1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
|
<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<title>Argyll Overview</title>
<meta http-equiv="content-type"
content="text/html; charset=ISO-8859-1">
</head>
<body>
<h2><u>Source Code Documentation</u></h2>
Some general comments about how different software works.<br>
<br>
<b>TARGET</b><br>
<br>
Nearly all current Color correction systems generate test charts
(or device characterization target charts) by laying out a regular
rectangular grid of test points in device space (Targen will do this if
you feed it a non-zero <span style="font-weight: bold;">-m</span>
option). On some consideration, this approach is far from optimal. Not
only is a regular grid inefficient in packing the multidimensional
device space but if the points are spaced evenly in device space, they
will probably not be optimal in determining the underlying device
characteristic, sampling too finely where the device behavior is
smooth, and too coarsely where the device behavior changes rapidly.
Some commercial color systems tackle
the latter problem by "pre-linearizing" the device, which amounts to
distorting
the regular device space grid points with a perceptual inverse per
device
channel lookup curve.<br>
<br>
The approach I have taken with Argyll, is a little different. Both a
device independent and device dependent approach are available. In the
device independent mode, test points are distributed so as to minimize
the distance from any point in the device space to the nearest test
chart value. When the device dependent approach is used, test points
are chosen so as to minimize the expected error between the resulting
profile and the underlying device response.<br>
<br>
For higher dimensional spaces, where the aim is to create a more
approximate device profile, I've used an "incremental far point" point
generator, that starting with a previous test point as a seed, use a
minimization algorithm to locate another point that is as far as
possible from the nearest existing test point in perceptual space,
while remaining in gamut at all tines. This means that ideally each
point "fills in" the gaps in the existing distribution.<br>
<br>
Another issue with laying test points out in regular grids, is that
this means that the device response is poorly sampled (since the grids
are usually coarse), and this can make it impossible to create detailed
device linearisation "shaper" curves from the resulting data ! Ideally,
in any colorspace (input or output), when viewed from any possible
angle, none of the test data points should appear to line up. The
Argyll target generator seems to achieve this goal.<br>
<br>
<br>
target/printtarg.c:<br>
<br>
printtarg simply generates a PostScript file containing the patches
laid out for an Xrite DTP51/DTP41/SpectroScan. It allows them to be
laid out on
a choice of paper sizes, with the appropriate contrasting color spacers
between
each patch for the strip reading instruments. Unlike other charts,
Argyll
charts are generated as required, rather that being fixed. Also unlike
most
other strip reading charts, the spacers may colored, so that the
density contrast
ratio is guaranteed, even when two patches are about 50% density.<br>
<br>
Another feature is the pseudo random patch layout. This has three
purposes. One is to try and average out any variation in the device
response in relationship to the location of the patch on the paper.
Color copiers and printing presses (for instance), are notorious in
having side to side density variations.<br>
<br>
Another purpose of the random patch layout, is that it gives the
reading program a good mechanism for detecting user error. It can guess
the expected values, compare them to the readings, and complain if it
seems that the strip is probably the wrong one.<br>
<br>
The final purpose of the random patch layout is to optimize the
contrast
between patches in a strip, to improve the robustness of the strip
reading.
Using this, small charts may be even be generated without any gaps
between
the test patches.<br>
<b><br>
RSPL</b><br>
<br>
A lot of core Argyll algorithms are bound up in the RSPL (Regular
Spline) class. The RSPL class is closely related to the CLUT table
structure used in the ICC profile format, and is a general purpose way
of mapping one multi-dimensional value to another, using interpolation
to reduce the mapping function to a manageable size.<br>
Regular splines are not directly related to other types of splines, and
do not generally suffer from the sort of "wiggles" and "overshoot"
characteristic of other spline systems.<br>
<br>
The RSPL class adds three basic methods to the underlying data
structure:<br>
<br>
1) Interpolation<br>
<br>
2) Creation from scattered data<br>
<br>
The regular spline implementation was inspired by the following
technical reports:<br>
<br>
D.J. Bone, "Adaptive Multi-Dimensional Interpolation Using Regularized
Linear Splines," Proc. SPIE Vol. 1902, p.243-253, Nonlinear Image
Processing IV, Edward R. Dougherty; Jaakko T. Astola; Harold G.
Longbotham;(Eds)(1993).<br>
<br>
D.J. Bone, "Adaptive Colour Printer Modeling using regularized linear
splines," Proc. SPIE Vol. 1909, p. 104-115, Device-Independent Color
Imaging and Imaging Systems Integration, Ricardo J. Motta; Hapet A.
Berberian; Eds.(1993)<br>
<br>
Don Bone and Duncan Stevenson, "Modelling of Colour Hard Copy Devices
Using Regularised Linear Splines," Proceedings of the APRS workshop on
Colour Imaging and Applications, Canberra (1994)<br>
<br>
see <http://www.cmis.csiro.au/Don.Bone/><br>
<br>
Also of interest was:<br>
<br>
"Discrete Smooth Interpolation", Jean-Laurent Mallet, ACM Transactions
on Graphics, Volume 8, Number 2, April 1989, Pages 121-144.<br>
<br>
3) Inversion<br>
<br>
Simplex interpolation is normally done in Baricentric space, where
there
is one more baricentric coordinate than dimensions, and the sum of all
the
baricentric coordinates must be 1.<br>
To simplify things, we work in a "Simplex parameter" space, in which
there are only <dimension> parameters, and each directly
corresponds with a cartesian coordinate, but the parameter order
corresponds with the baricentric order.<br>
For example, given cartesian coordinates D0, D1, D2 these should be
sorted from smallest to largest, thereby choosing a particular simplex
within a cube,
and allowing a correspondence to the parameter coordinates, ie:<br>
<br>
D1 D0 D2 Smallest -> Largest cartesian sort<br>
P2 P1 P0 Corresponding Parameter coordinates
(note reverse order!)<br>
<br>
B0 = P0 Conversion to Baricentric
coordinates<br>
B1 = P1 - P0<br>
B2 = P2 - P1<br>
B3 = 1 - P2<br>
<br>
The vertex values directly correspond to Baricentric coordinates,<br>
giving the usual interpolation equation of:<br>
<br>
VV0 * B0<br>
+ VV1 * B1<br>
+ VV2 * B2<br>
+ VV3 * B3<br>
<br>
where VVn is the vertex value for each of the 4 vertices of the simplex.<br>
<br>
Reversing the Parameter -> Baricentric equations gives the<br>
following interpolation equation using Parameter coordinates:<br>
<br>
VV0 - VV1 * P0<br>
+ VV1 - VV2 * P1<br>
+ VV2 - VV3 * P2<br>
+ VV3<br>
<br>
It is this which is used in rev.c for solving the reverse interpolation
problem. Within a simplex, only linear algebra is needed to compute
inverses. To deal with the 4D->3D nature of a CMYK profile, the SVD
(Singular Value Decomposition) approach is used to solving a CMYK for a
given CIE value, the result being the equation of the line of
solutions. The lines from adjacent simplexes form a solution locus,
that can be resolved into a single point solution once a black
generation rule is applied.<br>
<br>
Outline of inversion processing:<br>
<br>
Basic function requirements: exact,
auxil, locus, clip inversion.<br>
Fwd cell - reverse cell list lookup<br>
Basic layout di -> fdi + auxils + ink limit<br>
Basic search strategy<br>
Sub Simplex decomposition & properties<br>
How each type of function finds solutions<br>
Sub-simplex dimensionality &
dof + target dim & dof<br>
Linear algebra choices.<br>
How final solutions are chosen<br>
<b><br>
XICC</b><br>
<br>
Profiling using concatenated shapers to augment thin
plate splines for per channel curves.<br>
<br>
<b>IMDI</b><br>
<br>
This module generates C code routines which implement an integer
multi-channel transform. The input values are read, passed through per
channel lookup tables, a multi-dimensional interpolation table, and
then a per channel output lookup table, before being written.<br>
<br>
<b> MPP<br>
</b> <br>
Outline model parameter optimization using slope accelerated error
metric, and initial parameter speedup for adjacent spectral bands etc.<br>
<br>
<br>
<br>
</body>
</html>
|
