A VISION SYSTEM FOR TEXTILE FABRIC DEFECT DETECTION

Remus Brad* and Raluca Brad*** Dept. of Computer Science, ** Dept. of Textile
Technologies "Lucian Blaga" University of Sibiu,B-dul Victoriei 10, 550024 Sibiu,
Romaniaremus.brad / raluca.brad@ulbsibiu.ro
ABSTRACT
Existent industrial applications of computer vision systems are directly connected
to quality assurance requirements. The task of fabric defect detection is carried out
by human visual inspection, in most of the traditional textile industry. The
possibility of automated defect detection is investigated and a solution leading to
improved productivity and high quality in the weaving process is proposed. We are
introducing a defect detection using dwt and Neural network for the inspection of
textured materials .Discrete wavelet transform is employed for extracting the filter
coefficient. An appropriate thresholding of the filtered image followed by
segmentation accomplishes the defect detection. Real image tests shows that our
algorithm is robust and computationally efficient for the inspection of textured
materials.
Keywords:
Quality inspection, textile industry, computer vision, defect detection, DWT, NN
1. INTRODUCTION
Quality assurance systems have been developed in the aim of providing the client
with a high level of trust in the producers capacity to maintain permanently the
product specifications according to standards and original technical design. The
efficiency of these systems has already been confirmed by several international top
corporations such as Motorola, Hewlett Packard, etc., whos specific
implementation have become de facto standards (for instance Six Sigma). These
aspects have persuaded client companies to migrate into producers certified for
their efficacy and efficiency by quality assurance certification organizations.
Automatic production control is an important phase of quality assurance and its
major scope is to prevent client supply with defective products. In the textile
industry, due to the fact that all operations are hand-made, high productivity and
quality can be achieved only by intensive quality inspection before and between
the manufacturing stages. The high production speed and the large flexibility
required by customers urge to automated defect detection of the quality assurance

system. In the traditional textile industry, this task is carried out by human visual
inspection. Manual inspection is usually a difficult task due to the small scale of
detects and the large scale of inspected surface. In the case of the weaving sector,
inspection is performed at the end of the manufacturing stage. Large batches of
fabric rolls are manually inspected and actions are performed off-line of the
production system. Employing computer vision automation directly on the
production stage will improve the on-line reaction of the manufacturing staff and
reduce the number of defects. Besides the high processing speed, computer vision
systems can offer robust detection and large flexibility. Automation based on
image processing does not suffer of human limitations and could entirely replace
traditional methods. Automated visual inspection relies on material properties as
texture. Texture analysis techniques for fabric defect detection allows determining
texture features and statistically segment defects. Gray-levels texture properties
and statistics (Liu et al. 1998) are employed for local segmentation. Characterizing
the fabric texture using a Markov random field model in zdemir et al (1996),
detection is derived by hypothesis testing. In Campbell et al. (1996), model-based
clustering is used for woven fabric defects. Since texture can be defined as a
function of spatial variations in pixel intensities and possess a high level
of periodicity, Fourier transforms and Fourier-domain analysis (Tsai et al. 1999),
(Chan et al 2000) are convenient tools for discriminating texture variations. Multiresolution approaches are decomposing fabric images in several scales using a
bank of filters.
Defects:
Horizonatal fault
Vertical fault
THE VISION SYSTEM TEST UNIT
CONTROL SYSTEM DESIGN:
The efficiency of a manufacturing control system depends on its accuracy and
processing speed. A typical regulation system indicator is the reaction time,
between the defect detection and the corrective action. The reaction time can be
shortened by the use of efficient techniques for detection, analysis and data
communication. These goals can be achieved by installing control points in
specific manufacturing stages. The manufacturing control system comprises the
following operations:
Process / product control for deviation detection
control data analysis for the determination of deviation causes and requested
corrective actions

information of the process operator on corrective actions (feed-back)

process adjustmentSuch a modern concept of manufacturing control system
offers all the benefits of a short time reaction and theself-consciousness of quality
among the manufacturing personnel. Thus, the activities must be carried out
asclose as possible of manufacturing points.Figure 1. The vision system
schematicThe unit designed for testing and implementation of a textile
manufacturing control system is presented infigure 1. It consists of a PC based
computer, equipped with an image grabbing card and interfaced via the parallel
with a loom motion control driver. Only one B/W camera was employed in our
test, but the use of multiple units is overviewed for a complete coverage of the
weave.

2.2. VISION SYSTEM DESIGN

The Gabor filters are resulted from a modulation product of a gaussian and
sinusoidal signals. Gabor hasintroduced these elementary signals as optimal
transmitional signal in telecommunication. Application of Gabor filters starts from
edge detection and ends with texture classification and image compression. In the
two-dimensional plane, the Gabor function has the following general form:
Texture defect detection can be defined as the process of determining the location
of various defects based on the textural properties of the input image. A quality
control system implies an unsupervised process for defect detection, dealing
especially with unknown texture defective patterns. As presented above, Gabor
filters havea special feature that allow to texture tuning and consequently to
respond in a different manner to texture irregularity. Unsupervised inspection is
achieved by both local and global investigation. A bank of Gabor filters processes

the input image. Each filter is chosen in order to respond on a small frequency and
scaledomain. In our method, the filter bank is generated using equation (1) and (2).
The filter having the highest cost function will be selected for defect detection and
the original image will be subject to filtering. Then, a thresholding operation is
required for the final segmentation of texture defects.Segmentation of the filtered
image will perform defect detection. Choosing the appropriate filter representsthe
key to correct results. An unsupervised algorithm can be developed for the
selection of the filter that performs large outputs in the case of defect texture and
small output for defect-free. A cost function (Kumar et al. 2002) is required for the
discrimination of the filters bank results. Figure 3. Samples of fabric defect
imagesSelecting the threshold has been another challenging task. Some authors
recommend using the maximumvalue resulted on the filtered image, having as
original a defect free sample. In order to eliminate the pre- processing of a defectfree image, we are proposing an original method using as a threshold, the median
value
4. CONCLUSIONS AND FUTURE WORK
This paper presents an application of the Gabor function to texture inspection and
especially to fabric defectdetection. The main goal of the proposed method is the
improvement of quality assurance techniques. A bank of Gabor filters have been
generated and a cost function based on maximum filter response for the
giventexture is used. In order to eliminate pre-processing of a defect-free image, a
novel method for thresholdselection is introduced. Based on a region-growing
algorithm, the segmented binary image provides defectidentification and
localization.The results have shown successful detection of a variety of fabric
defects. Even if it covers most of thecommon defects from the weaving and
knitting industry, tests will be conducted on other possible situations.Being
computationally feasible, PC-based implementations of on-line fabric inspections
can be developed.

fabric faults are responsible for major defects found by the garment industry.
Due to the increasing demand for quality knitted fabrics high quality requirements
are today greater since customer has become more aware of Non-quality
problems.
In order to avoid fabric rejection, knitting mills have to produce fabrics of high
quality, constantly. Detection of faults during production of knitted fabric with
circular knitting machine (CKM) is crucial for improved quality and productivity.
Any variation to the knitting process needs to be investigated and corrected. The
high quality standard can be guaranteed by incorporating appropriate quality
assurance. Industrial analysis indicate that product quality can be improved, and
defect cost minimised, by monitoring of the circular knitting process (Jearranai and
Tiluk, 1999).
Fine gauge knitted fabric faults are very different in nature and
appearance and are often superimposed. They can be attributed
not only to the knitting, but also to the quality of yarns, dyeing
and finishing (Iyer et al, 1995). Some faults can be easily avoided
by respecting some fundamental pre-requisites on the circular
knitting machine such as the use of positive yarn feeders and the
respect of rigorous machine maintenance and cleaning schedule.
Other faults are much more difficult to expect because they are
not related to just one cause.
New generation circular knitting machines are conceived with auxiliary equipment
that ensure less fabric faults during knitting such as filter creel, lint removal, thread
survey, precise oiling and fabric faults detector devices. Nevertheless, some fabric
faults are not detectable with these equipment and fabric has to be inspected after
knitting. Fabric fault detector is able to detect holes and dropped stitches, but it is
sometimes not enough reliable and have to be disconnected especially when a
structured fabric is knitted because special fabric structures could be confused with
faults by the sensor. Other faults cannot be detected during knitting but only after
fabric relaxing or finishing such as fabric spirality and colour mismatching.
Many researchers have applied computer vision to improve inspection method of
human vision in textile products. In most of them, the image of a knitted garment
had been considered to specify the fault features (Shady et al, 2006, Kuo and Su,
2003, Palmer and Wang, Celik et al, 2005, Hemdan and Ayatallah, 2008). Other
works aimed to classify defects in knitted fabric by using image analysis and
neural network algorithm or Fuzzy logic (Shady et al, 2006, Slah et al). All theses
methods are not completely reliable because image analysis of knitted fabrics

involves difficulties due to the loop structures and yarn hairiness, compared to
woven fabrics consisting of neat warp and weft yarns. Knitted fabric faults can also
be detected by inspecting yarn input tension and loop length (De Araujo and
Catarino, 1999, Semnani and Sheikhzadeh, 2007) but only few types of faults are
concerned by theses methods.
Human inspection by using knitted fabric inspection machines remains today the
most used way to classify faults after knitting and after finishing. Generally, faults
are classified by type and by frequency in the inspected knitted roll. The inspection
assessment permits to appreciate fabric quality. The judgement of fabric quality
depend on faults tolerance levels fixed by each knitter and could be in some cases
subjective because it is often based only on the number of faults and not on fault
size and gravity.
References
1. Celik O, Ucar N and Ertugrul S (2005): Determination of Spirality in Knitted
Fabrics by Image Analysis, Fibers and Textiles in Eastern Europe, Vol 13, pp 4749.
2. De Araujo M, Catarino A and Hong H (1999): Process Control for Total Quality
in Circular Knitting, AUTEX Research Journal, Vol 1, pp 21-29.
3. Hemdan A A T and Ayatallah M S (2008): Online Fabric Defect Detection and
Full Control in Circular Knitting Machine, AUTEX Research Journal, Vol 8, pp
21-29.
4.Iyer C, Mammel B and Schach W (1995): Circular Knitting, Meisenbach GmbH,
Bamberg, Germany.
5. Jearranai L U and Tiluk D (1999): Quality Assurance in Circular Knitting, Part I:
Theoretical Analysis, Thammasat International Journal of Science and Technology,
Vol 4, pp 72-81.
6. Kuo C F J and Su T L (2003): Gray Relational Analysis for Recognizing Fabric
Defects, Textile Research Journal, Vol 73, pp 461-465.
7. Palmer S and Wang X (2004): Objective Classification of Fabric Pilling Based
on the Two-dimensional Discrete Wavelet Transform, Textile Research Journal, Vol
74, pp 140-145.
8. Semnani D and Sheikhzadeh M (2007): Online Control of Knitted Fabric
Quality: Loop Length Control, International Journal of Electrical, Computer, and
Systems Engineering, Vol 1, pp 213-218.

9. Shady E, Gowayed Y, Abou-iiana M, Youssef S and Pastore C (2006): Detection

and Classification of Defects in Knitted Fabric Structures, Textile Research
Journal, Vol 76, pp 295-300.
10. Slah M, Amine H T and Faouzi S (2006): A New Approach for Predicting the
Knit Global Quality by Using the Desirability Function and Neural Networks,
Journal of The Textile Institute, Vol 97, pp 17-23.
Identification of fabric defects based on discrete wavelet transform and backpropagation neural network

Liu Jianli

Zuo Baoqi
Journal of the Textile Institute (Impact Factor: 0.73). 01/2007; 98:355-362.
DOI: 10.1080/00405000701547193

ABSTRACT Detection of fabric defects can be considered as a texture segmentation and

identification problem, since textile faults normally have textural features that are different from
features of the original fabric. A feasible approach for the recognition of fabric defects based on
discrete wavelet transform and back-propagation neural network is proposed in this article, the
indispensable processes of which are defect image preprocessing, wavelet transform, feature
extraction, principal component analysis of the extracted feature parameters, and defect
identification. Under the experimental condition, the average recognition accuracy of defects and
nondefects are 99.2% and 100%, respectively. Experimental results show the advantages with
high identification correctness and high inspection speed.
0 BOOKMARKS

63 VIEWS

Citations (23)

Cited In (9)

Source

Article: Developing an Algorithm for Defect detection of Denim Fabric: Gabor

Filter Method
Lale Canan Dlger
[Show abstract] [Hide abstract]
ABSTRACT: ABSTRACT In this study, an algorithm is developed by using
Gabor filtering and double thresholding methods for fabric defect detection
automatically. The defective area of the fabric image is accentuated, and the
defect-free texture is attenuated as result of convolution operation with the
Gabor filter. The image is then converted into binary form for by using double
thresholding method. The noises are removed and the defective area is
determined clearly by using dilation and erosion morphological operations.
The boundary of the defect is labeled. A fabric defect image database
consists of five defect types; warp lacking, weft lacking, hole, soiled yarn and
knot is formed. The database includes 30 different images for each type of
defect and defect-free fabric samples. Thus, the algorithm is applied over 180
fabric images. All defective areas are detected with high success rates. The
performance of the algorithm is given statistically. Key Words: Gabor filter,
Fabric defect inspection, Image processing, Quality control, Denim fabric.
ZET
Tekstil ve Konfeksiyon 01/2013; 23(2):265-270. 0.30 Impact Factor

Article: Measurement of cloth fell position using image analysis

Bekir Yldrm, Gngr Baer
[Show abstract] [Hide abstract]
ABSTRACT: In this paper, a computeraided imageanalyzing method for
measuring cloth fell position directly in a noncontact manner is introduced.
The method measures the cloth fell position from an image taken by means
of a camera which is mounted over the woven fabric and warp region
including cloth fell on a weaving loom. Laser light lines are used as reference
lines. The image containing the reference lines, the fabric and warp system is
decomposed using wavelet transforms. Horizontal detail coefficient matrix is
filtered with local standard deviation of a kernel for textural feature extraction
and grouping achieved by means of kmeans clustering in order to determine
the cloth fell position. Due to the fact that reference line positions remain the
same on the fabric, cloth fell position alterations can be calculated by
measuring the distance between the reference line and the cloth fell. The
conversion of the distance between the cloth fell position and the reference
line, measured by the number of pixels, can be accomplished by using the
known distance between reference lines as a means of calibration.

Journal of the Textile Institute 11/2011; 102(11):905-916. 0.73 Impact Factor

Source
Article: Developing a hybrid intelligent model for constructing a size
recommendation expert system in textile industries
J Shahrabi, E Hadavandi, MS Esfandarani
[Show abstract] [Hide abstract]
ABSTRACT: PurposeIn shopping, for selecting the appropriate garments,
people have to try on multiple garments. This problem is due to lack of a
sizing system based on updated anthropometric data and the classification
system that introduces the appropriate size from the sizing chart to each
person. To solve this problem, as a first study in the literature, a hybrid
intelligent classification model as a size recommendation expert system is
proposed. The paper aims to discuss these issues.
International Journal of Clothing Science and Technology 01/2013; 25(5).
0.46 Impact Factor

Show more

Data provided are for informational purposes only. Although carefully collected,
accuracy cannot be guaranteed. The impact factor represents a rough estimation of
the journal's impact factor and does not reflect the actual current impact factor.
Publisher conditions are provided by RoMEO. Differing provisions from the
publisher's actual policy or licence agreement may be applicable.
Request full-text
Similar Publications

Multiscale texture segmentation via a contourlet contextual hidden Markov

model
Zhiling Long, Nicolas H. Younan

A fast Discrete Wavelet Transform algorithm for visual processing

applications.
Chih-Hsien Hsia, Jing-Ming Guo, Jen-Shiun Chiang

WAVELET-BASED SALIENT ENERGY POINTS FOR UNSUPERVISED TEXTURE

SEGMENTATION

M. K.bashar, N.ohnishi

http://www.researchgate.net/publication/232827808_Identification_of_fabric_d
efects_based_on_discrete_wavelet_transform_and_backpropagation_neural_network