Dyes and Pigments 147 (2017) 544e551

Contents lists available at ScienceDirect

Dyes and Pigments

journal homepage: www.elsevier.com/locate/dyepig

Color prediction model for pre-colored fiber blends based on modified

Stearns-Noechel function
Chun-ao Wei, Xiaoxia Wan*, Junfeng Li
School of Printing and Packaging, Wuhan University, No. 129 Luoyu Road, Wuhan 430079, China

a r t i c l e i n f o a b s t r a c t

Article history: The goal of this work is to propose a color prediction model for pre-colored fiber blends with high ac-
Received 28 April 2017 curacy. The transfer function plays a vital role in an additive color mixing model. The better linearity
Received in revised form between the transfer function and mass proportion, the higher accuracy of the model. However, the well-
19 June 2017
known Stearns-Noechel transfer function does not always hold good linearity, causing inaccurate color
Accepted 9 August 2017
matching in many cases. Aiming at compensating the poor linearity, a new transfer function was
Available online 12 August 2017
established by minimizing the linear deviation. The proposed transfer function was applied to the ad-
ditive model for color prediction of pre-colored fiber blends. The prediction accuracy of the proposed
Keywords:
Color prediction
model was assessed by 44 samples. The average color difference was 0.63 CIEDE2000 unit, which was
Additive model significantly better than the results of the Stearns-Noechel model (~1.23) and the two-constant Kubelka-
Pre-colored fiber blends Munk model (~1.11). These results indicate the proposed model has higher color prediction accuracy and
Stearns-Noechel model can better satisfy the color requirement of practical production.
Transfer function © 2017 Elsevier Ltd. All rights reserved.
Linearity

1. Introduction
Pre-colored fiber blending is an important coloration method
and is considered to be the most environmentally friendly method
in the textile industry. A limited number of pre-colored fibers (or
primaries) are blended together to create any desired color by
formulating recipes. In this field, one of the most important prob-
lems is to seek an appropriate recipe for a standard color. Therefore,
it is fundamental to establish color prediction models to describe
the corresponding relationship between the recipes and the
blended colors.
There has been a long history of work to derive color prediction
models for pre-colored fiber blends. The existing models can be
divided into two categories, namely theoretical models and
experimental models. The former attempt to model the behavior of
light propagation within the mixture, while the latter tend to derive
a transfer function from specific experiments so that an additive
model can be established. It is desirable to describe the color
mixing mechanism in theory, but the fiber blends system is so
complicated that the assumptions of theoretical models are hard to
be satisfied, resulting in inaccurate color prediction. The Kubelka-
* Corresponding author.
E-mail address: wan@whu.edu.cn (X. Wan).
Munk theory [1e3] has been reported to be applied in the field of
pre-colored fiber blends for many years. Yet, it has been proved the
single-constant Kubelka-Munk model yields a poor accuracy [4]
since the scatter coefficient S which is assumed to be constant
varies greatly from one primary to another primary fibers. Burlone
[4e6] suggested the application of the two-constant Kubelka-
Munk model (two-constant KM model) from a mathematical point
of view. A larger number of parameters (K and S for each primary
within a blend) increased mathematical flexibility of the model,
and the color prediction accuracy was improved significantly with
an average of 1.6 CIELAB color difference in Burlone's study [4]. The
Kubelka-Munk theory is derived from a basic assumption of infi-
nitely thick and opaque layer, however, the fibers commonly used
in the textile industry are neither fully opaque nor transparent.
Amirshahi et al. [7] tried to give a different theoretical explanation
from the simplified color formation process of pre-colored fiber
blends. They pointed the color mixing mechanism consisted of two
independent process acting simultaneously. The color of each point
on the surface was a function of the color of the lower layers, and
the colored dots were mixed spatially by averaging when the sur-
face was measured or viewed from a distance. This explanation was
built on the assumptions of regular arrangement of the fibers and
homogenous thickness of the blends, which were not actually true,
resulting in the inaccurate color prediction for some of the test
samples [7]. Friele model [8,9], a theoretical method specially

http://dx.doi.org/10.1016/j.dyepig.2017.08.012
0143-7208/© 2017 Elsevier Ltd. All rights reserved.
 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551
designed for pre-colored fiber blends, was established from the
perspective of light propagation paths. According to the Friele
model, the reflectance of the blends could be expressed as the sum
of all outgoing light on the surface, and the Beer-Lambert Law [10]
was adopted to calculate the intensity of outgoing light considering
the various propagation paths. Taking light propagation into ac-
count is very promising to explain color mixing mechanism in
optical nature. Unfortunately, the mediums described by the Beer-
Lambert Law are expected to be transparent and uniform, while the
fiber blends mediums are proved to be semitransparent and non-
uniform [6]. Therefore, the application of the Friele model in
many studies [4,9,11] tends to obtain large color difference.
Since this is the case, the experimental models that derived
directly from empirical data may be better solutions for color
prediction of pre-colored fiber blends. In experimental models, a
transfer function [12] can always be found to provide a simple
additive relationship between the color of a mixture and the mass
proportion of primaries used in the mixture. The models can be
formulated as





F Rl;m ¼ c1 F Rl;1 þ c2 F Rl;2 þ ::: þ cn F Rl;n
Where Fð$Þ denotes the transfer function; Rl;m is the reflectance of
P
the mixture at wavelengthl; Rl;i and Ci (subject to ci ¼ 1) repre-
sent the reflectance and mass proportion of the ith i (i¼1,2,3, …n)
primary within the mixture respectively. Obviously, the transfer
function plays a vital role in such a color prediction model. In other
words, the linearity between the transfer function and mass pro-
portion is the basis of the additive model to be established. The
better the linearity is, the more accurate the model is. In Stearns-
Noechel model (S-N model) [13,14], Fð$Þis derived from seven
black-and-white fiber blending samples with specific mass pro-
portion and it is a hyperbolic function with the property of
monotony decreasing, given by the following formula,
FðRl Þ ¼ ð1  Rl Þ=½bðRl  0:01Þ þ 0:01
where b is a dimensionless parameter and is expected to vary for
various mixing cases, such as the fiber types and the physical forms
of samples. By introducing the variable parameter, the Stearns-
Noechel function is endowed with the ability of self regulation to
achieve the linearity for various mixing cases. Even so, the accuracy
of the S-N model is still unsatisfactory [4,11]. In Burlone's study [4],
42 nylon fiber blending samples were used to test the S-N model,
and the average color difference was about 2.4 CIELAB unit. In the
following exploration of the S-N model, Philips-Invernizzi et al. [14]
proposed a modification of the transfer function that the parameter
b was assumed to vary with the wavelength, leading to much better
results with 1.66 averaged CIELAB color difference for 234 cotton
fiber blending samples. Essentially, the modified transfer function
by Philips-Invernizzi [14] could achieve adaptive linear adjustment
at different wavelengths in one sample, consequently bringing
more accurate results. Nevertheless, the color prediction accuracy
is still unacceptable since the colorimetric threshold of the textile
industry is usually set as 1, referring to the Chinese National Stan-
dards GB/T 21898-2008. In addition to the S-N model, other addi-
tive experimental models have also been developed, but they are
either inconvenient or too restrictive [14]. For example, Davidson
et al. [15] proposed a completely empirical transfer function that
must be re-derived for different cases rather than sharing a general
form. It has been reported that Artificial Neural Network (ANN)
based models [11,16,17] have also been applied in the field of pre-
colored fiber blends, which can be seen as another category of
experimental models. ANN is a system able to map an input vector
545
to an output vector via a hidden layer which has to be determined
by a training set of input-output samples with extensive repre-
sentation and adequate amounts. Unfortunately, the two basis
conditions required for the training set are difficult to achieve in an
independent test set due to the diversity of fiber blends and the
multi-dimension of spectral output vector, resulting in poor
generalization ability of ANN-based models [11].
Based on all above considerations, an accurate color prediction
model for pre-colored fiber blends is still needed today to satisfy
the colorimetric demand of the textile industry. This would enable
a computer-aided system to be developed, helping colorists to find
an appropriate recipe to reproduce a target color efficiently.
The aim of this study is to propose a color prediction model for
pre-colored fiber blends with higher accuracy. To achieve this goal,
a new transfer function for the additive model has been established
by improving the linearity between Stearns-Noechel function and
mass proportion. In the present study, the fiber blends were limited
to cotton fibers. A group of samples with known spectral reflec-
tance and recipes were used to test this proposed model, and the
color prediction results were compared with those obtained by the
S-N model and the two-constant KM model.
(1)
2. Materials and methods
The cotton fibers (1.67 dtex and average 37 mm long) used in
this study were dyed by four reactive dyes. These four pre-colored
fibers, together with the raw undyed fibers, constituted the five
primaries used in this study. The reflectance of these five primaries
is shown in Fig. 1, in which primary #1 is the raw undyed fibers.
The samples used in the present study were prepared in the
form of knitted fabrics to minimize the measurement error caused
by the instability of the fiber samples, e.g., avoiding the difference
of fiber compactness and pressure on the fiber samples during the
measurement. A small-sample carding machine was utilized to
obtain the pre-colored fiber mixtures. In order to obtain homoge-
neous mixtures, the fibers were fed to the carding machine three
(2) times. Then the mixtures were spun into yarns using open-end
spinning with the count of 29.2 tex and twist coefficient of 450
atex. The yarns were eventually knitted into single jersey fabrics
with 24 threads/inch as the experimental samples.
The X-rite Ci7800 spectrophotometer was chosen to measure
the spectral reflectance of the samples, and the readings were
recorded in the visible wavelength range (400e700 nm) at 10 nm
intervals, providing 31-dimensional data. The optical geometry was
Fig. 1. Reflectance of the primaries used in this study.
546 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551
a d/8 system, diffused illumination and 8 viewing angle, and the
specular component of the reflectance was excluded. The samples
to be measured were folded in thick-enough layers to avoid the
translucent effect and no glass cover or sample holder was attached
to the samples. To ensure that enough different areas over the
whole surface of the samples could be measured, the largest
aperture of 25 mm was selected and nine readings at different lo-
cations and rotational positions were taken randomly. Then the
average value was calculated as the true reflectance of the samples.
A total of 54 pre-colored fiber blending samples were prepared
and measured as described above, as shown in Fig. 2. The samples
were divided into two groups. One group included 10 samples, as
shown in Fig. 2(a). Primary #3 and #5 were separately blended
with the raw fibers #1 at the proportion of 20%:80%, 40%:60%,
50%:50%, 60%:40% and 80%:20% respectively. The 10 samples in this
group were used to assess the linearity between transfer function
and mass proportion for establishing a new transfer function with
better linearity. The other group included 44 samples, as shown in
Fig. 2(b). They were used to test the color prediction accuracy of the
proposed color prediction model. The recipes of these 44 samples
were randomly designed using combinations of 2e4 primaries.
There were 23 two-primary, 10 three-primary and 11 four-primary
fiber blending samples respectively.
3. Experimental
3.1. Linearity between the Stearns-Noechel function and mass
proportion
As mentioned above, the linearity between the transfer function
and mass proportion is the basis of the additive model to be
established. In this section, the linearity of the Stearns-Noechel
transfer function has been assessed.
Assuming that two primary fibers are mixed together at the
mass proportion of c: 1-c and the transfer function Fð$Þholds good
linearity with mass proportion c, the additive model can be written
as




F Rl;m ¼ cF Rl;q þ ð1  cÞF Rl;p
where subscripts p and q indicate the two different primaries; m
indicates the mixture; l indicates a specific wavelength. By trans-
position and collection of terms, Eq. (3) can be rearranged as:
Fig. 2. Pre-colored fiber blending samples prepared for the present study. (a) 10 samples for establishing the new transfer function. Left: primary #3 blending with #1; Right:
Primary #5 blending with #1. (b) 44 samples for testing the proposed color prediction model. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)



F Rl;m  F Rl;p
c¼

 (4)
F Rl;q  F Rl;p
If ½FðRl;m Þ  FðRl;p Þ=½FðRl;q Þ  FðRl;p Þis plotted against c in
Cartesian coordinate system, it will be found that a straight line
with a slope equal to 1 is obtained.
To test this relation, primary #3 and #5 were blended with raw
fibers #1 at five various mass proportions respectively to prepare
the test samples, as mentioned in the Section 2. FðRl;m Þ, FðRl;P Þand
FðRl;q Þwere calculated by using the actually measured reflectance
in the Stearns-Noechel transfer function under all the wavelengths.
The variable parameter b was optimized for each wavelength at the
value range of 0.05e0.3 [14] to adjust the linearity to be optimum.
Fig. 3(aeb) shows the actual curves defined by Eq. (4) at all the
wavelengths. It can be seen that the desired straight lines are not
obtained but are actually replaced by the concave curves at middle
wavelengths in Fig. 3(a) and at long and short wavelengths in
Fig. 3(b). In other words, Eq. (3) is not valid at these wavelengths
and should be replaced by





F Rl;m > cF Rl;q þ ð1  cÞF Rl;p ¼ F Rl;pre (5)
where Rl;pre indicates the predicted reflectance value at wavelength
l. Therefore, the poor linearity or the approximate concavity would
result in inaccurate color prediction of the S-N model, and the
predicted reflectance value would be always greater than the actual
value (Rl;pre > Rl;m ) since the Stearns-Noechel transfer function
possesses the property of monotony decreasing.
To verify above inferences, four pre-colored fiber blending
samples, for which primary #1 was blended with other primaries
(#2~#5) at the proportion of 60%:40%, were adopted to test the
color prediction accuracy of the S-N model. The results are shown
in Fig. 4. The inaccurate color prediction occurs at the wavelengths
with higher reflectance, where the predicted reflectance curves are
(3) above the actually measured ones as expected. This phenomenon
also appeared in the Rong Li et al.’s study [18]. These results proved
the poor linearity or the approximate concavity of the Stearns-
Noechel transfer function at the wavelengths with higher
reflectance.
 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551 547

Fig. 3. Linearity between the Stearns-Noechel function and mass proportion. (a) Five samples of primary #3 blending with #1. (b) Five samples of primary #5 blending with #1.

Fig. 4. Color prediction results of S-N model for 4 test samples. (a) Sample1 of primary #2 blending with #1. (b) Sample2 of primary #3 blending with #1. (c) Sample3 of primary #4
blending with #1. (d) Sample4 of primary #5 blending with #1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this
article.)

3.2. Proposed transfer function

( " #1=2 ),
X
m n

X 2
The poor linearity or the approximate concavity of the Stearns- K½FðRl Þ ¼ ci  cl;i m (6)
Noechel transfer function has been proved in the Section 3.1. In this l¼1 i¼1
section, a new transfer function is proposed to compensate the
linear deviation and to achieve better linearity. At first, the linear WhereK½FðRl Þis the linear deviation function with the transfer
deviation is defined as function FðRl Þ as the independent variable; ci indicates the sam-
pling points of the ideal straight line with the slope equal to 1; ci,l
indicates the sampling points of the actual curve at wavelength l
548 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551
and the actual curve is obtained by the FðRl Þ in the same way as the
Section 3.1; n indicates the number of the sampling points; m in-
dicates the number of the wavelengths. Then based on the linear
deviation function K½FðRl Þ, the problem to seek the new transfer
function becomes an optimization problem,
Fnew ðRl Þ ¼ argminK½FðR l Þ
  
s:t: FðRl Þ2 FðRl ; aÞjð1  Rl Þa ½bðRl  0:01Þ þ 0:01; a2ð1; ∞Þ
WhereFnew ðRl Þis our new transfer function which minimizes the
objective functionK½FðRl Þ; FðRl ; aÞis the feasible region of the
Fnew ðRl Þ. The feasible region is created by the parameter a, where a
is set to be greater than 1 to enhance convexity of the Stearns-
Noechel function based on the fact that the poor linearity to be
compensated is concave. In this study, the parameter a is traversed
ranging from 1 to 4 with the interval of 0.1 to create potential
transfer functions. The ten samples used in the Section 3.1 are re-
adopted to plot curves for the potential transfer functions in the
same way. In order to obtain the globally optimal solution of the
new transfer function, the objective function values are calculated
for all the potential transfer functions. The results are plotted in
Fig. 5 and certain numerical results are shown in Table 1. It can be
seen that the objective function is optimized to nearly to 0 by the
potential transfer functions, indicating that very good linearity can
be achieved among the potential transfer functions. Moreover, the
minimum of objective function appears at the domain of the
parameter between 2.5 and 3, therefore our new transfer function
is picked up from this domain,
ð1  Rl Þ3
Fnew ðRl Þ ¼
bðRl  0:01Þ þ 0:01
where the parameter a equal to 3 is determined for the sake of
convenient calculation. The linearity between the proposed trans-
fer function and mass proportion is shown in Fig. 6. The linearity is
improved significantly compared with that of the Stearns-Noechel
function (in Fig. 3), indicating that the compensation of the linear
deviation to the Stearns-Noechel function is reasonable and bene-
ficial. As a result, the inaccurate color prediction of the S-N model
(in Fig. 4) is removed effectively, as shown in Fig. 7.
Fig. 5. Objective function values with varied parameter a.
Table 1
Objective function values with varied parameter.
Samples #3 blending with#1 #5 blending with#1 Average
Parameter
a¼1 5.13 9.16 7.15
a ¼ 1.5 3.05 5.37 4.21
a¼2 0.81 1.96 1.38
a ¼ 2.5 0.06 0.31 0.19
a¼3 0.11 0.08 0.10
a ¼ 3.5 3.94 2.19 3.06
(7) a¼4 9.30 7.31 8.30
4. Results and discussions
The proposed transfer function was applied to the additive
model to establish a better color prediction model for pre-colored
fiber blends. In this section, 44 samples specified in Section 2
were used to test the color prediction accuracy of the proposed
model. For the sake of comparison, the S-N model [14] and the two-
constant KM model [5] were also tested by the same 44 samples.
For the S-N model and our proposed model, the parameter b was
trained by ten samples as shown in Fig. 2(a) and was assigned by
the median of the 10 optimal values at each wavelength [17]. For
the two-constant KM model, the constants K and S for each primary
were calculated from two samples, including a masstone of 100%
primary fibers and a tint mixed by the primary with raw fibers [5].
Two criterions were adopted to evaluate the accuracy of the
color prediction model. The Root Mean Square Error (RMSE) be-
tween the actually measured and predicted reflectance was
calculated as the spectral metric. The CIEDE2000 color difference
formula [19] under the CIE standard illuminant D65 and the CIE
1931 standard observer (DE*), which is the most uniform color
difference formula so far, was selected as the colorimetric metric. In
(8)
addition, a box-plot [20] was used to visualize the distributions of
the results and the inter quartile range (IQR) expressed as the
height of the rectangle in the box-plot could illustrate the disper-
sion of data.
Fig. 8 shows the color prediction results obtained by the S-N
model, our proposed model and the two-constant KM model (a) for
the 44 test samples in aspects of the spectral metric RMSE, and the
box-plot (b) visualizes distributions of the three data sets. Table 2
shows the statistic and distribution of the color prediction results.
As it can be observed, the RMSE of our proposed model (mean
RMSE ¼ 0.0037) is slightly less than that of the two-constant KM
model (mean RMSE ¼ 0.0047) and far less than that of the S-N
model (mean RMSE ¼ 0.0114). Besides, the dispersion of the results
from our proposed model is similar to that of the two-constant KM
model and more concentrated than that of the S-N model. On the
statistical data, 43 out of 44 samples (the percentage is 97.7%) have
the RMSE less than 0.01 for our proposed model, which is slightly
better than the two-constant KM model with the percentage of
90.9% and far better than the S-N model with the percentage of
52.3%. It could be concluded our proposed model showed consid-
erable improvements compared with the S-N model and was also
superior to the two-constant KM model in terms of the RMSE
metric.
Fig. 9 shows the color prediction results obtained by the S-N
model, our proposed model and the two-constant KM model (a) for
the 44 test samples in aspects of the colorimetric metricDE*, and
the box-plot (b) visualizes distributions of the three data sets.
Table 3 shows the statistic and distribution of the color prediction
results. As it can be observed, the color difference of our proposed
model is the smallest on the whole (mean DE* ¼ 0.63), followed by
that of the two-constant KM model (mean DE* ¼ 1.11) and that of
 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551 549

Fig. 6. Linearity between the proposed transfer function and mass proportion. (a) Five samples of primary #3 blending with #1. (b) Five samples of primary #5 blending with #1.

Fig. 7. Color prediction results of the proposed model for 4 test samples. (a) Sample1 of primary #2 blending with #1. (b) Sample2 of primary #3 blending with #1. (c) Sample3 of
primary #4 blending with #1. (d) Sample4 of primary #5 blending with #1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

the S-N model (mean DE* ¼ 1.23), respectively. Moreover, the
dispersion of the results of our proposed model is more concen-
trated than that of the other two models in that it’s IQR in the box-
plot is the smallest. On the statistical data, 39 out of 44 samples (the
percentage is 88.6%) have the color difference DE*less than 1 for our
proposed model, which is far better than the S-N model and the
two-constant KM model with the percentage of 50.0% and 56.8%
respectively. It could be appreciated that our proposed model
behaved very well in terms of the colorimetric metric. Especially,
the color difference of most samples was less than 1 which met the
550 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551

Fig. 8. Color prediction results of the spectral metric RMSE. (a) Results of three models. (b) Distributions of the results in a box-plot. (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.)

Table 2
Statistic and distribution of the color prediction results in RMSE.

RMSE Statistic Distribution

Model Mean Min Max <0.002 <0.004 <0.010 <0.020

S-N 0.0114 0.0014 0.0444 13.6% 34.1% 52.3% 79.6%

Proposed 0.0037 0.0008 0.0190 22.7% 63.6% 97.7% 100%
Two-constant KM 0.0047 0.0017 0.0239 6.82% 65.9% 90.9% 97.7%

Fig. 9. Color prediction results of the colorimetric metricDE*. (a) Results of three models. (b) Distributions of the results in a box-plot. (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)

Table 3 Table 4
Statistic and distribution of the color prediction results inDE*. Statistical results of the proposed model for 2-, 3- and 4-primary blends respectively.

DE* Statistic Distribution Metric RMSE DE*

Model Mean Min Max <1.0 <1.5 <2.0 <2.5 Samples Mean Min Max Mean Min Max

S-N 1.23 0.11 3.48 50.0% 63.6% 86.4% 90.9% 2-primary 0.0046 0.0013 0.0189 0.63 0.14 2.42
Proposed 0.63 0.15 2.43 88.6% 95.5% 95.5% 100% 3-primary 0.0027 0.0008 0.0051 0.44 0.18 0.84
Two-constant KM 1.11 0.14 4.63 56.8% 77.3% 88.6% 93.2% 4-primary 0.0025 0.0007 0.0056 0.78 0.22 2.23

colorimetric demand in the textile industry.
Furthermore, in order to check the suitability of the proposed
model for samples blended from different number of primaries, a
breakdown of the color prediction results for the 2-, 3- and 4-
primary blends was provided in Fig. 8(a), Fig. 9(a) and Table 4. It
can be seen that the number of primaries has very limited effect on
the prediction results. The average color difference is 0.63, 0.44 and
0.78 respectively for 2-, 3- and 4-primary blends, indicating the
proposed model are applicable for samples blended from different
number of primaries.
5. Conclusions
In this paper, the linearity between the Stearns-Noechel transfer
 C.-a. Wei et al. / Dyes and Pigments 147 (2017) 544e551
function and mass proportion was assessed by specific samples
with known recipes and spectral reflectance, and a new transfer
function aiming at achieving better linearity was proposed by
minimizing the linear deviation. Then the proposed transfer func-
tion was applied to the additive model and the color prediction
accuracy was assessed by 44 randomly designed pre-colored cotton
fiber blending samples. From the results, it can be concluded that
the color prediction accuracy of the proposed model was signifi-
cantly better than that of the S-N model and the two-constant KM
model and could meet the colorimetric demand of the textile in-
dustry in that the color difference of the most samples was less
than 1.
It should be noted that there is still an average color difference of
0.63 although the linearity of our proposed transfer function is very
good. Further investigations will examine the additive mixing law
for fiber blending medium and establish a nonlinear additivity
model concerning the interactions between the primaries.
Acknowledgments
This work was supported by the National Natural Science
Foundation of China [grant numbers 61575147, 61275172].
References
[1] Kubelka P, Munk F. An article on optics of paint layers. Z Tech Phys 1931;12:
593e601.
[2] Kubelka P. New contributions to the optics of intensely light-scattering ma-
terials. J Opt Soc Am a 1948;38(5):448e57.
[3] Duncan RD. The colour of pigment mixtures. Proc Phys Soc 1940;52(3):
390e401.
[4] Burlone DA. Theoretical and practical aspects of selected fiber-blend color-
551
formulation functions. Color Res Appl 1984;9(4):213e9.
[5] Burlone DA. Formulation of blends of pre-colored nylon fiber. Color Res Appl
1983;8(2):114e20.
[6] Burlone DA. Effect of fiber translucency on the color of blends of precolored
fibers. Text Res J 1990;60(3):162e7.
[7] Amirshahi SH,Pailthorpe MT. Applying the Kubelka-Munk equation to explain
the color of blends prepared from precolored fibers. Text Res J 1994;64(6):
357e64.
[8] Friele LFC. The application of colour measurement in relation to fibre-
blending. J Text Inst 1952;43(8):604e11.
[9] Philips-Invernizzi B, Dupont D, Caze C. Formulation of colored fiber blends
from Friele's theoretical model. Color Res Appl 2002;27(3):191e8.
[10] Beer A. Bestimmung der Absorption des rothen Lichts in farbigen Flüssig-
keiten. Ann Phys Chem 1852;86:78e88.
[11] Hemingray C, Westland S. A novel approach to using neural networks to
predict the colour of fibre blends. Color Technol 2016;132(4):297e303.
[12] Furferi R, Governi L, Volpe Y. Methods for predicting spectral response of fi-
bers blends. In: Murino V,Puppo E, editor. Lecture notes in computer science;
2015. p. 79e86. Genoa, ITALY: ICIAP 2015 Workshops.
[13] Stearns EI, Noechel F. Spectrophotometric prediction of color of wool blends.
Am Dyes Rep 1944;33(9):177e80.
[14] Philips-Invernizzi B, Dupont D, Jolly-Desodt AM, Caze C. Color formulation by
fiber blending using the Stearns-Noechel model. Color Res Appl 2002;27(2):
100e7.
[15] Davidson HR, Taylor M. Prediction of the color of fiber blends. J Opt Soc Am a
1965;55(1):96e100.
[16] Furferi R, Governi L. Prediction of the spectrophotometric response of a carded
fiber composed by different kinds of coloured raw materials: an artificial
neural Network-Based approach. Color Res Appl 2011;36(3):179e91.
[17] Shen J, Zhou X, Ma H, Chen W. Spectrophotometric prediction of pre-colored
fiber blends with a hybrid model based on artificial neural network and
Stearns-Noechel model. Text Res J 2017;87(3):296e304.
[18] Li R, Song Y, Gu F, Pan W. A spectrophotometric color matching algorithm for
precolored fiber blends. Color Res Appl 2009;34(2):108e14.
[19] Luo MR, Cui G, Rigg B. The development of the CIE 2000 colour-difference
formula: CIEDE2000. Color Res Appl 2001;26(5):340e50.
[20] Massart DL, Smeyers-Verbeke J, Capron X, Schlesier K. Visual presentation of
data by means of box plots. Lc Gc Eur 2005;18(4):215e8.