1980 - Buchsbaum. A Spatial Processor Model For Object Colour Perception
1980 - Buchsbaum. A Spatial Processor Model For Object Colour Perception
Nomenclature
L(A) illuminant
R(A) reflectance
{k(A)l?=l
three dimensional colour space basis that spans L(A) as a
linear combination of the basis members
{C(h)l?=3=I
three dimensional colour space basis that spans R(A) as a
linear combination of the basis members
a=(cqa*aJT vector of coefficients of the linear combination representing
the reflectance, R(A) = C?= f airi
"=(vl%%)T
vector of coefficients of the linear combination representing
the illuminant L(A) = CFeI viZi(A)
C,(A)(i = 1,2,3) visual system trichromatic colour mechanisms
Y = (YIY2YJT vector of trichromatic responses corresponding to a single
subfield in the scene yi AJ,,, C,(A)L(A)R(A) dh (vs = visual
spectrum)
z= (z*zzzJT vector of trichromatic responses corresponding to the entire
field. zi AJs,, C,(A)z(A).C, b,,,R,(A) dh, where b, is a
weighting factor associated with the m th subfield
h = (h,h,h,)* internal system standard of field average
Kijk ’ j,,Ci(A)lj(A)rk(A) dA
D(v) 3 X 3 matrix, the entries are defined by dik(v) = ~~=, ViKijk
Q(h) 3 X 3 matrix, the entries are defined by qij(h) = ~~=, hkrcijk
f,(A)(i = L&3) photographing filters defined in the projection experiment
+ = (&dW$JT 4 AJ5,, fi(AK(A)R(A) dA
pt(A)(i = 1,2,3) projection filters defined in the projection experiment
I. Introduction
The large degree of independence of perceived object colour on the il-
luminating spectrum was already noted by Helmholtz in 1866. He made a clear
distinction between surface (object) colour and film colour. The first term
refers to colour as a property of a surface and constitutes most of natural
colour experience. The latter term corresponds to colour perceived as an
expanse of light in a two-dimensional space without objective reference. The
term colour-constancy is used in this context to define the ability of the visual
system to estimate an object colour transmitting an unpredictable spectrum to
the eyes. Some of the best known earlier explanations will be mentioned here
briefly. However, before going into further details an important comment
should be made. The term constancy is somewhat misleading. The definition
(from Beck (6)) is: the approximate constancy of the perceived colour of an object
despite changing illumination that alters the intensity and spectral composition
of the light stimulating the eye. What this means is that it is not constancy that
we are dealing with but rather inconstancy. The different accounts, theories,
explanations and quantifications are for the changes in perceived hue and other
attributes that are caused under different illuminations and context conditions.
Helmholtz (1)explained the phenomenon attributing surface colour percep-
tion to the presumed ability of the visual system to discount the illuminant. The
interpretation of this property in the colour space is that it represents a shift of
the achromatic point defined by the average reflectance of the scene. In other
words, the scene itself establishes a reference for the visual system. No exact
knowledge of the illuminant is needed, the context of the scene is important.
The problem of how the visual system arrives at object reflectance without
prior knowledge of the illuminant is one of the central issues in the various
approaches. The implicit estimate of the illuminant on that basis is paramount
in the present theory and a very significant issue in the discussion.
One of the outstanding formulations is Helson’s (10,11)adaptation level
theory. The major assumption is that colour is detected with respect to a single
reference level common to the entire complex field, that is referred to as an
adaptation leuel. The adaptation level is the perception that corresponds to an
average gray and is a weighted function of the reflectances constituting the
complex visual field. The adaptation level is different for different scenes. The
very substance of Helson’s explanation is stated in the following principle:
samples whose reflectances are above the adaptation level increasingly take on
the hue of the illuminant; samples below the adaptation level increasingly take
on the hue of the complementary afterimage of the illuminant, and samples at
A Spatial Processor Model for Object Colour Perception
and vice versa) a reversal condition appeared. That is, red appeared green, etc.,
although the reversals were not necessarily the complementary hues.
The results of Land and his associates are mentioned here in detail because
they provide an extensive experimental basis for the model we present later.
Land found it necessary to introduce a new theory-the Retinex theory of
colour vision-to account for his results. The novelty is that the visual system
incorporates in addition to the classical filter transformations, a strategy for
processing colour. The system incorporates certain normalization and scaling
algorithms that allow it to correctly detect hue independently of the transmit-
ted spectrum. The Retinex theory does not require the ability of the visual
system to compute the illuminant or make assumptions on its spectral composi-
tion. In an additional series of experiments, the Mondrian experiments, Land
(19) and McCann et al. (22) clearly demonstrated the constancy of perceived
hue under different illuminations giving an account in terms of the Retinex
principles. In the Mondrian experiment, a scene composed of different col-
oured rectangles similar to a Mondrian painting was illuminated by coloured
projectors. Technically, it was arranged that, in different presentations, differ-
ent rectangles having different reflectances will transmit to the eye the same
radiance as measured through fixed filters corresponding to the retinal colour
mechanisms. The subjects in these demonstrations were able to detect the true
colours of the rectangles. An important result is that when the rectangular
subfields in question were viewed through a narrow tube isolating them from
their context they did not retain their object hues (19). Experiments concerning
surface brightness constancy and effects of spatial arrangement are reported by
Wallach (29) and earlier [reviewed in Beck (6)] and in recent demonstrations,
see Gilchrist (7,s). Land’s results raised a fierce debate in scientific literature,
see Judd (15), Land (18), Walls (30) and Wheeler (31), the basic issue being
whether a new theory of colour perception was needed. Jameson and Hurvich
(12) attribute many of the results to simple effects of colour-contrast. We will
not enter into this debate here, but rather leave room for it in the discussion
section. However, we should state here that our view is that a large gap in
understanding underlies each of the different approaches discussed. A perspec-
tive from the systems engineering point of view may clarify the difference
between a geometric transformed presentation and an algorithmic processing
approach. Models of visual processing that incorporate spatial characteristics
of the human visual system and relate visual models to current views on
image processing are presented by Hall and Hall (12) and Stockham (27). The
theory we present lacks, as yet, the back-up of appropriate experimental
data quantified in terms of its parameters. Therefore, although it presents
mathematical derivations and calculations, it is qualitative in nature. But it
incorporates and employs principles of communication engineering that allow
some generalizations and possible narrowing of the differences and gaps among
the views. It demonstrates that under some stated general conditions concern-
ing the scene and the illuminant, colour constancy is achieved. This provides a
unified understanding for many results, particularly those of Land, and offers a
basis for a comprehensive quantititive theory of object colour.
A Spatial Processor Model for Object Colour Perception
The basic assumption of the theory presented here is that the central visual
processor is recognizer of patterns supplied by the priphery. Given a basis in a
three-dimensional colour space and an internal standard field-average, the
system arrives at the illuminant spectrum which is subsequently used to
compute the reflectance of each subfield in the complex scene. The general
idea of pattern and template matching as a basis to account for other colour
vision phenomena has been demonstrated earlier [Buchsbaum (3), Buchsbaum
and Goldstein (4,5)]. The present report is in line with this general approach.
h(A)= f ai,ri(A),
i=l
FIG. 1. General block-diagram scheme of the model for surface colour estimation. The
stimulus as specified in the text and described at the left side of the figure feeds into
three colour filters. These colour filters constitute the basic trichromatic unit channel for
the model. Different subfields may stimulate different such colour channels according to
their spatial arrangement, that are, however, of the same wavelength characteristics.
Phase 1 of the processing exploits a weighted field average of the subfield to arrive at the
illuminant estimate 8. This is done under the assumption that the entire field has a
standard average h, that may be different from its true average a. In the second phase of
processing, the response of each single subfield together with the estimated illuminant 8
[are evaluated to get an estimate of the subfield reflectance &I.
A Spatial Processor Model for Object Colour Perception
R,(h)= jJ&,r,(A).
i=l
Note that in the above and all following formulae the integration is over the
wavelength visual spectrum. The index m corresponds to a specific mth
subfield, the other indices indicate vector elements or basis members.
Also, the integrals of the form j C,(A)lj(A)r,(A) dhd K~,~ are fixed system
constants and do not depend in the stimulus. The only variable quantities are
the illuminant and reflectance vectors v, &,,, and their corresponding estimates i
and ai,.
Equation (5a) may be put in matrix form
Ym = W+Xn (5b)
where di, (v) = I;= r uj~ijk are the matrix entries.
The trichromatic responses are internally estimated by an equation of
identical form
Ym = D(i)&. (ob)
The actual trichromatic response y,,, is equated to the estimate Y,,,
z= 1 b,Y,;
m
similarly the field average reflectance vector, a, is
a = 1 La,, (8b)
m
fi = I C,(h)i(A)R,(h) dA
hkrk
(A
= f f, ijiKijkhk, (94
j=l k=l
4 = C bmytm = C bm I? f UjakmKi,k
m m j=l k=l
2 = Q(a)v,
where
and
ak - c bmem.
m
Note that a is an actual average reflectance vector for the complex field, while
h is an internally assumed standard average.
Equating (9) and (10) gives the equation for estimating the
Q(h)e = z. (11)
Equation (11) can be solved for v which is then substituted into (7) to obtain
&, and forms the basic operation for phase 1 (Fig. 1).
What we have here is an adaptation of one procedure (set of equations) to
the solution of another procedure. Step 1 is the solution of the set 11. Step 2 is
the solution of the set 7, where certain coefficients adjust themselves according
to the solution of step 1. It is important to realize that (7) and (11) can be
solved to eliminate 9. The resulting formulation would be a single phase
operation that relates the vector & to stimulus quantities v and (Y only. It could
then be seemingly claimed that no estimate of illuminant is made. However, we
prefer to give the formulation explicit physical meaning rather than to avoid
the critical issue of how the system arrives at the illuminant. (See also the
Discussion section.)
Figure 1 summarizes the structure of the model. The stimulus as defined by
the model is passed through the filters. The information concerning the entire
field (a certain average of the single subfields) is processed to get i. This in
turn with the information concerning the sole subfields is processed in a second
processing phase to get & for each subfield. Different subfields are analyzed by
similar model units spatially arranged that are not drawn on the figure for
reasons of simplification.
The two results listed below are theoretically derived; in fact, the model is
designed to comply with them analytically: (1) If h= a, i.e. if the external
complex visual field in discussion has a spatial reflectance average that matches
the internal standard spatial average, then no matter what the illuminant is, the
solution for &, will be equal to ca,.
The proof is trivial, in Section II h = a+ 6= v and therefore in &, = cr,
irrespective of the illuminant vector v. One has to be careful here to exclude
some ill-defined singular special cases, but these do not correspond to any
actual practical experiments. The case h = a is basic, but most of our discussion
and investigation of the model is dedicated to deviations and perturbations
from the normal case.
(2) Let us assume a case where two different illuminants, say v and v’,
illuminate the same scene (not simultaneously) or illuminate two different
scenes. It may occur then that two different reflectances (each under a different
illuminant) will transmit to the eye spectra that will have the same filter outputs
(the same vectors y), in terms of the model this means
cu (Y,, one can find v’ such that the output of the filters will be equal (for
di%erent subfields) but still the estimates &,, bL, will not be affected, or
equivalently perceived hue will not change.
The vectors that correspond to the total field filter response in the two
scenes, namely z, z’ are of course not the same as
Equation (12), however, does not hold if the field consists of only one subfield.
This condition can be achieved by viewing the subfield in question through a
narrow tube isolating the specific subfield from the context (19). Then the two
subfields will have the same estimate &. Furthermore, the estimated reflectance
vector, &, of scenes comprised of one subfield will equal the internal standard
h. This follows immediately substituting a=a (which is the case when a sole
subfield constitutes the scene) in (7) and (11).
The experiments implied by these constraints have already been performed.
These are the Land “Mondrian” experiments [Land (18, 19) and McCann et al.
(21)] demonstrating hue constancy under different illuminants, with particular
subfields having the same integrated absorbed response, and the loss of hue
when viewing through a narrow tube.
K 112=K121=K2~~=0.1
All other index combinations vanish. The internal standard average h is taken
as h = (50 50 50)T: that is, the standard average has equal reflectance from
the short-wave, middle-wave and long-wave regions respectively. In the follow-
ing figures, we employ the model using different vectors v for the illuminant
and different vectors a for the true field average. Using different such vectors
we could check the outcome substituting an illuminant vector with a dominant
or suppressed term corresponding to an actual illuminant with a dominant or
suppressed spectral band. In addition, using different vectors a, we could
investigate the influence of the total field average on the estimation of a single
subfield.
We have used values that are quite extreme. Variation between a and h in
some regions goes up to 50% under different vectors v. Also, some of the fixed
parameters values for each figure are taken from the extremes. However, we
a2=40 ___
a,=60
_--- a,=20
‘I,, _ _ _ _ - a,=60 -
----..t. ......_......((.._.._....,......
a2 = 80 __....._
as*20
I_
, aI
25 50 75
al = 40 ___
.= y\,___: g: -
-.
a,=20
.98
t
I , a2
42.ah 25 50 75
I - al ~40
__,.,..._...,.,.,_..._......._._.___.......... an s 60 ---
a,.20 _
99
t\ ----___-
_e--
.
a2*80
a, =90
.96 1 ao=20
I , aa
25 50 75
FIG. 2(a)-(c). Model calculation of the cosine of the phase angle between the true
reflectance vector OLand the estimated vector &, plotted for different illuminants, 8, and
field averages, a. The ordinate indicates the cosine value. The abscissa corresponds to
a continuous change of one of the elements of the vector or, while the two other entries
are fixed parameters. Extreme (Y values are chosen in some cases to show that in a
broad range of conditions the correlation between & and cx is very high. In this figure
and in each of the following, the illuminant vector v and field average vector a are
indicated on the figure. Also are indicated the ratios of the corresponding entries in a
and h. h is always taken as (50 50 50)T.
A Spatial Processor Model for Object Colour Perception
conclude that no critical distortion in the vector orientation results from these
deviations of the true average a from the standard average h [Buchsbaum (3)].
Only a capsule of the numerical calculations is presented in the following
figures.
The ordinate on Fig. 2 corresponds to the cosine of the phase angle between
(Y and & in different conditions. The abscissa indicates a continuous change in
one of the entries of CY,(there are three figures for each condition) while the
other two serve as fixed parameters. In most cases, the two entries that serve as
fixed parameters have three different value-pairs so in total each condition has
nine curves representing it. On top of each figure are given the vector a (true
field average vector) which differs from the vector h (standard field average
vector) which is kept constant, h = (50 50 50)T. Only a different from h is
considered, for otherwise OLand & would be equal. The relation between a and
h essentially governs the distortion and if these two are reasonably close then
the correlation in orientation of cx and & is very high [Buchsbaum (3)].
Figure 3 presents the actual outcome of the model on two-dimensional maps.
This is necessary since from the former figures we could learn that the
estimated vectors are not too biased in orientation, however, we could not
learn in what direction the bias goes. It would be of use to know whether the
bias goes in the direction of the field average, a, or against it.
60 -
Y400,loo,l00?
40 - p’ wo,70.60)T
Q1=20
20 -
0 I 1 I , Qw
20 40 60
cap"
FIG. 3(a)-(c). Model calculations demonstrating the mapping of true reflectances (Y to
their biased estimates & resulting from differences between the true and standard
average reflectances a and II. Two of the reflectance vector entries and their correspond-
ing estimates are indicated on the axes, the third entry is a fixed parameter. The open
circles indicate points corresponding to OL,the filled circles indicate points corresponding
to h. The former points are connected by broken lines, the latter set of points by full
lines. The figure with the lines connecting the points demonstrates the distortion of the
entire structure imposed by the field conditions. In this figure a = (50 70 50)’ has the
second term pronounced relative to h = (50 50 50)T; the deflection as can be seen on
the figure is almost entirely in the direction of supressing the second term of &. This
means that the estimated reflectance & is biased in a direction that is opposite to the
terms inducing and dominating the true field average. The third entry, which is not
plotted, stays almost unchanged.
A Spatial Processor Model for Object Colour Perception
A
8,.
a0 -ai0 ~_--_p--__p--__-9
I I I
I I I I
I I I I
I
60 -
y'1100.100.100~
Q=C70.S0.S0,T
as=20
2c-
a,.
0 I I , 020
20 40 60 80
FIG. 3(b)
al=20
Bg.
0 , / , a20+
20 40 60 80
FIG. 3(c)
V. The Projection Experiment and its Duality with the Reflection Case
three coloured filters to generate three black and white slides [Land (16, 17)
refers to these slides as records]. The slides are reprojected simultaneously
through three coloured filters to reproduce the scene. The photographing and
projection filters may (but need not) be the same. A restriction, of course, is
that such a set of filters can serve as a basis for the colour space.
Let us denote the photographing filters f,(A), f2(h), f3(A) and the projection
filters pi(A), p2(h), p,(A). Each photograph is processed to a black and white
slide where each subfield of the scene is transformed to a density of the
corresponding subfield on the slide.
This density of the mth subfield on the ith slide has the form
The vector z depends on the average field density of each slide, say &j
(15)
With the inputs y,,, and z defined, the system equations are formed as in
Section II.
If fi(A) = pi(A), and the illuminant and reflectances can be represented by the
three dimensional bases, then all that is done is a projection (in functional
analysis terms) of a vector in a three-dimensional space on a basis in that space
and reconstructing it as a linear combination of the same basis. The same set of
equations will then solve this case and the normal case discussed earlier. The
duality between the two cases can be comceptualized as follows. In the simple
reflection case we assumed a common fixed illuminant for the whole scene
while the various subfields had different reflectance vectors. In this case the
projection spectrum is varied from subfield to subfield according to the density
of the slides; however, these projections are reflected from a uniform reflecting
screen. Mathematically, the processor solves both cases with the same equa-
tions, and so we will find that the projection experiment is most powerful in
simulating certain conditions. The projection procedure allows much experi-
mental freedom. The slides can be manipulated in various ways and, in fact,
this is the basis for Land’s two primary experiments which we reviewed in
detail in the introduction section. The manipulation of the slides, or the vector
4, can be represented mathematically by multiplying this vector by elementary
matrices. If we denote the elementary matrix as U then the system will solve its
equation with the vector U& and U4 controlling its inputs y and z.
Journal
of The Franklin Institute
16 Pergamon Press Ltd.
A Spatial Processor Model for Object Colour Perception
a
0
Lo
0
b
0
Type 1
aEl[&B91
!I LB Type 2 Type 3
(16)
- QslQ,=W.Q&Q
Q, (I-I&S 1
40 60 60
- Qs(QI.Qs ~60)
.95
I I 1
b
20 40 60 60
FIG. 4. Cosine of the phase angle between & and OLis plotted for some conditions of the
projection experiment. The ordinate indicates a continuous change in one of the entries
of 01 while the other two stay as fixed parameters. The interference matrix U is
indicated together with the rest of the.field conditions. Here the interference matrix
corresponds to placing neutral density filters in front of two of the projectors. It can
be seen that the system does well enough under this distortion and the cosine correlation
between & and et is still very high.
ii,.
0 , 1 , a20+
20 40 60 I#)
4
a2=
a20
60 -
60 -
~=cloo,loo.loolT
4- ~=t50.50.50~’
a,=00
20 -
a 5.
0 I 1 , aso,
20 40 60 60
FIG. 5. Bias in estimated reflectance caused by projecting only two records. Two entries
(Y* and (~3 are shown on the two dimensional map. The open circles present the true
entries and the filled circles the estimated entries. The third entry is a fixed parameter.
It seems that with two records only, the system does well; no major deflection is
detected. But here one should be careful and bear in mind that the third entry is always
dependent on the other two and not independent as earlier.
000
( >
lJ= 001
010
Y=(loo.loo.loo)T
Q*t60.so.600,T
a,- 20
&'--__,____d
/’
n
/’ a39
/’
0 1’ I I , a3o
20 40 60 60
A
22.
azo
60-
u=
60 -
y=mO.lOO,lOo~T
ap=80
2O-
/'
/' A
/' a6m
0,' I I aso
20 40 60 60
FIG. 6. Migration map of entries of (Y to entries of & when the interference matrix LJ
corresponds to the reversal of two records (the two records are interchanged). The open
circles correspond to the true vector entries, the filled circles to the estimated vector
entries. The estimated entries constitute a slightly deflected but almost exact mirror
image of the true entries relative to an axis that is a 45” straight line through the
origin. Some points on this line are also indicated with their image. In numerical
terms, this mirroring means that the true entries just interchange their numerical
in their image.
mental possibilities are discussed following a brief review of the model princi-
ples and results.
The basic assumptions of the theory are:
(a) The visual system implicitly computes the illuminant of the field and the
reflectance of each subfield in the scene by matching them to linear combina-
tions of three dimensional bases. The computed (or estimated) values are the
scalar coefficients of the linear combination of the basis. The basis members are
fixed and the basic constraint imposed on the basis is that it actually spans the
colour-space.
(b) The illuminant is assumed common to the entire field. The illuminant
estimate is made under the assumption that the visual field in question has a
fixed spatial reflectance-average that follows a certain standard, i.e. the system
presumes that its standard average reflectance is equal to the actual spatial field
average.
(c) A set of linear equations is formed for the illuminant and each subfield;
the equations are based on the inputs to the system. These generalized
equations [(7) and (ll)] are solved for the reflectance vector of each subfield in
the various experimental paradigms.
Under these assumptions we have shown using also numerical examples:
1. If the average of the visual field follows the standard, colour constancy
is completely maintained. By design of the model, in case of equality between
the actual average and the standard, irrespective of the illuminant, the coefficients
in the subfield equations adapt themselves such that the solution will be
unbiased. All the necessary information on the illuminant exists in the visual
field.
2. A bias between the computed reflectance-vectors and the actual reflec-
tances, is introduced in cases of mismatch between the true field average and
the standard. However, under a broad range of pertubations from the standard,
the bias introduced in the solution is small, Within certain tolerance perceptual
constancy will be maintained over a large domain of arbitrary scenes and
illuminants. Furthermore, the bias that is introduced when the average and the
standard do not match is in the opposite direction of the true field average.
This means that an inducing field will cause a perceptual effect on the other
subfields away from the inducing field reflectance. This is consistent with colour
contrast phenomenology.
3. The model is compatible with observations and findings in the Mondrian
experiments by Land (19) and McCann et al. (22). It is shown that two different
subfields in the same scene may transmit the same information to the eyes in
terms of integrated reflectances (under different illuminations) yet their com-
puted reflectance vector stays unaltered under the overall field processing. The
case when a sole subfield is viewed through a narrow tube isolating it from the
rest of the context is also consistent and discussed.
4. Under the same set of assumptions, properties of Land’s (16, 17) two-
primary projection experiment are accounted for. Mathematical duality is
proved to exist between the Mondrian experiment and the projection experi-
ment, thus a unified understanding in terms of the same basic engineering
Journal
of‘IheFranklin Institute
22 Pergamon Press Ltd.
A Spatial Processor Model for Object Colour Perception
The Retinex theory of Land (l&19) and McCann et al. (22) is inherently
different. It is assumed that independent records of the scene are taken through
each of the three colour mechanisms. The records are normalized with respect
to the brightest spot in each of them. Then the visual system scales each axis of
independent triplets computed for each subfield and finally arrives at the
reflectances of each subfield. The comparison among the subfields is done by
taking the response ratio of adjacent subfields in the short-, middle- and
long-waveband records. The scaling of the records in (22) involves a transfor-
mation of the filters output through a compressive nonlinear psychophysical
functions, namely Glasser et al.‘s (9) approximation to Munsel value. The
significant advantages of the model are two. The first is theoretical, the
Retinex model overcomes the requirement of uniform illuminance. The only
additional assumption is that the illuminance is spatially continuous; spatial
discontinuities are taken as changes in reflectance. The second advantage is
that a systematic realizable algorithm is proposed for the model, that can be
implemented choosing an arbitrary sequence of subfields in the scene. How-
ever, the scaling algorithm assigns the subfield with maximum lightness in the
scene to be white. The existing Retinex examples do not elaborate on cases
where this spot varies among the Retinex records or on cases where the
maximum lightness spot is not actually white. In the present approach it does
not matter whatsoever whether there actually exists a white subfield in the
scene or not.
Another account (for some of the phenomena) is given by Woolfson’s (32)
analysis of the two primary projection experiment. His idea is that the system
measures the ratio between light reflected from an arbitrary spot in the scene
and that reflected from a white spot. This corresponds to measuring the
reflectance of an arbitrary spot.
Other approaches that are based on simultaneous contrast effects are also
considered (13). These are based on the relations among adjacent subfields in a
complex scene. Recalling that the Retinex algorithm is a series-product of a
sequence of adjacent subfields, it may be debated as to whether the Retinex
algorithm is not an implicit generalization of simultaneous contrast effects, and
so implicitly bridging the gap between two highly opposing points of view. A
possible link between the present approach and the Retinex model can also be
found. The set of integrated reflectances [McCann et al. (22)] used for a
quantified calculation of the Retinex examples can be shown to be a linear
combination of the reflectance vector 01 [Buchsbaum (3)]. The present model is
an idealized processor that can be implemented in various ways. This systems-
model does not prefer any one possible algorithm or implementation to
another. However, it demonstrates how such an implementation or principle
can be unified under a general systems pattern recognizer, constrained by the
physical limitations imposed on the visual system.
The issue of estimating the illuminant is very delicate in this context. The
computed illuminant is an implicit operation for the visual processor, In fact,
the solution for & can, in terms of mathematical formality, be made indepen-
dent of the illuminant extimate; i.e. the estimated reflectance will be a function
of the light reflected from the visual scene (the actual stimulus) and of the
processor fixed properties and constant. This is accomplished by substituting
(7) and (11) in each other, eliminating +. As mentioned earlier, this mathemati-
cal maneuver may be the basis for arguing that the visual system does not
compute the illuminant. However, we find that an explicit presentation of the
operations the external spectrum undergoes is advantageous to conceptually by-
passing this intriguing issue. Some theories, particularly the Retinex theory, do
not require the system’s ability to arrive at the illuminant. We maintain that
determining the reflectance of a subfield is essentially the measurement of a
transfer function. The reflectance of an object (which determines its colour) is
essentially in basic systems engineering terms the transfer function between the
illumination as input and the light transmitted to the eyes as output. It is
inconceivable that deriving the transfer function on the basis of its output will
be independent from any assumption on the input. This means that an
algorithm that derives reflectances without explicitly deriving the illuminant
nevertheless must have this calculation implicitly built in. Conceptually, we find
this in compliance with the old basic principle of Helmholtz, though again it
should be emphasized, not its geometrical colour space interpretation. The
present model does not “discount” the illuminant but estimates it.
The system arrives at the illuminant estimate assuming a certain standard
common spatial spectral average for the total field. It seems that arbitrary
natural everyday scenes composed of dozens of colour subfields, usually none
highly saturated, will have a certain, almost fixed spatial spectral reflectance
average. It is reasonable that this average will be some medium gray, which
comes back to Helson’s principle.
However, we do not rely on the vague hypothesis that this can be done on
the basis of reflectance-information from the whole visual field but rather show
how this can be done and define a pattern-recognizing technique to do it. The
unified explicit mathematical formulation we propose solves the problem in
terms of spatial complex field analysis.
Acknowledgment
References
(1)H. L. F. von Helmoltz, “Treatise on Physiological Optics”, (translated by J. P. C.
Southall), Optical Society of America, 1924.
A Spatial Processor Model for Object Colour Perception
(2) P. K. Brown and G. Wald, “Visual pigments in single rods and cones of the human
retina”, Science, Vol. 144, pp. 45-52, 1964.
(3) G. Buchsbaum, “Models of Central Signal Processing in Colour Perception”,
Ph.D. Thesis, Tel-Aviv University, 1978.
(4) G. Buchsbaum and J. L. Goldstein, “Optimum probabilistic processing in colour
perception, Part I: Colour Discrimination”, Proc. R. Sot. Vol. B205, pp.
229-248, 1979.
(5) G. Buchsbaum and J. L. Goldstein, “Optimum probabilistic processing in colour
perception, Part II: Colour vision as template matching”, Proc. R. Sot. Vol.
B205, pp. 249-266, 1979.
(6) J. Beck, “Surface Color Perception”, Cornell University Press, Ithaca, New York,
1972.
(7) A. L. Gilchrist, “Perceived lightness depends on perceived spatial arrangement”,
Science, Vol. 195, pp, 185-187, 1977.
(8) A. L. Gilchrist, “The perception of surface blacks and whites”, Scient. Am., Vol.
240, pp. 112-125, 1979.
(9) L. G. Glasser, A. H. McKinney, C. D. Reilly and P. D. Schnelle, “Cube-root color
coordinate system”, _I. opt. Sot. Am., Vol. 48. pp. 736-740, 1958.
(10) H. Helson, “Fundamental problems in colour vision”, Parts I and II, J. exp.
Psycho/., Vol. 23, pp. 439-476 and Vol. 26, pp. l-29, 1938; 1940.
(11) H. Helson, “Some factors and implications of color constancy”, J. opt. Sot. Am,,
vol. 33, pp. 555-567, 1934.
(12) C. F. Hall and E. L. Hall, “A nonlinear model for the spatial characteristics of the
human visual systems”, IEEE Trans. Sys. Man. Cyber.. Vol. SMC-7, pp.
161-170, 1977.
(13) D. Jameson and L. M. Hurvich, “Theory of brightness and color contrast in human
vision”, Vision Res., Vol. 4, pp. 135-154, 1964.
(14) D. B. Judd, “Hue, saturation and lightness of surface colors with chromatic
illumination”, J. opt. Sot. Am., Vol. 30, pp. 2-32, 1940.
(15) D. B. Judd, Appraisal of Land’s work on two primary color projections”, J. opt.
Sot. Am., Vol. 50, pp. 254-267, 1960.
(16) E. H. Land, “Color vision in the natural image. Part I”, Proc. Nam. Acad. Sci.,
U.S.A., Vol. 45, pp. 116-129, 1959.
(17) E. H. Land, “Color vision in the natural image. Part II”, Proc. N&n. Acad. Sci.
U.S.A., Vol. 45, pp. 636-644, 1959.
(18) E. H. Land, “The retinex”, Am. Scient, Vol. 52, pp. 247-264, 1964.
(19) E. H. Land, “The retinex theory of colour vision”, Proc. R. Instn. Gr. Br., Vol. 47,
pp. 23-58, 1974.
(20) E. H. Land, “The retinex theory of color vision”, Scient. Am. Vol. 237, pp.
108-128, 1977.
(21) W. B. Marks, W. H. Dobelle and E. F. MacNichol, “Visual pigment of single
primate cones”, Science, Vol. 143, pp. 1181-1183, 1964.
(22) J. J. McCann, S. McKee and T. H. Taylor, “Quantitative studies in retinex
theory-a comparsion between theoretical predictions and observer responses to
the color Mondrian experiments”, Vision Res., Vol. 16, pp. 445-458, 1976.
(23) J. J. McCann, “Rod-cone interactions: Different color sensations from identical
stimuli”, Science, Vol. 176, pp. 12551257, 1972.
(24) J. J. McCann and J. L. Benton, “Interactions of long wave cones and rods produce
color sensations”, J. opt. Sot. Am., Vol. 59, pp. 103-107, 1969.
G. Buchsbaum
(25) S. P. McKee, J. J. McCann, and J. L. Benton, “Color vision from rod and
long-wave cone interactions: Conditions in which rods contribute to multicol-
ored images”, Vision Res., Vol. 17, pp. 175-185, 1977.
(26) D. E. Pearson, D. B. Rubenstein and G. J. Spivack, “Comparison of perceived
color in two primary computer generated artificial images with predictions based
on the Helson-Judd formulation”, J. opt. Sot. Am., Vol. 59, pp. 644-658, 1969.
(27) T. G. Stockman “Image processing in the context of a visual model”, Proc. IEEE,
Vol. 60, pp. 828-842, 1972.
(28) G. Wald, “The receptors of human color vision”, Science, Vol. 145, pp. 1007-
1017, 1964.
(29) H. Wallach, “Brightness constancy and the nature of achromatic colors”, .I. exp.
Psychol., Vol. 38, pp. 310-324, 1948.
(30) G. L. Walls, “Land! Land!“, Psychol. Bull., Vol. 57, pp. 29-48, 1960.
(31) L. Wheeler, “Color matching responses to red light of varying luminance and
purity in complex and simple images”, J. opt. Sot. Am., Vol. 53, pp. 978-993,
1963.
(32) M. M. Woolfson, “Some new aspects of color perception”, IBM Jl Res. Dev. Vol.
3, pp. 312-325, 1959.
(33) G. Wyszecki and W. S. Stiles, “Color Science”, John Wiley, New York, 1968.