INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4, pp. 661-667 APRIL 2015 / 661

© KSPE and Springer 2015

Controlled Trigger and Image Restoration for High SpeedProbe Card Analysis

Bonghun Shin1, Soo Jeon1,#, Jiwon Lee1, Chung Su Han2, Chang Min Im3, and Hyock-Ju Kwon1

1 Department of Mechanical and Mechatronics Engineering, University of Waterloo, 200 University Avenue West, Waterloo, Ontario, N2L 3G1, Canada2 Sedicon, 522 Dangjeong-dong, Gunpo-si, Gyeonggi-do, 435-833, South Korea

3 SDA Co. Ltd., 38-16 Ojeon-dong, Uiwang-si, Gyeonggi-do, 483-817, South Korea# Corresponding Author / E-mail: soojeon@uwaterloo.ca, TEL: +1-519-888-4567 (ext.) 38898, FAX: +1-519-885-5862

KEYWORDS: Machine vision, Real-time processing, Image restoration, Wafer probing

Latency and image blurring are major limitations of machine vision processes, which often require every target object to pause for

a moment for capturing and processing a still image. When a large number of objects are to be inspected, such a stop-and-go

approach may significantly degrade the test efficiency due to a long inspection time. This paper investigates the performance and

error analysis of dynamic imaging approach where the image is captured and processed on-the-fly while the target object is still

moving. Taking images of a moving object can substantially enhance the inspection speed but intensifies latency and image blurring.

To overcome these issues, firstly, we implement the controlled trigger, i.e., to operate the machine vision in synchrony with the position

sensing while the target object is moving. Then, we attempt to restore the blurred pixel data through advanced image restoration

techniques. The main ideas are applied to a semiconductor test process called the probe card analysis and its performance is

experimentally verified.

Manuscript received: April 8, 2014 / Revised: August 4, 2014 / Accepted: January 4, 2015

1. Introduction

Machine vision is widely used in industrial applications as a

primary means for inspection and testing. Examples include textiles,

printed circuit boards (PCB’s), integrated circuits (IC’s), labels,

machine tools, fruits, etc.1 Visual signal represents one of the most

informative sensory input, yet it has some drawbacks compared to

other sensing modalities. First of all, the image data available at the

present time step report the status of the object in the previous time

step, i.e. an activity’s initiation and its result occur at different time

instances. This is called latency. Another limitation of the vision

sensing is that the object has to stay still for a certain amount of time

duration, called the exposure time. If the object moves during the

exposure time, the image products will be blurred degrading the

integrity of inspection. Due to these reasons, typical machine vision

processes take the stop-and-go approach where the target object is held

stationary during image acquisition. If the machine vision is relatively

fast compared to other process steps that limit the manufacturing line

speed, it may not be a big issue to make frequent stops for visual

inspection. However, in many applications for high precision testing,

halting the process for every object for a still image can become a

major impediment for testing efficiency. One example is the inspection

of a probe card. The probe card is a component used to test

conductivities and functionalities of integrated circuits (IC’s) before

packaging.2

A probe card consists of a large number of probe pins aligned

within a small area of a single IC chip and provides an interface

between the tester and the IC.2 Fig. 1(a) shows an example of probe

cards. Electrical connection from each pin to the tester (i.e., the wafer

prober) is made through the wire branching out from the epoxy center

ring. The cantilever pins are stacked around the center ring and make

a physical contact with wafer dies. The magnified view of the center

ring and pins is shown in Fig. 1(b). During the probe card manufacturing,

the card manufacturer needs to inspect it for a number of specifications

on mechanical (alignment, planarity, tip radius, missing tips, etc.) as

well as electrical (leakage, capacitance, etc.) properties. The test

equipment used in this process is called the probe card analyzer. Fig.

2 shows the main inspection module of a probe card analyzer. The

probe card is placed on top of the module with the probe pins pointing

downward so that the camera can take their images from below. The

camera is mounted to the x-y stage and can be moved to different pin

locations. Typically, the inspection time for the probe card analysis

DOI: 10.1007/s12541-015-0088-z ISSN 2234-7593 (Print) / ISSN 2005-4602 (Online)

662 / APRIL 2015 INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4

increases in proportion to the number of pins. Therefore, the inspection

time is one of critical factors in evaluating the performance of probe

card analyzers.

Among items to be inspected by a probe card analyzer, there are two

mechanical properties that are essential but time-consuming to

measure: the planar alignment (x and y coordinates) and the tip

diameter of each probe pin. As is done in typical machine vision

applications,3 such mechanical properties are often optically (thus non-

contact) inspected using the digital imaging system of the probe card

analyzer where the stop-and-go strategy is taken to inspect one pin at

a time. The target point for the stage to be positioned for each pin is

provided by the predefined location map of probe pins. For each pin to

be accurately inspected (requirement: ±1.0 µm for each coordinate), the

stage needs to come to a complete stop before an image is taken and

analyzed. In this way, the speed of 1 probe tip per second will result in

the total inspection time close to three hours for a 10,000-pin probe

card.

Motivated by the need for enhancing the inspection speed in the

above mentioned tasks, this paper addresses the approach to capture the

vision image while the object is moving and to process it on-the-fly.

Such a dynamic imaging approach will clearly manifest two main

drawbacks of machine vision, i.e., (non-real-time) latency and image

blurring, as mentioned in the beginning. To overcome these issues, we

adopt controlled sampling of visual data and image restoration

technique. Firstly, vision images are sampled in a time-critical way

using a real-time trigger such that the visual sensing is synchronized

with the position measurement of the object. With some simple

calibration of event timings based on stage motion data, we can

establish a consistent latency to coordinate the timing of position data

with that of the vision image. Such a hard real-time image capturing

technique has also been used in some of recent motion control and

tracking applications.4,5 Secondly, to recover the clear vision data from

the imperfect image, the advanced image restoration techniques or

deblurring6-8 has been employed. The deblurring technique has long

been used in various industrial applications, yet its use in micro-scale

inspection is still rare.7

The remainder of this paper is organized as follows. The technical

background is explained in Section 2 which explains the real-time

machine vision and the image restoration technique. Section 3 and 4

present the demonstration of main ideas through experimental results

with the probe card analysis. The concluding remarks are summarized

in Section 5.

2. Technical Background and Main Issues

In this section, we briefly review relevant issues of high-speed real-

time machine vision applications and provide some technical

background on the basic approach we take.

2.1 Real-time coordination of vision and motion data

To use camera images as reference data, we first need to keep track

of the exact moment when each image is taken. This can be achieved

by real-time triggering and compensation of latency. The latency of a

vision system mainly comes from the exposure time tex (accumulating

light into electrical charges), the readout time tro (converting charges into

digital data), the transfer time ttr (transferring data into the processor)

and the processing time tpr. The total latency TL is thus given by

(1)

where δt stands for the jitter, i.e., the uncertainty in latency. The amount

of jitter should be minimized in applications requiring high determinism.

Fig. 3 shows the timing diagram of a vision system synchronized with

the measured position (x) of the target object. Using carefully configured

triggering operation and proper machine vision protocols (e.g., the

Camera Link or the GigE vision), tex, tro, and ttr (which constitute the

hardware latency) can be kept almost constant. On the other hand, the

software latency due to the processing time tpr, typically varies with

TL tex tro ttr tpr δ t+ + + +=

Fig. 1 A probe card. (SDA Technology Co., Ltd.)

Fig. 2 A probe card analyzer (SDA Technology Co., Ltd.)

INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4 APRIL 2015 / 663

each image sample. Let us denote the maximum and the minimum

values for tpr by and , respectively, (i.e., ≤ tpr ≤ ). As

shown in Fig. 3, the image processing starts right after the completion

of the transmission. As long as the completion of image processing (i.e.

the falling edge of tpr pulse) occurs before the transmission of the next

image data (i.e. the falling edge of the next tpr pulse), image processing

does not cause any backlog. In other words, two consecutive tpr pulses

do not overlap in the timing diagram of Fig. 3. Therefore, the period of

the trigger pulse denoted by Tg(=tk+1−tk) must be chosen as follows to

guarantee the real-time image processing:

(2)

If we send a trigger pulse at t = tk, the electronic shutter immediately

starts to integrate the image data. Any information drawn from this

image will be available at t = tk+TL. Hence, by knowing the latency and

the exact time of trigger, we can synchronize the vision image with the

motion data from the stage. On the other hand, movement of the

camera during the exposure time (tex) will cause motion blurring.

Denoting the stage position at time t = tk by xk, the motion blurring

occurs for the period of tk ≤ t ≤ tk+tex, which corresponds to the

displacement δxk as illustrated in Fig. 3.

2.2 Restoration of image blurred by linear motion

In our application, the motion blur is the result of relative motion

between the camera and the probe tips during camera exposure. From

a mathematical point of view, the blurred image can be considered as

a convolution of the ideal image with the blur function:6

(3)

where g(nx, ny) and f(nx, ny) denote the blurred image and the ideal

image, respectively, for each pixel at (nx, ny). M and N are the pixel

numbers for x (horizontal) and y (vertical) directions on the image

plane, respectively. d(nx, ny) is called the point spread function (PSF)

which represents the convolution kernel that describes the blurring

mechanism. w(nx, ny) is any additional noise affecting the captured

image, e.g., electrical noise of image sensors, illumination noise, etc. If

we have a prior knowledge on how the image is blurred, then we can

operate the deblurring (or the image deconvolution) by inversely

applying d(nx, ny) to g(nx, ny), the process known as non-blind image

restoration.8 The specific condition of measuring the probe tips meets

the non-blind image restoration because the blurring occurs due to the

movement of the x-y stage. The PSF d(nx, ny) can thus be described by

the stage motion. If we assume that the stage is running at a constant

speed, the blurring distance δxk in Fig. 3 becomes independent of time,

i.e., δxk=δx for all k = 0, 1, ... In this case, the relative motion between

the camera and the probe tips is linear and d(nx, ny) becomes spatially

shift-invariant, i.e., d(nx, ny)=d(nx−k, ny−k) for all k = 0, 1, ....

Assume that the vertical axis of the image plane is aligned with the

y coordinate of the stage and that the stage moves along y axis with a

constant speed v µm/sec. Then it becomes linear motion blur with the

length of motion δx=vtex. Denoting the spatial resolution of image

sensor along the y axis as s pixels/ìm, the PSF can be given by the

moving average filter

(4)

where L = [vtex s] denotes, in the number of pixels, the length of motion

during exposure. By applying the discrete-time Fourier transform to Eq.

(3) and denoting the transformed signals by capital letters, we get

(5)

for the spatial frequency (ωx, ωy), where the moving average filter is

written as

(6)

Therefore, the image restoration in this case is to find the ideal

image F(ωx, ωy) from the acquired one G(ωx, ωy) with a priori

knowledge on D(ωx, ωy) in the form of Eq. (6). The PSF D(ωx, ωy) is

non-minimum phase. Hence, the inverse filtering will amplify the noise

W(ωx, ωy) making it impractical.6 A more practical way is to formulate

the above mentioned deblurring problem as an optimal filter to

minimize the mean square error: i.e., given some observation of g(nx,

ny), say g1, find which minimizes the mean squared error (MSE):

(7)

Note that the blurring model in Eq. (5) is a linear system. If we

assume that the noise signal w(ωx, ωy) is zero mean, Gaussian distributed

and independent from other signals, the optimal solution to our deblurring

problem will be given by the minimum mean square error (MMSE)

estimator or Wiener filter:

(8)

(nx, ny) denotes the recovered image and the overline ( ) represents

the mean value of the corresponding signal. Note that if w is

zero mean. The effect of the mean value terms in Eq. (8) will show up

as an offset in the deblurring process and can be adjusted during the

calibration. According to the MMSE solution, the Fourier transform of

tpr tpr tpr tpr

Tg max tpr tex tro ttr+ +,( )≥

g nx ny,( ) d nx ny,( ) f nx ny,( )× w nx ny,( )+=

d i j,( )f nx i– ny j–,( ) w nx ny,( )+j=0

N 1–

∑i=0

M 1–

∑=

d nx ny,( )1

L--- if nx 0= 0 ny L 1–≤ ≤

0 otherwise⎩⎪⎨⎪⎧

=

G ωx ωy,( ) D ωx ωy,( )F ωx ωy,( ) W ωx ωy,( )+=

D ωx ωy,( ) 1

L---

ωyL

2---------⎝ ⎠⎛ ⎞sin

ωy

2------⎝ ⎠⎛ ⎞sin

----------------------e

jωy

L 1–( )2

---------------⎝ ⎠⎛ ⎞

–

=

f̂

minE f f ˆ–2

g g1

=[ ]

f̂ nx ny,( ) h nx ny,( ) g nx ny,( ) g nx ny,( )–( )× f nx ny,( )+=

f̂.

f g=

Fig. 3 Timing diagrams of vision events and the object position

664 / APRIL 2015 INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4

the convolution kernel h(nx, ny) for the MMSE estimator is computed as

(9)

where D*(ωx, ωy) is the complex conjugate of D(ωx, ωy), and K is the

additive factor given by

with Sw(ωx, ωy) and Sf(ωx, ωy) denoting the power spectrum of the noise

w and the ideal image f, respectively. Since both power spectrums are

unknown in practice, K is usually treated as a parameter that tunes the

effect of sharpening and noise.9 In summary, the image restoration is

achieved by applying the Wiener filter h(nx, ny) to the blurred image

g(nx, ny) to obtain (nx, ny) which represents the best estimate of the

original image f(nx, ny) in the sense of minimum mean square error.

3. Experiments

A tabletop probe analysis system has been designed and built to

demonstrate the main ideas given in the previous section. The test bed

comprises three main parts: the micro vision system, the x-y-z stage

with a probe card mount, and a real-time controller. The overall

configuration of the micro vision system and the stage is shown in Fig.

4. The micro vision system is attached to the stage so that the camera

scans the probe tips through the x-y planar motion of the stage during

the inspection of probe pins. Each axis of the stage is equipped with a

linear motion system that consists of a lead screw, a linear guide, and

a stepper motor. The z axis is fixed to the focal length of the lens and

does not move during the measurement. A linear encoder is attached

beside the y axis frame and provides the position and the velocity of the

camera along the y axis. The basic specifications of the x-y stage and

the linear encoder are listed in Table 1.

The real-time controller can communicate with both the vision

camera and the stage independently so that the motion data and the

image data can be synchronized in the way explained in the previous

section.

The designed machine vision system is shown in Fig. 5. The main

functional requirement of the vision system is to capture the image of

the end tips so that we can gauge the (x, y) position and the radius of

each pin tip. For this, the vision camera needs to magnify and to focus

only on a specific spatial point in the air. A 20× objective lens has been

used in our test bed for magnification. The LED (light-emitting-diode)

illumination is adopted as a light source and is placed on top of the

probe card with a diffuser. The camera is oriented with 90o angle with

respect to the LED. A beam splitter is mounted between the LED and

the lens. It transmits the light from the LED and projects the reflected

image to the camera.

Fig. 6 shows a sample still image of probe pins with 20× objective

lens where the direction of the stage movement (i.e., y direction) is

H ωx ωy,( )D* ωx ωy,( )

D* ωx ωy,( )D ωx ωy,( ) K+-----------------------------------------------------------=

KSw ωx ωy,( )Sf ωx ωy,( )------------------------=

f̂

Fig. 4 Experiment of the probe analyzer

Table 1 Specifications of the x-y stage

Specification Value

Microstep (step size of the stepper) 47.625 nm

Repeatability ≤ 1 µm

Encoder resolution (for y-axis) 0.5 µm

Fig. 5 Configuration of the machine vision component

Fig. 6 A sample still image of pin tips with 20× lens

INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4 APRIL 2015 / 665

indicated by the arrow. The pin tip has its radius around 3~8 µm and

the pins are located about 60 µm apart from each other. The camera

model is STC-CL202A manufactured by Sensor Technologies America,

Inc. Specifications of the camera and its image acquisition system are

listed in Table 2. The exposure time has been chosen to 10 ms based

on the intensity of the light source. The transmission time can be

computed using the size of the image data (=1620×1236×10 bits) and

the speed of the camera link communication protocol (=255 MB/s).

As explained in section 2.1, we need a deterministic real-time

platform that can process the vision data and the motion data in

synchrony. For this purpose, we adopted the RTOS (realtime operating

system) provided by National Instruments Inc. and converted a regular

PC into an RT processing machine. The overall configuration of the

real-time controller and the associated interface hardware is shown in

Fig. 7 in more detail. The PC houses three separate interface boards to

communicate with the camera (i.e. the frame grabber), the x-y stage and

the linear encoder, respectively. Especially, the interface board for the

linear encoder serves as the main DAQ (Data Acquisition) hardware

and sends out the deterministic trigger signal to the frame grabber such

that the exact timing of image capturing is controlled precisely by the

real-time controller. The original image has been converted to 8 bit data

to be compatible with the Labview image processing library functions.

4. Results and Discussion

The experimental test has been conducted with running the stage at

five different speeds: 0.2, 0.4, 0.6, 0.8 and 1.0 mm/s. The trigger period

of Tg = 130 ms for 1.0 mm/s of stage speed is chosen considering the

data size of each image to be processed. Determining the actual size of

pixel data for image processing needs some consideration and will be

explained in the next paragraph. Fig. 8 shows the flowchart of each

experimental trial. As soon as the stage starts to move, images are

captured every trigger period during which the pixel data get processed

with the synchronized encoder signal for computing x and y coordinates

as well as the radius of each pin tip. The computed pin information is

then stacked up to a predefined array to be provided to the user at the

end of the test.

As shown in Fig. 8, the basic image processing algorithm for the

probe tip analysis consists of deblurring (Wiener filtering), thresholding

and finding circles to compute their center coordinates and radii.

Although these are relatively simple image processing algorithms, the

total processing time for 1620×1236 pixel image takes more than 7

seconds in a standard PC, which is not acceptable for our purpose. The

simple way to reduce the image processing time is to reduce the size

of image data. From Fig. 6, we can first note that the pin tips are aligned

near the vertical center line and most of the side area is irrelevant to the

test. Secondly, for the range of speed we tested (up to 1 mm/s) and the

trigger period (130 ms), at least half of the total image data turns out

to overlap between two consecutive images. Therefore, we can trim

down some pixels along both x and y directions and, as a result, 255×

535 pixels are taken from the original image of 1620×1236 pixel size.

Another point considered in selecting the 255×535 pixel size for

Table 2 Specifications of the machine vision system

Camera Specification Value

Communication protocol Camera Link

Resolution 1620(H) × 1236(V) pixels

Bit depth 10 bits

Maximum frame rate 15 fps

Sensor type Progressive Scan CCD

Latency Value

Exposure time (tex) 10 ms

Readout time (tro) 65 ms

Transmission time (ttr) 9.87 ms

Fig. 7 Configuration of real-time controller

Fig. 8 Real-time image processing algorithm

Fig. 9 Image overlapping for a cropped pin

666 / APRIL 2015 INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4

online processing is that a small length along y direction actually needs

to overlap with the next image. This is because the image of a pin may

end up sitting on the boundary with some part of it being cut off. Fig.

9 shows such a situation where the lower pin of the left-hand side

image shows in its entirety while the same pin is clipped in the right-

hand side image. As long as the overlap length (denoted by ∆Ly) is

large enough to cover the entire length of a pin (including when it is

blurred), we can process the pin regardless of clipping.

As a result of image reduction explained above, the maximum image

processing time has been reduced to around 70 ms. Note that the

trigger period of our choice Tg = 130 ms satisfies Eq. (2) since 130

max (70, 10+55+9.87) = 84.87. Table 3 summarizes the parameters for

experimental results. When the stage speed gets lowered, the trigger

period can be increased accordingly to keep the overlap length ∆Ly

similar for all speeds. For our experiment, the trigger period has been

increased by 65 ms as v is reduced by 0.2 mm/s.

In order to evaluate the inspection error of the method proposed in

this paper, still images are first taken for all probe pins to have their

position and radius values computed and spared as reference data. The

pixel resolution is around 0.25 µm/pixel and the repeatability of the x-

y stage is 1 µm at maximum as shown in Table 1. Thus, the base error

margin of the reference data itself can be considered around 1.25 µm

at maximum. Fig. 10 compares three different images of the same pins:

the still image for reference data (Fig. 10(a)), the blurred image captured

during the stage motion (Fig. 10(b)) and the restored image (Fig. 10(c)).

For each pin, the (x, y) coordinates and the radius r are printed in µm

scale on Figs. 10(a) and 10(c) for comparison. For three pins shown

here, the x coordinate data from the restored image coincide with those

of still image while the y coordinate data and the radius of the bottom

pin introduce some errors. Note that the perfect image restoration is not

possible in practice due to the unknown noise w(nx, ny) (See Eq. (3)).

Experimental results have been collected from a batch of about 80

pins on the probe card. All the pins are inspected at once for each test

speed listed in Table 3. The inspection error for each pin is then

computed using the reference data a priori obtained from still images.

The mean and standard deviation of inspection errors are presented in

Fig. 11 using bar graphs. The circle mark for each speed denotes the

mean value of the inspection error and the line segment between the

circle and the upper (or lower) bar indicates the standard deviation.

Overall, the measurement error for the tip size r shows the smallest

standard deviation which is around 0.3 µm for all test speeds. The

standard deviations for position errors are a bit larger than that of r,

specifically, 0.5 µm for x axis and 0.8 µm for y axis. As expected, the

y direction that causes linear motion blur introduces the largest

inspection error. Nevertheless, standard deviation is comparable to the

base error margin (1.25 µm). The inspection error for y coordinate may

tpr

Fig. 10 Comparison of sample images for three different cases

Table 3 Experimental test conditions

Parameter Value

Image size for online processing 255(H) × 535(V) pixels

Max. processing time 70 ms

Camera speed v 0.2/0.4/0.6/0.8/1 mm/s

Trigger period Tg 390/325/260/195/130 ms

tpr

Fig. 11 Mean and standard deviation of measurement errors of tip

position and radius

Fig. 12 Histograms that show error distribution of x, y and r for two

different speeds of camera movement

INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 16, No. 4 APRIL 2015 / 667

also be attributed to the resolution of the encoder which is 0.5 µm. (See

Table 1). One interesting aspect of the results shown in Fig. 11 is that

the standard deviation of inspection errors is not necessarily

proportional to the test speed. This suggests that the Wiener filtering

used for deblurring may perform with similar mean square errors for

the range of speeds tested in this paper.

To take a more detailed look, the actual distributions of inspection

errors are plotted in Fig. 12 by histograms for two test speeds: 0.8 mm/

s and 1.0 mm/s. ex, ey and er denote the estimation errors for the x, y and

r values, respectively. The frequency data are normalized by the total

number of pins. We can see that the inspection errors are close to the

normal (Gaussian) distribution. For the number of pins tested, the

maximum error for r and x coordinate is around ±1 µm while that of

y coordinate is around ±1.5 µm.

To compare the inspection speed, a separate test algorithm has been

created and tested based on the primitive stop-and-go approach. The

maximum speed in this case is observed to reach 2 pins per second at

maximum for the same probe card. The probe pins are about 60 µm

distance apart, so the proposed probe card analysis technique operated

with 1 mm/s of stage speed will result in the inspection speed

equivalent to 16.7 pins per second which corresponds to more than 8×

improvement compared to the stop-and-go approach.

5. Concluding Remarks

We studied the dynamic imaging approach for fast and high speed

visual inspection. The main idea is to operate the vision-based inspection

on-the-fly while the camera (or the target object) is continuously

moving. In doing so, the position measurement from the encoder is first

synchronized with the image data that is captured by a controlled trigger

signal under the real-time setting. Capturing images from a moving

camera creates blurring in the image. The deblurring technique has

been employed to restore the original still images from blurred ones.

Experimental results with the probe card analysis show that the

inspection speed can be significantly increased compared to the

conventional stop-and-go approach with the inspection error still

contained within the range of base error margins corresponding to the

quantization level of encoder and the pixel resolution. The

implementation shown in this paper can be useful to other high density

visual inspection processes where the inspection time is a critical issue.

It should be mentioned that the image blurring can also be improved

by other ways such as the use of a high speed camera with the impulse

lighting. However, this will be at the expense of hardware modification

and the increased cost. The image restoration technique introduced in

this paper can be considered as a simple means to improve the

performance of a vision-based inspection system without additional

hardware and system modification. In fact, it can be applied to any

existing vision-based inspection system for further improvement in the

inspection speed by allowing the system to run at higher speeds.

ACKNOWLEDGEMENT

The work in this paper has been supported by the National Research

Foundation (NRF) of Korea under the Global Research Network

Program.

REFERENCES

1. Sun, T. H., Tseng, C. C., and Chen, M. S., “Electric Contacts

Inspection Using Machine Vision,” Image and Vision Computing,

Vol. 28, No. 6, pp. 890-901, 2010.

2. Mann, W. R., Taber, F. L., Seitzer, P. W., and Broz, J. J., “The

Leading Edge of Production Wafer Probe Test Technology,” Proc. of

Internationl Test Conference, pp. 1168-1195, 2004.

3. West, P. C., “Machine Vision in Practice,” IEEE Transactions on

Industry Applications, Vol. IA-19, No. 5, pp. 794-801, 1983.

4. Tao, Y., Hu, H., and Zhou, H., “Integration of Vision and Inertial

Sensors for 3D Arm Motion Tracking in Home-based Rehabilitation,”

The International Journal of Robotics Research, Vol. 26, No. 6, pp.

607-624, 2007.

5. Jeon, S., Tomizuka, M., and Katou, T., “Kinematic Kalman Filter

(KKF) For Robot End-Effector Sensing,” Journal of Dynamic

Systems, Measurement, and Control, Vol. 131, No. 2, Paper No.

021010, 2009.

6. Bovik, A. C., “Handbook of Image and Video Processing,” Academic

Press, pp. 71-257, 2010.

7. Banham, M. R. and Katsaggelos, A. K., “Digital Image Restoration,”

IEEE Signal Processing Magazine, Vol. 14, No. 2, pp. 24-41, 1997.

8. Nayar, S. K. and Ben-Ezra, M., “Motion-based Motion Deblurring,”

IEEE Transactions on Pattern Analysis and Machine Intelligence,

Vol. 26, No. 6, pp. 689-698, 2004.

9. Grimaldi, D., Lamonaca, F., and Macrì, C., “Correction of the

Motion Blur Alteration in the Human Lymphocyte Micro-Nucleus

Image based on Wiener’s Deconvolutionl,” Proc. of 16th IMEKO

TC4 Symposium, pp. 22-24, 2008.