III.

TEXTILE DEFECT DETECTION 

Inspection is the process of determining whether a product
has deviated from a given set of specifications. Texture defect (4.2)
detection [8] can be defined as the process of determining the ­ 1 if fi = f8
location and/or extend of a collection of pixels in a textured β11 = ®
image with remarkable deviation in their intensity values or ¯ 0 otherwise (4.3)
spatial arrangement with respect to the background texture.
­ 1 if fi = f 4
3.1 Manual Defect Detection β12 = ®
In the textile industry, inspection [2], is done to assure the
¯ 0 otherwise (4.4)
fabric’s quality before any shipments are sent to customers, γ1 = γ11+ γ12+ γ13 (4.5)
because defects in fabrics can reduce the price of a product by
45% to 65%.Currently, the quality assurance of web processing
­ 1 if fi = f1 = f8
is mainly carried out by manual inspection. However, the γ 11 = ®
reliability of manual inspection is limited by ensuing fatigue
and inattentiveness. Indeed, only about 70% of defects can be ¯ 0 otherwise (4.6)
detected by the most highly trained inspectors. ­ 1 if fi = f 2 = f 4
γ 12 = ®
3.2 Automated Defect Detection ¯ 0 otherwise (4.7)
Textile industries are facing increasing pressure to be more
­ 1 if fi = f5 = f 6
efficient and competitive by reducing costs. Therefore,
γ 13 = ®
¯ 0 otherwise
automated detection of defects in textile fabrics [2], which
results in high-quality products and high-speed production, is (4.8)
definitely needed. ξ = ξ 11 + ξ 12 + ξ 13+ ξ 14 (4.9)
Some of the advantages of automated detection [8] of
­ 1 if fi = f1 = f 2 = f8
defects are: ξ11 = ®
1.100% inspection can be conducted
2.It is a more consistent process when compared with ¯ 0 otherwise (4.10)
manual inspection ­ 1 if fi = f 2 = f3 = f 4
3.It is a non-contact inspection, thus avoiding problems that ξ12 = ®
arise as a result of using some contact inspection devices ¯ 0 otherwise (4.11)
4. It can usually result in
™ Lower labor costs ­ 1 if fi = f 6 = f 7 = f8
™ Improved quality ξ13 = ®
™ Faster inspection ¯ 0 otherwise (4.12)
™ Increased reliability
­ 1 if fi = f 4 = f5 = f 6
IV. MRCSF ξ14 = ®
MRCSF is a combination of first order, second order ¯ 0 otherwise (4.13)
statistical features along with spatial frequency for multi Similarly the other features are also extracted in same
resolution analysis [10, 11]. manner [10].

4.1 MRF Matrix 4.2 Multi resolution Markov Random Field Matrix
MRF Matrix is constructed from the 9 MRF parameters. (MRMRFM)
Nine different MRF parameters are extracted from the eight In fact it is difficult to fit a single MRFM to each texture
neighborhood system [9] shown in Figure 5. pattern, since MRF models are only suitable for the fine
textures [4, 9]. If there is a coarse texture, it is more suitable to
i1 i2 i3 look the texture as a series of MRFs on different scales and
directions. This method can be used in texture classification
i8 i i4 and texture segmentation effectively. However, the important
information in the high-pass filtered part is ignored. In this
section a new approach to model the textures will be proposed.
i7 i6 i5
The sub-bands are down sampled with the discrete wavelet
Figure 5 Pixel i and its eight neighbors in the second order transform [4]. Thus the texture structure represented by the
neighborhood system information of two far away pixels in the original image may
become the one represented by immediate neighbors in the
M = [ β1 , β 2 , β3 , β 4 , γ 1 , γ 2 , γ 3 , γ 4,ξ ] sub-band images on the higher levels [1]. This leads to a new
(4.1)
model which makes use of all sub-bands in different scales and
The procedure for determining the other values is given directions. An original image I is decomposed into a series of
below. sub-bands. If each sub band is modeled as an MRFM, then this
procedure can be called as Multiresolution MRFM
(MRMRFM) modeling [9].

164 2010 IEEE 2nd International Advance Computing Conference

Authorized licensed use limited to: Sona College of Technology. Downloaded on May 20,2010 at 11:48:39 UTC from IEEE Xplore. Restrictions apply.
4.3 Gray- Level Co-occurrence Matrix (GLCM) 5.2 Capturing and Feature extraction of test sample
The gray-level co-occurrence matrix [4] C [i, j] is defined This part comes under the classification stage where the test
by the first specifying a displacement vector d=(dx,dy) and samples are captured using a digital camera which is attached
counting all pairs of pixels separated by d having gray levels i to a shaft which moves over the entire sample. The movement
and j. The gray-level pairs for the specified distance vector d. of the shaft is controlled by embedded system which employs a
The features of mean, standard deviation, contrast, total energy, microcontroller. After capturing the sample images the feature
local homogeneity, cluster shade, cluster prominence are also are extracted in the same way as in the case of original image
defined using the gray-level co-occurrence matrix. and stored in the library.

4.4 Spatial Frequency 5.3 Comparison with Library

Frequency measures the overall information level in the In this stage the stored features of the original image and
regions of an image. The spatial frequency is the root mean the test sample are compared using the nearest neighborhood
square value of row frequency and column frequency [10, 11]. algorithm. The test samples are classified as defective or non-
SF for an M×N block of an image is calculated as follows: defective based on the comparison results.

SF = ( RF )2 + ( CF )2 (4.14)
5.4 Indication of the Defects
The obtained defect is analyzed for its type using the
1 N M available database of defects and hence the defect type is
CF = ¦¦ [ F( m,n ) − F( m − 1,n )] 2
MN n=1 m =2
displayed on the screen. The location of the defect is also
(4.15) displayed on the screen for the ease of the user.
M N
1
RF =
MN
¦¦ [ F( m,n ) − F( m,n − 1 )]
m =1 n = 2
2
VI. RESULTS AND DISCUSSION
(4.16) Gray Scale Fabric database contains 60 monochrome
images from Brodatz album [7] having the size of 512×512
where RF and CF are the row frequency and column
images. It contains 30 reference images and 30 defective
frequency respectively.
images. Training was done for 30 reference images only. The
classification done for all the images in the database. The fabric
V. DEFECT DETECTION SYSTEM patterns are checked whether defective or non defective and
The defect detection system uses computer vision and also determine the location of the defective based on sub
intelligence which controls the other peripherals such as motor, window selection. The sub window sizes are 256×256,
printer, etc. The test fabric is placed in such a way that, it is 128×128 and 64×64. Correct classification rate of 100 % for all
suitable for the camera to capture the image. the size was achieved.
The digital camera is made mobile by using the motor Color Fabric database contains 130 color images having the
which controls the shaft movement which in turn is controlled size of 512×512 images. It contains 65 reference images and 65
by the computer. The camera is interfaced to the computer. The defective images. Training was done for reference images only.
image samples are collected using the Data acquisition toolbox The sub window sizes are 256×256, 128×128 and 64×64.
of the MATLAB software. The collected image samples are Correct classification rate of 100 % was achieved.
then given to the image processing algorithm, which classifies The Real time fabric database contains 100 images. It
the test sample as defective or non-defective after comparing contains 17 real time fabrics and its corresponding 83 defective
the test image features with the database. If a defective sample images. Training was done for 17 reference real time fabric
is detected, the location and the type of defect is printed on the images. The classification rate depends on sub window size. If
screen for the user. the size of the sub window reduces the classification rate also
The entire process can be summarized as follows: reduces. Table 1 shows the classification rate for Textile Fabric
1. Feature Extraction of the original image using MRCSF. using MRCSF.
2. Capturing and Feature extraction of the test sample Table1 Classification rate for Textile Fabric using MRCSF.
3. Comparison with Library
4. Indication of the Defects Type of the No of Sub-window Classification
No
Fabric Images size Rate
5.1 Feature Extraction of original image
256×256 100%
This is the initial task in which the original non-defective 1 Gray Scale 60 128×128 100%
reference samples are collected and their features are extracted 64×64 100%
using appropriate algorithm and stored in a database. Before 256×256 100%
feature extraction the sample images are wavelet transformed 2 Color 130 128×128 100%
so that the samples are localized in both time and frequency. 64×64 100%
The features include Mean, Standard Deviation, Energy, 512×512 100%
256×256 98%
Entropy, Spatial Frequency, MRMRFM, and GLCM 3 Real time 100
128×128 97%
All the above mentioned steps are done using MATLAB 64×64 95%
Image Processing toolbox and Database Toolbox. Figure 6 shows the classification rate of real time fabric.
The rate of correct classification is purely based on size of the
images, when the size of the image is increased the
classification rate also increases. The work has succeeded in
classifying the fabrics with repeated patterns as defective or

2010 IEEE 2nd International Advance Computing Conference 165

Authorized licensed use limited to: Sona College of Technology. Downloaded on May 20,2010 at 11:48:39 UTC from IEEE Xplore. Restrictions apply.