See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322944038

Local neighborhood analysis for fabric defect detection

Conference Paper · August 2017

DOI: 10.1109/ICOMICON.2017.8279095

CITATIONS READS

13 147

3 authors, including:

Maheshwari Shivanand Biradar Kishor Bhangale

Padmashri Dr. D.Y patil institute of technology, Pimpri, Pune Pimpri Chinchwad College Of Engineering and Research
15 PUBLICATIONS 84 CITATIONS 45 PUBLICATIONS 677 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Kishor Bhangale on 25 January 2020.

The user has requested enhancement of the downloaded file.

,(((,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQ&RPPXQLFDWLRQ,QVWUXPHQWDWLRQDQG&RQWURO ,&,&,&ದ 
3DSHU,G

Local Neighborhood Analysis for Fabric Defect

Detection
Namita Kure1, Maheshwari S. Biradar2, Kishor B. Bhangale
Department of Electronics & Telecommunication, Department of Electronics & Telecommunication,
Siddhant College of Engineering, Sudumbare, D. Y. Patil College of Engineering, Akurdi,
Pune, India. Pune, India.

Abstract—This paper analysis the detection of defects in Some techniques used for defect detection are Fourier
fabric based on Local Neighborhood Analysis. In the proposed analysis [8], wavelet transform [9], Gabor filtering [10, 11]
algorithm full image is inspected by moving Local Neighborhood one is to use optimal filter [12, 13] and another is to use filter
Window over it. For the homogeneity measure, Coefficient of bank [13].
variation is used. The value of coefficient of variation is
maximum if there is defect in the fabric. Defected image is Fourier transform gives only frequency information,
segmented by using thresholding. Morphological filtering on limitation of wavelet transform is computational penalty and
defect segmented image is used to enhance defective region. The requires feature extraction algorithm, which is time
computational efficiency increased by using integral image consuming. Gabor filtering requires special algorithm and has
technique. It avoids complicated spectral decomposition. Image drawback of computational complexity.
database include TILDA database and in-house database.
Experiments are performed for defects likes slack end, loose weft, Different types of fabric defect studied are displayed in Fig. 1.
drop stitch, holes, broken end, missing plush loop etc.

Keywords—Local neighborhood analysis, Homogeneity

measure, Morphological filter, Coefficient of variation.

I. INTRODUCTION

Fabric means cloth produced by weaving or knitting
textile fibers. The elements used to produce fabric are wool, (a) Bar (b) Slub (c) Dye mark
cotton or composite of wool and nylon or polyester. Two types
of fabric and they are pattern fabric and non-pattern fabric.
Non-pattern fabric is also known as non-motif based fabric.
The Local neighborhood algorithm is applicable for finding
defects in non-motif based fabric. This technique of fabric
defect detection is non-supervised based. Which does not (d) Holes (e) Pin marks (f) Burst selvedge
requires training of database. Fig. 1. Defects in fabric

A flaw on fabric surface means it is defective fabric and

may leads to loss in revenues [1, 2] and market value of
Here in the paper, we developed defect detection algorithm
company, so automatic fabric defect detection technique is
for fabric in which homogeneity of pixels which are present in
beneficial.
given fabric image is calculated.
In most fabric manufacturing industries visual inspection
The paper contains five sections. Section II includes
methods is used for defect detection in which visual inspection
explanation of proposed methodology. Description of Local
is done by human. Manual inspection demands skilled
neighborhood algorithm is given in section III. Section IV
workers whose performance is inadequate and inefficient.
shows experimental results of defects in fabric. Conclusion
Routine job degrades the accuracy of human visual inspection.
and future scope is discussed in section V.
This results in slow, overprice, unreliable inspection. To solve
the above problem instead of using visual inspection
automated inspection is preferred. The automated system will
upgrade the product quality and raise the productivity, reduce
the labor cost and increase the accuracy of inspection [3, 4].
Machine faults, yarn problem, excessive stretching, poor
finishing leads to defect in fabric [5]. For defect detection in
textile industry image processing is used [6, 7].
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQ&RPPXQLFDWLRQ,QVWUXPHQWDWLRQDQG&RQWURO ,&,&,&ದ 
3DSHU,G
II. PROPOSED METHOD III. DISCRIPTION OF LOCAL NEIGHBORHOOD ALGORITHM
A. The images captured from the camera then processing is Images having defect consist of defect free region and
done on captured image which includes filtering and defective region. The defective region is non homogeneous
histogram equalization. Filtering removes the noise from and defect free region is homogeneous. Homogeneity is the
the image and histogram equalization is used for constrast quality or condition of being all equal or all of the same kind.
enhancement. Features extracted are mean and standard Homogeneity means sameness. The faulty region is unusual
deviation. Coefficient of variation is calculated. To detect region, its presence pixel gray mutation or unequally
the defect adaptive thresholding is used [14,15]. The steps distributed. The distribution of data is shown by coefficient of
for defect detection in fabric is shown in Fig. 2. variation. For measuring homogeneity of pixels coefficient of
variation is used. Small coefficient of variation means less
dispersion. Compare to faulty region a fault free region will
Read fabric Image Mean generate a smaller value of deviations.
image processing calculation I (x, y) is gray image with MൈN size. Where, N specifies
number of rows and M illustrates number of columns. Where
x=1,2,3…..M and y=1,2,3…..N. P(x, y) gives the local
homogeneity measure (LHM) [18,19]. Window size is W × W
centering on the pixel P(x, y), window is calculated by using
Adaptive Calculation Standard
formula W = 2w + 1 for some integer w. Calculate the
threshold of deviation
coefficient ‫ܥ‬௩ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ by using Equation (4),
value calculation
calculation of variation
σ௪ ௪
௜ୀି௪ σ௝ୀି௪ ‫ܫ‬ሺ‫ ݔ‬൅ ݅ǡ ‫ ݕ‬൅ ݆ሻ
ߤ௫ǡ௬ ൌ ሺʹሻ
ܹൈܹ
Defect
detection
Where, ߤ௫ǡ௬ is the mean gray value of pixels which are present
Fig. 2. Steps for fabric defect detection. in the neighborhood window and WൈW is the window size.

B. The coefficint of variation

Coefficient of variation is also noted as dispersion σ௪ ௪
௜ୀି௪ σ௝ୀି௪ ሺ‫ܫ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻ െ ߤ௫ǡ௬ ሻ
ଶ

coefficient. The coefficient of variation is a measure of spread ߪ௫ǡ௬ ൌ ඨ ሺ͵ሻ

ܹൈܹ
which represents the amount of changeability relative to the
mean. The coefficient of variation has no unit. Instead of the
Where, ߪ௫ǡ௬ is standard deviation and size of window is
standard deviation coefficient of variation can be used to
compare the spread of data sets which have different means or represented as WൈW.
different units. To measure homogeneity the coefficient of
variation is used. The homogeneity analysis is used to detect ߪ௫ǡ௬
discontinuities [16, 17]. A fault free region generates a smaller ‫ܥ‬௩ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ ൈ ͳͲͲΨሺͶሻ
value of variation coefficient than that of a damaged region. ߤ௫ǡ௬
The ratio of standard deviation to the mean gives coefficient
of variation. The coefficient of variation is given following Where, ‫ܥ‬௩ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ means the local homogeneity measure value.
Equation (1), Regions with large homogeneity will have small variance.
Variance means quality of being different. For a
ߪ homogeneously texture gray distribution on any local
‫ܥ‬௩ ൌ  ൈ ͳͲͲΨሺͳሻ neighborhood is uniform. Its LHM value will be
ߤ
comparatively small. Applying neighborhood window of size
Here, standard deviation is denoted as ߪ and ߤ means mean. WൈW, the LHM value of each pixel in the given image can
The spread of a data distribution is measured using standard be calculated with the help of (4). In defective image grey
deviation. More the spread out a data distribution higher is distribution is not uniform and LHM value of defective region
standard deviation. Standard deviation is always positive. A is comparatively high. Defect free region is having small value
standard deviation is close to zero if the data points are close of LHM and LHM for defective image is high. Fig. 3 shows a
to mean. The further the data points are from the mean, the fabric image with having the hole defect. The two peaks in
higher the standard deviation. Fig. 4 represents defective region. The defective region shows
greater LHM values of the pixel. LHM value is compared with
threshold of an image. LHM value of a pixel above the
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQ&RPPXQLFDWLRQ,QVWUXPHQWDWLRQDQG&RQWURO ,&,&,&ದ 
3DSHU,G
threshold indicates image has defective region and LHM value IV. EXPERIMENTAL RESULT
below the threshold value reveals image has defect-free The proposed algorithm is implemented the following
region. So to identify the defective area and the defect-free system configuration. The operating system is Windows 10,
area LHM value is discriminating feature. The threshold is Intel (R) Core(TM) i3-2.64GHz Dual Core Processor CPU
given by Equation (5), 4GB 3.10GHz RAM. Software used for implementation of
results is MATLAB R2014a. TILDA Textile Texture
ܶ௖௩ ൌ ߤ௖௩ ൅ ‫ݓ‬ሺͷሻ Database [20] and in-house database is used to get various
defective images of fabric. Experiments are performed by
Where,ܶ௖௩  is threshold value of image, ߤ௖௩ specifies the mean using these defective fabric images. The algorithm is applied
of LHM values of total pixels present in an image, Control to fabric samples with slack end, loose weft, drop stitch, holes,
variable marked as w. Grey or color images are input images broken end and missing plush loop. For various types of
for adaptive thresholding and gives binary images as output defects in fabric, defect detection results are given in
presenting the segmentation. Threshold has to be calculated following Table 1. On basis of detection success rate the
for each pixel in the given image. Adaptive thresholding helps effectiveness of algorithm is measured, also known as cross
to modify the threshold value dynamically over the image. validation accuracy and compared with popular plain fabric
Peaks and valleys of image histogram can help in deciding the defect detection methods like wavelet transform and Gabor
proper value of threshold. Different images have different transform, is given as,
threshold values. It means threshold ܶ௖௩ is not a fixed value.
Control variable takes different values at different object —„‡”‘ˆ•ƒ’Ž‡•
detection. The proposed method gives high accuracy and high ‘””‡ –Ž›†‡–‡ –‡†
speed and also it is insensitive to uneven illumination.  —”ƒ › ൌ  ‫ͲͲͳ כ‬Ψሺ͸ሻ
‘–ƒŽ—„‡”‘ˆ•ƒ’Ž‡•

TABLE 1. Cross Validation Accuracy (%) for different techniques.

Database Cross Validation Accuracy

Wavelet Gabor Local

transform Transform Homogeneity
Analysis

TILDA 93.20% 94.50% 96.40%

Fig. 3. Defective fabric image (in-house database)

The performance of algorithm is also evaluated by

varying the local homogeneity window size (W), smaller value
of W results in detection of defect edges only and larger value
of W smoothen the image which ignores small defects.
Medium size window i.e. 7ൈ7 performs better than 3ൈ3,
5ൈ5, 9ൈ9, 13ൈ13 etc. as shown in Table 2.

TABLE 2. Cross validation accuracy for different Local Homogeneity

window size

Window size (W ൈ W) Cross Validation Accuracy

3ൈ3 80.30%
5ൈ5 86.30%
7ൈ7 96.40%
Fig. 4. Surface plot for distribution of LHM
9ൈ9 84.00%
13 ൈ 13 76.00%
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQ&RPPXQLFDWLRQ,QVWUXPHQWDWLRQDQG&RQWURO ,&,&,&ದ 
3DSHU,G
The experimental results for TILDA data base are shown in V. CONCLUSION & FUTURE SCOPE
Table 3. In this paper, homogeneity of fabric texture is analyzed
TABLE 3- Experimental results
using local neighborhood analysis for defect detection.
Defected region in fabric has larger homogeneity value than
Type of Gray image Histogram Detected
non- defected region. The experimental results on TILDA
defect equalization defect
dataset shows that this method acquires higher cross validation
accuracy than existing algorithms.
This algorithm still faces challenges in detecting defect in
pattern fabric and motif based because of large texture
Holes
variation.

References

Slack 1. P. Sengottuvelan, A. Wahi, A. Shanmugam “Automatic

end fault analysis of textile fabric using imaging sustems,”
Research Journal of Applied Sciences, vol. 3, pp. 26-31,
2008.
2. T. Vikrant, S. Gaurav, “Automatic Fabric Fault
Detection Using Morphological Operations on Bit
Loose Plane,” International Journal of Engineering Research
weft & Technology, vol. 2, pp. 856-861, 2013.
3. Xianghua X (2008), “A review of recent advances in
surface defect detection using texture analysis
techniques,” Electron Lett Comput Vis Image Anal
7(3):1–22
Drop 4. Yadong L, Peihua G (2004), “Free-form surface
stitch inspection techniques state of the art review,” Comput-
Aid Des 36(13):1395–1417
5. Hong Kong Productivity Council, Textile Handbook
2000, The Hong Kong Cotton Spinners Association,
2000.
6. R Drobino, M. S. Machnio, ‘‘Application of image
Holes
analysis technique for textile identification,’’ AUTEX
Research Journal. Vol. 6, No. 1, pp. 40-48, 2006.
7. Y. F. Zhang, R. R. Bresee, ‘‘Fabric defect detection and
classification using image analysis,’’ Textile Research
Journal, Vol. 65, No. 1, pp.1-10. 1995.
8. C. H. Chan and G. Pang, “Fabric defect detection by
Broken
Fourier analysis,” IEEE Trans. Ind. Appl., vol. 36, no. 5,
end
pp. 1267–1276, Sep./Oct. 2000.
9. D.M. Tsai and B. Hsiao, ‘‘Automatic surface inspection
using wavelet reconstruction,’’ Pattern Recognition, vol.
Missing 34, pp. 1285–1305, 2001.
plush 10. Jing L, Tingting D, Dan S et al (2014), “A review on
loop surface defect detection,” J Front Comput Sci Technol,
8(9):1041–1048.
11. M. Li, R. C. Staunton, ‘‘Optimum Gabor filter design
and local binary patterns for texture segmentation,’’
Pattern Recognition Letters, pp. 664-672, 2008.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQ&RPPXQLFDWLRQ,QVWUXPHQWDWLRQDQG&RQWURO ,&,&,&ದ 
3DSHU,G
12. R. Han, L. M. Zhang, “Fabric defect detection method neural network,’’ Journal of photonics, vol 2015, ID
based on Gabor filter mask,” Global Congress on 376163, 2015.
Intelligent Systems, pp. 184-188, 2009. 18. A. Rebhi, S. Abid, F. Fnaiech, “Fabric Defect Detection
13. A. Kumar and G. K. H. Pang, ‘‘Defect detection in Using Local Homogeneity and Morphological Image
textured materials using Gabor filters,’’ IEEE Trans. Ind. Processing,” Image processing, Applications and
Appl, vol. 38, no. 2, pp. 425-440, 2002. Systems (IPAS), 2016.
14. A. Kumar and G. Pang, ‘‘Fabric defect segmentation 19. Chengfei Li, Xinhua Chen, “Surface Defect Detection
using multichannel blob defectors,” Opt. Eng, vol. 39, Algorithm Based on Local Neighborhood Analysis,”
no. 12, pp. 3176-3190, 2000. Springer international Publishing Switzerland,2017.
15. Y. Han, P. Shi, ‘‘An adaptive level-selecting wavelet 20. Workgroup on Texture Analysis of DFG, TILDA textile
transform for texture defect detection,’’ Image and texture database,
Vision Computing, vol. 25, no. 8, pp. 1239-1248, 2007. http://lmb.informatik.uni-freiburg.de/research/dfg-
16. A. Rebhi, S. Abid, F. Fnaiech, “Fabric defect detection texture/tilda.
using line variances of the local homogeneity images,’’
Computer System Science & Engineering, Vol 30, No 4,
pp. 1-11, 2015.
17. A. Rebhi, I.Benmhamed, S. Abid, F. Fnaiech, “Fabric
defect detection using local homogeneity analysis and

View publication stats