argyllcms/gamut/gammap.txt

Author: Graeme W. Gill
Date:   2000/10/28
Updated: 2006/1/17

Discussion of the gamut mapping algorithm used in Argyll:


	Jam Morovic provides an extensive summary of previous Gamut Mapping Algoriths
	(GMA) in his thesis "To Develop a Universal Gamut Mapping Algorithm".

	Mark Fairchild and Gustav Braun also discuss some interesting aspects of
	gamut mapping in "General-Purpose Gamut-Mapping Algorithms: Evaluation
	of Contrast-Preserving Rescaling Functions for Color Gamut Mapping".

	One thing that is striking in reading the current Gamut Mapping
	literature, is the schism between gamut mapping, and gammut clipping.

	A number of studies have indicated the satisfactory results from performing
	gamut clipping by mapping out of gamut points to the closest point
	within gamut, using a minumim delta E criteria. Ideally this would be
	in a perceptually uniform space, and in practical spaces (ie L*a*b* space),
	it appears a mapping that weightes luminance errors twice as much as
	hue and saturation errors is the most satisfactory
	(ie. Katoh & Ito 1996, Ebner & Fairchild 1997).
	This approach makes perfect sense if the perceptually linear
	delta E spaces is working as its should :- providing a means of
	computing the minimal perceptual error between two colors.

	Almost all the gamut mapping algorithms on the other hand, take a
	completely different approach. They are dominated by algorithms with
	appeal for their mathematical and algorithmic simplicity, rather than
	any objective sense. Clipping along a line of constant hue, in a
	particular direction, with some sort of compression along line
	length curve seems to be a favourite.
	My own experience with such an approach used for gamut clipping
	shows up the logical flaws in this approach.  In printed media, for
	instance, the yellow dye tends to have a very sharp cusp in (say) Lab
	space.  When mapping along a line towards the neutral axis, it is almost
	impossible to come up with a choice of line direction that
	doesn't severly distort the clipped color. It is often extremely
	de-saturated at the point the line hits the gamut boundary. Line
	directions that take some account of the equivalent yellow cusp
	in the target gamut improve the situation, but it seems highly
	illogical to choosing an algorithm that gives a result with such
	high delta E's, measured in any colorspace !

	My conclusion is this: If we are working in a good perceptually
	uniform space (ie. L*a*b*, Luv or CIECAM02), then we want to minimise
	the delta E of the gamut mapped points at all times (this is the point
	of using a perceptually uniform space !). This means that points
	on the source gamut surface, should map to the target surface in
	a very similar way to which they would be mapped using gamut clipping,
	ie. minimum delta E. The distinction between gamut mapping and clipping
	only becomes evident then, when we consider points that fall within
	both gamuts. For pure gamut clipping, these points should not be
	modified. For gamut mapping they should be modified in a way that
	trades off increased absolute delta E, for reduced delta E relative
	to surrounding colors.

	[ Saturation intent brings some factors other than minimum
	  delta E into the picture, but minumum delta E is a good
	  starting point. ]

	Gamut clipping should not be something separate to gamut mapping,
	but should just be one extreme in a continuum of gamut mapping
	choices.

	A consideration revealed by Jam Morovic's work, is that it may be
	desirable to treat colors on and close to the neutral axis a little
	differently that those saturated colors near the gamut surface.
	It seems desirable to align and compress the neutral axis to
	give a good gray-scale mapping, as well as preserving the
	relative colorimetry of low saturation colors.

	Mark fairchild's work also indicates the desirablility of
	trying to maintain the relative contrast of image data after
	compression. This can be achieved by using a sigmoidal
	mapping during compression (often called a soft knee compression
	characteristic), rather than linear compression. Gamut
	clipping can be considered to be an extreme example of
	knee compression, where the knee is "hard".

	A final consideration is how the various user intents are going
	to be accomodated. One of the nice features of a consistent
	clipping/compression approach is that many of the distinctions
	between different intents seems to disappear.

 Description of algorithms used:

	The algorithm chosen here is (as far as I am aware) a new one,
	that tries to combine all the considerations made above.

	The gamut mapping is divided into two parts, mapping the luminance
	axis orthogonally to other considerations, and then dealling
	with any remaining gamut compression or expansion.

	The basic transform mechanism is broken down into three part;
	Aligning the neutral axes, mapping the luminence coordinate,
	performing an overall 3D -> 3D mapping to compress and/or
	expand the gamut surface. The rotation uses a matrix, the 1D -> 1D
	luminence mapping uses a RSPL regular spline that is available
	in Argyll, and the 3D -> 3D mapping also uses a RSPL.
	By controlling the number, extent and strength of the sample
	point mapping used to create the RSPL, it is possible to arrive
	at a smooth mapping that has (controllable) areas of influence
	between different regions, and can achieve sigmoid like "soft"
	clipping curves.


    Aligning the neutral axes.

	The first step is to align the neutral axes. If the
	colorspace used is an appearance space (Such as CIECAM02 Jab),
	then the white points will already be close together, but may not
	exactly cooincide. The black points will also probably differ.
	There are a number of options for dealing with this. One
	option would be to assume that people adapt to black points
	the same way they adapt to the white points of a colorspace
	or image. Research and experiemnce indicates that this
	may not be the case though. If we assume that people
	do not in general adapt to the black points of a colorspace,
	then a consequence is that it is difficult to fully exploit
	the full dynamic range of the colorspace, since a
	source black that is misaligned with the destination black
	will probably not be able to achieve as low a J value,
	and this can noticably affect the percieved quality of
	an image.

	The compromise I deemed the best in Argyll, is to
	assume that people do not adapt to black points,
	and that therefore only the white points should be
	aligned by rotating the source space around 0,0,0
	to line them up. The minimum J value on the other hand,
	is mapped as if the black points were being fully
	adapted to, and at the point that the source
	neutral axis would leave the destination gamut,
	it is clipped to the destination.

	This gives a neutral axis in the destination that looks
	the same as that of the source, while fully exploiting the
	dynamic range of the destination colorspace. Since
	the departure of the mapped neutral axis from neutral
	only occurs in very dark colors, its deviation is not
	usually visible.

	[ The code allows for 4 possible black point/neutral
	axis mapping algorithm, set by the value of bph:
    Adapt source black point to destination, giving full black range.
    Don't adapt black point to destination, giving compromised black range.
    Don't adapt black point, bend it to dest. at end, giving full black range.
    Don't adapt black point, clip it to dest. at end, giving full black range.
	]

	Note that the black point mapping is performed by manipulating
	J independently of the color components (a & b). In adapting
	color from one space to another, this proves to be the correct
	approach, maintaining the appearance of an image when
	transformed to different device colorspaces.
	This should not be confused with the situation where
	optical flare or haze is being removed or simulated,
	in which case the black point mapping should be done in a
	linear light colorspace (such as XYZ), and will have an
	effect simulaniously on lightness and saturation.


	Mapping the luminence range.

	The luminence mapping is acheived by created a small number of
	mapping points, and then created a detailed RSPL (Regular Spline)
	based on the points. The end points (which are heavily weighted
	to ensure the overall range mapping is useful) are created
	simply from the known J values of the source and destination
	white and black points (taking into consuideration what
	J values will result from possible gamut boundary clipping).

	Middle points are introduced to either give a linear mapping (in J),
	or to introduce a slight range compression aimed at preserving
	the contrast ratio of the image. 


	Maping 3D gamut points.

	The mapping points created for determinining the 3D compression or
	expansioin are created from the source gamut surface verticies
	that lie outside the destinatio gamut. The raw mapping for the
	outside of the gamut is created by itterative optimisation,
	that strikes a weighted balance between three different
	factors: Mapping to the closest point on the destination surface,
	Mapping in a way that is smoothly consistent with neighbour
	mapping rays, and Mapping radially to the center of the colorspace.
	By manipulating the weighting factors (held in a gammapweights
	structure), it is possible to control to some degree the
	nature of the gamut mapping. 

	Some compensation is applied to the vectors to try and counteract
	the effects of the vector and (subsequent) RSPL smoothing.
	This smoothing may cause the mapping to end up outside the
	destination gamut, so we try and make sure that such smoothing
	effects over rather than under-compress.
	
	In addition to the suface mapping points, the 1D neutral axis 
	mapping is maintained by introducing a string of points
	down the 3D neutral axis, to ensure that the 3D surface mappings
	do not alter it. 

	Closer to the surface, optional "knee" mapping points
	are also added, in an attempt to cause the inner
	3D mapping to be more linear, with the compression (or
	expansion) being concentrated near the surface of the
	gamuts ("Sigma" compression).
	
Limitations and future challenges:

	After re-working the gamut mapping in December 2005/January 2006,
	the following challenges remain:

	1) For saturation mapping in particular, the handling of hue
	mapping using the "nearest" weighting isn't very flexible.
	It's not possible to map RGB Cyan to CMYK Cyan without
	introducing unacceptable side effects.

	To solve this, I would have to add some extra mapping controls
	and gamut information. For each gamut the 6 major "cusps" would
	have to be identified (either directly from the colorant
	combinations, or indrectly by locating the point on the gamut
	surface with the largest C (?) in each hextant), and
	then a means of warping the hue angle (and Luminence ??)
	in each hextant. 

	The gammapweights structure would be expanded to add a weight
	as to how much the source hue cusps would be distorted.

	[Added to the code by Feb '06]


	2) There is not much flexibility in what the gammapweights
	weightings can achive. Some controls simply don't work
	very well. Often increases in saturation in one area (at
	the transitions between major colors) are bought at the
	expense of the middle of the major colors becomming
	very "flat" and desaturated. This is generally when we
	are in a region in which the source gamut is within the
	destination gamut. 

	3) Geting smoothness and saturation is very difficult
	for some reason. The "relative error" weights do spread
	things out, but at the cost of obvious banding (why ? -
	possible answer - reflects the roughness of the gamut surface.
	Solution is to use better algorithm for gamut surface extraction ?? ).
	Increasing the 3D RSPL smoothness seems to have no
	effect (why ?). Reducing absolute chroma weight improves
	smoothness, but at the cost of reduced saturation and increased
	lightness.