See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3171009

Fabric Defect Detection by Fourier Analysis

Article in IEEE Transactions on Industry Applications · October 2000

DOI: 10.1109/28.871274 · Source: IEEE Xplore

CITATIONS READS

510 2,105

2 authors:

Chi-Ho Chan Grantham Kwok Hung Pang

University of Surrey The University of Hong Kong
53 PUBLICATIONS 1,960 CITATIONS 192 PUBLICATIONS 6,115 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Chi-Ho Chan on 13 December 2012.

The user has requested enhancement of the downloaded file.

 Title Fabric defect detection by Fourier analysis

Author(s) Chan, CH; Pang, GKH

The 34th IEEE - I A S Annual Meeting Conference

Citation Record Industry Applications Society, Phoenix, Arizona,
USA, 3-7 October 1999, v. 3, p. 1743-1750

Issue Date 1999

URL http://hdl.handle.net/10722/46161

©1999 IEEE. Personal use of this material is permitted.

However, permission to reprint/republish this material for
advertising or promotional purposes or for creating new
Rights collective works for resale or redistribution to servers or
lists, or to reuse any copyrighted component of this work
in other works must be obtained from the IEEE.
 Fabric defect detection by Fourier analysis
Chi-ho Chan and Grantham Pang
Dept. of Electrical and Electronic Engineering
The University of Hong Kong
Pokfulam Road
Hong Kong
Fax: (852)-2559-8738
Phone: (852)-2859-2689
Email: [email protected]

Abstract - Many fabric defects are very small and Zhang et. al. [3] have introduced two approaches to detect
undistinguishable, which are very difficult to detect by only defects: gray level statistical and morphological methods.
monitoring the intensity change. Faultless fabric is a repetitive Lanes [4] has defined a number of convolution masks to
and regular global texture and Fourier transform can be applied detect the defect. These methods, which depend on intensity
to monitor the spatial frequency spectrum of a fabric. When a
defect occurs in fabric, its regular structure is changed so that
change on the fabric image, can only capture the significant
the corresponding intensity at some specific positions of the defects such as knot, web and slub. However, faultless fabric
frequency spectrum would change. However, the three- has a periodic regular global structure. The occurrence of a
dimensional frequency spectrum is very difficult to analyze. In defect in the fabric means that the regular structure has been
this paper, a simulated fabric model is used to understand the destroyed. Therefore, the fabric defect can be detected by
relationship between the fabric structure in the image space and monitoring fabric structure. C.Ciamberllini and his
in the frequency space. Based on the three-dimensional colleagues have used the optical Fourier transform to monitor
frequency spectrum, two significant spectrum diagrams are the fabric structure. Their methods can be classified as
defined and used for analyzing the fabric defect. These two follows: fixed masks or structured detectors, adaptable filters
diagrams are called the central spatial frequency spectrums.
The defects are broadly classified into four classes: (1) double
[5], electronic elaboration of the image [6], and binary
yarn; (2) missing yarn; (3) webs or broken fabric; and (4) yarn histogram [7]. Fixed masks, structured detectors and
densities variation. After evaluating these four classes of defects adaptive filters are the fixed optical filters placed over the
using some simulated models and real samples, seven photodetector active area. They are composed of periodic
characteristic parameters for central spatial frequency array transparent and opaque zones, and their locations
spectrum are extracted for defect classification. depend on the peak of the fabric diffraction image. When a
defect occurs, some localized peaks in the diffraction image
will be passed to the sensor. However, this type of operation
I. INTRODUCTION requires a careful alignment and different fabric type requires
different filter. The method of electronic elaboration is based
In the textile industry, before any shipment to customers, on the subtraction of the reference image of faultless fabric
inspection is needed for maintaining the fabric quality. from that of the fabric under inspection. The binary
Srinivasan et. al.[l] have stated that the price of second- histogram method is based on a summation of the gray level
quality fabric is only 45 to 65% of that of first-quality fabric. pixels, which have gray level greater than a threshold. The
However, the current inspection process still depends mainly threshold is defined according to the fabric type and system
on human sight. This nature of work is very dull and sensitivity. Anyhow, optical processing is very expensive
repetitive. Moreover, there could be many human errors in and requires careful calibration. Also, this method could not
this process. According to some studies, human visual recognize fabric with higher fabric density.
inspection can only catch around 60% to 75% of the
significant defects [2]. Therefore, in order to lower the cost Another class of method is dependent on digital image
of the inspection process and to increase the competitive processing. Wood [8] has used Fourier and associate
advantage of the products, it is necessary to automate the transform to characterize carpet patterns. Ravandi and
inspection process. Toriumi [9] have used Fourier transform analysis to measure
fabric appearance. Fabric surface characteristics, fill and
In this study, the central spatial frequency spectrum warp yarns for plain-weave cotton fabric were also discussed.
approach is introduced and examined. This method would Escofet et. al. [lo] have used the angular correlation of the
reduce the computational time for defect detection and Fourier spectra to evaluate fabric web resistance to abrasion.
provide more parameters for defect classification. Before Sari-Saraf and his colleagues [ 113 have used Fourier
introducing this method, the characteristics of fabric structure transform to detect fabric defects. Their approach examines
in frequency spectrum will be examined and some defect and performs a one-dimensional signature in the two-
examples will be described. After that, the procedures of the dimensional spectrum. The signature is obtained by
method and experimental results will be discussed. integrating the points within each ring in the two-dimensional
spectrum. The rings are concentric with different radii, and
they are used to monitor the fill and wrap densities. The idea

0-7803-5589-X/99/$10.00 0 1999 IEEE 1743

 of this approach is similar to the concentric ring filter in
optical processing [ 5 ] . The main advantage of this approach
is that it is less sensitive to the background noise (intensity
source variation). Also, it is more effective for revealing a
defect due to dimensional changes in the structure of the
fabric. This method can recognize the high fabric densities
by using the zoomed camera, but long computational time is The simulated pattern of a faultless fabric is shown in
expected. Figure l(c). The faultless fabric image pattern, f(x,y), is
modeled as shown in (4) below.
11. BACKGROUND
KNOWLEDGE

A. Fourier Transform

Fourier theorem states that any signal can be represented

by the sum -of the sine and cosine wave with various rect(u, v ) =
{A I U IS 0.5 and I v IS 0.5
otherwise .
amplitudes and frequencies. That is, the relationship between
a repetitive, regular and uniform fabric pattern in the image where 6 (u,v)= 6 (u)6 (v), which is the two-dimensional Dirac
space and its spectrum in the spatial frequency can be linked 6 function.
by operating two-dimensional Fourier transform. Let a two-
dimensional image be f(x,y), which is a real function The first part is a function that describes the fabric
representing the gray level in x, y spatial coordinates, and let structure gap (Figure l(a)), where s, is the width of the gap
the image width and image length be N . Then F(n,m) denotes and sy is the height of the gap. Let w, be the wrap period and
the Fourier. transform of f(x,y) with n and m spatial wv be the fill period. The second part is the impulse function
frequencies. The general equation of two-dimensional that models the location of the gap in the fabric (Figure 1(b)).
discrete Fourier transform is shown below.
The Fourier transform of the convolution of two functions
F(n, m ) = -
1
c
N-1

N Z y=o
N-1

x=o
f(x, y ) * e~'z"''+y"'~'N in an image space is equal to the product of their respective
Fourier transforms as shown in ( 5 ) below.

The computational time for Fourier transform is generally 1 FK ,fy ) I=Is, . sy .sinc(23, .s, ). s i n c ( 2 .~sy~) +

long. For two-dimensional discrete Fourier transform, it is

proportional to the second order of the image size. In order
to reduce the computation time, Fast Fourier transform (FFT)
is used. Fast Fourier transform is a discrete Fourier transform
with some reorganization that can save enormous amount of The frequency spectrums of the corresponding figures of
time. For one-dimensional FFT, the computation time is Figure 1 are shown in Figure 2. The envelope of the faultless
NZogzN. Because of the separable transform [ 121 being used fabric in frequency space shown in Figure 2(c) is proportional
to perform the two-dimensional transform, the computation to the envelope of the Fourier transform of the fabric
time is proportional to 2N2Zog2N. One of the advantages for structure gap as shown in Figure 2(a). In fact, the shape of
the spatial frequency spectrum approach is the translation the envelope of Figure 2(a) is affected by two parameters: s,
property of Fourier transform [ 121 that is shown in (2), which and s,,. For example, if the structure gap s, and s,, are
means that the magnitude of frequency spectrum does not increased, the envelope shape of Figure 2(a) will be narrower.
- change when the fabrick moved up. The spectrum is only Referring to Figure l(b), the location of the impulse in the
varied by the change of fabric structure. image space is inversely proportional to the impulse location
in frequency space in Figure 2(b). The distance between two
impulses in Figure 2(b) should be I/w,in the horizontal axis
and I/w, in the vertical axis. These properties can aid to
. B. Characteristics of Fabric Structure in the Magnitude of monitor the yarn densities in the two directions. For
Frequency Spectrum ~
example, the density in the horizontal direction is equal to the
distance between two impulses in the horizontal axis (i.e.
A simulated frequency spectrum should be examined for I/w,) multiplies with the image width.
understanding the characteristics of fabric structure in the
magnitude of frequency Bpectrum. Two-dimensional discrete C. Simulation of Fabric Defect
convolution theorem is involved for constructing a simulated
fabric. Let h and & be the input and output image In this paper, the defects are broadly classified into four
respectively, and t is a. convolution mask. Therefore, the classes: (1) double yam; (2) missing yam; (3) webs or broken
convolution denoted by 60 is the computation of weighted fabric; and (4)yarn densities variation. Examples are shown
sums of the image pixels with the convolution mask, which is in Figure 3. The double yam (fill) is a change of spatial
shown in (3). periodicity on the vertical axis. The spectrum on thef,-axis
(fill direction) denotes the corresponding change of spatial

1744
frequency. In a simulated model of the double yarn, D(x,y), the crosses. Both Figure 6(b and c) show that the central peak
shown in Figure 4(c), the defect can be regarded as a value (Pl) of the defect fabric is lower than that of the
subtraction from a faultless fabric (Figure 4(a)) to a series of faultless fabric. This is because one or more vertical (Wrap)
rectangle function, d(x,y) (Figure 4(b)). Because of the threads are added in the faultless fabric, which leads to be
distributivity property of the Fourier transform, lower light intensity in the defect image. The double wrap
defect is a change of spatial periodicity in the horizontal axis,
FJD(x3 Y)) = Fcf(x,Y) - d(X1-Y))= Flf(x,YI/' - F/d(XI Y)} (6) and therefore the IF(0,fJI diagram is changed mostly. In this
which means that the defect in the frequency space can be diagram (Figure 6 (c)), the first peak value of the defect at&,
formed by subtracting the faultless fabric frequency spectrum =35 are lower than the faultless fabric first peak values and
to the Fourier transform of an irregular structure function ripples occur. So P5 should be lower and P7 should be
d(X,Y). higher. The first peak location (P6= J$ =35) is a
fundamental yam frequency, which means that the peak
D. Central Spatial Frequency Spectrum location is proportional to the yarn density. In this example,
P6 is almost unchanged. A summary is shown in the second
Due to the nature of the fabric structure, many defects row of table 1. With a similar interpretation, the double fill
would occur along the x and y-axes, which means that those defect is a change of spatial periodicity in the vertical axis.
characteristics would appear on the wrap G) and fill V;) Hence, the parameters P2 and P4 are changed because the
direction in the frequency spectrum. In addition, a three- defect only affects the lF&,O)\ diagram.
dimensional graph of the frequency spectrum is very difficult
to analyze. The method of central spatial frequency spectrum The difference between missing yam and double yam is
is therefore proposed in this paper. This method extracts two their fabric threads count. Average light intensity P1 can
diagrams along the fx and fy direction (IF&, 0) I and IF(0,fy) I) show this characteristic. For missing yam, higher P1 is
from the three-dimensional graph, which are shown in Figure expected because there is less thread in this defect. Broken
5. fabric and yam densities variation are a change of periodicity
in both x and y axis, and both IF&, 0) I and IF(0,fy) I diagrams
Seven significant features can be extracted in these two are mostly changed. P3 and P6 values are not changed in the
diagrams for describing defect characteristics. The equations broken fabric. This is because this defect is a instant change
of these parameters are shown below. of the fabric density, and it only affects P1, P2, P4, P5 and
P7 values. The P1 is high because the fabric is broken, and
leads to increase of light intensity. Details of the expected
results are summarized below.
p3 =&I
TABLE 1
DIFFERENCE
IN PARAMETERS PREDATED BETWEEN THE FABRIC
AND ITS
DEFECT

p7 '1OOx
[f;:o I j
C F ( O , f , , ) F(0,O)

whereLl andhi correspond to the first harmonic frequency in

their respective diagrams that are also shown in Figure 5.
The first feature P1 is the average light intensity of the image,
which is used to characterize the yam density. Higher yarn
density decreases the light intensity and P1 is decreased, and
vise versa. P5, P6, P7 are used to monitor the wrap (vertical)
threads structure whereas P2, P3 and P4 are for detecting the
fill (horizontal) threads structure. When defects occur, the
amplitude of harmonic frequencies plus other changes would H. Higher, L. Lower, NC. no change.
appear in the central spatial frequency spectrum. Features P2,
P4, P5 and P7 are used to describe and detect these 111. EXPERIMENTAL
SETUP
characteristics. Feature P3 and P6 are used to monitor the
wrap and fill threads density in the image. Those features are A. Image Acquisition
more concentrated on analyzing the region between the
central peak and first peak (first harmonic frequency) because In this study, plain white fabrics are used. Defects with
higher harmonic frequency components are significantly double yarn, missing yarn, web and fabric density variation
distorted in real environment. were inspected and compared with the faultless fabric. The
image acquisition system includes a personal computer
A defect such as double wrap is shown in Figure 6(a). A (Pentium-200Mm), frame Dabber, cmos-imager and a
comparison of the defect spectrum and faultless fabric system monitor. The fabric image is captured by a cmos-
are shown in ~i~~~~6(b and c), where the defect is imager, and a frame grabber digitizes the video signal into a
denoted by the solid line and the faultless fabric is denoted bj, 768*576 pixel with eight-bit gray level resolutions image and

1745
 stores it into computer memory. This image data is then (Wrap) (HI (NC) (NC) (NC) (L)
processed by the defect detection procedures as shown in Yarn
Broken 33.4 -14.9 0 32.95 -24.2
Figure 7. Backlighting illumination is used to achieve the Fabric (H) (L) (Nc) (H) (L)
high contrast fabric images. (Web)
Low 32.2 -17 -8 -0.8 -253
B. Histogram Equalization fabric (H) (L) (L) (L) (L)
density
Histogram equalization is performed to obtain a uniform TABLE 3
density image histogram [121. This process extends the DIFFERENCE
M PARAMETERS OBTAINED BETWEEN A REALFABRICAND ITS
dynamic range of gray levels and increases the image DEFECT.
contrast. The aim is to standardize the brightness and
contrast of the images. Figure 8 gives an example of using
equalization to improve the overall brightness and contrast of
an image.

C. Fast Fourier Transform

A Fast Fourier Transform (two-point) transform is used

[12] for fast computation, which means that the image size is
cut to 512x512 pixels in our experiment. This is because the
image length and width should be a power of two. A Missing Wrap is an example for describing these
software package, Matlab, is used for this experiment. After parameters. Since this defect is missing one or more vertical
the Fourier transform, the central spatial frequency spectrum threads, it only occurs in IF(OfJ1 diagram. Therefore, the
diagram is extracted from the three-dimensional diagram. A significant parameters should be PI, P5, P6 and P7. Since
real sample of double fill and its central spatial frequency the fabric thread count is lower than the faultless fabric, the
spectrum diagrams are shown in Figure 9(b, c and d). average light intensity of defect is higher. Due to this reason,
P1 should therefore be higher. In addition, due to the
By observing these two diagrams (Figure 9(c and d)) and irregular texture in the wrap direction, the first peak value is
comparing them with the simulated diagrams of double wrap decreased and ripple occurs, which would cause P5 lower and
in Figure 6(b and c), the orientation of the defect mainly P7 higher. However, because of the noise effects on the real
affects the particular diagram. For example, double wrap samples that are mentioned in the previous section, the
affects the parameters in IF(Of,,I diagram, and double fill parameters in table 3 are distorted. Anyway, those changed
only affects the parameters in IF&,O)l diagram. However, parameters from the real samples are quite similar to the
the high spatial frequency peaks in Figure 9(b and c) are simulated parameters. Moreover, the bolded parameters in
loosely localized and embedded with some noise. A number table 2 and table 3 show that different classes of defects have
of reasons can explain this result. For example, the surface their own changed parameters, which means that these seven
tufts give a random textured component in the fabric images parameters can be used to classify the fabric defect type.
that distort the periodic structure. In addition, the
illumination fluctuation gives high frequencies noise V. CONCLUSIONS
background. Because of these effects, only parameters
around the central peaks and the first peaks can be extracted The Fourier transform approach is described to detect the
for defect characteristics. structural defect. The simulated models are used to
understand the behavior of frequency spectrum. Since the
IV. RESULTS three-dimensional frequency spectrum is very difficult to
analyze and many defects occur along the horizontal and
The negative parameter in the table means that the vertical axes, the central spatial frequency spectrum approach
particular characteristic of the defect spectrum pattern is has been proposed to increase the efficiency of the analysis
lower than that of the faultless fabric. In this experiment, process. Seven significant characteristic parameters can be
four defect models and the four corresponding real samples extracted from the central spatial frequency spectrums for
are used to examine this approach. Table 2 gives the describing the defect type. Experiments have shown that the
difference in parameters obtained between simulated fabric extracted parameters can be used to classify fabric defects.
and its defect. The results in table 2 follows our expectation
summarized earlier in table 1. REFERENCES
TABLE 2 [ 11 KSrinivasan, P.H. Dastoor, P.Radhakrishnaiah and S. Jayaraman,
DIFFERENCE IN PARAMETERSOBTAINED BETWEEN A SIMULATED FABRIC “FDAS: A Knowledee-based Framework for Analvsis of Defects in
MODEL AND ITS DEFECT. Woven Textile Stru&res”, J. Text. Inst., vol. 83, Part 1, no. 3, pp.431.
I n -PI- n -~2- I -~3- I _ ~4 . 11 -PS- I _~6_ I P7 II .l A.A.-R 11993\
\-,--,

Double -2.1 0 0 0 -14.4 0 66.7 [2] K Schicktanz, “Automatic fault detection possibilities on nonwoven
(Wrap) (L) (NC) (NC) (NC) (L) (NC) (HI fabrics,” Mellrand Textilberrchie,pp 294-295, (1993)
yarn
Missing 2.9 0 0 0 -5.3 0 71.8

1746
X. F. Zhang and R. R. Bresee, “Fabric Defect Detection and [8] E. J. Wood, “Applying Fourier and associated transforms to Pattern
classifiction using Image Analysis,” Textile Research Journal, vol. Characterization in Textiles,” Textile Research Journal, pp.212-220,
65(1), pp.1-9, (1995). (1990).
J. S. Lane, S.C. Moure, “Textile fabric Inspection System,” US Patent [9] S.A.H. Ravandi and K. Toriumi, “Fourier Transform Analysis of plain
No. 5,774,177 (1998). Weave Fabric Appearance,” Textile Research Journal, vol. 65(1 I), pp.
C. Ciamberlini, F. Francini, P. Sansoni and B. Tiribilli, “Defect 676-683 (1995).
detection in textured materials by optical filtering with structured [IO] J. Escofet, M.S. Millan, H. Abril and E. Torrecilla, “Inspection of
detectors and self-adaptable masks”, Opt. Eng., vol. 35(3), pp. 838-844, fabric resistance to abrasion by fouries analysis,” Proc. SPIE , vol.
(1996). 3490, pp. 207-210 (1998).
C. Castellini, F. Francini, G. Longobardi and B. Tiribilli, “On-line [l I ] H. Sari-Saraf and J. S. Goddard, “On-line optical measurement and
Textile Quality Control using Optical Fourier Transforms,” Optics and monitoring of yarn density in woven fabrics,” Automated Optical
Lasers in Engrneering,vol. 24, pp. 19-32 (1996). Inspectionfor Indushy, SPIE, vol. 2899, pp. 444-452 (1996).
C. Ciamberlini, F. Francini, G. Longobardi, P. Poggi, P. Sansoni and B. [I21 R. C. Gonzalez and R.E. Woods, “Image transforms”, Chap.3 and
Tiribilli, “Weaving defect detection by Fourier Imaging,” Vision “Image enchancement”, Chap.4 in Digital Image Processing, Addison-
systems in applications SHE, vol. 2786, pp. 9-18 (1996). Wesley, pp.81-128, 166-189 (1993).

(b) (c)
Figure l(a) Fabric structure gap. (b) Fabric gap location. (c) Simulated faultless fabric

(a) (b) (c)

Figure 2. Magnitude of frequency spectrum (Central portion): (a) Fabric structure gap, (b) fabric gap location, (c) simulated faultless fabric

(4 (b) (c) (4
Figure 3 Fabric defect samples: (a) Double yarn; (b) Missing yarn; (c) Broken fabric; (d) Variation of yam.

(a) (b) (c)

Figure 4 (a) Faultless fabricf(x,y), (b) Rectangle series function d(x,y), (c) Double fill defect

1747
 F(h.01 spccwm

(c)
Figure 5 (a) Portion of faultless fabric frequency magnitude spectrum (b) Fill direction frequency spectrum, (c) Wrap direction frequency spectrum

I
Fabric Image Acquisition - - - Histogram
Equalization
Fast Founer
Transform
Central Spatial
Frequency
Spectrum Analysis

7000
7000
6000 6000
5000 5000

4000 4000

3000 3000

2000 zoo0

1000 1000

0 0

0 02 04 06 08 1 0 02 04 06 08 1

8 (c) (4
Figure 8 Example of an image with gray level histogram.
(a) original image and (b) gray level histogram (c) image after equalization and (d) gray level histogram after equalization.

1748
 F(fx,O) Spectrum of Double Wrap
110

100

90

80
>
.-
U
U]
2
+
70
.-S
2 60
.-m
-
7s
w 50
-
.-N
E
Tt
40
0
Z
30

20

10

0
-250 -200 -150 -100 -50 0 50 100 150 200 2 0
Fill direction
F(0,fy) Spectrum of Double Wrap
110

100

90

> 80
.-U
VI
6 70
+
U 4
.-S
2 60
-25)
.-
U
w 5c I I
-
.-N
E
0
4c
z
3c

2c
~

It -t

c I
0
Wrap direction
Figure 6 (a) Double yarn, (b) its fill direction spectrum, (c) its wrap direction spectrum

1749
 F[fx,O) direction
120 I I I I I I I I I

100

20 +

0
-1
U+&+,
. ...
,..
-80 -60
,.
-40
/

-20 0
:@
40 60 80 20
Fill direction
\-I

F(ik.0) direction
120

100

80

;
._
8
._
E
60
U
._
N

-fz"
40

20

-60 -40 -20 0 20 60 80

Wrap direction
Figure 9 (a) faultless fabric, (b) double fill, (c) their fill direction spectrum, (d) their
wrap direction spectrum.

1750

View publication stats