0% found this document useful (0 votes)
27 views11 pagesUsing Singular Value Decomposition in Textile Production Quality Control
Stjsjtxkyckhfutdjrdithfzkhxjfxjgxlhckgdjtxjtxkhxjtxjgxjfxjgxmhckhckgxkhckhckyxjgxjtsjtsitsykydykyditsutzutdutsiyxkhckhckgxigxydjgxjgxjtzjgxjgxkhzyeajrautxclvjvlvhchxjtdjtditsugzjrsjgxitsutdofukxgxjtsutditsutsutsjtsjtsjtsjtsjtdjtsitditditsitditsitsutsutsursurshrzhdfy
Uploaded by
Hasan sonjayaCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
27 views11 pagesUsing Singular Value Decomposition in Textile Production Quality Control
Stjsjtxkyckhfutdjrdithfzkhxjfxjgxlhckgdjtxjtxkhxjtxjgxjfxjgxmhckhckgxkhckhckyxjgxjtsjtsitsykydykyditsutzutdutsiyxkhckhckgxigxydjgxjgxjtzjgxjgxkhzyeajrautxclvjvlvhchxjtdjtditsugzjrsjgxitsutdofukxgxjtsutditsutsutsjtsjtsjtsjtsjtdjtsitditditsitditsitsutsutsursurshrzhdfy
Uploaded by
Hasan sonjayaCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 11
‘Theoretical and Applied Informatics
ISSN 1896-5334
Vol 20 (2008), no. 3
pp. 167-177
Using singular value decomposition in textile production quality control
Piotr BEDNARCZUK *
“Institute of Computer Science,
Lublin University of Technology, Poland
[email protected]
Received 1 October 2008, Revised 20 October 2008, Accepted 20 November 2008.
Abstract: The article presents a possibility of using singular value decomposition in textile quality
control. This approach consists in using only first singular values in the analysis and examining their
standard deviation from the referential values. The examined object is textile, represented in the form of an
image and saved as a data matrix. The article includes comparison of the efficiency and speed of different
data analysis algorithms and methods. SVD method, in regard to defect detection, shows a high efficiency.
Furthermore, the speed of the proposed solution is comparable with the fastest algorithms and is the best
from among the methods with the same efficiency. The results of the conducted and described experiment
consisting in examining five different textile defects confirm the potential of the method chosen.
Keywords: Textile quality control, singular value decomposition, SVD
1. Introduction
Quality control in today’s production systems is as important, as the control process
itself. These elements often work together in a feedback system. Indeed, this solution
has been applied for a long time, but nowadays, in quality control systems computers
and complex data analysis algorithms have been widely used. It stems mainly from the
quality examination of more and more complex objects or processes. Depending on
the type of the examined object, this complexity may be so high that the measurement
consists in the analysis of data saved as images obtained from vision systems. Nowadays,
we use analyses of various data, whose images are textures and these are e.g.: paper,
textile, wood, pottery, flames, clouds, astronomical objects, DNA code and many others.
Depending on the objects analysed, various methods of image processing are applied.
From the simplest ones, statistic or stochastic, through methods based on transforms
and decompositions, to the most advanced methods using artificial intelligence. The 168
former ones are characterised by high speed, but they have low efficiency. Methods
based on transforms or decomposition are efficient, but because of their complex numeric
algorithms, they are too slow. Methods based on artificial intelligence were to bring a
satisfactory efficiency and sufficient speed. These methods, however, require learning,
which, depending on the complexity of the object takes time. And vice versa, the time of
learning influences efficiency, so those methods gain a satisfactory efficiency only after
some time. The following article proposes a solution characterised by high efficiency
and relatively short processing time. The chosen algorithm is considered to be one of
the most efficient methods, based on data decomposition. Reduction of processing time
was obtained using an SVD decomposition feature, consisting in using only the first
singular value. This procedure results in loss of efficiency, but because of the properties
of SVD decomposition, this loss is slight.
2. Singular Value Decomposition (SVD)
Singular value decomposition consists in saving one data set as another data set,
preserving most of the information about this set. For each real matrix A with the di-
mensions m x n, where m is the number of rows and n is the number of columns, there
are such orthogonal U matrices with the dimensions m by m and V wit the dimensions
n by n, for which the following dependence is true::
qa
where / = min(m,n).
If the size of A matrix is r, then: 07 2072 2.02 0} > O17 p41 = Ors2 a=0
U and V matrices are orthonormal, i.e. orthogonal U~' = U7 and V~' = V7 and the
length of each vector equals 1. Those matrices are called right and left singular matrices
respectively.
Matrix 5 = diag(o'1, 2, .-.'n) is a diagonal matrix of singular values o with the dimen-
sions m x n set in a descending order. The first singular value 0; located in Sigma matrix
in 511 position has the greatest importance and highest sensitivity for any disturbances
in the analysed data [4].
There are many algorithms of numeric singular value matrix decomposition. One of
them, proposed by G.H. Golub and C. Reinsch is a numerically stable algorithm, based
‘on orthogonal transformation of the input matrix using Householder transform to a bidi-
agonal matrix, which is subject to diagonalization, resulting in obtaining singular val-
ues. [6].
The basic feature of SVD is that by using only the first singular values, it is pos-
sible to reconstruct the original data preserving their most relevant features using the
dependence 2: 169
A=USV™ Q)
With such limitation, the information content, entropy and energy of the reconstructed
image still remain on a very high level. Using this most important feature of singular
value decomposition, the possibilities of the method in defect detection according to its
efficiency and speed of analysis of textile images was examined.
Other methods of image analysis
There are many methods of quality image analysis, so there are also many ways of
their classification. It seems reasonable to divide texture analysis methods according to
image processing into:
© statistical, stochastic
© histogram examination [24]
© LBP local binary pattern [7] [18]
© GLCM -co-occurrence - event matrix method [19] [9]
© transforms
© Gabor filters [23] [24]
© DCT — discrete cosin
© DWT - discrete wavelet transform [3]
nsfonn [8]
decompositions
o PCA - eigen values analysis [11] [20].
© KLT —Karhunen-Loeve transform [22]
o SVD — singular values decocoposition [4]
methods based on artificial intelligence [13] [14]
Since the 70s, those methods have been constantly developed and combined, thus mak-
ing the image analysis more and more effective [10].
‘Comparative study of the efficiency of image analysis methods
Those methods may be used in qualitative image analysis, we just have to choose the
right classifier, which will make a classification according to predefined quality thresh-
‘olds. Some of these methods are very efficient, the others very quick. It is difficult to 170
find a method that would be both very efficient and very fast. There are many publi-
cations comparing efficiency and speed of algorithms [17]. The author of one paper
observed that the best efficiency is obtained by the image analysis methods based on
decompositions, PCA and KLT, obtaining a result over 85%.
Comparison of the speed of image analysis methods
To compare SVD algorithm speed to other decomposition-based algorithms, com-
parative analysis in Matlab software was conducted. The same methods, analysed be-
fore in regard to their efficiency, were chosen for the study. The speed of algorithms was
compared for images in 8-bit colour mode and the size of 512 x 512 and 102 x 1024
pixels in the same hardware and software conditions: computer AMD Sempron 2800
MHz 1GB RAM, operating system Windows 2000. The results are shown in table 1.
Because of large differences in computational time, a logarithm scale on a time axis was
used to visualise the results. It enables us to show the relationship between the speed of
specific algorithms clearly.
Methods Method Time processing, ms
Matrix size | 512x512 | 1024 x 1024
histogram 47 180
Statistical LBP 1970 8090
‘GLOM 34 370
Transforms DcT 140 660
DWT 8 230
DWHT 15 16
PCA 12440 96550
Decompositions | KLT 2750 24440
SVI6 410 1680
Tab. 1. Comparison of the processing time of various image analysis algorithms 171
7” seo0n02s
7 s02
“yiegam UP Gc ocT. DWT COWHT PCA KUT. SVD
Fig. 1. Comparison of the speed of image analysis algorithms
On the basis of the measurements conducted we may say that among from the com-
pared methods, those characterised by high efficiency are the slowest. In the figure 1 it
was shown that singular value decomposition method using only the first singular value
is at least one order of magnitude faster than other decomposition-based methods, e.g.:
KLT and PCA based on eigenvalues. The proposed approach is also comparable with
the fastest methods based on transforms and even faster than some statistical methods.
3. Using singular value decomposition method to analyze textile images
Textile images have some common qualities, which we can use in defect analysis. These
are, for example:
¢ Uniform textile colour with different possible hues, according to structural param-
eters, construction of the yarn, materials used or light reflection coefficient for
specific fibres.
Uniform surface in the same areas, resulting from the structure or workmanship.
Identical effect on the reflected light, in which images of the same textile were
taken.
Preserving identical method of observation for all the samples it is possible to analyze
textiles with regard to the occurrence of differences between its separate elements. 172
To analyze images of large textile surface area, we propose to divide into smaller frag-
‘ments and each fragment should be analyzed individually. Textile image shall be repre-
sented by a matrix of data, whose elements shall constitute values of separate pixels of
this image. The complexity of the analyzed images determine whether it is enough to
analyze a black-and-white or colour image with complex RGB or CMYK colour modes,
the analysis of which shall be conducted similarly to black-and-white images, the differ.
ence being that we do not analyze just one component of the image, but 3 of 4, respec-
tively.
Using one of the SVD decomposition features, which says that in a diagonal matrix E,
singular values are arranged in descending order and the first singular value 1 situated
on the diagonal has the largest weight and sensitivity, much larger than the following
singular values, so we use only this one to analyze textile images [4]. It was therefore
assumed that the first singular value is enough to preserve the information on textile
defect. In the figure 2, images reconstructed using the first singular value are presented,
both the fragment without defects and the one with defects. The picture 2.d shows that
after image reconstruction using the first singular value, defect information is not lost.
Fig. 2. Sensitivity of the first singular value to image changes: a) textile image without defects; b)
reconstructed textile image without defects using first singular value; c) textile image with a pattern defe
4) reconstructed textile image with a pattern defect using first singular value
The shape, texture, surface and colour of a textile have the largest influence on singular
values in textile images. On the basis of the above, the first singular values of textile 173
images without defects and images of the same textiles with defects differ significantly,
To use this method to textile production quality control, it is enough to determine the
range of singular values for textiles without defects. In other words, it is enough to
determine the maximum standard deviation of the first singular values for RGB colour
mode. All the images of textile fragments, whose singular values are not included in
the determined range, or for which the standard deviation is higher than the maximum
deviation for images without defects should be treated as images with defects or images
that do not meet quality standards.
4. The experiment results
The algorithm using SVD method to detect textile defects was designed in Matlab
environment. Textile image was loaded to memory and saved in the form of a matrix.
SVD matrix decomposition was made using SVD function. In the result of the decom-
position, two matrices U and V and one real singular value vector S, whose first element
is the basis for detecting textile defects were obtained. During the experiment, such
quality classification rule was employed that the images with standard deviation of at
least twice the size of the maximum permissible standard deviation for images without
defects were classified as images with defects. This standard deviation may be expressed
using formula 3:
& .
x bi - 31
D=-=— G)
where
xj — first singular values of textile image for R, G, B colour mode, respectively.
n — number of images without defects
¥ — mean singular values expressed by formula 4:
