Proceedings of the 2005 IEEE/ASME
International Conference on Advanced Intelligent Mechatronics
Monterey, California, USA, 24-28 July, 2005
Automated Tiny Surface Defect Detection Using DCT Based
Enhancement Approach for Statistical Textures
Hong-Dar Lin and Duan-Cheng Ho
Abstract—This research proposes a novel approach that
applies DCT based enhancement for the detection of pinhole
defects on SBL chips. A two-stage decomposition procedure is
proposed to extract an odd-odd frequency matrix after a digital
image has been transformed to DCT domain. The cumulative
sum algorithm is then proposed to detect the transition points of
the gentle curves plotted from the odd-odd frequency matrix.
After the transition points are determined, the best radius of the
cutting sector is computed and the high-pass filtering operation
is implemented. The filtered image is then inversed and
transformed back to the spatial domain. Finally, the restored
image is segmented by an entropy method and some defect
features are calculated. Experimental results show the proposed
pinhole defect detection method can reach the pinhole defect
detection rate by 90% and decrease the deviation of the defect
areas by 90%.
I. INTRODUCTION
S URFACE defects affect not only the appearance of
electronic devices but also their functionality, efficiency
and stability. Large and obvious surface defects such as
indents, scraps and scratches are usually inspected by
automated visual inspection systems. But tiny surface flaws
such as dust, cavities and pinholes are very difficult to detect
because of their extremely small sizes. Artificial inspection
systems, commonly used today to detect tiny surface flaws,
are vulnerable to wrong judgments owing to inspectors’
subjectivity and eye fatigues. Nevertheless, it is also hard to
precisely inspect tiny flaws by machine vision systems.
When product images are being captured, the area of a tiny
flaw could expand, shrink or even disappear due to the
uneven illumination of the environment, complex texture of
the product surface, and so on. Thus, the inspection accuracy
of the machine vision system is greatly reduced.
Pinhole defects, a common kind of surface defects, occur
frequently in the manufacturing process and incur significant
losses and damages. They exist on the surfaces of industrial
parts such as casting and steel as well as on those of high-tech
products such as wafer and electronic chip. Pinhole defects
Manuscript received February 23, 2005. The authors thank the National
Science Council of Taiwan, R.O.C., for the financial support through the
Grant NSC 92-2212-E-324-001. The authors also express their gratitude to
Wan-Ting Lin for her careful proofreading and revision.
H. D. Lin is with the Department of Industrial Engineering and
Management, Chaoyang University of Technology, Taichung County,
Taiwan 413 R.O.C. (phone: 886-4-2332-3000 Ext 4258; fax:
886-4-2374-2327; e-mail: hdlin@cyut.edu.tw).
D. C. Ho was with the Department of Industrial Engineering and
Management, Chaoyang University of Technology, Taichung County,
Taiwan 413 R.O.C.
0-7803-9046-6/05/$20.00 ©2005 IEEE. 373
MC4-04
cause much greater harms and impact, when they appear on
high-tech products than on industrial parts. Although pinhole
defects occupy only a small area on the product surface, they
could affect the appearance, functions and security of the
product.
This research uses disk-like Surface Barrier Layer (SBL)
chips as the testing samples. SBL chips are ceramics
capacitors which are one kind of electronic passive
components commonly used in many electronic appliances.
With a width of 0.005 mm and a diameter of 7 mm, SBL chips
are small in size, light in weight, and suitable for mass and
large-lot-sized production. Pinhole defects often occur on the
surfaces of SBL chips and occupy only an extremely small
area. For a SBL chip image of 256 x 256 pixels, the size of its
pinhole defect falls within the range of 1~15 pixels and
occupies 0.0015%Д0.0229% of the image area. Pinhole
defects of this magnitude are defined as tiny defects in this
research. To improve the inspection accuracy of pinhole
defects, we propose a Discrete Cosine Transform (DCT)
based enhancement approach to overcome the difficulties of
traditional spatial domain methods in detecting tiny defects.
This research presents a global approach for the automatic
inspection of pinhole defects on SBL chips. As common
computer vision systems are not good at detecting tiny flaws,
we transform a digital image to DCT domain and analyze the
energy trends of the frequency matrix. After some frequency
components of the normal regions are removed, the rest of the
components are transformed back to spatial domain to
produce a restored image with enhanced pinhole defects.
II. LITERATURE REVIEWS
A. Defect detection
In the aspect of defect detection, Kim et al. [1] presented a
new image contrast enhancing method based on the piecewise
linear approximation of Cumulative Density Function (CDF).
Their method utilizes simplified histogram information
obtained by the piecewise linear approximation to attain the
controllability with much less effort. The major feature of
this method is the controllability of the enhancement extent
and partial adjustment. This method is an efficient and useful
contrast tool with much reduced hardware for digital image
display devices. Sari-Sarraf and Goddard [2] presented a
robust segmentation algorithm for the detection and
localization of woven fabric defects. The essence of the
presented segmentation algorithm is the localization of those
defects in the input images that disrupt the global
homogeneity of the background texture. The overall
detection rate of this approach was 89% with a localization
accuracy of less than 0.2 inches and a false alarm rate of
2.5%.
Kim [3] utilized mean preserving bi-histogram
equalization method to do image enhancement. The mean of
gray levels of an image is calculated as a threshold value to
separate two regions for applying different histogram based
enhancement methods. This method can avoid the loss of
color information due to over-enhancement resulted from the
large accumulation of a gray level. Ruzon and Tomasi [4]
used color distributions to detect edge, junction, and corner.
This research models a neighborhood as a distribution of
colors. Their goal is to show that the increase in accuracy of
this representation translates into higher-quality results for
low-level vision tasks on difficult, natural images, especially
as neighborhood size increases. Wu et al. [5] developed an
automated visual inspection system for printed circuit boards
(PCBs). It utilized an elimination-subtraction which directly
subtracts the template image from the inspected image, and
then conducts an elimination procedure to locate defects in
the PCB. This two-stage PCB automated inspection method
is efficient because only simple features are used to classify
the defects.
B. Image processing in frequency domain
Lee [6] presented a novel approach to achieve spatial
thresholding using only frequency domain operations. The
original spatial thresholding is re-modeled such that the
substitute operations can be carried out in the Fourier
frequency domain. This permits manipulation of the extended
dimension only to achieve a thresholding effect. Tico et al. [7]
applied the wavelet transform of frequency domain to the
field of finger-print recognition. This method is more
efficient and has the same effect of performance evaluation
by KNN classifier compared with other techniques applied in
the same field. Tsai [8] presented a global approach for the
automatic inspection of defects in directionally textured
surfaces which arise in textile fabrics and machined surfaces.
Tsai [9] also proposed a global approach for the automatic
inspection of defects in randomly textured surfaces which
arise in sandpaper, castings, leather, and many industrial
materials. Both of the proposed methods are based on a
global image restoration and reconstruction schemes using
the Fourier transform.
C. Applications of image processing using DCT
Use of the DCT in a wide variety of applications has not
been as extensive as its properties would imply due to the lack
of an efficient algorithm. Thus, many algorithms and VLSI
architectures for the fast computation of DCT have been
proposed [10]-[12]. Chen [13] developed new and fast
algorithms for edge enhancement of remote sensing image
data in the DCT domain and implemented in three steps:
high-pass filtering, adding back full or part of gray levels to
the original image, and contrast stretching. Experimental
374
results show that the quality of images generated by the new
algorithms are comparable with that of images generated by
the corresponding methods in the spatial domain.
Zhong and Jain [14] proposed an algorithm for object
localization using shape, color, and texture. The texture and
color features are directly extracted from the DCT
compressed domain and are used to find a small number of
candidate images in the database, and regions in the candidate
images which share similar texture and color as the query can
be identified. Experimental results demonstrate performance
of the algorithm and show that substantial computational
savings can be achieved by utilizing multiple image cues.
Ngo et al. [15] presented approaches for indexing shape,
texture, and color features directly in the DCT domain by
exploiting ten DCT coefficients. For color and shape
indexing, the proposed methods achieve significant speed up
compared to the same approach operating in the
uncompressed image domain. Overall, the retrieval results
are competent since most of the top retrieved images are
relevant.
III. RESEARCH METHODS
This research applies DCT and cumulative sum algorithm
to detect pinhole defects in electronic components. A product
image is transferred from spatial domain to frequency domain
by discrete cosine transform. Then, a two-stage
decomposition procedure is proposed to analyze the energy
trend of a frequency matrix. The first stage decomposes a
two-dimensional discrete cosine spectrum into three
one-dimensional frequency matrices. The second stage
further decomposes the three frequency matrices to produce a
principal frequency matrix for energy trend analysis. The
Cumulative SUM (CSUM) algorithm is applied to find the
transition points of the gradual curves plotted by the principal
datum of the decomposed frequency matrix, and to determine
the best cutting sector radius. After frequency components
within the best radius of the cutting sector are removed, the
frequency domain image is transferred back to the spatial
domain and the pinhole defects are apparently enhanced.
Then, an image segmentation technique, entropy method, is
applied to separate the defects from normal regions.
A. DCT and Frequency domain analysis
Since digital images are two dimensional discrete data
arrays, two dimensional DCT transform is needed to proceed
the transformation. The DCT of an image dx,y of size M x N is
given by the Eq.(1) [16]. This expression must be computed
for values of u=0, 1, 2, …, M-1, and also for v=0, 1, 2, …, N-1.
Similarly, given Du,v, we obtain dx,y via the inverse DCT
transform, given by the Eq.(2) for x=0, 1, 2, …, M-1 and y=0,
1, 2, …, N-1. Eqs. (1) and (2) comprise the two-dimensional,
DCT pair. The variables u and v are frequency variables, and
x and y are spatial variables.
M 1N 1
ª (2 x  1)uS º ª (2 y  1)vS º
Du , v U (u ) U (v) ¦ ¦ d x, y cos « » cos « » (1)
x 0y 0 ¬ 2M ¼ ¬ 2N ¼
 M 1N 1
ª (2 x  1)uS º ª (2 y  1)vS º
d x, y ¦ ¦ U (u) U (v) Du ,v cos« » cos«
¬ 2M ¼ ¬ 2N
u 0v 0
where,
­ 1 ­ 1
° , u 0 ° , v 0 °v 0, 1, 2, ..., N  1
U (u ) ®
° M
,
° N
U (v ) ® ,
° 2 ° 2
° M , u 1, 2, 3, ..., M  1
¯
° N,
¯
v 1, 2, 3, ..., N  1
Based on the fluctuations of the energy trend and the
properties of the low and high frequency zones, we design a
high-pass sector filter centered at the origin of the 2-D DCT
spectrum to filter out the major frequency components of the
target. An adequate radius is first determined for the sector
filter in the spectrum space. Frequency components (low
frequencies) within the radius of the sector filter are then set
at zero, and those outside the filter (medium and high
frequencies) are retained. Finally, the inverse DCT is used to
enhance the tiny pinhole defects. Fig. 1 demonstrates a 2-D
DCT spectrum diagram with a high-pass sector filter.
Fig. 1. Adding a high-pass sector filter into a 2-D DCT spectrum diagram.
As the high-pass sector filter is derived from the decreasing
energy trend of the DCT spectrum diagram, conducting a
proper analysis on the fluctuation trend of the energy
spectrum carries great importance. It takes little effort to
observe the energy trend from 2-D spectrum diagrams. But, a
formal algorithm deduction must be carefully developed if
the frequency spectrum diagram is to be analyzed from an
objective and theoretic view. To construct the analysis steps
of the DCT spectrum, this research analyzes the 2-D DCT
spectrum and proposes a two-stage decomposition procedure.
Due to the difficulty of analyzing the frequency fluctuation
trend in a 3-D spectrum diagram, this research decomposes a
2-D DCT spectrum into 1-D frequency arrays to discover the
regular patterns of the spectrum in the frequency domain.
B. Frequency trend analysis of DCT domain
Let G* be a waveform sign of 1-D frequency arrays and
axis be a 1-D coordinate vector. Then, G*(axis) is a
waveform plot along the axis coordinate and can be denoted
as:
­axis Du#,10 ~ ( N 1) or D0#~1 ( M 1),v
» (2)
¼ °°
where ®u 0, 1, 2, ..., M  1 .
°v 0, 1, 2, ..., N  1
­u 0, 1, 2, ..., M  1 °¯
°
® The magnitude of frequency components far away from the
° x 0, 1, 2, ..., M  1
°¯ y 0, 1, 2, ..., N  1 top left origin in the DCT domain falls rapidly and
approximates zero. Because most image information gathers
around the DC location, the waveforms fluctuate most widely
there. And, the waveforms of G*(D#10~(M-1),v) remain more and
more stable as the values of the u axis increase. The same
phenomenon also happens to the waveform plots of
* #1
G (D u,0~(N-1)).
The complexity of the 2-D DCT domain makes it difficult
to comprehend. But, by applying Eq. (3), we can decompose
a 2-D DCT domain into 1-D waveform matrices, in which
systematic and regular patterns can be found. Hence, to take
advantage of decomposition and dig out more useful
information, this research further decomposes the 2-D DCT
domain to detect the changes of frequency trends.
C. Decomposition of DCT domain
The waveforms of G*(D#10~(M-1),v) of 2-D DCT frequency
components fluctuate as they approach the DC location. And,
the closer the waveforms are to the DC location, the more
widely they fluctuate. We decide to further decompose the
2-D DCT domain D#1u,v into two matrices: the odd matrix
DOu,v' and the even matrix DEu,v', whose definitions are given
below:
DOu ,vc Du#,12vc (4)
DE u ,vc Du#,12 vc1 (5)
­u 0, 1, 2, ..., M  1
where ® .
¯v c 0, 1, 2, ..., ( N 2)  1
Based on the definitions of the odd and even frequency
matrices (Eqs. (4) and (5)), we use the waveform plots
* *
G (DO0~(M-1),v') and G (DE0~(M-1),v') to examine their respective
1-D waveforms. We find some waveform regularities among
the frequency fluctuations in the decomposed odd and even
frequency matrices. Taking the waveform plots G*(DO0,v')
and G*(DE0,v') as examples, we present the waveform plots of
D#10,vΕDO0,v' and DE0,v' in Fig. 2.
As shown in Fig. 2, after D#1u,v is decomposed into DOu,v'
and DEu,v', the frequencies of the waveform plots G*(DO0,v')
and G*(DE0,v') become more stable and steady; they do not
fluctuate as much as before the decomposition. This
phenomenon occurs not only in G*(DO0,v') and G*(DE0,v'), but
also in waveform plots G*(DO1~(M-1),v') and G*(DE1~(M-1),v').
These occurrences lead to the conclusion that the coefficients
of the DCT domain have a lot to do with whether the basic
cosine waveform matrices have complete or incomplete
periods. Examining the waveforms of G*(DOu,0~(N/2)-1) and
*
G (DEu,0~(N/2)-1), we find no similar phenomenon. The
G * axis  Du#,1v (3) frequencies do not become more stable or steady because the
u axis of the DCT frequency matrix has not been decomposed.

375
Therefore, the original frequency matrix D#1u,v can be further in Fig. 3. The purpose of the first stage is to simplify the
decomposed into odd-odd DOOu',v', odd-even DOEu',v', matrices and use waveform plot G* to examine how the
even-odd DEOu',v', and even-even DEEu',v' matrices under the frequencies of the simplified matrices fluctuate to identify the
alternating decomposition of odd and even parts. These four main properties of the 2-D DCT domain. Then, according to
matrices can be denoted as follows, respectively: the properties identified, three frequency matrices that
DOO u ',v ' D 2#u1 ', 2 v ' (6) represent best how the frequencies fluctuate are extracted
from the 1-D arrays. Take for example a 2-D DCT matrix
DOE u ',v ' D 2#u1 ', 2 v '1 (7) with size M by M. After the first stage of decomposition,
there will be only (M x 3) sets of 1-D frequency arrays left.
DEO u ',v ' D 2#u1 '1, 2v ' (8)

DEE u ',v ' D2#u1 '1, 2v '1 (9)

2D
FDCT

­u ' 0, 1, 2, ..., ( M 2)  1
where ® .
¯v' 0, 1, 2, ..., ( N 2)  1
˔ʳ̇˸̆̇˼́˺ʳ˦˕˟ʳ˼̀˴˺˸ Spectrum analysis of
2-D arrays
Frequency

Power
u
(a) 3-D spectrum diagram of DCT domain
-6113 1082 5225 184 2125 59 1133 -17 628 0

740 -62 72 47 9 -9 24 0 -13

v=0 255
(a) Waveform plot G *(D #10,v ) 5433 -137 838 -346 -1749 -198 -1449 -43 -958

v
42 28 48 15 -29 8 -25 1 -27

2209 -105 -1718 -10 -1498 203 31 208 557 (a) G*(v=0) axis spectrum diagram of D u,v
Frequency

48 4 29 -33 -37 -7 -2 -1 23

1193 -72 -1465 92 -4 142 916 -28 474

58 4 -1 -15 -35 14 9 15 14

657 -38 - 996 97 564 19 486 -139 -311 -6113 1082 5225 184 2125 59 1133 - 17 628 0
v'=0 127
(b) Waveform plot G*(DO0,v')

(c
˄ˀ˗ʳ˹̅˸̄̈˸́˶̌ʳ˴̅̅˴̌ʳ Du, v=0ʳ˼̆ʳ˸̋̇̅˴˶̇˸˷ʳ˴˿̂́˺ʳ ऎ ̈ʳ
(b)

)
˴̋˼̆
0 0
Frequency

(b) G*(u=0) axis spectrum diagram of Du,v (c) G*(DA) axis spectrum diagram of DAw

v'=0 * 127
(c) Waveform plot G (DE0,v') -6113 740 5433 42 2209 48 1193 58 657 0 -6113 -62 838 15 -1498 -7 916 15 -311 0

Fig. 2. Three waveform plots of the D#10,v, DO0,v', and DE0,v'. ˄ˀ˗ʳ˹̅˸̄̈˸́˶̌ʳ˴̅̅˴̌ʳD u=0, vʳ˼̆ʳ˸̋̇̅˴˶̇˸˷ʳ˴˿̂́˺ʳ ̉ʳ ˄ˀ˗ʳ˹̅˸̄̈˸́˶̌ʳ˴̅̅˴̌ʳ DAwʳ˼̆ʳ˸̋̇̅˴˶̇˸˷ʳ˴˿̂́˺ʳ
˴̋˼̆ ˷˼˴˺̂́˴˿ʳ˴̋˼̆

Fig. 3. Example of the first stage of the proposed decomposition procedure in

Besides, to obtain the right sector radiuses that will help
DCT domain.
improve the analysis results, the negative 45 degree diagonal
coefficients of the frequency matrix can be decomposed into D. CSUM control scheme
a 1-D array DAt when M and N are equal and the top left DC Suppose that Z sample sets with sample size n = 1 are
location is the origin. This diagonal matrix can be denoted as collected, and Xz is the observation of the z-th sample, where
follows: z = 1, 2, 3, …, s, …, Z. The cusum scheme works by
accumulating derivations from µ0 that are above target with
DAt Dt#,t1 (10) one statistic Cs+ and accumulating derivations from µ0 that are
where t=0, 1, 2, …, M-1. below target with another statistic Cs-. The statistics Cs+ and
Cs- are called one-sided upper and lower cusums, respectively.
The frequencies of the waveform plot G*(DAt) of the matrix
DAt fluctuate in the same way as those of the waveform
They are computed as follows [17]:
matrices G*(DO0~(M-1),v') and G*(DE0~(M-1),v'). So, we can further
decompose the matrix DAt into DAOt' and DAEt' matrices. C s >
max 0, X s  ( P 0  K )  C s1 @ (13)

These two matrices can be denoted as follows:

C s max >0, ( P 0  K )  X s  C s1 @ (14)
DAOt ' DA2t ' (11) G
where C 0 C 0 0, K V.
DAEt ' DA2t '1 (12) 2
In Eqs. (13) and (14), K is usually called the reference
where t ' 0, 1, 2, " , ( M 2)  1 .
value, and it is often chosen about halfway between the target
Based on the above theoretic deduction and the µ0 and the out-of-control value of the mean µ1 that we are
calculations on the testing images, we propose a two-stage interested in detecting quickly. Thus, if the shift is expressed
decomposition procedure. In the first stage, a 2-D DCT in standard deviation units as µ1 =µ0 + įı, then K is one-half
frequency matrix is broken into multiple 1-D matrices shown the magnitude of the shift. Note that Cs+ and Cs- accumulates

376
deviations from the target value µ0 that are greater than K, method.
with both quantities reset to zero on becoming negative. If
F. Image restoration and defect segmentation
either Cs+ or Cs- exceed the decision interval H, the process is
considered to be out-of-control. A reasonable value for H is Fig. 6 shows the whole process of the proposed DCT based
four or five times the process standard deviation ı [17]. enhancement approach for pinhole defect detection on SBL
chips. A testing image with pinhole defects which are
E. Reverse order CUSUM method (R-CUSUM) difficult to detect in the spatial domain is processed by the
To detect the transition points in the gentle curves, this proposed DCT high-pass filtering operation. Then, taking the
research proposes the CUSUM algorithm, which is inverse DCT of the filtered result, we can have the tiny
commonly used in statistical process control to detect the defects easily segmented by the entropy method.
slight shift or deviation from the normal production process.
Generally, the CUSUM method processes data, that are ʻʳ˃ʿʳ˃ʳʼ
Pө
smooth in the beginning periods and that deviate slightly in u'
the later periods. However, since the curves fluctuate sharply
in the beginning periods and then turn smooth in the other
periods, our algorithm applies the reverse order CUSUM
Ѕ˅ PԀ PԀ
method, which processes data in the reverse direction.
The data sequence should be arranged in a reverse order PӪ
before the reverse-order CUSUM method is applied.
ˇˈ̓
According to the definitions of the CUSUM method (Eqs. (13)
and (14)), we assume Cz+ and Cz- as the one-sided upper and

w'
v'
lower cusums of the sequence Xz, respectively. We must
Fig. 4. A sector radius diagram obtained from transformation of the three
substitute the data of the three decomposed frequency
principal frequency arrays.
matrices DOOu',0, DOO0,v', and DAOt' into the sequence Xz*,
respectively, before applying the proposed reverse-order ˄˃˃
P 0*=0.945
CUSUM method. Then, from the situations of being within
ˋ˃
Frequency
ˉ˃
V *=6.342
or beyond a decision interval H, we determine the status of ˇ˃

˅˃ P2( DOOu , 0 )=7

the process variation and identify the transition points ˀ˅ ˃

˃

P(DOOu',0), P(DOO0,v'), and P(DAOt') of the three frequency ˀˇ ˃

˄ ˄˄ ˅˄ ˆ˄ ˇ˄ ˈ˄ ˉ˄ ˊ˄ ˋ˄ ˌ˄ ˄˃˄ ˄˄˄ ˄˅˄ u'
matrices. ˄ˇ˃

The first value in each of the three frequency matrices ˄˅˃

˄˃˃
P 0*=0.961
Frequency

ˋ˃

DOOu',0, DOO0,v' and DAOt' is the DC value in a DCT domain. ˉ˃

ˇ˃
V *=10.534

P2( DOO0, v )=4

This value represents the average gray level of an image and
˅˃
˃
ˀ˅ ˃

is usually the extreme value (the largest or the smallest) in a

ˀˇ ˃
ˀˉ ˃
ˀˋ ˃

frequency matrix. Hence, this research sets the DC value at ˄ ˄˄ ˅˄ ˆ˄ ˇ˄ ˈ˄ ˉ˄ ˊ˄ ˋ˄ ˌ˄ ˄˃˄ ˄˄˄ ˄˅˄ v'

zero when detecting the transition points in the gentle curves ˅ˈ

˅˃ P0*=0.567
˄ˈ

to determine the best radius of the sector filter.

Frequency

˄˃
V *=3.538
ˈ

Setting the DC value at zero will avoid significant ˃

ˀˈ
P2( DAOw )=6
ˀ˄ ˃

variations among the frequency components. After the ˀ˄ ˈ

ˀ˅ ˃
2 ) |7
AVG=average(7,4,6 Ѕ

calculations in Eqs. (13) and (14), three transition points will

*
ˀ˅ ˈ R2 =2AVG+1=15
˄ ˄˄ ˅˄ ˆ˄ ˇ˄ ˈ˄ ˉ˄ ˊ˄ ˋ˄ ˌ˄ ˄˃˄ ˄˄˄ ˄˅˄ w'
be identified in the gentle curves by the R-CUSUM method. Fig. 5. Calculation of the filter radius R2* by the R-CUSUM method.
The best radius of the sector filter R* can be denoted as:
x u

R* 2 AVG ( PD , PE , PZ )  1 (15)
y v
Uftujoh!
where AVG ( PD , PE , PZ ) average( PD , PE , 2 PZ ) . jnbhf! FDCT R-CUSUM
dx,y

The calculation of the best radius R* is presented in Eq.

(15), where three transition points are used. Fig. 4 shows (a) d#1x,y (b) D#1 u,v
points PĮ, Pȕ, and PȦare the three transition points of the x x u

gentle curves along u', v', and diagonal axes, respectively. y y v

The point PȦ must be multiplied by a weight of 2 because IR m
IDCT
+

this point comes from the diagonal matrix DAOt' and must be Entropy

adjusted to be in the same scale as PĮ and Pȕ.

And then, the three transition points are substituted into Eq. (e) Result (d) d#2x,y (c) D#2u,v
(15) to compute the best filter radius R*. Fig. 5 shows the Fig. 6. Pinhole defect detection procedure diagram of the DCT based
calculation of the best filter radius R* by the R-CUSUM enhancement approach.

377
 IV. IMPLEMENTATION AND EXPERIMENTS
In this section, the precision and rotation-invariant
properties for the proposed approach to inspect the surface
pinhole defects of the SBL chips are evaluated. Real SBL
chips are used as testing samples in this section. The image is
256 x 256 pixels wide with 8-bit gray levels. In the
experiments with large samples, 120 testing SBL images with
size 256 x 256 pixels are tested for detecting pinhole defects.
In the testing images, 100 images are defective chips with one
or more than one pinhole defects and the other twenty images
are regular chips without any pinhole defect.
In the processing time of the proposed approach, the time
of taking forward and inverse DCT needs about 4 seconds on
average without doing any parameter optimization in the
transformation. This processing time will be significantly
shortened after the parameter optimization is conducted and
the DCT implemented in a single integrated circuit. If we do
not count the time of taking forward and inverse DCT, the
time of taking the R-CUSUM method is less than 10 ms and
the processing time of each method is very time-saving.
Therefore, to compare the performance of the proposed
method, the experimental results are shown in Table I based
on the number of total real defects. The proposed pinhole
defect detection method can reach the pinhole defect
detection rate by 90% and decrease the deviation of the defect
areas by 90%.
TABLE I
AN EXPERIMENTAL RESULT ANALYSIS TABLE
(BASED ON THE NUMBER OF TOTAL REAL DEFECTS)
Method R-CUSUM
Number of total real defects 114
Number of correctly detected defects 102
Number of erroneously detected defects 12
Number of missing defects 12
Type I error (Į) 0.1
Type II error (ȕ) 10.526%
Area deviation rate of detected defects 10.836%
V. CONCLUSION
This research proposes a novel approach that applies DCT
based enhancement for the detection of pinhole defects on
SBL chips. Real SBL chips are used as testing samples and
experiments with large samples are conducted in a practical
inspection environment to verify the performance of the
proposed approach. Experimental results show that the
proposed method is precise and rotation-invariant. Besides,
the proposed passive component inspection algorithm is easy
to implement and does not require a primitive-matching
process. Therefore, the proposed approach can be further
applied to detect surface tiny defects of various types of
electronic components that are similar to SBL chips.
While spatial domain techniques can hardly detect defects
of such small sizes, the proposed approach is able to locate
378
the ranges of the defects and a color analysis could be
conducted to strengthen the detection accuracy in the future.
This color analysis step can help increase the detection
accuracy rate, reduce type I and type II errors, and decrease
detection processing time. If the proposed approach can be
implemented in a single integrated circuit, the processing
time will be significantly decreased to meet the requests of
on-line automatic inspection.
REFERENCES
[1] Sang-Yeon Kim, Dongil Han, Seung-Jong Choi and Jong-Seok Park,
“Image contrast enhancement based on the piecewise-linear
approximation of CDF,” IEEE Transactions on Consumer Electronics,
Vol. 45, No. 3, pp.828-834, 1999.
[2] Sari-Sarraf Hamed, and J. S. Goddard, “Robust defect segmentation in
woven fabrics,” Proceedings of Computer Vision and Pattern
Recognition, pp.938-944, 1998.
[3] Y. T. Kim, “Contrast enhancement using brightness preserving
bi-histogram equalization,” IEEE Transactions on Consumer
Electronics, Vol. 43, pp.1-8, 1997.
[4] Mark A. Ruzon and Carlo Tomasi, “Edge, junction, and corner
detection using color distributions,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, Vol. 23, No. 11, pp.1281-1295,
2001.
[5] Wen-Yen Wu, Mao-Jiun J. Wang and Chih-Ming Liu, “Automated
inspection of printed circuit boards through machine vision,”
Computers in Industry, Vol. 28, pp.103-111, 1996.
[6] Zhenghong Lee, “Thresholding implemented in the frequency domain,”
IEEE Transactions on Image Processing, Vol. 10, No. 5, pp.708-714,
2001.
[7] M. Tico., P. Kuosmanen and J. Saarinen, “Wavelet domain features for
fingerprint recognition,” Electronics Letters, Vol. 37, No. 1, pp.21-22,
2001.
[8] D. M. Tsai and C. Y. Hsieh, “Automated surface inspection for
directional textures,” Image and Vision Computing, Vol. 18, pp.49-62,
1999.
[9] Du-Ming Tsai and Tse-Yun Huang, “Automated surface inspection for
statistical textures,” Image and Vision Computing, Vol. 21, pp.307-323,
2003.
[10] Wen-Hsiung Chen, C. Harrison Smith and S.C. Fralick, “A fast
computational algorithm for the discrete cosine transform,” IEEE
Transactions on Communications, Vol. 25, No. 9, pp.1004-1009, 1977.
[11] Nam Ik Cho and Sang Uk Lee, “Fast algorithm and implementation of
2-D discrete cosine transform,” IEEE Transactions on Circuits and
Systems, Vol. 38, No. 3, pp.297-305, 1991.
[12] Rafael C Gonzalez. and Richard E. Woods, Digital Image Processing,
2nd Ed., Prentice Hall, 2002.
[13] Biao Chen, Shahram Latifi and Junichi Kanai, “Edge enhancement of
remote sensing image data in the DCT domain,” Image and Vision
Computing, Vol. 17, No. 12, pp.913-921, 1999.
[14] Yu Zhong and Anil K. Jain, “Object localization using color, texture
and shape,” Pattern Recognition, Vol. 33, pp.671-684, 2000.
[15] Chong-Wah Ngo, Ting-Chuen Pong and Roland T. Chin, “Exploiting
image indexing techniques in DCT domain,” Pattern Recognition, Vol.
34, pp.1841-1851, 2001.
[16] N. Ahmed, T. Natarajan, and K. R. Rao, “Discrete cosine transform,”
IEEE Transactions on Computer, Vol. 23, No. 1, pp.90-93, 1974.
[17] Douglas C. Montgomery, Introduction to Statistical Quality Control,
4th Ed., John Wiley & Sons, pp.406-419, 2001.