obvious when the bri
31. d Lab work as intermediate spaces to connect the profiles which is why in the context of the ICC standard they are termed profile connection spaces PCS The use of a PCS allows a maximum of flexibility in combining different profiles The profile format is complemented by a header containing meta data such as a description or the definition of a profile class The most common profile classes are input profiles which are used for scanners or digital cameras display profiles used for display devices such as monitors and output profiles for devices like printers There is one other class of profiles that requires special mentioning device link profiles They are equal in structure to the profiles described before but do not translate from or to a PCS Instead they translate directly between two device colour spaces In the previous example a device link profile would convert straight from the scanners colourspace to the monitor s colourspace without the intermediate step of converting to XYZ Device link profiles are created by melting two ordinary profiles into one which means multiplying their conversion matrices or combining their LUTs Device links are created for efficiency If for example a display has to be refreshed often one table look up can be considerably less expensive than two The drawback is obviously that device link profiles are tailored to one specific conversion and therefore less flexible 23 PCSs device link profiles
32. d is mainly due to the underlying CinePaint rendering engine which is at the moment far less then satisfactory It was intended to run a final test against Adobe Photoshop as a reference implementation It turned out difficult however to sensible compare the implementations since the details of Photoshops implementation were obviously not unknown Many factors influence the final outcome such as the chosen ranges of the input parameters the weighting curves the colour models Often there is more than one valid choice depending on the intentions The test was therefore rather informal but it was almost always possible to recreate a given manipulation in the other program even if with slightly different settings Look profiles finally were simple to test All that had to be done was to create two copies of an image The first one was manipulated and the result stored A look profile was created from the manipulation and applied to the second copy The two manipulated images were inspected for equality using the colour picker tool to inspect individual colour values 51 52 Conclusion This chapter summarises the achievements of the project and presents directions for future work 7 1 Achievements The goal to equip CinePaint with a display colour management system has been fully met CinePaint now is the second open source application after Scribus to provide such as system The handling of profiles was realised as efficiently as poss
33. does not match the workspace profile CinePaint can either ignore the mismatch and keep the image profile automatically convert the image to the workspace or consult the user on whether to convert or not Which action is taken depends on the preference setting When image space doesn t match workspace Finally you can define a default rendering intent to use colour management As explained in chapter 2 4 the rendering intent defines the way colours are approximated that are not in the display device s gamut For a detailed description of the options available refer to any introductory text on colour management such as 10 or 11 The intent selected here is the default option You will later be able to change the rendering intent for each display individually A final note For each of the profiles settings it is possible to select no profile at all Also you can choose to have no profile assigned to a newly opened image Notice however that CinePaint will only be able to colourmanage an image if you have set a display profile in the preferences and have assigned a profile to the image Should you forget to do so CinePaint will remind you 4 2 2 Use Now that you are all set up you can start to have your images colourmanaged Most of the colourmanagement functionality unfolds in the background without the need for any direct interaction just as it should As the user you can get on with your work and rely on the colours to be correct with
34. e by colourists The bottom line of all colour correction is the definition of a transfer function The input are the original colours of the image the output are the manipulated colours To change brightness for example one could use a transfer function that simply adds a little intensity to each pixel such as Drew X Dod x 5 The application could just prompt the user for a transfer function apply it and be done In a way this is what the most powerful tool for colour correction the curves tool does It allows the visual manipulation of a transfer curve that is then applied to the image It does require a thorough understanding of colours though to achieve the desired effects Other tools are less flexible but more intuitive In a way they hardcode a certain transfer function and let the user specify the input parameters Colour wheels are an example They allow changing the colour balance of an image to remove or deliberately create colour casts The colour wheel is useful to predict how a change in one color component affects other colors If the colourist detects a colour cast in an image he finds the colour on the colour wheel and pushes the balance towards the colour that is opposite on the wheel Since opposite colours are complementary this will balance the cast There are many different mathematical models to describe colours and to allow operations on them such as changing their lightness or mixing them The quality of the same co
35. e converting from the 43 weight functions applying profiles creating profiles sample strip input output input profile QUEM _PCS X display profile look profiles space Figure 5 6 The complete profile chain image space to a PCS It continues through one or several look profiles that manipulate the data but leave it in the PCS It concludes with the display profile converting the data into the display s colourspace Figure 5 6 illustrates this Look profiles are kept in singly linked list structures that are assigned to each GDisplay This makes it possible to have two displays showing the same image with different look profiles applied Look profiles blend seamlessly into the existing colour management framework with one exception look profiles should still work when colourmanagement is switched off for a display or even when no image or display profile is defined In that case the image profile and the display profile in the chain of figure 5 6 are replaced by a standard RGB profile As a result data is transformed from RGB to the profile connection space is manipulated in the PCS using the look profiles and then converted back into the original RGB space to be displayed The display is in this case not corrected since colour management is switched off but the look profile will still be applied Committing profiles to an image uses exactly the same transformation only that the result is not sent to t
36. eeping us from noting them a phenomena called colour constancy but the camera will pick them up and make them visible If they are not evened out the light changes will be noticeable to the viewer Even subtle shifts will make the editing appear surprisingly uneven and the flow of the scene will be broken The viewer will be very aware of the filming and the film will have failed to create the illusion it wanted Obviously a lot of effort is put into matching lighting conditions and cameras but the final touch in matching the different parts of the scene often has to be done in colour correction Another different example A film needs to be shot in a restricted environment Access can only be granted for short times during the day The scene however is supposed to take place in the evening The solution is to shoot it at day and attempt to create a colour impression of evening twilight by using colour correction The colourist and director might even want to go beyond reality To symbolise a characters distorted vision of his surroundings they could manipulate the colours in strange ways They might want to make a film appear like an old technicolor film Or they might want scenes in an epic to appear as if they were immense wall painting a prominent example is the Lord of the 19 Colour correction tools Look Management Rings in particular the last part All these are applications for colour correction a highly artistic task don
37. efficient by avoiding duplication when several images use the same profile Unlike Windows and MacOS Linux does not yet have builtin colour management functionality Therefore there is no standardised directory to hold colour management profiles in the same way as there are for example directories for fonts After consultation of other CMS related projects such as littleCMS and Scribus it was decided to try and set an example hoping that others will go along with it and turn it into the standard Linux profile directory CinePaint s profile directories were set to usr share color icc for system wide profiles and color icc for user profiles following a suggestion by Marti Marias the developer of littleCMS These directories are scanned for profiles and the findings are presented whenever the user has to select a colour management profile To make opening many profiles efficient the implementation will have to include a caching mechanism which will be based on hashtables and reference counters The mechanism ensures that no profile is opened more than once Transformations will always be pre computed in the form of device links Again there will be no duplication Since pre computation is time intense about 2 seconds on a PII 300 computed transformations will be cached in temporary files on disc if they are evicted from memory The mechanism will be described in more detail in chapter 4 3 1 3 1 4 A format for look profiles ICC abstract p
38. en and blue light overlap to form a second patch The subject of the experiment is asked to adjust the intensities of the three overlapping sources such that the two colour patches look the same This is repeated setting the reference patch one by one to all wavelengths of the visible spectrum Repeating the experiment with many subjects lead to the derivation of what is known as the rgb colour matching curves Figure 2 1 The curves express how much red green and blue have to be used to mix a given colour The experiments discovered however that certain colours could not be matched at all For certain colours of the reference patch it was impossible to find settings of the three mixing light sources such that the mixed patch would be identical The only way of achieving identity was to modify the reference patch itself to add a little red to it Mathematically this corresponds to adding negative red to the mixed patch which is reflected in the curves This discovery poses a serious problem for computer graphics Colours on a monitor are generated by mixing red green and blue But we just saw that it is impossible to mix all colours the HVS is capable of perceiving Even worse different sets of primaries of base light sources will generate different sets of colours when combined The set of all colours that can be mixed from a given set of primaries is called their colour space or colour gamut 2 3 Physical limits problems with matching
39. existing code base of CinePaint which a couple of lines will be spent on Firstly to point out some details important to the implementation of this project and secondly because the process resulted in another feature addition to CinePaint that is not directly linked to colour management or colour correction though still to do with colour handling but nonetheless interesting and worth mentioning 21 Justification of the decision The ICC format 3 1 Fundamental design decisions Some of the initial design decisions that usually have to be taken in the process of developing a software application were predetermined by CinePaint by the fact that this project is set within an already existing project CinePaint uses C as programming language although a gradual transition to C is planned and gtk 1 2 as GUI toolkit Other decisions had to be taken by this project These were e a storage format for colour space conversions as described in the previous chapters This format will be the ICC profile standard alibrary to read the ICC profile format which will be litteCMS e a way to load and manage bigger numbers of profiles efficiently This will be implemented using a simple caching mechanism based on hash tables and reference counters astorage format for look profiles which will be abstract ICC profiles 3 1 1 A format for colour management conversions ICC profiles The first decision that had to be taken was how to store t
40. g a great deal of different pictures with the same profile combination never resulted in more than one transformation being created Reducing the maximum number of unused transformations in memory to one and continuously switching between images with different profiles which causes transformations to be requested and released resulted in transformations being correctly swapped between disc and memory A test of the actual correction of the image display itself was difficult to set up since effective colour management depends as much on good profiles as it does on the underlying system Profiles depend not only on the monitor model but also the graphics board the ambient light and much more Creation of good profiles is an art in itself and often requires special measurement hardware One such device is available at the King s Buildings Graphics Centre an EyeOne colorimeter and I spent a day trying to work with it I got nice results for some images but did not manage to generate one generally useful profile Since the actual conversion is driven by littleCMS however it seems valid to argue that it is 48 correct since littleCMS has been tested and tried in many projects and by many eyes over the last couple of years Also my implementation is already in use by one other CinePaint developer who focuses on colour related developments and no quality problems have come up so far 6 2 Colour Correction 6 2 1 Interface Justification Colo
41. ghtness is decreased again In the first case all over white values were lost They have all become ordinary white Reducing their brightness will turn them now into an ugly grey without any detail In the latter case the values have been 27 preserved and as they slip below the white limit again the detail reappears on the screen Another use case is scanning film Film captures subtle differences in very bright colours that on a monitor would all appear as white So when scanning the white limit is set to the monitor s white but values beyond it are stored anyway If brightness is later decreased detail will appear that was always there on the film just could not be seen on the screen To be able to store over white values requires a suitable data format Unsigned integer is not suitable since white is already defined to be the maximum value There is no 8 bit integer value beyond 255 float could be used but is in practise currently clipped to 1 0 in CinePaint Therefore 16 bit binary fixed point BFP was introduced as the first format that can store over white In 16 bit BFP the highest order bit is interpreted as the single digit before the decimal point 0 or 1 and the remaining 15 bits as the decimal places Thus BFP ranges from 0 0 to 1 99 and can store over white To represent over white on the screen where it normally cannot be seen required allowing the user to choose a strong warning colour that is used whenever colours are be
42. he colourspace conversions needed for colour management It was somewhat predetermined by the fact that there really only is one standard format used for this purpose that is adopted by virtually all colour managed applications from Apple ColorSync to Adobe Photoshop It is a file format based on a standard by the International Colour Consortium ICC Some applications use proprietary colour space transformations that are hard coded into the application but there is no other open profile standard available The flexibility profiles offer over hard coding and the acceptance without exception of the standard make supporting the standard almost compulsory for any colour management system Also there is an increasing number of applications that allow generation of ICC profiles tailored to a given working environment This is important because colours do not only depend on the model of a monitor but also on factors like ambient light the graphics adapter or the ink cartridges used in the printer Profiling applications therefore often use specialised hardware so called colorimeters to read the spectra of test colour patches that are output on the printer or the screen that is profiled The profiles are then generated on the basis of the data from these devices Unfortunately support is still limited under Linux but with an increasing number of applications supporting colour management this will hopefully change soon ICC profiles relate colour data
43. he display but stored back into the image Creating a look profile essentially means filling its multi dimensional look up table LUT with appropriate values Functionality for profile creation is provided by the littleCMS API The creation starts by completing the information in the profile header such as profile class and input and output colour space Since look profiles will later be used between input profile and display profile as depicted in figure 5 6 they have to use a profile connection space PCS as both input and output colour space The LUT is filled by providing a callback function that is repeatedly invoked by littleCMS once for each sample point in the LUT The function receives input colour values from littleCMS and is expected to return the corresponding output colours Input and output colours are sampled from the image before and after the manipulation As seen in chapter 3 1 1 the LUT will not contain all possible values but interpolation will be used when the profile is later applied Therefore it is important to provide evenly spaced samples during the creation of the profile Based on the normal image data this is not necessarily possible A picture of the sea will have a lot of blue samples in it but little red A picture of a forest will be dominated by green The sample process itself would already have to use interpolation which is not desirable due to the introduced error A better alternative is to generate a h
44. he first graphics package to provide full colour management functionality The users however had not gone through the same process of realisation as the developers The release of Photoshop 5 caught them unprepared and caused a lot of confusion often resulting in users turning colour management off Today however at least in the professional world colour management has finally become the serious issue it should have been from the outset In the domain of open source software colour management is still underdeveloped Linux does not provide a CMS on operating system level However there is a very mature and reliable library for colour conversions LittleCMS www littlecms com Scribus a desktop publishing program www scribus net is at the time of this writing the only open source application to fully implement colour management Users want predictable and consistent colours They want the colours they see on the screen to 8 match the colours on any screen or the colours they finally print Modern graphics applications are based on the principle that they provide the user with a preview of the final result This should include colours Without reliable colour preview on monitors proof prints are an absolute necessity and cause extra costs The more reliable colour preview on monitors is the less proof prints will be needed There is a need for more applications to adopt colour management especially in the open source sector both to
45. ible assuring that no profile or transformation is loaded or precomputed more than once irrespective of how often 1t is used The only feature that is commonly found in colour management systems but was not realised during this project is the handling of embedded profiles It has been explained in chapter 4 2 1 that this is not a priority for work in film studios However the necessary infrastructure to load 53 CMS colour corrector embedded profiles into the CMS cache has been provided Loading and storing profiles from various file formats such as TIFF and JPEG is currently being implemented by another CinePaint developer The goal to provide a colour wheels colour corrector has equally been met It may even be argued that the corrector implemented is slightly beyond the original requirements as it presents the user with a range of fine tuning options The only non satisfactory aspect is performance but it has been explained that this is due to the underlying CinePaint rendering engine Look profiles have been implemented as anticipated They are the truly novel part of the development in this project It is a feature that is not commonly implemented yet in standard applications and is only starting to be explored by developers Hopefully look profiles will be a good foundation stone to develop new exciting CinePaint features on As an additional CinePaint feature outside the scope of this project binary fixed point has been impleme
46. idden reference strip of samples It consists of n coloured pixels evenly spread over the range of possible colour values n is the number of sample values per channel of the input colour space as entered by the user when the profile is created 3 is the 44 number of input channels Look profiles are working in a PCS which both in case of x Pha in case of Lab has 3 channels The samples are generated by three simple nested Joos Bras over the range of the three e colour channels The sample strip is generated when the colour correction dialog is opened and disposed of when it is closed The sample strip is affected by all the same manipulations as the image itself and thus provides the data to base the look profile on Artificially generating samples instead of relying on the data contained in the image achieves a set of samples that is perfectly adjusted to the structure of the LUT The reference strip contains precisely the colours that will later be stored in the LUT No interpolation is required during the sampling process 45 46 Occitan This chapter describes the steps taken to evaluate the usefulness and correctness of the implemented system As before the chapter is divided into two parts one for colour management and one for colour correction Both parts start by giving a justification of the interface design A full user based evaluation was unfortunately not possible due to reasons hinted earlier Several attempts
47. improve quality and reliability and to further the understanding of colour handling amongst users and developers of such applications Ideally reliable colour reproduction would go without mentioning for any professionally used graphics application Their real focus lies on providing means of manipulating images One aspect of this is manipulating their colours In film the work that is centred around colour is carried out by a colourist and called colour correction or colour grading His task is two fold One is to make the film look consistent The shooting of a scene may take place under changing light conditions For example it may be partly shot outside and partly on set If in the story the scene is meant to take place in a single location the colourist has to even out differences in lighting otherwise the editing will appear uneven The other task is more creative It is to give the film a look and to create colour effects A blueish futuristic look of Minority Report with Tom Cruise Epic scenes that look like great wall paintings like in The Lord of the Rings The impression of an old technicolor film Faded colours very strong colours or overexposed colours to express the particular state of mind of a character The examples are endless It is often not easy to spot the particular problems or needs for correction of a given image The difference between a dull photo and a vivid one that brings back all the memories of where it was taken
48. in one colour space to data in another colour space by specifying a conversion function The conversion can be expressed as a matrix as a multi dimensional look up table LUT or as a combination of both Which approach is chosen depends on the complexity of the conversion Converting between RGB and XYZ can be expressed in a 3x3 matrix because XYZ values are merely linear combinations of RGB values Conversion is carried out by multiplying the R G B triple with the matrix The conversion from XYZ to CMYK a printers colour space however is highly non linear It can only be captured adequately in a big LUT The LUT has one dimension per channel of the input colour space e g 3 in case of RGB A conversion is carried out by looking up the table cell that corresponds to the input channel values The cell contains the corresponding colour in the output space Usually only a subset of all possible input values is stored in the table and interpolation is used as appropriate 22 For every profile either the source colourspace or the destination colourspace has to be CIE XYZ or CIE Lab in order to be able to flexibly connect profiles to conversion chains to convert between two colour specifications e g between the RGB of a scanner and the RGB of a monitor two profiles are necessary The profile of the scanner converts from scanner RGB to XYZ or Lab the profile of the monitor takes the data from there and converts from XYZ to monitor RGB XYZ an
49. ired effect 2 6 Colour Correction Colour Correction is altogether different from Colour Management Colour Management is concerned with reproducing colours correctly according to a given specification It is a rather technical endeavour something that ideally would be taken care of by the system itself without any user intervention at all Colour correction another commonly used term is colour grading however is a user feature Colour correction in film is the artistic process of giving an image a certain appearance a look a feeling The goal can be to remove unwanted colour casts on images thus the term correction or to deliberately provoke them to create a certain atmosphere such as warm evening light Colour Correction is best understood by examples Assume a scene was partly shot on location in front of a magnificent mountain landscape and partly on set as the camera swings around and shows an old medieval village that could not be built on location due to it being a natural reserve Shooting on location takes a whole day from the early morning until the afternoon The final edited version will have to create the impression of one scene taking place during five consecutive minutes in one fixed location examples The artificial light on set contains more blue than daylight and appears colder Afternoon light contains more orange and is warmer than light at midday The brain is very good at evening out these differences at k
50. ive manipulation of the colour balance of an image Manipulation is done individually for shadows the wheel marked Sh midtones Mid and highlights Hi in the image The terms refer to the brightness range of the pixels that are most affected by the transformation A division into three ranges is standard in graphics applications since it is often found that certain problems only appear in certain parts of the image such as a black suit exhibiting a slightly green colour cast 38 r Colour Wheels i curves H 49 40 3s 3073 31 3333 3 H 286 46 3s s038 4Lf6276 3 Hf0 00 4sfo00 50 00 4 OK Cancel Options Save Prome Figure 5 1 The colour wheels colour corrector The wheel display is based on the HLS colour model which was briefly explained in chapter 2 5 The wheel you see is a slice through the HLS double cone The slider on the right of each wheel allows you to change the lightness moving up and down the cone Using the handle inside the wheel you can change the colour balance of the image Dragging the handle in any one direction will tilt the colour balance in that given direction The further you move it out of the centre the stronger the more saturated the colour cast will become The wheels allow easy and very intuitive adjustment of the image s colour balance If you note a colour cast in the image the right step to take is shifting the corresponding wheel handle in the opposite direction to balance it If fo
51. lays an important part in the world above all in art and design Colours are chosen specifically a costumer wants an arm chair in azure blue not in royal blue or in sky blue The difference between a dull photograph and one with vivid colour that brings back the memories of when it was taken may be subtle but it makes all the difference In short colour matters Colour reproduction in computer graphics however has for a long time been of very low quality Before the advent of affordable desktop computers colour reproduction used to be the task of trained specialist using expensive equipment But then it moved into the reach of people who had no background in colour theory and who were given the impression that they had the freedom to use any colour they wanted and that the print would be an exact match of the image on the screen Many developers probably believed it themselves However it has long been known that this must be an illusion Colorimetry the physics of colours made great advances over the last century The response of the human visual system to colour was quantified through the colour matching experiments Sophisticated colour models were developed by the CIE the Commision International de L Eclairage International Commission on Illumination www cie co at cie an international body devoted to cooperation and exchange in the science and art of lighting Here is the bad news that comes as a surprise to many Your monitor canno
52. le selects Y as her workspace sets when opening image to assign default profile and makes CinePaint convert automatically to her working space Whenever she now opens one of her images it will automatically be converted into her working colour space and you can edit it Example 2 The user has a range of images in different colourspaces He needs to edit them all in his workspace Y In this case he selects Y as your workspace sets when opening image to prompt for profile and again makes CinePaint convert automatically Now whenever he opens one of the images he will be asked for the appropriate profile and CinePaint will convert the image to his workspace automatically If he only needed some of the images converted he could make CinePaint prompt you each time on whether to convert Since colour management can be a serious performance drawback it made sense not to force it onto the user but to leave it at her discretion to switch it on and off 6 1 2 Testing the Implementation The most critical part of the implementation is the caching mechanism which was tested by adding debug messages specifying whether a profile or transformation was loaded from the cache or from from disc The swapping of transformations into temporary files could easily be observed from the creation of files in the tmp directory while the program was running From the tmp directory and the debug messages the correct operation of the cache was tested Openin
53. lice of one of the hexcones with the edges smoothened into a circle Again colours are ordered such that hue changes with angle and saturation with the distance from the centre Hue and saturation are essentially the polar coordinates of a point on the wheel Depending on whether the wheel is derived from the HSV or HLS model either lightness or value are assumed to be constant across the wheel An important feature of the colour wheel is that complementary colours colours that cancel each other to become white are at opposite positions This makes it useful as a tool to remove colour casts from images as we shall see in section 2 6 Figure 4 in the appendix shows an HLS colour wheel Flaws of Colour wheels HSV and HLS however are seriously flawed with respect to the physical wheels reality of the human visual system The regular shape of the wheel suggests that the distance HLS amp HSV between minimum and maximum saturation was the same for every hue It is not There is no standardised conversion to obtain HSV or HLS coordinates from RGB But most formulae such as those provided by Foley van Dam et al 6 are based on the assumption that the luminance of a pixel was just the average of its RGB intensity values R G B 3 The human visual system however perceives some wavelengths as brighter than others which in reality leads to a weighted sum of about 0 3R 0 6G 0 1B with the exact coefficients depending on the exact primaries HSV and HLS
54. ll if a program is deleted during its execution it will eventually crash too 36 Implementation II Colour Correction This chapter continues with the detailed description of the work done in this project Whilst the previous chapter dealt with colour management this one will cover colour correction It follows the same structure It starts out by pointing out the major challenges that were faced in the implementation of this project It then describes the implementation from a user s point of view much in the style of a user s manual It concludes by describing the implementational details Note In spite of CinePaint offering a range of different bit depths for storing image data Colour Correction was only implemented for 8 bit This is because the implementation of different bit depths in CinePaint is currently rather inefficient Adding new bit depths requires a substantial amount of code duplication of cutting and pasting and replacing pointer types which is neither a very challenging nor a very interesting task However colour correction will be readily extensible especially if the CinePaint bit depth implementation is re factored in the future as is hoped for 37 gaining insight interface algorithm the wheels 5 1 Challenges The central difficulties in implementing colour correction were clearly specifying the requirements and defining an underlying algorithm for the colour wheels The problem with specifying the
55. lour correction tool can greatly vary with the underlying abstractions it uses in its implementation to achieve the desired effect For instance there is more than one way of manipulating lightness but some ways are more in accordance with the nature of the HVS and will lead to better results With an increasing number of films being edited digitally the possibilities of artistic expression are increasing too very recent concept is look management After a colourist has adjusted one individual frame of a scene or a film he stores his manipulations in a look profile He can then view the whole film through this profile without changing the actual image data He can create different looks and experiment with them before actually committing them to the image Look profiles are a technique studios already use manually on the basis of their own tools or scripts They essentially compare the original and the manipulated image colour by colour and create colour conversion tables to relate the original to the new colour values These can then be applied to other images 20 d 9 Design Colour management and colour correction in CinePaint are based on a couple of fundamental components such as formats and libraries This section will explain their choice and briefly describe them Of course one of the major components that this project was built upon is CinePaint itself This required a substantial amount of familiarisation with the
56. n nicely with the first part of the project on colour management As it turns out the ICC profile standard already provides for this type of profiles There is a class of profiles called abstract profiles In structure they are no different from ordinary profiles Their function however is to apply colour manipulations to an image instead of converting between different colour representations They usually use data in a PCS as both input and output and manipulate it A profile could for example increase the image brightness by increasing all input colour values LitteCMS provides the API to fill the LUT of these profiles with sample data 3 2 Aside familiarisation with the code base binary fixed point The familiarisation with the codebase took up a considerable amount of time during the first weeks of this project The codebase inherited from The GIMP comes with hardly any documentation not even the central data structures and building blocks are documented So the code itself and information gathered on the developer s mailing list were all there was to work from Since code is not as readable as a novel it seemed a good idea to combine the familiarisation with an initial implementational task that was suggested by Robin Rowe To implement binary fixed point as a data format What follows is a short explanation Colour data in CinePaint is stored as RGB triplets in four possible formats 8 bit unsigned integer with 0 0 0 representing bla
57. n terms of computation time since data in CinePaint is already stored in RGB If at a later point it was desired to do the same manipulation in other colour spaces such as CMYK one could convert loss free to XYZ using the colour management system carry out the calculations in XYZ and finally convert back to the original space Changing the colour balance towards a given hue then amounts to interpolating between the current colour of each pixel and the hue to shift towards If a pixel has colour 120 140 55 and is to be shifted towards blue at 0 0 255 the two values are interpolated giving 60 70 100 which is considerably bluer than the original value The interpolation is weighted by the saturation selected on the wheel as a measure of how far to shift towards the given hue The results above would have been for a saturation of 5096 A saturation of 25 would have resulted in 30 35 50 If the user decided to not keep lightness constant this already is the new pixel colour Otherwise the algorithm proceeds with the second step of adjusting the lightness of the new colour The goal is to assure that the lightness before and after the manipulation are the same RGB does not allow easy manipulation of lightness HLS and L a b however do in the L and L component respectively So original pixel colour and the new pixel colour are converted to one of the two spaces depending on the user setting The lightness of the new colour is then set 42
58. ndency Frequently Asked Questions about Color known as The ColorFAQ Charles A Poynton http people ee ethz ch buc brechbuehler mirror color ColorFA Q pdf Frequently Asked Questions about Gamma known as The GammaFAQ Charles A Poynton http people ee ethz ch buc brechbuehler mirror color GammaFAQ pdf The standard reference material on colours in computing Computer Graphics Principles and Practice Foley van Dam Feiner and Hughes 2nd ed Addison Wesley The standard book on computer graphics Chapter 13 is on colour and colour models It contains the HLS RGB conversion algorithm used for the implementation of the colour wheels colour corrector Colour Space Conversions Adrian Ford and Alan Robert 1998 www poynton com PDFs coloureq pd Descriptions of many colour spaces and colour models and conversions between them Color Order Systems RGB HLS HSB Gernot Hofffmann www fho emden de hoffmann hlscone03052001 pdf Introduces Hoffmann HLS a revised and presumably improved version of HLS For the application in this project no significant improvement could be observed however efg s Color Reference Library Earl F Glynn II http www efg2 com Lab Library Color index html Probably the most complete link and reference collection on colours colour theory and colours in computing that there is 57 Colour Management 10 Why Colormanagement James C King Adobe Systems Incorporated http www color org whyc
59. nted as a new image data format 7 2 Scope for future work The direction for future work on the colour management system has already been set The next step will be to handle embedded profiles An extension to printer colour management should also follow which would turn the current system into a full colour management system For the colour corrector there is a wealth of possible extensions The logically next step will most likely be the integration with the curves tool that is already hinted in its interface Importance should then be given to tackling the performance issue The most exciting colour manipulations become a lot less satisfactory if they cannot be previewed interactively This will however require thorough refactoring of the CinePaint rendering engine To give just one example Even though an image may be displayed on the screen at only a fraction of its real size every preview involves processing the full sized image data instead of a down sampled and adjusted version Other possible extensions to the colour corrector are split screen which allows to show half the image before and half the image after the transformation to enhance the possibilities of visual comparison an automatic matching feature The user selects two colour samples from two different images The algorithm will automatically adjust the colour balance of the two images such that the two selected samples match This can be used to adjust different
60. oddity that was encountered was that white is affected when L a b is used to preserve lightness Using HLS white is never affected if lightness is preserved White should be the lightest colour there is and cannot be changed without a loss of lightness Any correction that preserves lightness should therefore not affect white I therefore manually computed some example conversions between RGB and L a b It turns out that after readjusting lightness in L a b and converting back to RGB some RGB values can end up with an intensity greater than the maximum As an example the attempt to keep lightness constant might result in a colour of 210 200 260 where 255 is the maximum intensity In L a b this will correspond to a perfectly well defined colour In RGB however the 260 will be clipped to 255 which leaves the formerly white area with a blue cast Defining ways of handling such exceptional cases would further improve the algorithm but the perfect 50 colour correction algorithm would be a project of its own Giving the user the option to S8 e E between L a b and HLS seemed a good compromise O8 5 o During the tests performance turned out to be a serious drawback On an older machine like the PII 300 that development was done on the user can virtually watch how the preview is rebuilt block by block after each change of settings This is however a problem the colour corrector shares with other CinePaint tools an
61. olormanagement pdf 11 Inside Colormanagement The latest on what color management is how 1t works and what tools are available David Broudy Electronic Publishing 2001 http www rit edu lwlppr trends ColMan pdf Two introductory texts explaining the necessity and functionality of colour management systems to users I prefer King s text It has more depth Broudy s text is slightly polemic 11 Color Management and Color Science Introduction Norman Koren http www normankoren com color_management html Another introduction to colour management but with more focus on explanations and aspects of colour sciene A good link collection 12 Color Management Jonathan Sachs 1999 2000 Digital Light amp Colour http www dl c com Color Management pdf there is a space between Color and Management In depth explanation of a full colour management setup with focus on MS Windows and Digital Light amp Colour s Picture Window Pro 13 Colour Management Workflow Analysis S ren Winslow 2000 The Graphic Arts Institute of Denmark http www color org sorenwinslow pdf An analysis of how nine different newspapers incorporate colourmanagement into their daily work The ICC standard 14 File Format for Color Profiles Version 4 0 0 Specification ICC 1 2001 12 2002 International Color Consortium http www color org newiccspec pdf the official specification of the ICC profile standard 15 Building ICC Profiles the Mechanics
62. on This is because they have no high priority in movie studio applications An embedded profile can increase the file size noticeably A digitised movie consists of hundreds of thousands of individual files one per frame of the movie They all use the same profile since they were scanned using the same scanner It would be an incredible waste of space to embed a profile into every single one of them Instead you would keep one single profile in a separate file and define it as the default image profile Embedded profiles are useful in pre press applications to exchange individual images with a printer Having explained the display profile and the default image profile this still leaves the third profile setting in the preferences The workspace profile defines your preferred colourspace your in house colourspace Most professional users such as newspapers agencies or movie studios will do all their work in one fixed colourspace because standardisation greatly helps coherent workflows and avoids confusion 32 display profile image profile embedded profiles workspace profile default behaviour rendering intent assigning and converting The workspace profile is the profile that CinePaint applies to every newly created image and that it may enforce for opened image if you choose so As explained above CinePaint assigns a profile to a newly opened image according to the rule selected in the user preferences If this profile
63. ours from colours Artists however have long known that you can create new colours by mixing existing ones Biology discovered that the human retina contains three different kinds of colour sensors the so called cones that exhibit sensitivity to different wavelength with peak sensitivity at approximately red green and blue respectively the so called tristimulus theory From these three sensor responses the brain is able to create all possible colour sensations This suggests an attractive way of specifying colours as a weighted mixture of base colours as a colour recipe if you like take two measures of red three of The base colours are referred to as primaries and the most commonly used primaries are red green and blue due to how the HVS is built The process of defining a colour by a weighted sum of three primaries is called colour matching and the science around it is called trichromatic meaning three coloured 14 Color Ratio 300 400 500 600 700 800 Wavelength nm Figure 2 1 The rgb colour matching curves colorimetry All modern display devices such as TVs or monitors are based on this principle They mix colours from varying proportions of red green and blue Experiments were conducted to further quantify the responses of the three cones In these matching experiments a light source emits light of a single wavelength producing a coloured reference experiments patch on a wall Next to it three sources of red gre
64. out having to think about it Since colourmanagement causes a certain loss of performance it is at your discretion whether to use it or not Notice the toggle entry View gt Colormanage Display in each display s menu It allows you to switch colourmanagement on or off for each display individually You can change the rendering intent used by CinePaint for a particular display using the menu View gt Rendering Intent By default this is set to whatever you selected in the preferences The title bar of each display shows you which profile is currently assigned to an image If you want to assign a different profile you can do so by using the menu Image Assign ICC Profile You will be prompted for the new profile to be assigned If you choose to assign no profile colourmanagement will be switched off for all displays of this image Remember that CinePaint always needs to know an image and a display profile to be able to colourmanage Finally you can convert an image s data into another profile s colourspace Obviously this also requires a profile to be assigned to the image so that CinePaint knows where to convert from In contrast to ordinary colourmanagement which does not affect the image data this will actually manipulate it to transfer it to the new colourspace Again you are prompted for a profile to convert to and in this case also for a rendering intent to use in the conversion 33 4 3 Behind the scenes When colour management i
65. physical specifications a simpler colour specification tagged with a colour name and a list of flowers or insects that exhibit it Today the same principle is still widely used in printing and design where colours are often specified as references to standardised and often patented colour charts The pantone system is a common example Artists searched for more generic ways of describing colour in their quest to gain control over them and to use them for their purposes One such way of describing the visual sensation of a colour is by the three attributes hue saturation and lightness Hue is the most familiar attribute It refers to the basic quality of a colour It is what leads us to label a colour with a common name such as orange green or blue Saturation is a measure of a colours purity or intensity Saturated colours such as royal blue appear intenser whilst unsaturated colours such as sky blue appear paler Finally lightness expresses how light or dark a colour appears Both colourcharts and descriptions in terms of hue saturation and lightness are subjective ways of specifying colours based on visual impressions They are appropriate for art and design in a technical context however more precise specifications are needed Such specifications started to develop with the beginning of the scientific study of light and the the human visual system HVS by Newton in 1704 which eventually spawned a branch of physics known as colorimet
66. proprietary contexts of editing studios and software companies and the techniques that make one implementation of a correction tool superior stay secret CinePaint is the first package that is open and accessible while aiming at achieving the same high quality standard in image editing as professional commercial packages It makes them available to a wide audience offering new ways of artistic expression If it succeeds in its goal to provide high quality implementations of standard image manipulation techniques it will also be an interesting source for study A graphics package is a complex application and much can be learned from it At the moment that information is difficult to access 1 2 CinePaint CinePaint is used by a range of studios and played a part in the post production of films such as Harry Potter The Last Samurai Stuart Little II and 2 Fast 2 Furious It is a painting and image retouching program which started in 1998 as a branch of the GIMP package which was known as filmgimp It was however never built into a GIMP release and almost forgotten about In July 2002 Robin Rowe wrote an article for the Linux Journal linuxjournal com about its use at the Rhythm amp Hue film studios Subsequently Rowe decided to revive the project as a sourceforge project and became its project leader Filmgimp became CinePaint CinePaint focuses on professional features such as high colour depth or high dynamic range and on film specific features
67. r example the image exhibits a yellow colour cast you shift the handle away from yellow towards blue To allow exact adjustment you can enter values for Hue Saturation and Lightness into the three boxes underneath the wheel Lightness and saturation range between 0 00 and 100 0 Hue stretches from 0 00 to 360 00 Note that you can also resize the dialog which will result in the wheels growing to give you more precision in using the wheel handle if you need it The effect of each of the wheels can be switched on and off individually using the toggle button in the top left corner of each wheel which is labeled Sh Mid and Hi respectively The overall effect of all wheels can be enabled and disabled using the toggle button on the Colour Wheels tab You will notice that there is a Curves tab too This is still without function It merely indicates how colour wheels and curves should be integrated in the future to create a single power colour correction tool Clicking the Options button will open the options dialog see figure 5 2 which allows you to fine tune the algorithm that is used to apply your colour changes Changing the colour balance of an image is achieved by manipulating the intensities of individual colour components This is likely to result in an overall change of the lightness of the image Normally the algorithm will attempt to counteract any lightness changes apart from those explicitly requested by moving the lightness
68. r the pixels in its drawing area and calculating the appropriate RGB values based on the HLS colour model as defined by Foley van Dam et Al 6 First the wheel s centre coordinates and radius are determined from the width and height of the drawing area Afterwards coordinates can be translated such that the origin is at the wheel s centre Hue and saturation then simply correspond to the polar coordinates 0 and r such that translating from a position in the circle to hue and saturation is equivalent to translating from Figure 5 4 hue and saturation as polar coordinates 41 preserving lightness shifting the balance Cartesian to polar coordinates Figure 5 4 illustrates this If the resulting value for r is greater than the radius the pixel lies outside the circle and is ignored Otherwise r is scaled to a range of 0 100 to give the saturation 0 directly determines the hue Given hue saturation and the lightness determined by the lightness slider which is constant for all x y on the wheel we can convert the HLS values into monitor RGB values using Foley van Dam et al s algorithm The pixel is then coloured in the resulting colour To make the drawing more efficient and to avoid flickering the wheel is obviously drawn onto an off screen buffer that is copied onto the screen when a refresh is requested 5 3 2 Colour correction What the colour corrector essentially does is shifting the image s colour balance towards
69. rofile Adobe RGB En Workspace profile ColorMatch RGB m Display profile iiyamaA702H HT zd Default rendering intent Perceptual zd Default Behaviour When opening image Assign default image profile 4 When image space doesn t E match workspace Keep Agente a F Colormanage new displays by default OK Save Cancel describes the image data and the profile that describes the display device The display profile is selected directly via the setting of the same name The display profile describes the display device you are using such as your monitor and is a system wide setting It will be used in displaying all colour managed images CinePaint currently does not provide for systems using two different display devices at the same time The image profile is assigned to the image when it is opened In the preference dialog you can determine how this is done You can choose to always assign a selected default image profile to never assign a profile or you can choose to be prompted for the profile to be assigned whenever you open an image You can see which profile is currently assigned to an image from the description in the image s title bar In principle there is another way in which a profile can be assigned to an image by embedding it into the image file itself Several file formats such as TIFF and JPEG allow embedding of profiles However detection of embedded profiles is not provided for in the current implementati
70. rofiles The question of which format to use arose again in the design of look profiles As explained in the previous two chapters look profiles capture the colour corrections done to an image in order to preview them on other images or to easily apply them to other images There are two principal ways of achieving this The corrections could be captured in a LUT relating the original pixel colours to the new pixel colours This could work in just the same way it works for colour management profiles as described in section 3 1 1 The other option would be to store the inputs of the tools used to achieve the correction for example the settings of the colour wheels These inputs could then be used to reapply the same 26 littleCMS a storage directory profile cache choices abstract profiles data formats whiter than white computations to other images It seems obvious that the first method is likely to turn out faster the latter to be more precise Ideally there would and in case of CinePaint it is likely that there will be in the future both options at the discretion of the user Due to the limited amount of time for this project and the fact that look profiles are only one part of it this project adopts the first approach It is likely to be the more useful of the two Considering that look profiles will often be applied on the fly to preview a scene through a given profile efficiency is important Besides it ties i
71. rred to HLS model as HSV and HLS models there is room for confusion between the abbreviation HSV and the abbreviation of the human visual system HVS In this section I will therefore always spell out human visual system 17 Green Hue Saturation Cyan Yellow Gren Yellow al Saturation DE cd m He Cyan Ke 7 Blue A Red a B R Magenta Figure 2 2 Visualisations of RGB HSV and HLS space from left to right The models are best understood from their visualisations They can all be derived from the RGB model The RGB model is visualised as a cubic three dimensional space with one dimension for each red green and blue The coordinates refer to the varying mixtures of the three base colours HSV space can be visualised as a hexagonal cone The base of the cone is obtained from the RGB cube by looking at the cube from the corner that represents white as depicted in figure 2 2 and flattening it onto two dimensions Hue is measured around the perimeter of the cone Saturation is the distance from the centre and value decreases along the centre line towards the apex Figure 3 in the appendix shows a coloured picture of the HSV space HLS is then derived from HSV by grabbing white at the HSV base and pulling it up to form a double hexcone colour A common artistic tool based on the HSV or HLS colour model is the colour wheel The colour wheels wheel orders colours on a circle based on hue and saturation It is approximately a s
72. ry Physicists discovered that light fundamentally is electromagnetic waves with wavelengths ranging from 400nm to 700nm that most light contains a wide spectrum of different wavelengths and that it is the wavelengths that determine the colour impression They realised that there were physical correspondences to hue saturation and lightness The hue is the dominant wavelength in a colours spectrum Saturation corresponds to what is called the excitation purity of a spectrum measuring how little or how many different wavelength it contains Lightness arises from the luminance of a light source Luminance is the light energy a source emits weighted by a spectral weighting curve The weighting curve reflects the fact that the HVS perceives some wavelength as brighter than others A blue light will appear darker than a yellow light even though the light energy emitted may be the same Luminance and lightness are easily confused They are however not the same Luminance refers to the physical brightness of a light source lightness to how this brightness is perceived by the HVS This perception is non linear A light source with only 18 luminance of a reference light source will appear to the HVS as half as bright Since dominant wavelength excitation purity and luminance are physical quantities they allow a precise specification of a colour Unfortunately this is not practical since these quantities are hard to measure 2 2 Matching mixing col
73. s switched on the standard control flow as explained in chapter 2 applies The image data is sent through the profile assigned to the image and through the display profile The image profile translates the data to the profile connection space PCS which is either CIE Lab or CIE XYZ the display profile takes it from there to the display device s space Each GlImage structure contains a pointer to the image s respective profile The display profile is a global setting and stored in a global variable 4 3 1 Efficiency I The Profile cache A lot of effort in the implementation went into providing an efficient profile handling It is based on a two layered caching mechanism using hash tables and reference counters As a first layer profiles are cached No profile is opened more than once if it is used by more than one display The implementation provides a function to request a profile identified by its filename Internally profiles are identified by a combination of information from their header This assures that equality between profiles is detected even if identical profiles are stored in different files or in the future are embedded inside image files If the requested profile is found to be open already it is returned from the cache A reference counter keeps track of how often a profile is in use The counterpart is a function to return the profile to the CMS The reference counter is decreased and as it reaches zero the profile is closed
74. s the formats and libraries that were used in the design of this project and explains their choice It also gives some detail on the existing codebase of CinePaint Chapter 4 and 5 explain the work done in this project in detail Chapter 4 focuses on Colour Management whilst chapter 5 covers the colour wheels colour corrector and look profiles Chapter 6 outlines the steps that were taken to ensure the usefulness and correctness of the implementation of this project Chapter 7 finally summarises the achievements and sets the scope for possible future work The Appendix contains a few colour illustrations 11 12 ZBackground This chapter gives the necessary background on colour theory and colour models and explains why colour reproduction cannot be reliable without a colour management system It concludes with an overview of colour correction 2 1 Describing colours In order to be able to store information about colours and to reproduce colours from this information reliable ways of describing them are necessary The earliest approach to classifying colours was to use nature as a reference by naming colours after organic compounds or plants that exhibited them most distinctively This is still reflected in common colour names such as orange rose violet or azure The foremost colour researches of these days were natural scientists who developed colour atlases containing reference patches 13 colour charts colour attributes
75. shot that are cut into one scene e a visualisation of brightness ranges To visualise which parts of the image are considered as shadows midtones and highlights the pixels of the given range or highlighted or all other pixels are blacked out e a visualisation of the image s colour data For every colour that is contained in the image on point will be plotted in the appropriate position on the colour wheel This gives a good indication of whether there is a strong imbalance towards one particular hue in the image that may have to be corrected An appropriate extension or even replacement of colour wheels It has been mentioned before that colour wheels are colorimetrically incorrect Often this is irrelevant when they are only used as interface elements In some cases however they will give wrong 54 indications to the user An example was in chapter 6 3 Adjusting the highlights of an image towards blue will have a lot less effect than adjusting towards yellow if lightness has to be preserved The wheel however indicates that the adjustment was the same for all hues and thus conveys wrong ideas to its users that can lead to difficulties and unexpected behaviour An extension to look profiles was already mentioned in chapter 3 1 4 There should be the alternative of storing the exact input values to the manipulation functions instead of only sample data obtained from the image before and after the manipulation Their use would be slower
76. t a profile describing how to convert the RGB specifications in the image into XYZ This profile could for example be the profile of the scanner that the image was sampled on We then have a second profile describing the RGB specifications of our monitor The two profiles can be used to covert from image RGB to XYZ using the first profile and then from XYZ to monitor RGB using the second profile The two sets of RGB values before and after the complete conversion are likely to be different They are different because they take into consideration the differences between the scanner and the monitor but thanks to the conversion they correspond to the same XYZ colour specification to the same physical colour This is the essence of colour management 2 5 Visual Colour Models and colour wheels In practice there is a whole wealth of colour models many more than only RGB and XYZ Both RGB and XYZ are models oriented on hardware and physics They are not very intuitive for artistic use In the introduction we saw that artists often describe colours on the basis of attributes such as hue or lightness But given a certain yellow it is not obvious which of R G or B needs be adjusted to make it lighter Models that express such visual concepts more directly would be preferable Two common examples of such models order colours according to Hue Saturation and Value HSV and luminance and Hue Lightness and Saturation respectively They are commonly refe
77. t display all the colours you can see The set of colours a monitor can reproduce differs from model to model and the same colours might appear differently on a different monitor The same colour is most likely to appear differently on a printer This arises from the difficulty to communicate colours clearly Different devices such as printers or monitors use different ways of reproducing colours It is as if they were speaking different languages Still they all pretend to understand each other The result are many misunderstandings a more factual explanation can be found in chapter 2 Eventually however software developers realised that they had to provide solutions to the problem in the form of colour management systems CMS Apple was first to release its ColorSync www apple com macosx features colorsync colour management system in 199 as part of its operating system System 7 In 1993 the International Colour Consortium ICC www color org founded by several computer manufacturers Adobe Agfa Gevaert Apple Kodak Microsoft Silicon Graphics Sun Microsystems Taligent and FOGRA the German graphics arts research institute took up work with the goal to create and promote the standardisation of open colour management systems Microsoft introduced a first version of its colour management system with the release of Windows95 CM http www microsoft com whdc device display color icmwp mspx And finally in 1998 Adobe Photoshop 5 was t
78. tation of the HLS model can be seen from the correct display of the colour wheels in the interface The correctness of L a b is based on the correctness of littleCMS which it is provided by The models were further tested by observing their particularities as they are used to preserve lightness Their behaviour coincided with the expectations which was further confirmation for their correctness In many cases you will see that HLS causes the resulting colours to be too light such as in figure 7 of the appendix This is caused by the wrong assumptions about luminance that are made by the HLS model as was explained in chapter 2 5 There is a modified version of HLS called Hoffmann HLS 8 that is claimed to produce better results However implementing and testing it against ordinary HLS gave no noticeable improvements so Hoffmann HLS was disregarded That L a b models lightness more correctly can be best observed when correcting highlights Increasing blue has a lot less effect than increasing green see figure 8 of the appendix This is because increasing blue would make the image darker a lot quicker since intensities of blue are perceived a lot darker than equal intensities of red or green In HLS the effect is always the same This is an example where the colour wheel is inappropriate as an intuitive interface element Its even shape suggests that the magnitude of the effect was the same for any hue as much for yellow as for blue The only
79. they are merely mathematical constructions and do not correspond to real light sources One could not build a light that emits X Y or Z light Their definition is still useful because any combination of real light sources can be translated into XYZ components So they provide a way to identify a colour unmistakably and independently of any specific device or light sources However to actually display an XYZ colour it obviously has to be converted it into a mixture rendering of any real R G and B light sources This may mean that the conversion is not possible intents because the particular colour specified in XYZ cannot be matched by the red green and blue of a given monitor In this case a way of approximating the colour is needed This may mean clipping colours outside the gamut to the nearest colour inside Or modifying the colour trying to maintain certain qualities like its saturation The different approaches to approximation are known as rendering intents in the context of colour management To convert between RGB and XYZ a conversion function is needed This function depends on conversion the primaries of the particular device that is used for colour reproduction In colour with profiles management such functions are usually stored in a file called a colourmanagement profile Commonly two profiles are needed in a complete colourmanagement conversion assume we had an image with colours specified in RGB With the image we have go
80. tions take place on the fly the actual image data stays untouched only as it is sent to the display it is transformed There is a variety of factors that can cause performance issues Profiles are based on big multi dimensional look up tables LUTs that can take up significant amounts of main memory Sizes range from 200K up to megabytes each depending on the number of sample points stored in the table and the number of colour channels in the image Merging the profiles involved in a transformation into a device link is a time intense task e Finally transforming big amounts of image data in itself is an additional performance burden Fault tolerance is important because colour management relies on external components namely the profile files What happens if between two runs of the program the needed profiles have been moved or deleted for example by a diligent sysadmin who could not make sense of them CinePaint has to be able to deal with this by using fail safe defaults And finally the user interface had to provide enough flexibility to provide for a lot of different environments and needs Some users may want to use the same profile for all their images Other users need different profiles and will have to decide on an image by image basis The images provided may not be available in the colourspace they are needed in which makes conversion necessary The difficulty was to define a workflow that was flexible enough without
81. to the implementation of look profiles described in chapter 5 The flipbook allows you to concurrently load a sequence of numbered images that are thought of as the frames of film to move through them and to play them like a film sequence 28 BFP Glmage amp GDisplay flipbook If a sufficient number of management layers are superimposed on top of each other it can be assured that disaster is not left to chance Norman Augustine US Author 29 Implementation I Colour Management This chapter and the following one describe the work done in this project in detail Whilst this chapter deals with colour management the next one will cover colour correction Both chapters follow the same structure They start out by pointing out the major challenges that were faced in the implementation of this project They then describe the implementation from a user s point of view much in the style of a user s manual They conclude by describing the implementational details 30 efficiency fault tolerance user interface storage path 4 1 Challenges The three central issues to resolve in the implementation of colour management were efficient handling of profiles fault tolerance and an accessible and understandable user interface Efficiency probably was the single most important issue This concerns efficiency both in terms of runtime and in terms of memory consumption It is particularly important because colour transforma
82. ture permits to reduce the task of colourmanagement to a single transformation of the display buffer such that afterwards the buffer contains the corrected data The speed increase compared to transforming every time data is copied from the buffer onto the screen is enormous Introducing one buffer for each display increases memory consumption but it is certainly made up for by the improved performance Besides opening more than two or three displays of the same image is a very uncommon usecase After all the space on the screen to view several displays at the same time quickly reaches its limits 4 3 3 Fault Tolerance Fault tolerance mainly refers to the CMS s resistance to stumbling over lacking or non defined profiles The user may delete or move profiles that were set as display or default image profile He might decide to not assign a profile to an image All this will make colourmanaging an image impossible Therefore the CMS examines the settings carefully each time a profile is needed It provides an initialisation function that is called from CinePaint s main initialisation function that sanity 35 checks all settings and deletes settings that have become invalid such as profiles that do not exist anymore The only hazard that CMS will stumble across is deleting an opened profile during the execution of CinePaint because profiles are only loaded into memory when needed This was not considered a serious problem however After a
83. ur wheels are a common colour correction tool Their basic structure differs little from implementation to implementation There are almost always three wheels There is almost always an alternative between visual input using the wheels and numerical input The day spent colour grading at the BBC brought a few interesting new insights in colour correction there are often no easy recipes to follow Which tool works best differs from situation to situation This led to the conclusion that it should be desirable to have options to fine tune a tool rather than a single predefined setting which the developer considered best Especially in cases where it is impossible to define a best option This led to the design of the option dialog It predefines the options that seem most suitable in most cases but leaves the user the freedom to fine tune as special cases arise Interface features that the BBC editor used a lot were 1 Comparing the image before and after the manipulation 2 Entering values numerically The colour wheel is a very intuitive tool but sometimes precisely the same settings as before are needed to see whether they will work again 3 Quickly applying an identical manipulation to other images without having to click through a forest of dialogs Point 1 is satisfied in the interface design by being able to toggle the overall effect as well as the effect of each individual wheel Point 2 is provided for by the numerical input fields
84. ure of a Colour Management System 25 is important for CinePaint which supports all three platforms Last but not least it is highly optimised and very small with a footprint of only 100K The core of littleCMS consists of functions to open profiles to create profiles and to apply profiles to a buffer of image data To apply a profile it has to be turned into what littleCMS calls a transformation which is essentially an ordered list of profiles If this transformation is then applied to image data profile by profile is taken from the list and applied to it If for instance we wanted to convert image data from scanner RGB to monitor RGB we would first create a transformation consisting of the scanner and monitor profile and then apply this transformation to the data LittlleCMS can make transformations more efficient by first melting the profiles in the transformation into a single profile into what was previously described as a device link That way only one computation per colour is necessary not one per colour and profile 3 1 3 Handling many profiles Storage and Caching Even though littleCMS as the CMM provides functionality for reading and applying profiles it is the colour management system s task to manage the profiles and to do so efficiently They will have to be stored in a place where CinePaint can find them Since many of them contain big look up tables that occupy considerable amounts of memory access will have to be made
85. ves tool in mind Integrating the two will equip CinePaint with a single powerful colour correction tool Finally the colour correction tool will be complemented by look profiles They will provide the possibility to store manipulations in a separate file called a look profile and allow previewing images through the profile without affecting the image data At the same time they will allow committing the manipulations to the image data meaning permanently changing the image data in accordance with the profile They should be seen as the first step for CinePaint towards providing its own full Look Management 1 4 Results The system specified above has been fully implemented The display colour management was implemented particularly efficiently using a profile cache and will form a robust foundation for further developments in the area such as printer colour management The exact requirements for the colour wheels were difficult to establish which led to the development of a range of smaller variations and user options room for interesting comparisons and fine tuning Look profiles were implemented as anticipated 1 5 Remaining Chapters The remainder of this report is divided into chapters as follows Chapter 2 gives the necessary background on colour theory and colour models and explains why colour reproduction cannot be reliable without a colour management system It concludes with an overview of colour correction Chapter 3 describe
86. ving causes the second problem f a colour can be reproduced on one monitor it does not mean it can be represented on any other monitor This is a very common problem in printing Printers use a colour model that is altogether different from monitors a substractive instead of an additive model I will not go into details since they are not relevant here The result is that many colours you can see on your screen can never be printed the same way Finally non linearity of device input signals influences colour appearance An example is a monitors gamma Applying red green and blue input signals to a monitor usually assumes that the relation between input value and brightness was linear If an input of 255 corresponds to maximum brightness 127 should be 50 brightness 64 should be 25 It is not The relation between a monitors input signal and the generated brightness is a power function The numerical value of the exponent of this power function is called the monitor s gamma Gamma for a typical CRT is around 2 5 which gives approximately intensity input This means that an input of 127 5096 actually only corresponds to a brightness of 1796 Evening out all these problems is the task of colour management 2 4 Colour Management The first step in tackling this is finding an appropriate colour space The RGB space is obviously inadequate because colours specified in this space depend on the device they are reproduced on Such gamuts
87. were made to receive feedback on the CinePaint mailing list in particular on the part of colour management which is hardest to test due to the difficulty of creating reliable profiles These attempts had little success however The chapter proceeds by outlining the tests that were done to assure the correctness of the implementation In the case of colour correction this includes a comparison of the alternatives in the implementation namely Gaussian curves and Bernstein polynomials and HLS and L a b 47 the cache cms in use 6 1 Colour Management 6 1 1 Interface Justification To develop an understanding of what the user interface had to provide existing systems with colour management functionality Apple ColorSync Adobe Photoshop and Scribus and a study on colour management workflows served as foundations 13 The definition of an image workspace and display profile has become a de facto standard that is used in all colour management systems The options for default behaviour are mainly inspired by Adobe Photoshop The fairly large number of settings to define the default behaviour may appear redundant at first They do however allow the maximum flexibility Two examples may illustrate this Example 1 The user has a whole range of images that she knows to be in colourspace X Her working colour space however is Y and she would like the images to be in Y before she edits them She selects X as her default image profi
88. yond white This first task gave a good overview of the core code and produced an encouraging first result 3 3 CinePaint internals A few short remarks about the CinePaint structure should be made to aid the further understanding of the implementation The different data formats have already been mentioned Unfortunately at the moment they are not implemented very efficiently The introduction of a new format requires rewriting almost all functions that manipulate image data The part of this project concerned with colour correction will therefore focus on 8 bit unsigned integer as a format and the extension to other formats will be left for the future Image data in CinePaint is currently always stored as RGB data or optionally as RGBA where A the alpha channel stores the opacity of a pixel Two of the most important data structures of CinePaint with respect to this project are GImage and GDisplay Glmage represents a physical image It contains its dimensions image data and other information GDisplay is a view onto the image displayed on screen Several GDisplays can refer to the same GImage For example the user might want to display the same image twice at different zoom levels or with different display filters applied The physical image behind the two views however is the same and manipulations of the image in one display will automatically affect the other CinePaint provides a feature called the flipbook which has significance

Colour management and colour correction in

Contents

Download Pdf Manuals

Related Search

Related Contents