Content uploaded by Jinlian Hu
Author content
All content in this area was uploaded by Jinlian Hu on Jan 29, 2015
Content may be subject to copyright.
978-1-4244-5961-2/10/$26.00 ©2010 IEEE 417
2010 Sixth International Conference on Natural Computation (ICNC 2010)
Neural Network Modeling of Polar Fleece
Fabric Appearance Evaluation
Binjie Xin
Fashion College
Shanghai University of Engineering Science
Shanghai, 201620, China
Jinlian Hu
Institute of Textiles and Clothing
The Hong Kong Polytechnic University
Hung Hom, Kowloon, Hong Kong
Xiaoxia Liu, Lantian Lin
Fashion College
Shanghai University of Engineering Science
Shanghai, 201620, China
Abstract — This paper presents the development of a fairly new
neural network based method that aims at characterizing polar
fleece fabric appearance for the purpose of objective quality
evaluation. Co-occurrence matrix analysis is used to give the
quantitative descriptions of fabric appearance properties; neural
network model is used to establish the relationship between these
essential features and the final rating grade of fabric appearance.
The experimental results demonstrate that good correlation can
be achieved between the actual rating grade and the predicted
rating grade and reveals the possibility of the development of
artificial intelligence system to simulate the functions of human
eyes and brain.
Keywords-Neural Network; Co-occurrence Matrix; Appearance;
Polar Fleece Fabric;
I. INTRODUCTION
Polar fleece is a soft napped insulating synthetic wool
fabric made from Polyethylene terephthalate (PET) or other
synthetic fibers. Fleece has some of wool's finest qualities but
weighs a fraction of the lightest available woolens. Polar fleece
designates fabric with very high specific volume, which makes
fabric soft, lightweight, warm and comfortable. It is
hydrophobic, holding less than 1% of its weight in water, it
retains much of its insulating powers even when wet, and it is
highly breathable. These qualities make it useful for making
clothing intended to be used during strenuous physical activity.
However, the fibers on the surface of the polar fleece tend to
bunch up and form regular beards on the fabric surface after
wearing or washing. As an important factor of polar fleece
fabric quality, evaluating wearing appearance of polar fleece
fabrics objectively, quickly and reliably becomes urgent for
both the purchaser and the supplier to achieve quality
agreement with the market expanding.
Traditionally, evaluating fabric appearance still depends on
the human eyes and brain and thus suffers the drawbacks of
reliability due to human fatigue, speed, and overall accuracy.
These years, image analysis techniques, instead of human eyes
and brain, have been applied to textile manufacturing and
inspection of textile surface characteristics [1-3], analysis of
cotton fiber maturity [4], characterization of nonwoven
structures [5,6], and evaluation of carpet aesthetic appearance
[7,8].
In this paper, we establish one set of objective evaluation
system based on the co-occurrence matrix analysis and neural
network method for the evaluation of polar fleece fabric
appearance after abrasion. Co-occurrence matrix is used to
characterize the appearance properties of polar fleece fabric
and BP neural network is used for the grade rating as artificial
intelligent classifier.
II. DATA ACQUISITION
Grade1 Grade2 Grade3
418
Grade 4 Grade 5
Figure 1. Standard Polar Fleece Appearance Photographs
According to the comfortable feeling of visual perception
when evaluating polar fleece appearance, we divide polar
fleece fabrics into five grades as illustrated in Figure 1. Grade
five is the most comfortable, which give a satisfied aesthetic
feeling. Grade one is the worst; it cannot satisfy the basic
aesthetic feeling requirement of consumers. It is obvious that
texture properties such as uniform, roughness and regularity
are the most important features for the quality evaluation. The
investigation of this phenomenon has been conducted in our
previous research papers. As the co-occurrence matrix is an
effective method for the characterization of texture image, it is
utilized in this paper to give a quantitative description of the
polar fleece fabric after abrasion.
A. System set-up
Figure 2. Digital System for the Fabric Appearance Evaluation
As illustrated in Figure 2, the instrument setup consists of a
digital camera to capture the image of the mounted polar
fleece fabric samples on the sample platform, one set of
lighting sources to illuminate the fabric sample at a low angle,
one enclosure box containing the fixed sample platform and
the lighting sources, a computer connected to the digital
camera for image data collection and further data analysis. The
digital camera is placed above the lighting box and mounted at
a distance sufficient to capture a well focused image of fabric
sample.
B. Sample Specification
In our research, 15 kinds of commercially available polar
fleece fabrics with different solid colors and different pilling
resistant abilities were used as samples. Each fabric sample
was cut into approximately 105mm (43/16 in.) square and
subjected to laboratory abrasion in a Random Tumble Piling
Tester for half an hour according to ASTM Test Method
D3512.
The image resolution of the digital camera used here is 1600
by1200 pixels. In this study, we cropped each fabric image
into size 512 by 512 pixels; each pixel has values of 256 gray
levels, with the zero value representing black and 255
representing white. After digitalization of polar fleece fabric
appearance image, a set of algorithms for the calculation of
Co-occurrence matrix parameters will be developed in section
III.
III. CO-OCCURRENCE MATRIX ANALYSIS
A. Basic Principle of Co-Occurrence Matrix[9,10]
Grayscale Co-occurrence matrix is a widely applied method
for texture analysis, whose indicators are usually used as
feature parameters for such analysis. Grayscale Co-occurrence
matrix contains valuable spatial organization information of
texture. A regular sharp texture is reflected by accrete matrix
concentrated around the diagonal, while a scattered Co-
occurrence matrix is an indication of irregular and fuzzy
texture. If the texture shows highly uniform orientation and the
displacement vector is consistent with texture orientation, the
accrete matrix would be on the main diagonal.
Assuming that the size of image ),( yxf is yx NN × and
grayscale is g
N, the distance between two random pixels
),(),,( nmkj and position angle
θ
should be:
dnkdmj
dnkdmjdR
dnkmjdR
dnkdmj
dnkdmjdR
nkdmjdR
LD
V
RD
H
−=−=−
=−−=−°=
=−=−°=
−=−−=−
=−=−°=
=−=−°=
,
,,:)(135
,0:)(90
,
,,:)(45
0,:)(0
θ
θ
θ
θ
(1)
Times of occurrence of grayscale pair ),( qp in Co-
occurrence matrix are:
q}p,f(m,n)(d),f(j,k)m,n))||R#{((j,k),()5P(p,q,d,13
q}p,f(m,n)(d),f(j,k)m,n))||R#{((j,k),()P(p,q,d,90
q}p,f(m,n)(d),f(j,k)m,n))||R#{((j,k),()P(p,q,d,45
q}p,f(m,n)(d),f(j,k)m,n))||R#{((j,k),()P(p,q,d,0
RD
V
LD
H
===°
===°
===°
===°
(2)
4
1
2
3 3
5
1 CCD Camera
2 Lighting Box
3 Lighting Source
4 Sample platform
5 Com
p
ute
r
419
Hereinto, # stands for number of elements in the
aggregation. As Co-occurrence matrix is symmetric matrix,
thus ),,,(),,,(
θθ
dijPdjiP =.
Figure2 demonstrates the co-occurrence matrix of five
standard polar fleece appearance photographs by the means of
2-D gray scale image, gray level represents the frequency of
occurrence of grayscale pair.
Grade5(θ=0,d=1) Grade4(θ=0,d=1) Grade3(θ=0,d=1)
Grade 2(θ=0,d=1) Grade 1(θ=0,d=1)
Figure 3. Co-occurrence matrix of five standard images
B. Feature Parameters in Co-Occurrence Matrix
The Co-Occurrence matrix is normalized for the
convenience of feature parameter definition before the
specification of features.
• Texture uniformity (Q1):
∑∑
=
==
gg
N
i
N
j
jipQ
00
2
1),( (3)
Uniformity (Q1) reflects the homogeneity of image’s
texture. A higher value of uniformity indicates a more
consistent texture.
• Texture contrast (Q2):
∑
⎪
⎭
⎪
⎬
⎫
⎪
⎩
⎪
⎨
⎧
∑∑
=
−
=
=−
==
1
000
2
2),(
ggg
N
n
nji
N
i
N
j
jipnQ (4)
Contrast (Q2) reflects the diversity of image’s texture. A
low value of Q2 indicates coarse texture.
• Texture entropy (Q3):
∑∑
−=
==
gg
N
i
N
j
jipjipQ
00
3)},(log{),( (5)
An ordered distribution of Co-occurrence matrix
corresponds with maximum value of entropy. A higher value of
entropy indicates better regularity in texture.
• Texture correlation (Q4):
yx
N
i
N
jyx
gg
uujipij
Q
σσ
∑∑ −
===00
4
),()(
(6)
Hereinto, yxyx uu
σσ
,,
,is mean value and variance
of yx pp ,, ∑
=
=
g
N
i
xjipp
0
),(, ∑
=
=
g
N
j
yjipp
0
),(.
Correlation coefficient of texture reflects the correlations
between textures in different orientations. Coarse texture shows
a higher correlation coefficient value than fine texture.
TABLE I. PARAMETERS OF CO-OCCURRENCE MATRIX
Grade Q1
(θ=0,d=1)
Q2
(θ=0,d=1)
Q3
(θ=0,d=1)
Q4
(θ=0,d=1)
1 0.079135 45.389 0.0292 -3.1401
2 0.051062 109.06 0.0839 -2.3635
3 0.039969 169.06 0.0927 -2.1544
4 0.024251 264.54 0.1129 -1.1701
5 0.012324 332.95 0.1381 -0.8302
IV. NEURAL NETWORK MODELING
Artificial Neural Networks (ANN) [11-13] have been applied
to different textile problems such as classifying patterns or
defects in textile textures, and predicting fabric parameters
[14-15]. Among many Artificial Neural Network schemes,
feed-forward neural networks with back-propagation learning
algorithms based on gradient descent have been widely used,
since they offer unlimited approximation power for non-linear
mappings. Therefore, we establish an ANN model based on
feed-forward back propagation network and investigate the
performance of its neurons function to simulate the human
judging behavior;
According to the essential texture features based on Co-
occurrence matrix analysis, a two layer ANN system with 4
inputs and 1 output as shown in Figure 3, four input variables
include Q1, Q2, Q3 and Q4, one output variable is the final
rating grade. To be more explicit, a two-layer feed forward
NN model consists of a stream of input vector X of 4 neurons,
i.e. },,,{ 4321 xxxx , to the input layer, and an output layer with
a single neuron. The network has the general form:
⎭
⎬
⎫
⎩
⎨
⎧∑++=Θ=
=
)(),(
0
10
n
iiii bxwfbfXfY (7)
where Yrepresents the generated output response variable;
Θdenotes the overall parameters space; i
wdenotes the
connecting weight of the ith input neuron; 0
b and i
b are the
bias nodes; 1
fand fare the activation functions of first layer
and second layer respectively.
420
Figure 4. Design of Neural Network
Here, we use LOGSIG as the activation function of four
neurons of first layer and PURELIN as the activation function
of second layer.
;)(
;)(
1xx
xx
ee
ee
xf
xxf
−
−
+
−
=
=
(8)
V. SYSTEM TRAINING
The neural network model for the fabric appearance evaluation
was trained using a training set that includes 50 samples
ranging from grade 1 to grade 5. The training process was
terminated at a point where all the samples in this training set
were correctly identified. We first input the training data to the
network and perform gradient descent using the Levenberg-
Marquardt Algorithm. Parameters are adjusted iteratively until
global error function converges to some specified minimum so
that good generalization is achieved. Some important
considerations are the number of training iterations, the
learning rate and the momentum coefficient.
Supposing the estimated grade y is governed by the underlying
function );( *
θ
xf , where
x
is the set of input variables and *
θ
represents the true values of the parameter vector
θ
from the
parameter space Θfor the function which models the process.
With nobservations, the simulation process is represented by:
Θ∈=+= ** ,~1,);(
θεθ
nixfy iii (9)
The least-square estimate of *
θ
is the
θ
, which is obtained by
minimizing the error function:
∑−=
=
n
iii xfyS
1
2
)];([)(
θθ
(10)
The Levenberg-Marquardt algorithm uses this approximation
to the Hessian matrix in the following Newton-like update:
eJJJ TT
kk
1
1][ −
+Ι+−=
μθθ
(11)
Where, Jis the Jacobian matrix that contains first derivatives
of the network errors with respect to the weights and biases,
and e is a vector of network errors. When the scalar µ is zero,
this is just Newton's method, using the approximate Hessian
matrix. When µ is large, this becomes gradient descent with a
small step size. Newton's method is faster and more accurate
near an error minimum, so the aim is to shift towards Newton's
method as quickly as possible. Thus, µ is decreased after each
successful step (reduction in performance function) and is
increased only when a tentative step would increase the
performance function. In this way, the performance function
will always be reduced at each iteration of the algorithm. For
this study, we used the mean square error (MSE) between the
predicted output of the model and the target actual value as the
following equation:
n
TY
MSE
n
iii
2
1
)(
∑−
== (12)
Where n is the number of training data set. We trained the
neural network model using the parameters extracted using co-
occurrence matrix, and the MSE is 0.002624.
Figure 5. Performance of the model on the training data set
Figure 6. The Training Performance of Neural Network Model
VI. VALIDATION
After we trained the neural network models, the neural
network system was tested using the validation data set of 50
samples. The experimental results show that the artificial
intelligence system developed in this project could achieve a
good consistent with the human evaluation results, the
predicted rating grade tracked the actual rating grades very
421
well, as the correlation coefficient was sufficiently high (R2 =
0.9828).
Figure 7. Validation of Neural Network Model
VII. CONCLUSION
An artificial intelligence experts system based on image
analysis and neural network method was developed for the
objective evaluation of polar fleece fabric appearance. Co-
occurrence matrix was used to characterize the polar fleece
fabric appearance and one set of two-layer BP neural network
model based was established for rating of the degree of polar
fleece fabric appearance. The testing results for validation
show that the artificial intelligent system developed in this
research has very good correlationship with the human being’s
behaviour. It proves that it is workable to replace the
traditional subjective evaluation method based on human eyes
and brain by using the modern artificial intelligence method.
ACKNOWLEDGMENT
The first author of this paper would like to thank the Hong
Kong Polytechnic University for supporting him to work at the
Institute of Textile and Clothing and Department of Computing
during of his PhD studies.
REFERENCES
[1] J.Amirbayat and M.J.Alagha, Objective Assessment of Wrinkle
Recovery by Means of Laser Triangulation, J.Textile Inst., 1996, Vol.87,
Part I, No.2, 349-354.
[2] B.Xu and D.F.Cuminato, N.M.Keyes, Evaluating Fabric Smoothness
Appearance with a Laser Profilometer, Textile Res. J. 68(12), 900-906,
1998.
[3] Z. Fazekas, J.Komilves, I. Renyi and L. Surjan, Towards Objective
Visual Assessment of Fabric Features, Image Processing and its
applications, Conference Publication No.465 ©IEE 1999,411-416.
[4] D.D.Thibodeaux, Cotton Fiber Maturity by Image Analysis, Textile
Res.J，1986,VoL 56,130~139
[5] R.H.Gong and A.Newton, Image-analysis Techniques, Part II: The
Measurement of Fiber Orientation in Nonwoven Fabrics,
J.Text.Inst.,1996, 87, No 2, 371-388.
[6] F.Thorr, J.Y.Drean, and D.Adolphe, Image Analysis Tools to Study
Nonwovens, Textile Res. J. 69 (3),162-168 (1999).
[7] B. Pourdeyhimi, B. Xu, and J. Sobus,, Evaluating Carpet Appearance
Loss: Surface Intensity and Roughness, Textile Res. J. 63(9), 523-
535(1993).
[8] B. Pourdeyhimi, B. Xu, and L. Wehrel, , Evaluating Carpet Appearance
Loss: Periodicity and Tuft Placement, Textile Res. J. 64(1), 21-22(1994).
[9] M. Robert, K.H. Shanmugam, Textural Features for Image
Classification, IEEE TRANSAXTIONS ON SYSTEMS, MAN, AND
CYBERNETICS, 1973, SMC-3(6): p. 610-621.
[10] L.S.Davis, S.A.J., J.K. Aggarwal, Texture Analysis Using Generalized
Co-Occurrence Matrixes, IEEE TRANSAXTIONS ON PATTERN
ANALYSISAND MACHINE INTELLIGENCE, 1979, PAMI-1(3): p.
251-259.
[11] P. Lipmann, an Introduction to Computing with Neural Nets, IEEE
ASSP Magazine, April, 1987.
[12] J.E. Dayhoff, Neural Network Architectures: An Introduction, Van
Nostrand Reinhold, USA, 1990.
[13] L. Fu, Neural Networks in Computer Intelligence, McGraw-Hill,
Singapore, 1994.
[14] P. Chen, T. Liang, etc., ‘Classifying Textile Faults with a Back-
Propagation Network using Power Spectra’, Textile Research Journal,
68(2), 121-126,(1998).
[15] I.S. Tsai, C. Lin and J. Lin, ‘Applying an Artificial Neural Network to
Pattern Recognition in Fabric Defects’, Textile Research Journal, 65(3),
123-130,(1995).
