You are on page 1of 6

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03

18

Automated Detection of Glove Defects


Using Vision Control
W. S., Chan, K. F. Thang, and N. S. Lai

Abstract An automated vision control system is a convenient


integrated platform to identify defect on objects in fast
production lines. Here, the proposed system is designed to detect
defective gloves during fast manufacturing process thereby
removing human error. The system is capable of detecting a few
unique types of glove defects based on region of interest (ROI)
and area of the gloves. This method has shown that the system is
able to achieve an overall accuracy of 81%.
Index Term Defects detection, gloves defects, region of
interest, vision control.

I. INTRODUCTION
IN modern industries, automation plays a very important role
in maintaining quality of their products at the highest quality.
In the past, man power is used to perform repetitive tasks
where accuracy and efficiency of the work are found
deteriorating over time. To solve this issue, automated systems
are usually used, in our case a computer vision based system.
This paper presents a vision-control system that has a fast and
accurate detection of defects. Here, the product that has been
chosen is the latex glove, which is manufactured at a high
speed where manual defect detection by human labor often
introduces plenty of mistakes.
At the current state of research, vision, laser, ultrasonic have
been used to detect defects on various products to improve the
production yield. Although there are many types of sensors,
the one that has proven to be the most reliable is the visionbased techniques. There are many types of vision-based
techniques used for defect detection such as neural network
[1]-[3], edge detection [4]-[5], morphology processing [6]-[8],
[10], Gabor transform [9]-[10], fuzzy inference system [11]
and valley-emphasis method [12]. The neural network is a
complex system that can memorize and learn from historical

images or data to the neural network as input, the more likely


the system will be working as expected [1] - [3]. On the other
hand, morphology techniques use dilation and erosion of the
image, which have the ability to increase or to reduce the outer
layer of the object. This method was mainly used to filter out
noises and to produce desired shapes which can be identified
by the vision system [6]-[8]. Edge detection has been adopted
where the outline of the desired object will be highlighted
while removing the background. This method can preserve the
shape of the object if used correctly [4]-[5]. Gabor transform is
a method to further filter the images before detecting the
defects [9]-[10]. And, the fuzzy inference was used to solve
classification problems like classifying similar defects on
similar circumstances [11].
II. METHODOLOGY
The methods used to detect defects in this project are based
on the area and the region of interest (ROI) of the image. With
these parameters three types of glove defects can be classified
namely, normal gloves, stuck gloves and the torn gloves. As
shown in Fig. 1, the video acquisition system, i.e. a costeffective webcam will be initiated to capture the live video
streams, which are subjected to further video processing that
encapsulates the developed algorithms. The system will also
record the date and time, as well as the quantity of the three
types of gloves detected.

data to solve the problems. By inserting a large amount of


W. S. Chan is with the School of Engineering, Asia Pacific University of
Technology & Innovation, Technology Park, Kuala Lumpur, Malaysia (email:
ff7_cloudviii@hotmail.com).
K. F. Thang is with the School of Engineering, Asia Pacific University of
Technology & Innovation, Technology Park, Kuala Lumpur, Malaysia (email:
dr.ka.fei@apu.edu.my).
N. S. Lai is with the School of Engineering, Asia Pacific University of
Technology & Innovation, Technology Park, Kuala Lumpur, Malaysia (email:
dr.nai.shyan@apu.edu.my).

Fig 1. Methodology of the proposed vision system.

Fig. 2 illustrates a normal glove without defects; this is the


desired glove for the industry. As compared to the other two
defects, the normal gloves have a moderate area which can be
used for the classification with the ROI. On the other hand,
Fig. 3 illustrates a stuck glove where it is stuck on the mold.

163203-5959-IJET-IJENS June 2016 IJENS

IJENS

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03


This happens when the glove might not be dried properly and
was sticking to the mold hence was not able to be pulled out
by the stripping machine, this defect is more common as
compared to the torn glove and it is not desirable. Since the
area of the glove is the smallest among the three gloves, this
makes the classification method a lot easier. Finally, Fig. 4
illustrates a torn glove where the glove is being torn off by the
stripping mechanism, this happens because the outer edge of
the glove is attached to the stripping mechanism which causes
the tearing of the glove. Since the area of the torn glove is the
largest, it makes it easy to be classified.

19

final step, only the glove remains in the image as the object to
be processed further.
With reference to the background removal, there are several
important factors to be considered namely the background
complexity, the lighting and also the location of the mounting
of the video acquisition system. In this case, external lighting
cannot be applied as it will introduce reflective shadows on the
glove surface. Secondly, the conveyor systems holding the
gloves would also introduce quite a complex background
which is difficult to be removed as.
Lastly, the location of the mounting of the acquisition
system is also very important such as to capture the image of
the gloves appropriately in a fast-pace manufacturing
environment. In here, blob analysis is used to reduce noise in
the video as well as to calculate the area of all detected objects
after filtering. By converting the pixels values to another unit,
the area of each object can be determined and used for
classification as illustrated in Fig. 5.

Fig. 2. Normal gloves.

Fig. 5. Area and bounding box using blob analysis

Here, a bounding box is obtained as a result of the blob


analysis which can then be used to determine the area and
centroid of the glove. The area of the bounding box as labeled
in Fig. 5 can then be used to determine the average range for
the area of the normal gloves. The final value is found to be
between 10000 and 20000 pixel values. Next, the cropping
function is used to crop a specific area and to focus on that
particular area for further processing. By using the cropping
method for the ROI, the algorithm can execute much faster as
it focuses only on one glove at a time as well as effectively
removes unwanted noise from the image or video.
A detailed step-by-step illustration on the image processing
algorithms being adopted to obtain the object of interest, in
this case the glove, is provided in Figs. 6 15 below.

Fig. 3. Stuck gloves.

Fig. 4. Torn gloves.

To determine the area of the object taken, the image


acquired in real-time will be subjected to processing by several
algorithms so that the desired object is extracted from the
background. The main objective of this algorithm first hand is
to isolate the glove image from the background so that in the

163203-5959-IJET-IJENS June 2016 IJENS

IJENS

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03

20

Fig. 11. Eliminating small objects: filtering.

Fig. 6. Original image.

Fig. 12. Filling holes.

Fig. 7. Greyscale image.

Fig. 8. Morphology: dilation.

Fig. 13. Noise removal filter.

Fig. 9. Thresholding image.

Fig. 14. Final processed image.

Fig. 10. Filling holes.


Fig. 15. Region of interest (ROI) extracted.
163203-5959-IJET-IJENS June 2016 IJENS

IJENS

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03


As illustrated in Figs. 6 to 15, there are a few repetitive
functions used in the algorithm. This happens because the
background of these gloves varies and due to reflection of
natural lighting. This means by adjusting threshold alone is not
sufficient enough to separate the gloves away from the
background. Dilation was first used in the filtering process to
create a small gap in terms of thresholding so that future
filtering will be made easier. After using morphology, the
image will be converted to binary form to further separate the
desired object from the image. Since the surface of the glove is
not flat, some parts of the glove does not reflect as much light
as the other parts which will form some gaps within the
glove parameters. In order to solve this, erosion and holefilling algorithms can be used to regain its original size as
much as possible. In order to eliminate the remaining noises, a
special filter was used which removes the noises if their pixel
size is small, this is more effective than the median, Gaussian
or morphology filter dilation because it requires less step and
immediately remove noises without changing affecting the
glove parameter. Lastly, the image was cropped to only focus
on one of the gloves in the system.
The manufacturing line for gloves are operating at high
speed hence the final algorithm to be implemented has to cope
with the speed of the manufacturing. With that, the proposed
system has been implemented on a real-time platform with
graphical-user-interface (GUI) developed to illustrate the
result of detection in real-time for quality checking purpose.
The overall process of the system is illustrated in Fig. 16. The
system is initiated via activation of the GUI, and with the time
and date stamp being displayed in the GUI. Next, the preprocessing of the videos will commence. This is where the
desired object will be extracted from the background image
while removing unwanted noises by using morphology, blob
analysis, thresholding, ROI and multiple filtering of the image.
When the gloves are passing through the video acquisition
system, there are a few conditions to either classify the gloves
or not to activate the processing system. The first condition is
that the extracted object (i.e. glove) must have a pixel area of
over 30000 pixels. This is the minimum threshold accepted by
the system to determine if the object is actually a desired
object or just a random noise.
Since the defects are classified using the ROI and the area
of the extracted objects, the normal gloves are found to have
an area between 10000 to 20000 pixels and the torn gloves
will have a larger area due to stretching and tearing which is
more than 20000 pixels. Finally, since stuck gloves are gloves
that stick to the mold, the area is the smallest. As for the area
of interest, anything besides the ROI will be removed from the
camera focus and therefore, certain x-axis values will be
adjusted to reduce error in the system. This is great for the
system because it improves processing speed by eliminating
the field of vision that needs to be processed and also reduces
the amount of filter used to obtain desired result. As shown in
Fig. 16, the value is 400 on the x-axis of the ROI. This means
that when the centroid of the object is moving to an area which

21

Fig. 16. Flow-chart of the video processing system.

has less than 400 pixels of the x-axis, the classification will
trigger when both requirements of the activation has been
fulfilled which is the area and the centroid value. When the
specific gloves have been detected, the specific counters will
be triggered and shown in the GUI so that the records can be
maintained for quality checking later on. Finally, when the
GUI is halted, the stop date and time stamp will be shown in
the GUI.
Owing to the fast production speed, there are times where
the gloves went past the video acquisition system without
processing being performed. To circumvent the issue, a unique
algorithm has been added where it will detect the gloves
passing the ROI multiple-times. However, if this were to
happen, one glove will be counted multiple times and this has
been solved by the subsequent algorithm to ensure that when
the counter was increased by one, it will not be added again
until the glove reaches a certain x-axis location within the
region of interest.
In summary, the glove enters the ROI will trigger the
counter and when the glove is near the end of the ROI, it will
release the trigger, allowing the second glove to be counted as
another glove and a new count. By doing that, the vision
control system that has been developed will have fast reaction
time which can cope with fast pace of the production typical in

163203-5959-IJET-IJENS June 2016 IJENS

IJENS

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03

22

gloves manufacturing industry. Fig. 17 to 19 illustrates the


three types of gloves detected which have been processed by
the video processing system. The GUI developed as the frontend and for the quality checking is illustrated in Fig. 20. The
GUI consists of Start and Stop button which will trigger the
date and time stamp and to stop/halt the vision control system
for gloves defects detection. For example, Fig. 20 shows that
the counter for the normal gloves is 43, whereas there are 2
torn gloves and 3 stuck gloves have been detected. Lastly, a
formula has been developed to calculate the efficiency of the
production line.

Fig. 20. Graphical front-end of the system.

III. RESULTS & ANALYSIS

Fig. 17. Normal glove as an object within the ROI.

In this section, results on actual testing in the production


line will be illustrated herein in order to show the performance
of the system. Table I illustrates the outcome of testing
performed on the same production line but at different date
and time. As shown in Table 1, average efficiency of the
system is 81%. This calculated percentage on efficiency will
be useful for the manufacturing plant to work on quality
improvement or to reduce the defects rate.
TABLE I

TESTING RESULTS FROM THREE PRODUCTION RUNS

Fig. 18. Stuck glove as an object within the ROI.

Production

Normal
Gloves

Stuck
Gloves

Torn
Gloves

Accuracy
(%)

46

82.14

64

87.67

43

89.58

However, there are a few different tests which were done


while conducting the project namely the results obtained from
a recorded video with and without a light source. This was
done to determine which situation is more suited to be used in
the live video algorithm.
TABLE II
TESTING RESULTS FROM RECORDED VIDEO WITHOUT LIGHT SOURCE
Gloves
Detected
Stuck up
Torn
Accuracy,
Gloves
Gloves
Gloves
%
22
20
1
2
70.5

Fig. 19. Torn glove as an object within the ROI.

TABLE III
TESTING RESULTS FROM RECORDED VIDEO WITH LIGHT SOURCE
Gloves
Detected
Stuck up
Torn up
Accuracy,
Gloves
Gloves
Gloves
%
26
--1
-----

163203-5959-IJET-IJENS June 2016 IJENS

IJENS

International Journal of Engineering & Technology IJET-IJENS Vol:16 No:03


In Table II, the live video was recorded without a light
source and after processing, the accuracy of the system is
around 70%. As compared to table III, there are no results
obtained by the vision system. This happens because the
created algorithm was an adaptive system which adapts the
natural light intensity as well as the unique background at the
production line. When there is an external light source added
to the system, in a hope to extract the desired objects away
from the background easier, the algorithm could not recognize
the object anymore because the light might be 1) Not
specifically focused on the gloves 2) The light somehow
enhances the background more than the object and 3) The
external light source is not compatible with natural lighting in
the production line.

23

[10] ESTER RESENDIZ, Automated Visual Inspection of railroad Tracks,


IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION
SYSTEMS, 14(2). June, 2013. p. 751-760, 2013.
[11] ALICE, VALENTINA, MARCO, V. & MARCO, V., A Fuzzy Inference
System Applied to Defect Detection in Flat Steel Production, Fuzzy
Systems (FUZZ), 2010 IEEE International Conference. Barcelona. pp. 16, 2010.
[12] MANIMOZHI & S.JANAKIRAMAN, Defect Detection in Pattern
Texture Analysis, International Conference on Communication and
Signal Processing, India. pp. 17, 2014.

IV. CONCLUSION
The proposed vision control system has successfully
detected three different types of gloves with an average
accuracy of 81%. The system can be improved further by
incorporating algorithms to detect the defects in gloves
according to different sizes and colors of the gloves, which
will further introduce more variety of gloves defects.
Nevertheless, this paper has illustrated and summarized an
investigation work that proves promising and provides a good
foundation for future work to continue in a bid to automate the
quality inspection process for the gloves manufacturing
industry.
REFERENCES
[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

L. Wang and Y. Shen, "Design of Machine Vision Applications in


Detection of Defects in High-Speed Bar Copper," E-Product E-Service
and E-Entertainment (ICEEE), 2010 International Conference on,
Henan, pp. 1-4, 2010.
M.S. BIRADAR & SHARDA P. JADHAV, Fabric Defect Detection by
using Neural Network technique, IOSR Journal of Electronics and
Communication Engineering (IOSR-JECE), e-ISSN: 2278-2834, pISSN: 2278-8735. 9(2), Ver.VIII. p.70-76, Aug. 2014.
P. BANUMATHI & G. M. NASIRA, Fabric Inspection System Using
Artificial Neural Networks, International Journal of Computer
Engineering Science (IJCES). 2(5). p.20-27, 2012.
ABDELHEDI, TAOUIL & HADJKACEM, Design of Automatic
Vision-based Inspection System for Monitoring in an Olive Oil Bottling
Line, International Journal of Computer Applications (0975 8887),
51(21). August, 2012. p. 39-46, 2012.
ESTER RESENDIZ, Automated Visual Inspection of railroad Tracks.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION
SYSTEMS, 14(2). p. 751-760, June. 2013.
BIN LIU, SHENGJIN WU & SHIFANG ZOU, Automatic Detection
Technology of Surface Defect on Plastic Products Based on Machine
Vision, Mechanic Automation and Control Engineering (MACE), 2010
International Conference on, Wuhan, pp. 2213-2216, 2010.
XIAOJUN JIA, Fabric Defect Detection Based on Open Source
Computer Vision Library OpenCV, 2nd International Conference on
Signal processing System(ICSPS). Jiaxing, China: IEEE. pp.342-345,
2010.
ANDERS & MATTHEW, Morphology-Based Crack Detection for
Steel Slabs, IEEE JOURNAL OF SELECTED TOPIC IN SIGNAL
PROCESSING, 6 (7). p.866-875, 2012.
DORIAN
SCHNEIDER,
TIMM
HLTERMANN,
FLORIAN
NEUMANN, ACHIM HEHL, TIL AACH & THOMAS GRIES, A
vision based system for high precision online fabric defect detection," 7th
IEEE Conference on Industrial Electronics and Application (ICIEA).
Germany. pp. 1494-1499, July. 2012.

163203-5959-IJET-IJENS June 2016 IJENS

IJENS

You might also like