![The schema of a single pixel of a TFT panel ([26])](https://www.researchgate.net/profile/Chi-Jie_Lu/publication/227299671/figure/fig1/AS:393818147966980@1470904971996/The-schema-of-a-single-pixel-of-a-TFT-panel-26_Q320.jpg){kind=tracking}
Content uploaded by Chi-Jie Lu
Author content
All content in this area was uploaded by Chi-Jie Lu on Dec 21, 2013
Content may be subject to copyright.
DOI 10.1007/s00170-003-1832-6
ORIGINAL ARTICLE
Int J Adv Manuf Technol (2005) 25: 53–61
Chi-Jie Lu · Du-Ming Tsai
Automatic defect inspection for LCDs using singular value decomposition
Received: 13 March 2003 / Accepted: 13 June 2003 / Published online: 2 June 2004
Springer-Verlag London Limited 2004
Abstract Thin film transistor liquid crystal displays (TFT-
LCDs) have become increasingly popular and dominant as dis-
play devices. Surface defects on TFT panels not only cause
visual failure, but result in electrical failure and loss of LCD
operational functionally. In this paper, we propose a global ap-
proach for automatic visual inspection of micro defects on TFT
panel surfaces. Since the geometrical structure of a TFT panel
surface involves repetitive horizontal and vertical elements, it
can be classified as a structural texture in the image. The pro-
posed method does not rely on local features of textures. It is
based on a global image reconstruction scheme using the singu-
lar value decomposition (SVD). Taking the image as a matrix of
pixels, the singular values on the decomposed diagonal matrix
represent different degrees of detail in the textured image. By se-
lecting the proper singular values that represent the background
texture of the surface and reconstructing the matrix without the
selected singular values, we can eliminate periodical, repetitive
patterns of the textured image, and preserve the anomalies in the
restored image. In the experiments, we have evaluated a variety
of micro defects including pinholes, scratches, particles and fin-
gerprints on TFT panel surfaces, and the result reveals that the
proposed method is effective for LCD defect inspections.
Keyw ords Defect inspection · LCDs · Machine vision ·
Singular value decomposition
1 Introduction
Thin film transistor liquid crystal displays (TFT-LCDs) have
become increasingly important in recent years due to their full-
colour display capabilities, low power consumption and small
C.-J. Lu · D.-M. Tsai (u)
Department of Industrial Engineering and Management,
Yuan-Ze University,
135 Yuan-Tung Road, Nei-Li, Tao-Yuan, Taiwan, R.O.C.
E-mail: iedmtsai@saturn.yzu.edu.tw
Tel.: +886-3-4638800
Fax: +886-3-4638907
size. In order to monitor the process stability and guarantee the
display quality of LCD flat panels, the inspection of defects on
the TFT panels becomes a critical task in manufacturing. Hu-
man visual inspection and electrical functional tests are the most
commonly used methods for LCD defect detection. However,
manual inspection is a time-consuming and tiresome task. The
manual activity of inspection could also be subjective and highly
dependent on the experience of human inspectors. The elec-
trical functional test is inherently limited to offline operations,
and generally can only be accomplished after the fabrication of
a TFT panel is complete. In this paper, we propose an automatic
visual system for LCD defect inspections.
Surface defects on the TFT panel not only cause visual fail-
ure but also electrical failure to operate an LCD panel. Appear-
ance defects on TFT panels can be roughly classified into two
categories: macro and micro defects [12]. Macro defects include
“MURA”, “SIMI” and “ZURE”. MURA means unevenness of
TFT panels, SIMI mean stains on TFT panels and ZURE means
misalignment of TFT panels. Micro defects include pinholes, fin-
gerprints, particles and scratches. The macro defects appear as
high contrast regions with irregular sizes and shapes. They are
generally large in size and, therefore, can be easily detected by
human inspectors. However, sizes of micro defects are generally
very small and can not be easily found by human inspectors or
detected with electrical methods. The proposed method in this
paper especially focuses on the inspection of micro defects by
utilising the structural features of TFT panels.
Regarding automatic inspection systems for LCDs, several
electrical or optical based inspection techniques have been de-
veloped for LCD manufacturing [3, 7, 18, 25]. Henly and Ad-
diego [5] used a 2-D electro-optic modulator to generate voltage
images, which measured the surface potential of the LCD panel
in a non-contact manner. Kido et al. [26] presented an optical
charge sensing technique for partially completed active-matrix
LCD panels. Surface reflection was used to sense optical changes
and to generate a map that shows the type and location of line
and point defects. Most existing methods of automatic inspection
systems for LCDs are based on conventional electrical methods
to detect the surface potential. Those electrical methods work
54
well for functional verification of a TFT panel. As aforemen-
tioned, they can only be accomplished after the fabrication is
completed. In-process inspection may not be possible with the
functional test approach.
A few vision-based techniques that use pattern-matching al-
gorithms were developed for LCD inspection. Nakashima [12]
presented an inspection system based on image subtraction and
optical Fourier filtering for detecting defects on an LCD colour
filter panel. The image subtraction method was utilised to detect
white and black defects, such as black matrix holes and particles,
and the Fourier filtering was applied for grain defects. Sokolov
and Treskunov [24] developed an automatic vision system for
final chech of LCD output check. For defect detection they com-
pared brightness distributions between a reference LCD image
and a test image. The existing vision-based techniques generally
need a pre-stored reference image for comparison. This requires
a large volume of data for reference and precise environmental
controls such as alignment and lighting for test images. More-
over, the techniques for LCD inspection mainly focus on final
appearance checks for defects such as nap or dark/bright spots
after the fabrication is completed.
A TFT panel generally involves repetitive horizontal gate
lines and vertical data lines. Since the geometrical structure of
a TFT panel surface involves these horizontal and vertical elem-
ents, it can be classified as a structural texture in the image. The
textural feature of a TFT panel surface results in a homogeneous
image that consists of an arrangement primarily of horizontal and
vertical lines appearing periodically on the surface. Since singu-
lar value decomposition (SVD) of a matrix involves horizontal
and vertical basis functions, it is well suited for representing the
structural features of TFT panels. In this study, we use the SVD-
based image reconstruction technique to detect the micro defects
on TFT panel surfaces.
The SVD method was first proposed in the 1970s and has
been applied in a wide range of computer vision applications,
such as image hiding [14, 21], image restoration [10, 11, 16], and
image compression and reconstruction [1, 2, 9, 15, 19, 20, 22, 27,
28]. For image compression and reconstruction applications, the
SVD-based methods were mainly applied to extract the signifi-
cant feature components of the image. The global, main informa-
tion of the image is mostly concentrated within a certain number
of singular values with related singular vectors. Only the rele-
vant parts of the singular values and related singular vectors need
to be retained as the compressed data for reconstructing the ori-
ginal images. The local, detailed information can be truncated to
eliminate the redundancy of image compression.
A few studies have been done with the use of the SVD for
texture analysis in computer vision. Luo and Chen [8] utilised
the SVD for texture discrimination. They used the proportion of
dominant singular values of an image matrix as textural features
to discriminate textures. Hatipoglu and Mitra [23] combined
Teager filters and the SVD for texture feature extraction. Teager
filters were first used to find the local energy values and provide
efficient feature vectors. The filter outputs were then combined
with the eigenvalues obtained from the SVD of local image par-
titions to form feature vectors. Kvaal et al. [13] used the SVD
for feature extraction from the image of bread. They applied the
SVD to extract the singular values of each inspected image and
to form an SV-spectra matrix. Then, the SV-spectra matrix was
used to classify the sensory porosity of wheat baguettes.
The aforementioned SVD-based methods for texture analy-
sis generally use singular values or singular vectors derived from
a textured image to characterise the textural features, and then
use complicated classifiers to segment or classify textures. How-
ever, different textures may need different textural features to
describe the textural patterns. The feature extraction process for
a best set of textural features is generally carried out by trial and
error, and may highly rely on human expertise.
In this paper, we propose a global approach that uses an
SVD-based image reconstruction technique for inspecting micro
defects including pinholes, scratches, particles and fingerprints
on the surface of TFT panels. The proposed method does not
rely on textural features to detect local anomalies, and does not
require a reference image for comparison. It alleviates all limi-
tations of the feature extraction schemes and template matching
methods just mentioned.
The SVD can be used to decompose an image and obtain
a diagonal matrix. The ordered entries of the diagonal matrix
are singular values. The main information, or the approxima-
tion, of an image can be represented by a few singular values
with large magnitude. The remaining singular values with small
magnitude provide detailed information of the image. Since the
TFT panels contain periodical horizontal and vertical structures,
the larger singular values retain the information of the repeti-
tive structural pattern of a TFT panel, and the smaller singular
values are associated with anomalies in the TFT panel. In the
application of LCD defect inspection, we can set the larger sin-
gular values to zero and preserve the smaller singular values to
reconstruct the image. The background texture will be removed
and anomalies can be distinctly enhanced in the restored image
accordingly.
The SVD is ideally suited for describing the orthogonal
pattern in a grey-level image. By considering an input image
as a matrix, the SVD process decomposes the image into the
eigenvalue-eigenvector factorisation. We first select the proper
number of larger singular values to represent the repetitive, orth-
ogonal structure features of a TFT panel. Then, we set the se-
lected singular values to zero and reconstruct the image. For
a faultless TFT panel, the reconstruction process will result in
a uniform image. For a defective TFT panel, the anomalies will
be preserved and the periodical patterns of the horizontal and
vertical lines will be eliminated on the restored image. Finally,
the statistical process control principle is used to set the threshold
for distinguishing between defective regions and uniform regions
in the restored image.
This paper is organised as follows: Sect. 2 first discusses the
structural characteristics of TFT panel surfaces, and describes
the properties of the SVD. The selection of the proper number of
larger singular values, and the image reconstruction scheme for
LCD defects inspection, are then thoroughly described. Section 3
presents the experimental results from a variety of LCD micro
defects. Sensitivity of changes in image rotation and the sensitiv-
55
ity of the selective number of singular values are also evaluated
in this section. The paper is concluded in Sect. 4.
2 The defect detection scheme
2.1 The structural pattern of a TFT panel
A TFT panel generally contains horizontal gate lines on one
plane and vertical data lines on the other plane. At each pixel, the
gate of the TFT is connected to the gate line and the source is
connected to the data line. Figure 1 shows the schema of a single
pixel of a typical TFT panel [26]. Since the TFT panel is com-
prised of horizontal gate lines and vertical data lines, it forms
a structural texture that contains horizontal and vertical line pat-
terns. The structurally textured image of a TFT panel is shown in
Fig. 2.
2.2 Singular value decomposition
Consider an input image of size M × N as a matrix X of di-
mensions M × N,whereM ≥ N. It is possible to represent this
image in the r-dimensional subspace, where r is the rank of X,
and r ≤ N. The SVD [6] is a factorisation of a matrix X into
orthogonal matrices,
X = USV
T
(1)
where U is an M × r matrix and consists of the orthonormalised
eigenvectors of XX
T
, V is an N × r matrix and consists of the
orthonormalized eigenvectors of X
T
X. S is an r × r diagonal
matrix consisting of the “singular values” of X, which are the
nonnegative square roots of the eigenvalues of X
T
X. These sin-
gular values, denoted by σ, are sorted in non-increasing order,
i.e., σ
1
≥ σ
2
≥···≥σ
r
≥ 0.
Fig. 1. The schema of a single pixel of a TFT panel ([26])
The SVD is based on orthonormal bases for decomposing
the matrix X[17]. The singular values (σ ) represent the energy
of matrix X projected on each subspace. The singular values
and their distribution, which carry useful information about the
contents of X, vary drastically from image to image. For an
image with orthogonal texture content such as horizontal and/or
vertical structures, only a very few larger singular values will
dominate, and yet all others have magnitudes close to zero. Fig-
ure 3 shows two artificial images and a TFT panel image and
their corresponding first ten singular values. Both artificial im-
ages in Figs. 3a,b contain well-structured lines with different line
spacing. Figure 3c is a real TFT panel image. All three images
contain horizontal and vertical lines patterns. From Fig. 3d, we
can observe that the first (the largest) singular value dominates
all other singular values, which decrease to zero rapidly.
In most of the cases, the larger singular values (with lager
magnitude) represent the global approximation of the original
image. All other smaller singular values provide the local, de-
tailed information of the image. Therefore, we can select the
proper number of larger singular values to represent the global,
Fig. 2. The surface image of a TFT panel
Fig. 3. a and b Two artificial lines images with different line spacing;
c A TFT panel image; d The plot of the corresponding first ten largest
singular values
56
repetitive textural feature of the image and remove such back-
ground texture by reconstructing the image without the use of
larger singular values.
2.3 SVD-based image reconstruction
In this study, we use machine vision to tackle the problem of de-
tecting micro defects including pinholes, scratches, particles and
fingerprints, which appear as local anomalies in TFT panels. The
SVD has desirable properties of orthogonal bases to deal with
the orthogonal textural feature of TFT panel surfaces. Therefore,
the SVD-based image reconstruction technique is used to remove
the orthogonal line patterns in TFT panel surfaces. With this ap-
proach, we do not have to define various features for different
types of defects. For defect detection purposes, the SVD-based
image reconstruction scheme simply eliminates all repetitive ho-
rizontal and vertical patterns of TFT panels. What is retained in
the resulting image can then be easily identified as defects on the
TFT panels.
The image reconstructed from the selective singular values is
given by
∧
X
=
r
j=k+1
U
j
σ
j
V
T
j
(2)
where
∧
X
is the reconstructed image, U
j
and V
j
are jth column
vectors of U and V , respectively; k is some selected number of
singular values; σ
j
is jth singular value of S,andr is the rank of
the matrix X.
For the image compression and reconstruction application,
we can preserve only the k largest singular values to closely
approximate the original image. Figure 4 shows an artificial
orthogonal-lines image and its reconstructed images. Figure 4a
shows the original image containing horizontal and vertical lines.
Figures 4b1,b2 show the images reconstructed from individual
σ
1
and σ
2
, respectively. It can be observed that the reconstructed
image from σ
1
or σ
2
alone can not sufficiently represent the tex-
ture feature of the original image. It can be seen from Fig. 4b3
that the reconstructed image from the first two largest singular
values (both σ
1
and σ
2
) can well represent the original image
(Fig. 4a). Note that the image size in Fig. 4 is 256 × 256, and
there are a total of 256 singular values. Only the first two largest
singular values, σ
1
and σ
2
, dominate the global, repetitive texture
of the line pattern.
On the other hand, in the defect inspection application, we
can set the proper number of larger singular values (from σ
1
to
σ
k
) to zero and preserve the smaller singular values to recon-
struct an image. The background texture will be removed and the
defects will be preserved if they exist. Figures 4c1–c3 illustrate
the reconstructed images by excluding the singular values with
larger magnitude. Figure 4c1 shows the reconstructed image by
excluding the largest singular value σ
1
, and Fig. 4c2 shows the
resulting image by excluding the second largest singular value
σ
2
. It can be seen that solely setting σ
1
or σ
2
to zero cannot suf-
ficiently eliminate the background texture in the reconstructed
Fig. 4. a The artificial horizontal/vertical lines image (the original image);
b1 the reconstructed image from σ
1
; b2 The reconstructed image from σ
2
;
b3 The reconstructed image from both σ
1
and σ
2
; c1 The reconstructed
image excluding σ
1
; c2 The reconstructed image excluding σ
2
; b3 The re-
constructed image excluding both σ
1
and σ
2
image. Figure 4c3 demonstrates the reconstructed image by ex-
cluding both σ
1
and σ
2
simultaneously. We can observe that the
resulting image is approximately a uniform black image for the
original faultless image.
2.4 Selecting the proper number of singular values
When using the SVD image reconstruction scheme for defect
detection, we first use Eq. 1 to decompose the image and ob-
tain a set of singular values. Then, we need to select the proper
number (i.e., the parameter k in Eq. 2) of singular values, which
can sufficiently represent the repetitive structural pattern, and use
Eq. 2 to reconstruct the image. In this study, the proper number
k is determined by the difference between two adjacent singu-
lar values σ
i
and σ
i+1
. The difference can be considered as the
degree of significance of σ
i+1
with respect to σ
i
. Since various
images have a different spread of ranges of magnitudes of sin-
gular values (see Fig. 3d), it is difficult to determine a suitable
threshold directly from the difference between σ
i
and σ
i+1
for
the selection of the proper number k. In order to find the proper
value of the threshold and to accommodate various images, sin-
gular values of each image under inspection must be normalised.
The normalisation proceeds as follows:
σ
i
=
σ
i
− µ
σ
s
σ
, i = 1, 2,... ,r (3)
where σ
i
is the ith normalised singular value, σ
i
is the ith singu-
lar value, µ
σ
is the mean and s
σ
is the standard deviation of all
singular values for a given image.
Let ∆σ
i
= σ
i
− σ
i+1
, which represents the normalised mar-
ginal gain of energy at singular value number i.If∆σ
i
is larger
than some threshold (T
∆σ
), the additional singular value σ
i+1
is considered to be significant. It must also contribute signifi-
cant energy of the repetitive horizontal and vertical background
texture. In that case, it should be excluded in the reconstruction
process.
57
Figure 5a represents an artificial structural image that con-
tains two scratch defects. Figure 5b shows the plot of the mar-
ginal gain (∆σ) of the normalised singular values of the image
in Figure 5a. It can be observed from Fig. 5b that the marginal
gains decrease rapidly to zero after the stable point and become
steady afterwards. Thus, the value at the stable point can be used
as a threshold (T
∆σ
) to determine the proper number of singular
values that contribute in a major way to the background texture.
For Fig. 5, the proper number is 4 (k = 4), i.e., σ
1
, σ
2
, σ
3
and σ
4
should be excluded for image reconstruction.
Once the proper number of singular values is selected, we
can eliminate the background texture and preserve the defects
by excluding the first k largest singular values in Eq. 2. The
reconstructed image in Fig. 5c shows that the resultant region as-
sociated with the repetitive line pattern becomes approximately
uniform and local anomalies of scratches are well preserved in
the reconstructed image.
Since the intensity variation in the background region is very
small in the reconstructed image, we can use the statistical pro-
cess control principle to set up the control limits for distinguish-
ing defects from the uniform region. The upper and lower control
limits for intensity variation in the reconstructed image are given
by
µ
∧
X
± t · s
∧
X
(4)
where, µ
∧
X
and s
∧
X
are the mean and standard deviation of grey
levels in the restored image
∧
X
;andt is a control constant. Ac-
cording to the Chebyshev’s theorem [4], the probability that any
random variable x will fall within t standard deviations of the
mean is at least 1 −
1
t
2
.Thatis,
p(µ
∧
X
− t · s
∧
X
< x <µ
∧
X
+ t · s
∧
X
) ≥ 1 −
1
t
2
.
In TFT panel manufacturing, the size of a micro defect is
generally small with respect to the whole sensed image. In this
Fig. 5a–d. The artificial orthogonal image with scratch defects: a The ori-
ginal image; b The plot of the marginal gain (∆σ) of normalized singular
values; c The restored image; d the resulting binary image for defect seg-
mentation
study, we set the control constant to t = 4, which corresponds to
93.75% for pixels falling within the control limits.
If a pixel with its grey level falls within the control limits,
the pixel is classified as a homogeneous element of the back-
ground region. Otherwise, it is classified as a defective element.
Figure 5d depicts the defect detection result of Fig. 5c as a binary
image. It shows that the two scratches in the original image are
correctly presented in the resulting binary image.
3 Experiments and discussion
3.1 Experimental results
In this section, we present experimental results from a variety
of micro defects including pinholes, scratches, particles and fin-
gerprints on TFT panel surfaces to evaluate the performance of
the proposed defect detection scheme. The test images are 256 ×
256 pixels wide with 8-bit grey levels. Figures 6a–c show re-
spectively three defective images containing pinhole, scratch and
particle blemishes on TFT panel surfaces under a fine image
resolution (60 pixels/mm). Figure 7 shows a fingerprint defect
under a coarser image resolution (20 pixels/mm). The pinhole,
scratch and particle defects can only be detected in images of fine
resolution, whereas the fingerprint defect can only be observed in
images of coarse resolution.
Figures 8a–d depict the plots of marginal gain (∆σ) of the
four test images shown in Figures 6a–c and Fig. 7, respectively.
It can be observed from Fig. 8 that if the marginal gains are
less than 0.05, they rapidly decrease approximately to zero and
become steady afterwards. The singular values with ∆σ>0.05
sufficiently represent the orthogonal structure pattern of a TFT
panel surface.
Table 1 summarises the detailed information about the nor-
malised singular values and their marginal gains of the four de-
fect images in Figures 6a–c and Fig. 7. It can be seen from the
table that the marginal gains greater than 0.05 are fluctuant and
drop rapidly, whereas those smaller than 0.05 will steadily de-
crease and approximate to zero afterwards. Therefore, 0.05 is the
threshold (T
∆σ
= 0.05) of the marginal gain (∆σ)usedinthis
study to determine the proper number of singular values for de-
fect detection in TFT panel surfaces. The numbers of singular
values selected for pinholes (Fig. 6a), scratches (Fig. 6b), par-
Fig. 6a–c. Three defective images under fine image resolution (60 pix-
els/mm): a Pinhole; b Scratch; c Particle
58
Fig. 7. A defective image with fingerprint under coarse image resolution (20
pixels/mm)
ticles (Fig. 6c) and fingerprints (Fig. 7) images are 5, 8, 4 and 6,
respectively, based on the resulting statistics in Table 1.
Figures 9a1, b1, c1 and d1 show the defect images of the
TFT panel surfaces in Figure 6a–c and Fig. 7, respectively. Fig-
Fig. 8a–d. The plots of marginal
gain (∆σ) of defective images:
a The defective image of pin-
hole in Fig. 6a; b The defective
image of scratch in Fig. 6b; c
The defective image of particle
in Fig. 6c; d The defective image
of fingerprint in Fig. 7
Defective image Pinhole Fig. 6a Scratch Fig. 6b Particle Fig. 6c Fingerprint Fig. 7
Singular value (σ
i
)σ
∆σσ
∆σσ
∆σσ
∆σ
σ
1
15.86 14.56 15.77 13.96 15.83 14.38 15.89 15.29
σ
2
1.30 0.85 1.81 0.98 1.45 0.90 0.60 0.10
σ
3
0.45 0.15 0.83 0.30 0.55 0.26 0.51 0.17
σ
4
0.30 0.13 0.53 0.21 0.30 0.04 0.34 0.05
σ
5
0.17 0.04 0.32 0.05 0.25 0.04 0.29 0.10
σ
6
0.13 0.04 0.27 0.10 0.21 0.04 0.19 0.02
σ
7
0.09 0.03 0.17 0.10 0.17 0.03 0.17 0.01
σ
8
0.06 0.03 0.07 0.02 0.14 0.02 0.16 0.04
σ
9
0.03 0.01 0.05 0.02 0.12 0.03 0.12 0.01
σ
10
0.02 0.02 0.04 0.01 0.09 0.01 0.11 0.02
Tab le 1. The normalised
singular values and their
marginal gains (∆σ)forthe
defective images in Figs. 6
and 7
ure 9a2 shows the reconstruction result by setting the first five
largest singular values (i.e., σ
1
,σ
2
, ··· ,σ
5
) to zero for the pin-
hole defective image with (Fig. 9a1). It can be found that the
repetitive structural texture becomes an approximately uniform
grey-level region and the abnormal pinhole is well enhanced in
the restored image. Figures 9b2 and c2 show the reconstruction
results of the defective images with scratch (Fig. 9b1) and par-
ticle (Fig. 9c1) by setting the first eight (i.e., σ
1
,σ
2
, ··· ,σ
8
) and
the first four (i.e., σ
1
,σ
2
, ··· ,σ
4
) singular values to zero, respec-
tively. They also reveal that the scratch and particle defects are
well preserved in the restored images. Figure 9d2 illustrates the
restored image of Fig. 9d1 by setting the first six singular values
(i.e., σ
1
,σ
2
, ··· ,σ
6
) to zero. The fingerprint is also distinctly en-
hanced in the restored image. Figures 9a3–d3 show the defect
detection results of Figures 9a1–d1 as binary images, of which
the control constant t = 4 is used for all test images. It can be
seen that the orthogonal texture patterns on the TFT panel sur-
faces are eliminated and defects are distinctly preserved.
In order to test the robustness of the proposed method, the de-
tection result of a faultless TFT panel surface is also evaluated.
59
Fig. 9. a1-d1 The defective images with pinhole, scratch, particle and fin-
gerprint, respectively; a2-d2 The respective restored images; a3-d3 The
resulting binary images for defect segmentation
Figure 10a shows a faultless version of the image in Fig. 6, and
Fig. 10b depicts its corresponding marginal gain (∆σ ). Table 2
summarises the faultless image’s detailed statistics for the nor-
malised singular values and their marginal gains. Based on the
threshold (T
∆σ
= 0.05) selected before, it can be observed from
Fig. 10b and Table 2 that the proper number of singular values
are six (k = 6). Figure 10c shows the restored image by set-
ting the first six largest singular values (i.e., σ
1
,σ
2
, ··· ,σ
6
) to
zero. The restored image of the faultless surface is approximately
a uniform grey-level image. As seen in Fig. 10d, the resulting bi-
nary image of the faultless surface is uniformly white. No defect
is claimed in the resulting image for the faultless sample.
3.2 Effect of varied number of singular values
The number of singular values, k, determines how many singu-
lar values will be used to represent the background texture. Too
many selected singular values will remove both background tex-
ture and local anomalies in the restored image, and may overlook
subtle defects. However, too few selected singular values cannot
completely remove the background texture in the restored image
Fig. 10. a A faultless version of the image in Fig. 6; b The plot of marginal
gains (∆σ); c The restored image; d The resulting binary image
and may result in false alarms. The test image in Fig. 7 is used
as the sample to evaluate the effect of varied numbers of singular
values on detection results.
As mentioned, we use the threshold T
∆σ
= 0.05 to determine
the proper number of singular values. In this experiment, we have
examined eight different numbers in the neighbourhood of the
selected number of singular values k, i.e., k ± i for i = 1, 2, 3, 4.
Figure 11a presents the restored image of Fig. 7 by excluding
the first six largest singular values (i.e., k = 6). Figures 11b–
e show the restored images of Fig. 7 with selected numbers 5,
4, 3, and 2, respectively. When the selected number of singu-
lar values is not sufficient, the background texture residuals may
be retained in the restored image. They have no effect on the
local defects. It can be seen from Figures 11b,c that the back-
ground texture can also be sufficiently removed in the restored
image for the numbers of singular values k − 1andk − 2How-
ever, Figs. 11d,e show that the residuals of repetitive horizontal
and vertical lines patterns along with the defects are retained in
the restored images.
60
Fig. 11a–i. The restored results of the fingerprint image in Fig. 7 from dif-
ferent selected numbers of singular values: a The result from k = 6; b–e The
results from different selected numbers k − 1, k − 2, k − 3andk − 4, respec-
tively; f –i The results from different selected numbers k + 1, k + 2, k + 3and
k + 4, respectively
Tab le 2. The normalised singular values and their marginal gains (∆σ )of
the faultless image in Fig. 10a
Singular value Faultless image in Fig. 10(a)
(σ
i
) σ
∆σ
σ
1
15.87 14.55
σ
2
1.31 0.87
σ
3
0.44 0.27
σ
4
0.17 0.09
σ
5
0.08 0.05
σ
6
0.03 0.03
σ
7
0.00 0.02
σ
8
−0.02 0.01
σ
9
−0.03 0.00
σ
10
−0.03 0.01
Figures 11f–i show the restored images of Fig. 7 with se-
lected numbers 7, 8, 9, and 10, respectively. When the selected
number of singular values is more than required, the local de-
fects will be blurred and better uniformity of the background
texture will be generated in the restored image. It can be seen
from Figs. 11f,g that the restored images of the TFT panel sur-
face can still well enhance the defects. However, Figs. 11h,i show
that the defects become blurred with the increasing numbers of
singular values k + 3andk + 4.
Based on the result from Fig. 11, we can find that k ± 2
will not affect the result of an LCD defect inspection. The re-
stored images of defective surfaces can still effectively remove
the background texture and well preserve defects. The selection
procedure for the number of singular values along with the tol-
erance of SVD image reconstruction make the proposed method
practical for defect detection in TFT panels.
Fig. 12a–f. The restored results of the scratch image in Fig. 6 from various
rotation angles: a The result from the original image; b –f The results from
the images with 1
◦
-, 2
◦
-, 3
◦
-, 4
◦
-and5
◦
-rotation, respectively
3.3 Effect of image rotation
The SVD is based on the orthogonal bases used to decompose
a matrix. The orthogonal bases are sensitive to the rotation of an
image. The test image in Fig. 6b is used as the sample to evalu-
ate the effect of rotation on detection results. The test image is
rotated by 1
◦
,2
◦
,3
◦
,4
◦
and 5
◦
.
Figures 12a–f present the restored images in varied angles of
the test image in Fig. 6b. Figure 12a shows the restored image of
the original image in Fig. 6b without rotational change. It can be
observed from Figures 12b,c that the restored images of the TFT
panel surface can still well enhance the defects with a few ran-
dom noisy points when the rotation angles are not larger than 2
◦
.
However, when the rotation angles are larger than 2
◦
, the resid-
uals of repetitive vertical line patterns along with the defects are
retained in the restored images, as seen in Figs. 12d–f. The larger
the rotation angles, the more residuals of structural lines are pre-
served in the resulting images.
Based on the results from Fig. 12, we find that 2
◦
is the
acceptable rotation for LCD defect inspection. When rotation an-
gles of a sensed image are smaller than 2
◦
, the restored image
of a defective surface can still effectively remove the background
texture. Conversely, if the rotation angles are greater than 2
◦
,the
restored images will contain many structural background noisy
points and may result in false rejection of a faultless image. In
LCD manufacturing, the TFT panels are generally well aligned.
The 2
◦
-rotation restriction of the SVD-based machine vision
scheme will not affect the detection performance in practice.
4 Conclusions
Surface defects on TFT panels not only cause visual failure,
but result in electrical failure and loss of LCD operational
functionally. In this paper, we have presented a global ap-
proach for automatic visual inspection of micro defects on TFT
61
panel surfaces. The proposed method does not rely on conven-
tional electrical and feature extraction methods to detect defects.
It is based on an image reconstruction scheme using singu-
lar value decomposition. The SVD approach decomposes an
image into the eigenvalue-eigenvector factorisation. The SVD
orthogonal bases can well represent the horizontal and verti-
cal structures of a TFT panel. By selecting the proper number
of singular values on the diagonal matrix and reconstructing
the image without the use of the selected singular values, we
can eliminate global repetitive patterns of the structurally tex-
tured image, and preserve local anomalies in the reconstructed
image.
In the experiments, we have evaluated a variety of micro de-
fects including pinholes, scratches, particles and fingerprints on
TFT panel surfaces. The experimental results have concluded
that 0.05 is a well-suited threshold of marginal gains (∆σ)to
determine the proper number of larger singular values that con-
tribute to the repetitive background texture. The experiments also
show that the selected number k of larger singular values can tol-
erate minor variation without affecting the reconstruction result.
The experiment on the effect of rotation has shown that the toler-
able rotation angles of the proposed method are 2
◦
. The proposed
SVD-based machine vision scheme has shown promising results
for micro defects inspection of TFT panels.
References
1. Cagnoli B, Ulrych TJ (2001) Singular value decomposition and wavy
reflections in ground-penetrating radar image of base surge deposits. J
Appl Geophys 48:175–182
2. Popesuc BP, Pesquet JC, Petropulu AP (2001) Joint singular value
decomposition-a new tool for separable representation of image. IEEE
International Conference on Image Processing 2, pp 569–572
3. Lin CS, Wu WZ, Lay YL, Chang MW (2001) A digital image-based
measurement system for a LCD backlight module. Opt Laser Technol
33:499–505
4. Anderson DR, Sweeney DJ, Williams TA (2002) Statistics for business
and economics. South-Western, Cincinnati, OH
5. Henly FJ, Addiego G (1991) In-line functional inspection and repair
methodology during LCD panel fabrication. 1991 SID Symposium Di-
gest Paper 32:686–688
6. Strang G (1998) Introduction to linear algebra. Wellesley-Cambridge,
Wellesley, MA
7. Jeff H (2000) Electro-optics technology tests flat-panel displays. Laser
Focus World 36:271–276
8. Luo JH, Chen CC (1994) Singular value decomposition for texture
analysis. Proceedings of SPIE-The International Society of Optical En-
gineering, vol. 2298, Applications of Digital Image Processing XVII,
San Diego, CA, pp 407–418
9. Wei JJ, Chang CJ, Chou NK, Jan GJ (2001) ECG data compression
using truncated singular value decomposition. IEEE Trans Inf Technol
Biomed 5:290–299
10. Konstantinides K, Yovanof GS (1995) Application of SVD based spa-
tial filtering to video sequences. IEEE International Conference on
Acoustics, Speech and Signal Processing 4, Detroit, MI, pp 2193–2196
11. Konstantinides K, Natarajan B, Yovanof GS (1997) Noise estimation
and filtering using block-based singular value decomposition. IEEE
Trans Image Process 6:479–483
12. Nakashima K (1994) Hybrid inspection system for LCD color filter
panels. Proceedings of the 10th International Conference on Instrumen-
tation and measurement Technology, Hamamatsu, pp 689–692
13. Kvaal K, Wold JP, Indahl UG, Baardseth P, Næs T (1998) Multivari-
ate feature extraction from textural image of bread. Chemometrics Intell
Lab Syst 42:141–158
14. Chung KL, Shen CH, Chang LC (2002) A novel SVD- and VQ-based
image hiding scheme. Patt Recogn Lett 22:1051–1058
15. Song LP, Zhang SY (1999) Singular value decomposition–based re-
construction algorithm for seismic travel time tomography. IEEE Trans
Image Process 8:1152–1154
16. Ibrahim MA, Hussein AWF, Mashali SA, Mohamed AH (1998) A
blind image restoration system using higher-order statistics and random
transform. IEEE International Conference on Electronics, Circuits and
Systems 3, Lisboa, Portugal, pp 523–530
17. Petrou M, Bosdogianni P (1999) Image processing. Wiley, New York
18. Chen PO, Chen SH, Su FC (2000) An effective method for evaluating
the image-sticking effect of TFT-LCDs by interpretative modeling of
optical measurements. Liq Crystals 27:965–975
19. Waldemar P, Ramstad TA (1997) Hybrid KLT-SVD image compression.
Proceedings of IEEE International Conference on Acoustics, Speech,
and Signal, Munich, Germany 4, pp 2713–2716
20. Steriti RJ, Fiddy MA (1993) Regularized image reconstruction using
SVD and a neural network method for matrix inversion. IEEE Trans
Signal Process 41:3074–3077
21. Liu R, Tan T (2002) An SVD-based watermarking scheme for protect-
ing rightful ownership. IEEE Trans Multimedia 4:121–128
22. Chandrasekaran S, Manjunath BS, Wang YF, Winkeler J, Zhang H
(1997) An eigenspace update algorithm for image analysis. Graph
Model Image Process 59:321–332
23. Hatipogly S, Mitra SK (1998) Texture feature extraction using Teager
filters and singular value decomposition. International Conference on
Consumer Electronics, Los Angeles, CA, pp 440–441
24. Sokolov SM, Treskunov AS (1992) Automatic vision system for final
test of liquid crystal display, Proceedings of the 1992 IEEE Inter-
national Conference on Robotics and Automation, Nice, France, pp
1578–1582
25. Kido T (1992) In-processing inspection technique for active-matrix
LCD panels. Proceedings of IEEE International Test Conference, Bal-
timore, MD, pp 795–799
26. Kido T, Kishi N, Takahashi H (1995) Optical charge-sensing method
for testing and characterizing thin-film transistor arrays. IEEE J Sel
Topic Quantum Electron 1:993–1001
27. Selivanov VV, Lecomte R (2001) Fast PET image reconstruction based
on SVD decomposition of the system matrix. IEEE Trans Nucl Sci
48:761–767
28. Hoge WS, Miller EL, Lev-Ari H, Brooks DH, Karl WC, Panych LP
(2001) An efficient region of interest acquisition method for dynamic
magnetic resonance imaging. IEEE Trans Image Process 10:1118–1128
