Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009, pp-18-29

A review of automatic fabric defect detection techniques

1 2 3
Mahajan P.M. , Kolhe S.R. and Patil P.M.
1
Department of Electronics and Telecommunication Engineering, J.T.Mahajan College of Engineering, Faizpur, [email protected]
2
Department of Computer Science, North Maharashtra University, Jalgaon, India, E-mail: [email protected]
3
Department of Electronics Engineering, Vishwakarma Institute of Technology, Pune, India, E-mail: [email protected]

Abstract- Quality inspection is an important aspect of modern industrial manufacturing. In textile industry
production, automate fabric inspection is important for maintain the fabric quality. For a long time the fabric
defects inspection process is still carried out with human visual inspection, and thus, insufficient and costly.
Therefore, automatic fabric defect inspection is required to reduce the cost and time waste caused by
defects. The development of fully automated web inspection system requires robust and efficient fabric
defect detection algorithms. The detection of local fabric defects is one of the most intriguing problems in
computer vision. Texture analysis plays an important role in the automated visual inspection of texture
images to detect their defects. Various approaches for fabric defect detection have been proposed in past
and the purpose of this paper is to categorize and describe these algorithms. This paper attempts to present
the survey on fabric defect detection techniques, with a comprehensive list of references to some recent
works. The aim is to review the state-of-the-art techniques for the purposes of visual inspection and decision
making schemes that are able to discriminate the features extracted from normal and defective regions.
Therefore, on the basis of nature of features from the fabric surfaces, the proposed approaches have been
characterized into three categories; statistical, spectral and model-based.

1. Introduction
In textile industry, inspection of fabric defects
plays on important role in the quality control. prevent these defects from reoccurring. There
However, the current inspection task is primarily are many kinds of fabric defects. Much of them
performed by human inspectors and this are caused by machine malfunctions and have
intensive labour cannot always give consistent the orientation along pick direction (broken pick
evaluation of products. Fabric Automatic Visual yarns or missing pick yarns), they tend to be long
Inspection (FAVI) system is an attractive and narrow. Other defects are caused by faulty
alternative to human vision inspection. Based on yarns or machine spoils. Slubs are often
advances in computer technology, image appeared as point defects; machine oil spoils are
processing and pattern recognition, FAVI system often along with the direction along the warp
can provide reliable, objective and stable direction, and they are wide and irregular. An
performance on fabric defects inspection. A good automated defect detection and identification
automated system means lower labor cost [1] system enhances the product quality and results
and shorter production time [2]. There are in improved productivity to meet both customer
numerous reported works in the past two needs and to reduce the costs associated with
decades during which computer vision based off-quality. Recently, the fault detection is done
inspection has become one of the most important manually after a sufficient amount of fabric has
application areas. The texture materials can be been produced, removed from the production
further divided into uniform, random or patterned machine and then batched into larger rolls and
textures. Brazakovic et al. [3] have detailed a then sent to the inspection frame. An optimal
model based approach for the inspection of solution for this would be to automatically inspect
random textured materials. The problem of from the fabric as it is being produced and to alert
printed textures (e.g. wall paper scanning, the maintenance personnel when the machine
ceramic flaw detection and printed fabric needs attention to prevent production of defects
detection) requires evaluation of color uniformity or to change process parameters to prevent
[4] and consistency of printed patterns. Ngan et automatically to improve product quality. This is
al. [5] have introduced the new regular bands done by identifying the faults in fabric using the
(RB) methods which is effective approach for image processing techniques and then based on
pattern texture inspection. This paper focuses on the dimension of the faults; the fabric is classified
the inspection of uniform textured materials and and then graded accordingly. Nickoloy et al. [8]
presents a survey on the available techniques for have shown that the investment in the automated
the inspection of fabric defects. fabric inspection system is economically
attractive when reduction in personnel cost and
2. Fabric defects associated benefits are considered.
Fabric faults or defects are responsible for nearly
85% of the defects found in the garment industry 2.1 Texture analysis techniques for fabric
[6]. Manufactures recover only 45-65% of their defect inspection
profit from second or off quality goods [7]. It is Texture is one of the most important
imperative therefore to detect, to identify and to characteristics in identifying defects or flaws.

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
 A review of automatic fabric defect detection techniques

Figure1.shows some example of defects in texture features extracted from fractal

various fabric materials. It provides important dimensions, first order statistics, cross
information for recognition and interpolation. In correlation, edged detection, morphological
facts, the task of detecting defects has been operations, co-occurrence matrix, eigenfilters,
largely viewed as a texture analysis problem. rank order functions, and many local linear
With reference to several texture analysis survey transforms have been categorized into this class.
papers [9-14], we categorized texture analysis
techniques use for visual inspection into four 3.1 Defect detection using co-occurrence
ways: statistical approaches, structural matrix features
approaches, filter based approaches and model The co-occurrence matrix method, known also as
based approaches. A variety of techniques for the spatial gray-level dependence method, has
describing image texture have been proposed in been widely used in texture analysis. It is based
the research literature. M. Tuceryan and Jain in repeated occurrences of different grey level
[15], on the other hand, defined five major configurations in a texture. Automatic visual
categories of features for texture analysis: inspection techniques for textured images
statistical, geometrical, structural, model based generally compute a set of textural features in the
and signal processing features. Geometrical and spatial domain or in the spectral domain. Siew
structural methods try to describe the primitives and Hogdson [16] presented the assessment of
and the rules governing their special organization carpet wear using Spatial Gray Level
by considering texture to be composed of texture Dependence Matrix (SGLDM), neighboring Gray
primitives .These two approaches have not been Level Dependence Matrix (GLDM), Gray Level
attempted in fabric defect detection, mainly due Difference method (GLD), and the Gray Level
to the stochastic variations in the fabric structures Run Length Method. Also it has been applied to
(due to elasticity of yarns, fabric motion, fiber wood inspection [17], surface defect detection
heap, noise etc.) which poses severe problems in [18], and fabric defect detection [19]. The original
the extraction of texture primitives from the real investigation into SGLDM features was pioneered
fabric samples. Therefore, in this paper the by Haralick et.al. [20]. Texture features, such as
proposed defect detection techniques have been energy, entropy, contrast, homogeneity, and
classified in three categories: statistical, spectral correlation are then derived from the co-
and model-based. Table 1 shows a summary list occurrence matrix. However only six of such
of some of the key texture analysis method that features have been used for the defect detection
have been applied to defect detection. on wood and fabric defect detection has been
shown with only two of these six features.
Several works have been reported using co-
occurrence matrices to detect defects, such as
[17, 19, 21, 22, 23] for example in [21] Iivarinen
et al. applied co-occurrence texture features to
detecting defects in paper web where the normal
texture have characteristic frequency. Conners et
(a) (b) (c) (d) al. [17] have used six features of co-occurrence
Fig. 1- Fabric defect samples: (a) Double yarn; matrix, to identify nine different kinds of surface
(b) Missing yarn; (c) Broken yarn; (d) Variation of defect in wood. Tsai et al. [19] have detailed
yarn fabric defect detection while using only two
features, and achieved a classification rate as
3. Statistical approaches high as 96%. Rosler [22] has also developed a
Statistical texture analysis methods measure the real fabric defect detection system, using co-
spatial distribution of pixel values. An important occurrence matrix features, which can maintain
assumption in this approach is that the statistics 95% of the defects as small as 1mm2 in size. In
of defects free regions are stationary, and these order to derive maximum texture information
regions extend over a significant portion of using co-occurrence matrix, the values parameter
inspection images. Statistical methods can be ө should agree with the orientation of the fabric
classified into first- order (one pixel), second- pattern and distance d should be equal to the
order (two pixels) and higher- order (three or pattern period [24]. Bodnarova el al. [25] have
more pixels) statistics based on a number of examined this issue on the optimal displacement
pixels defining the local features. The first -order vector d for the fabric defect detection. The co-
statistics estimate properties like the average and occurrence matrix features suffer from a number
variance of individual pixel values, ignoring the of difficulties. It appears there is no generally
spatial interaction between image pixels, second- accepted solution for optimising the displacement
and higher order statistics on the other hand vector [4, 26]. The number of gray levels is
estimate properties of two or more pixel values usually reduced in order to keep the size of the
occurring at specific locations relative to each co occurrence matrix manageable. For a given
other. The defect detection methods employing displacement vector , a large number of features

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 19

 Mahajan PM, Kolhe SR and Patil PM

can be computed, which implies dedicated limited experimental results which suggest 96%
feature selection procedure .This technique can detection on eight types of defects. The
be computationally expensive for the demands of localization accuracy of these defected defects is
a real-time defect inspection system , but it have very poor and high false alarm.
been exploited in many studies as highly
accurate technique. 3.4 Defect detection using edge detection
Edges can be detected either as micro edges
3.2 Defect detection using local linear using small edge operator masks or as macro
transforms edges using large masks [37]. The distribution of
Texture properties can be extracted by using amount of edges per unit area is an important
several bidimensional transform such as Discrete feature in the textured images. The amount of
Cosine Transform (DCT), Discrete Sine gray level transitions in the fabric image can
Transform (DST), Discrete Hadamard Transform represents line, edges, spots, ripples and other
(DHT), Karhunen –Loeve Transform (KLT) and spatial discontinuities. These features have been
eigenfiltering. Unser [27] tested different local used to detect fabric defects [38, 39, 40 ]. Conci
linear transforms for texture classification and and Proenca [39] have used sobel edge
found KLT as the best algorithm. Also Ade et al. detection to detect fabric defects and compared
[28] compared Laws filters, KLT, DCT and DHT the results with those based on thresholding and
for textile defect detection. In their experiments, fractal dimension. J.S.Lane [40] has detailed a
the KLT performance, particularly on larger systematic approach to detect fabric defect. [41]
window size, was amongst the best. Neubauer gives useful approach for the characterization of
[29] has detected fabric defect using texture low resolution web surface using facet model.
energy features from low mask on 10×10 These approaches using edge detection are
windows of inspection images. In his approach, suitable for plain weave fabrics imaged at low-
three 5x5 Laws masks corresponding to ripple, resolution.
edge, and weave features [30] are used to
extract histogram features from every window of 3.5 Defect detection using cross-correlation
the image. These features are then used for the Correlation is used for locating features in one
classification of the corresponding window into image that appear in another and therefore
defect-free of defect class, using a three-layer provides a direct and accurate measure of
neural network. Using eigenvalues of covariance similarity between two images. Any significant
matrix as a feature Őzdemir and Ercil [31] have variation in the values of resulting measure
implemented fabric defect detection using an indicates the presence of a defect. Bodnarova et
approach which is a variation of the Karhunen- al. [42] have used the correlation coefficient from
Loeve (KL) transform or eigenfilters method. A multiple templates to generate a correlation map
novel scheme of characterizing and classifying for defect declaration. The correlation approach
defects in woven textile fabrics has been in [43] yields satisfactory results when detecting
attempted in [32]. This back propagation based imperfections in regularly textured backgrounds.
neural network coupled with the DCT technique On the other hand, randomly textured
can lead to outstanding results for classification backgrounds do not correlate well and
of various fabric defects. In online fabric demonstrate a limitation of this approach.
inspection, the local transforms such as DCT or
DST could be preferable to eigenfilters or KL 3.6 Defect detection using bi-level
transforms, since DCT or DST can be directly thresholding
obtained from the camera hardware using Use of gray level thresholding enables to detect
commercially available chips that perform fast high contrast defects. The occurrence of a defect
and efficient DCT or DST transforms. causes the signal level to rise or fall locally; the
presence of a peak or trough then indicates a
3.3 Defect detection using fractal dimension defect. This defect is detected when the signal
(fd) crosses a decision threshold. This idea is used to
Fractal-based texture analysis was introduced by detect fabric defects [44, 45] on moving textile
Pentland [33]. Voss [34] refers to box counting as web. The defect detection can be effective even
the process of estimating the probability that m when web is covered by a fine and complex
points lie in the box. Keller et al. [35] proposed a pattern. Cho et al. [46] proposed algorithm for
modification of method due to Voss, which finding defect in textile fabrics with fine web
presents satisfactory results up to FD = 2.75. The surface which shows 80% recognition rate on
utilization of fractal dimension is investigated by warp and pick float. The fabric inspection system
Conci and Proenca [36] for discriminating that uses thresholding, proposed by Stojanovic et
defective areas. The decision for defect al. [47], gives high detection rate with good
declaration is based on the variation of FD. This localization accuracy and low rate of false alarm.
method is infact computationally enough to be
suitable for PC implementation, but presents very

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
20
 A review of automatic fabric defect detection techniques

3.7 Defect detection using morphological result shows very high accuracy rate of
operations recognition and thus provides decision support in
Zhang and Bresee [48] have detailed on defect classification. M. Shi et.al. [57] describes
morphological operations for detection of fabric an adaptive image-segmentation method based
defects. The practical utility of this approach is on a simplified pulse-coupled neural network
limited as most of the commonly occurring fabric (PCNN) for detecting fabric defects. They
defects will be missing from the binary image introduce a new parameter called the deviation of
generated from the simple thresholding the contrast (DOC) to describe the contrast
operation. Detecting defects morphologically on difference in row and column between the
spatially filtered images of fabrics produces better analyzed image and a defect – free image of the
results [49], particularly when the fabric is fine same fabric. The simplification of PCNN reduces
and contains defect of small size. In their the number of the network parameters by utilizing
experiments the morphological operations are the local and global DOC information for the
only performed on a periodic images defect, parameter selections. Castilho et al. [58] presents
unlike the case in [48] where the entire structure a real-time fabric defect detection based
of thresholded fabric image was utilized. intelligent techniques. Neural networks (NN),
fuzzy modeling (FM) based on product – space
3.8 Defect detection using neural networks fuzzy clustering and adaptive network based
Neural networks are one of the fastest most fuzzy inference system (ANFIS) were used to
flexible classifier used for fault detection due to obtain a clearly classification for defect detection.
their non-parametric nature and ability to Experimental results for real fabric defect
describe complex decision regions. A new detection, shows the usefulness of the three
approach for the segmentation of local textile intelligent techniques and they further stated that
defects using feed-forward neural network (FFN) NN has a faster performance. Online
and also a new low-cost solution for the task web implementation of the algorithms showed they
inspection using linear neural network is can be easily implemented and may be adapted
presented in [50]. The usefulness of these two to industrial applications without great efforts.
proposed approaches is shown by experimental Recently, another method of textile flaw detection
results obtained from the real fabric defects. and classification based on wavelet
Hung and Chen [51] have used back propagation reconstruction and BP neural network is detailed
neural network, with fuzzy logic, to achieve the in [59]. They detected two types of textile flaws,
classification of eight different kinds of fabric namely oil stain and hole. For the extraction of
defects along with defect-free fabric. A compact flaw features, histograms of hole and oil stain are
fabric inspection system using neural network is computed as the input of BP neural network.
presented in [52] but is not adequately detailed. Their experimental result shows that this method
Recently, H.M.Elragal [53] proposes an can effectively detect defects and classify the
automated visual inspection system using types of defects with high recognition correct
Adaptive Neural Fuzzy Interface System (ANFIS) rate.
that can detect and classify knitting machine
fabric defects. [54] Investigates two methods for 3.9 Defect detection using histogram
the detection of defects on textured surfaces Histogram and the rank function provide exactly
using neural networks and Support Vector the same information. Natale [60] has used rank
Machines (SVM). The real-time implementation order functions for the detection of artificially
of defect segmentation scheme using FFN is introduced defect in some Brodatz textures [61].
computationally costly. Although the real time H. Kauppines [62] detailed the parquet slab
computational complexity of SVM is also similar, grading using cumulative histogram. Defect
but do not suffer from the problem of local detection in ceramic tile using a set of adaptive
minimum and is computationally simple to train. rank order function is discussed in [63]. The
Another analogous work for texture defect colour information in textured images can also be
detection using cellular neural networks is used to extract colour histograms and this has
detailed in L.Occhipinti [55]. The FFN and SVM been used in [64, 65] to detect defects.
require training from the known classes of fabric
defects. Yuen et al. [56], presented a novel 4. Spectral approaches
hybrid model through integration of genetic These are robust and efficient computer-vision
algorithm (GA) and neural network to classify the approaches for fabric defect detection. In this
type of garment defects. They developed a approaches texture is characterised by texture
segmented window technique to segment images primitives or texture elements, and the spatial
into several classes using monochrome single- arrangement of these primitives [66] .Thus, the
loop ribwork of knitted garment. Four primary goals of these approaches are firstly to
characteristic variables were collected and input extract texture primitives, and secondly to model
into a back propagation (BP) neural network to or generalise the spatial placement rules. The
classify the sample images. Their experimental high degree of periodicity of basic texture

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 21

 Mahajan PM, Kolhe SR and Patil PM

primitives, such as yarns in the case of textile lenses and spatial filters. The fabric defect
fabric, permits the usage of spectral features for detection system using the measurements of the
the detection of defects. However, random first- and the zero-order intensities have been
textured images cannot be described in terms of developed [78, 79, 80, and 81]. Ciamberlini et al.
primitives and displacement rules as the [82] have described the design of spatial filters: a
distribution of gray levels in such images is rather fixed filter adaptable for different types of fabric
stochastic. Therefore, spectral approaches are and a universal spatial filter for the detection of
not suitable for the detection of defects in random defects in textured materials. Campbell and
texture materials. Various approaches for the Murtagh [83] have detailed a Windowed Fourier
detection of defects in uniform textured material transform based method to detect defect on
using frequency and spatial- frequency domain denim fabric samples.
features have been reported in the literature. In
spectral- domain approaches, the texture 4.2 Defect detection using gabor filter
features are generally derived from the Fourier The Fourier transform is an analysis of the global
transform [67, 68], Gabor transform [69, 70] and frequency content in the signal, it is not able to
Wavelet transform [71]. They are summarized in localise the defective regions in the spatial
the following sections. dependency into Fourier analysis is through the
windowed Fourier transform. If the window
4.1 Defect detection using discrete fourier function is Gaussian, the windowed Fourier
transform transform becomes the well known Gabor
The Fourier transform (FT) has the desirable transform, which can be achieving optimal
properties of noise immunity and enhancement of localisation in the spatial and frequency domain
periodic features. The FT characterizes the [84]. Jain and Farrokhnia [85] used it in
textured image in terms of frequency segmentation and classification of textures with
components. The periodically occurring features dyadic coverage of the radial spatial frequency
can be observed from the magnitude of range. The Gabor filter bank has been
frequency components. These global texture extensively studied in visual inspection. Kumar
patterns are easily distinguishable as and Pang [86] perform fabric defect detection
concentration of high-energy bursts in the using only real Gabor functions. Later in [87],
spectrum. Liu and Jernigan [72] reviewed a set of they used a class of self similar Gabor functions
28 textural features extracted in the Fourier to classify fabric defects. They also investigated
spectrum for texture analysis. Escofet et al. [73] defect detection using only imaginary Gabor
used the angular correlation of the Fourier functions as an edge detector. Bodnarova et al.
spectra to evaluate fabric web resistance to [88] applied a Fisher cost function to select a
abrasion. Chan and Pang [74] used the Fourier subset of Gabor functions based on the mean
analysis for fabric defect detection. Seven and standard deviation of the template feature
textural features extracted from the vertical and images to perform textile flaw detection. Shu et
horizontal frequency components in the Fourier al. [89] detailed a method of detecting the fabric
spectrum are used to discriminate four defect defects automatically based on multi-channel and
types including double yarn, missing yarn, webs multi-scale Gabor filtering. It is based on energy
and yarn densities. Later, in [75], an approach response from the convolution of Gabor filter
based Fourier transform has been used to detect banks in different frequency and orientation
the various types of fabric defects. The central domains. Experiments on various simulated
spatial frequency spectrum is used, from which defect fabric images have shown the
seven significant characteristic parameters are effectiveness of this method. This method has
extracted for detecting the type of defect. Further, accurate location and fine detection of fabric
they carried out experiments to detect only two defects. Han and Zhang [90] proposed a fabric
classes of defects namely double yarn and defect detection method based on Gabor filter
missing yarn which found to be consistent for a masks. The performance is evaluated off-line by
number of samples. In [76], the author used the using a group of fabric sample images containing
Fourier transform to reconstruct textile images for many kinds of fabric defects. Their experimental
the defect detection. The line patterns in the result shows accurate defect detection with low
textile images, supposed to be defects, were false alarms. In [91], Gabor filters are designed
taken out by removing high energy frequency on the basis of the texture features extracted
components in the Fourier domain using a one- optimally from a non-defective fabric image by
dimensional Hough transform. The difference using a Gabor wavelet network (GWN). Their
between the restored image and the original result exhibits accurate defect detection with low
image were considered as potential defects. A false alarms, showing the effectiveness and
similar idea was explored in [77], but low pass robustness of the scheme. Gabor wavelet
filtering was used to remove the periodic transform is applied to detect the defects in
information. The Fourier transform of textile fabric fabrics [92]. Gabor filter scheme that imitates the
can also be obtained in optical domain by using early human vision process is applied to the

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
22
 A review of automatic fabric defect detection techniques

sample under construction. The result obtained images without defects, where 98.2% detection
by proper thresholding ensures segmentation of rate and 1.5% false alarm rate were achieved in
the defect, which in turn confirms efficiency of defect detection. The detection of fabric defects
this method. Hou and Parker [93] investigate a using wavelet packet decomposition and
method for detecting defects on textured surfaces Independent Component Analysis has been
using a Support Vector Machines (SVM) investigated in [103]. Kumar and Gupta [104]
classification approach with Gabor wavelet have used mean and variance of “Haar” wavelet
features. Instead of using all the filters in the coefficient for the identification of surface defects.
Gabor wavelets, an adaptive filter selection The fabric texture can also be considered as
scheme is applied to reduce the computational noise and removed using wavelet shrinkage.
cost on feature extraction while keeping a Recently, Truchetet and Laligant [105] gave a
reasonable detection rate. Their experimental detailed review on wavelet analysis in industrial
result shows, this method can successfully detect application. On the basis of wavelet and singular
and segment defects in texture images. signal characteristic analysis, Guan et.al. [106]
presented a new defect detection method based
4.3 Defect detection using wavelet transform on wavelet characteristics. The detail signal
In the recent past, multiresolution decomposition feature after wavelet decomposition of fabric
schemes based on wavelet transform have image is extracted, and it is compared with the
received considerable attention as alternatives detail signal feature of normal fabric image
for the extraction of textural features. The decomposition to determine fabric defects. Their
multiresolution wavelet representation allows an experimental result shows the defect detection
image to be decomposed into a hierarchy of accuracy is over 92.5%. An extracted sub-image
localized sub images at different spatial features approach based on wavelet transition
frequencies. It divides the 2D frequency spectrum with one resolution level and Fourier transform is
of an image into a low pass (smooth) sub image presented in [107]. By using restoration scheme
and a set of high pass (detail) sub images. The of the Fourier transform, the normal fabric
textural features are then extracted from the textures of smooth sub-image in the spatial
decomposed sub images in different frequency domain are removed by detecting the high-
channels and different resolution levels. In [94], energy frequency components of sub-image in
Sari-Sarraf and Goddard have developed a fabric the Fourier domain, setting them to zero using
defect detection system that can detect small frequency-domain filter, and back-transforming to
defects with an overall detection rate of 89%. a spatial domain sub-image. Then, the smooth
Their defect detection scheme uses the low-pass and detail sub-images are segmented into many
and the high-pass ‘Daubechies’ D2 filter [95]. sub-windows, in which standard deviation are
Scharcanski [96] used the discrete wavelet calculated as extracted features. These extracted
transform to classify stochastic textile texture. features are compared with normal sub-windows
Rather using fixed scales, Kim et al. [97] features to determine whether there exists defect.
employed a learning process to choose the Ngan et.al. [108] have developed the method of
wavelet scales for maximising the defect ability of wavelet preprocesses golden image substraction
fabric defects. Latif-Amet et al. [98] extracted co- (WGIS) for defect detection on patterned fabric.
occurrence and MRF-based features from They concluded that the WGIS method provides
wavelet transform coefficients for fabric defect the best detection result. The overall detection
detection. Gray level diference-based features success rate is 96.7% with 30 defect-free images
from sub bands of the wavelet transform were and 30 defective patterned images for one
also applied in classifying fabric defects. Jasper common kind of patterned Jacquard fabric.
et al. [99, 100] have detailed the design of a
texture specific wavelet basis filter, which can be 5. Model-based approaches
tuned to a particular texture. The design of Model - based texture analysis methods are
adaptive orthonormal wavelet bases has been based on the construction of an image model that
shown [101] to achieve the best performance in can be used not only to describe texture, but also
the characterization of fabric defects. Later, in to synthesize it. Model-based approaches are
[102], the authors detailed the adaptive wavelet – particularly suitable for fabric images with
based methodology from the use of a single stochastic surface variations (possibly due to
adaptive wavelet to multiple adaptive wavelets. fiber heap or noise) or for randomly textured
For each class of fabric defect, a defect – specific fabrics for which the statistical and spectral
adaptive wavelet was designed to enhance the approaches have not yet shown their utility. The
defect region at one channel of the wavelet model parameters capture the essential
transform, where the defect region can be perceived qualities of texture. Markov random
detected by using a simple threshold classifier. fields (MRF) have been popular for modeling
This multiple adaptive wavelets method has been images. MRF theory provides a convenient and
evaluated on the inspection of 56 images consistent way for modelling context dependent
containing eight classes of fabric defects, and 64 entities such as pixels, through characterising

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 23

 Mahajan PM, Kolhe SR and Patil PM

mutual influences among such entities using therefore suggested for its practical usage.
condition MRF distribution [109]. Several Ozdemir et al. [116] compared six texture
probabilistic models of the textures have been features, consisting of MRF, KL transform, 2D
proposed and used for defect detection. In [110], lattice filters, Laws filters, Co-occurrence
Cohen et al. used Gaussian Markov Random matrices, and a FFT - based method, for
Fields (GRMF) to model defect free textile web. detecting textile defects. For each method, the
The inspection process was treated as a effects of various parameters have been
hypothesis testing problem on the statistics examined and concluded that, although many of
derived from the GMRF model. The images of the methods gave promising results, texture
fabric to be inspected are divided into small modeling using the 9th order Markov Random
windows in inspection process; a likelihood ratio Field model gave the best results. Also, by
test is then used to classify the windows as non– considering the results obtained with respect to
defective or defective. The testing image was speed and reliability, MRF approach seems
partitioned into non-overlapping sub-blocks feasible for a real-time factory implementation.
where each window was then classified as Also Bodnarova et al. [117] have concluded that
defective or non-defective. Baykut et al. [111] the optimal Gabor filters (optimized to detect five
implemented this method in a real-time types of defects) perform better than gray level
application with a dedicated DSP system. In co-occurrence matrix, correlation or FFT based
[112], the authors showed that MRF based approaches. However this comparison is very
methods were competitive in a comparative study limited on a set of 25 images and the information
against other statistical and spectral based about the image resolution is also missing. Lee
methods in defect detection. Brzakovic et al. Tin Chi [118], compares the performance of three
[113, 114] discuss a theoretical approach based methods which utilize matched masks, wavelet
on a poissonian model for inspection of web transform and neural network for fabric defect
materials. The inspection objective is to quantify detection. An evaluation of the performance of
the randomness and homogeneity across the the methods was conducted on eight classes of
material. Campbell et al. [115] detects an fabric defects (Broken End, Dirty Yarn, Mispick,
alignment pattern in preprocessed images via Netting Multiples, Stack End, Thick Bar, and
model based clustering and uses an approximate Wrong Draw). In the first method, a multichannel
Bayes factor to assess the evidence for the filtering bank equipped with five matched mask
presence of a defect. was used. Matched masks are 2-D filters that
characterize specific texture properties. Using
6. Comparative studies this method, 96% of fabric defects were
A classification of fabric defect detection successfully detected, and the false alarm rate
techniques is shown in Table 1. The statistical was 6%. The second method employed wavelet
and structural approaches have been in favour in transform to decompose fabric images into multi-
terms of the amount of research reported. It is scales and orientations. During the training stage,
also worth noting that the categorisation of the parameters to be optimized include the
texture analysis techniques used for fabric defect rotation angles and the two thresholds applied on
detection as describe above and listed in table 1 the horizontal and vertical transformed images.
is not a crisp classification. There are several Using this method, only 76% of fabric defects
comparative studies in the literature that evaluate were identified, and the false alarm rate was 7%.
texture analysis methods in applications to fabric The last method took advantage of the fault
defect detection. It must be noted that different tolerance and learning ability of neural networks.
studies use different datasets and possibly They explored the texture structure of defect-free
different parameter settings. Also resolution of images so that feature extraction was conducted
the acquired images is an important factor in on repeating units with proper selection of
selecting the suitability of an approach for the locations. For defect images, similar feature
defect detection. Therefore comments / vectors were extracted and passed to the neural
conclusions on the suitability of some network. Using this method, the detection rate
approaches, recently cited in the literature, based was as high as 92% and the false alarm rate was
on the image resolution, computational 6%. They further concluded that, the method
complexity, and performance would be useful. employing matched masks proved the most
The approaches discussed in this survey have effective in detecting fabric defects. The neural
been evaluated on image sample with various network method was next best. The wavelet
resolutions. The high resolution images are transform method was the least effective,
highly suitable for detecting defects with very because it was only able to detect effectively
subtle intensity variations, but their use will certain classes of fabric defects. Comparative
require a high volume of online computations for studies performed by Randen et.al. [119, 120]
unsupervised defect detection. However, and Chen et al. [121] indicate that the Gabor
supervised defect detection on these high- features in most of the cases outperform the
resolution images is a real possibility and is other methods regarding the complexity and

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
24
 A review of automatic fabric defect detection techniques

overall error rate. But the Gabor features suffer need of some standard datasets in order to carry
from a number of difficulties. A major difficulty of out fair comparative analysis.
this method is how to determine the number of
Gabor channels at the same radial frequency and 8. References
the size of the Gabor filter window in the [1] Chin R. T. and Harlow C. A. (1982) IEEE
application. Although a solid conclusion cannot Trans .Pattern Anal. Machine Intell.,
be drawn to determine the best method for defect 4(6), 557-573.
detection, it is clearly evident that filtering [2] Smith M. L. and Stamp R. J. (2000)
approaches, in particular Gabor filtering has been Computer Ind. 43,73-82.
more popularly applied in these areas. The [3] Chin R.T. (1987) Computer Vision,
recent texture inspection approach by Graphics, and Image Processing, 41(3),
Chetverikov and Henbury [122] using the 346-381.
measure of structural regularity and texture [4] Yang X. Pang G., and Yung N (2005) IEE
antisotropy gives quite convincing experimental Proceedings Vision, Image Processing,
results. Therefore, a combination of these two 152(6) 715-723.
approaches can offer the best performance for [5] Ngan H.Y.T. and Pang G.K.H. (2009) IEEE
textile web inspection and is suggested for future Trans. on Automation Science and
investigation and comparison. Engineering, 6(1).
[6] Sengottuvelan P., Wahi A. and Shanmugam
7. Conclusion A. (2008) Research Journal of Applied
The review of fabric defect detection Sciences, 3(1) pp 26-31.
methodologies using image processing [7] Srinivasan K., Dastor P. H.,
techniques gives us possible trend of this Radhakrishnaihan P., and Jayaraman
application area. These available techniques S.(1992) J. text. Inst., vol. 83, no. 3, pp.
were classified into three categories: statistical, 431-447.
spectral and model-based. Although the research [8] Nickolay B. N. and Schmalfub H. (1993)
on visual inspection is diverse and ever- Mellind Textilberichte, vol. 73, 33-37.
changing, the following observations can be [9] Harlick R. (1979) Proceedings of IEEE,
made. 67(5) 786-804.
1. The core ides of these methodologies along [10] Wechsler H. (1980) Signal processing, 2
with their drawbacks / critics were discussed 271-282.
whenever known. [11] Van Gool L., Dewaele P., and Oosterlinck
2. Filter bank based methods have been very A. (1985) Computer Vision, Graphics
popular in fabric defect detection. The filters can and Image Processing, 29, 336-
be manipulated and designed in all sorts of 357.
directions and scales to decompose textures in [12] Vilnrotter F., Nevatia R., and Price K. (1986)
order to highlight defects. However, it is notable IEEE Transactions on Pattern Analysis
that recent researches suggest contextual and Machine Intelligence, 8,76-89.
analysis which directly based on local [13] Reed T. and Buf J. (1993) Computer
neighbourhoods without dedicated filtering is a Vision, Image Processing and Graphics,
promising alternative approach. 57(3),359-372.
3. In order to understand the formation and [14] Kumar A. (Jan-2008) IEEE Trans. of
nature of the defects, it is important to be able to Industrial Electronics, 55(1).
accurately localise the defective regions rather [15] Tuceryan M. and Jain A. (1998) In
than classifying the surface as a whole. This can Handbook of Pattern recognition and
provide possibilities of classifying the defects and computer vision, 235-276.
further studies of the characteristics of the [16] Siew L. H., Hodgson R. M., and Wood E. J.
defects. (1988) IEEE Trans. Patt. Anal .Machine
4. Due to lack of uniformity in the image Intell., 10(6), 92-105.
database, performance evaluation and the nature [17] Conners R. W., McMillan C. W., Lin K., and
of intended application, is not prudent to explicitly Vasquez-Espinosa R. E. (May 1980)
declare the best available methods for fabric IEEE Trans .Patt. Anal. Machine Intell.,
defect detection. Therefore, the effective 3(6),204-222.
performance evaluation requires careful selection [18] Iivarinen J. (2000) Proceedings of SPIE
of data sets along with its clear definition of 4197, 140-145.
scope. [19] Tsai I., Lin C., and Lin J. (Mar 1995) Text.
5. The statistical, spectral and model – based Res. J., 65, 123-130.
approaches gives different results and hence the [20] Harlick R. M., Shanmugam K. and Dinstein
combination of the approaches can give better I. (1973) IEEE trans. Sys. Man Cybern.,
results, than either one individually and is 3, 610-621.
suggested for future research. There is also a

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 25

 Mahajan PM, Kolhe SR and Patil PM

[21] Iivarinen J., Rauhamaa J., and Visa A., [43] Bennamoun M. and Bodnarova A. (1998)
(1996) In International Conference on Proc. IEEE Conf. SMC, 4340-4343.
Pattern Recognition, 4, 356–360. [44] Norton-Wayne L., Bradshaw M., and Jewell
[22] Rosler R. N. U. (1992) Mellind A. J. (1992) Proc. British Machine Vision
Texilberichte, 73, 292. Conf., Leeds (U.K), 217-226.
[23] Amet A. L., Ertüzün A., and Erçil A. (1998) [45] Norton-Wayne L., Bradshaw M., and
Proc. IEEE Southwest Symposium on Sandby C. (1993) Proc. SPIE 1989, 2-
Image Analysis and Interpretation, 205- 13.
210. [46] Cho C.S., Chung B. M. and M-J. (Aug.
[24] Zucker S. W., and Terzopoulos D. (1990) 2005) IEEE Trans. Industrial
Image Modelling, A. Rosenfield (Ed.), Electronics, 52(4).
423-445. [47] Stojanovic R., Mitropulos P., Koulamas C.,
[25] Bodnarova A., Williams J. A., Bennamoun Karayiannis Y., Koubias S., and
M., and Kubik K. K. (1997) Proc. IEEE Papadopoulos G. (2001) Real-Time
TENCON’97 Conf., Brisbane (Aus.), Imaging, 7, 507-518.
307-310. [48] Zhang Y. F. and Bresee R. R. (1995) Text.
[26] Monadjemi A. (2004) PhD thesis, Res. J., 65, 1-9.
University of Bristol, UK. [49] Mallick-Goswami B. and Datta A. K. (2000)
[27] Unser M. (1986) Signal Processing, 11,61- Text. Res. J., 70, 758-762.
79. [50] Kumar A. (2003) Pattern Recognition, 36,
[28] Ade F., Lins N. and Unser M. (1984) In 1645-1659.
ICPR, 1, 428-431. [51] Hung C. C. and Chen I. C. (2001) Text.
[29] Neubauer Claus. (1992) Proc. 11th Intl. Res. J., 71(3), 220- 224.
Conf. Pattern Recognition, The Hague [52] Rohrmus D. (Jun. 2000) Proc. IEEE Conf.
(The Netherlands), 688-691. Computer Vision & Pattern Recogn., 2,
[30] David E. R. (1997) Machine Vision: Theory, 794-795.
Algorithms, and Practicalities, 2nd ed. [53] Elragal H. M. (2006) Proc.of 23rd National
New York , 569-572. Radio Science Conference.
[31] Őzdemir S. and Erçil A. (1996) Proc. IEEE [54] Kumar A. and Shen H. C. (2002) Proc. Intl.
Conf. Emerging Technologies and Conf. Image Process., ICIP-2002, 353-
Factory Automation, EFTA’96, 2, 697- 356,.
703. [55] Occhipinti L., Spoto G., Branciforte M., and
[32] Behra B. K., Mani M. P. (Dec. 2007) Indian Doddo F. (2001) Proc. IEEE Conf.
Journal of Fibre and Textile research, Circuts & Sys., ISCAS 2001, pp.481-
.32, 421-426. 484.
[33] Pentland A.P. (1984) IEEE Trans. On [56] Yuen C. W. M., Wong W. K., Qian S. Q.,
Pattern Analysis and Machine Chan L. K. and Fung E. H. K.. (2009)
Intelligence, 6 , 661-674. Expert Systems with Applications: An
[34] Voss R. I (1986) R.Pynn and A. Skjeltorp International Journal, 36(2), 2037-2047.
Eds. [57] Shi M., Jiang S., Wang H. and Xu B. (2009)
[35] Keller J., Crownover R. and Chen S. (1989) Machine Vision and Applications,
Computer Vision Graphics Image Proc., 20(2)131-138.
150-160. [58] Castilho H. P., Goncalves P. J. S., Pinto J.
[36] Conci A. and Proença C. B. (1998) R.C. and Serafim A. L. (2007) Image
Computer Networks and ISDN Systems, Analysis and Recognition, SpringerLink,
30, 1887-1895. 4633.
[37] Davis L. S. and Mitiche A. (1980) Computer [59] yin Y., Lu W. B., Zhang K. and Jing L.
Graphics and Image Processing, 12, (2009) Global Congress on Intelligent
25- 39. System, 4, 167- 171.
[38] Jasper W. J. and Potapalli H. (1995) Text. [60] De Natale F. G. B. (1986) Intl. J. Pattern
Res. J., 65, 683-692. Recognition and Artificial Intell., 10(8).
[39] Conci A. and Proença C. B. (2000) Text. [61] Brodatz P. (1966) Texture: A Photographic
Res. J., 70, 347-350. album for artists and designers, Dover,
[40] Lane J. S. (June 1998) US Patent No. New York.
5,774,177. [62] Kauppinen H. (2000) Proc. IEEE Conf.
[41] Abidi B. R., Sari-Sarraf H., Goddard J. S., Pattern Recognition, Barcelona (Spain),
and Haunt M.A. Scientific Literature 4, 803-806.
Digital Library, [63] Desoli G. S., Fioravanti S., Fioravanti R.,
http://citeseer.ist.psu.edu/284961.html and Corso D.,(1993) Proc. Intl. Conf.
[42] Bodnarova A., Bennamoun M.,and Kubik K. Industrial Electronics, 3, 1871-1876.
K. (1998) Proc. IEEE SMC’98 Conf.,
San Diego (U.S.), 4423-4428.

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
26
 A review of automatic fabric defect detection techniques

[64] Boukouvalas C., Kittler J, Marik R, and [88] Bodnarova A., Bennamoun M., and Latham
Petrou M., (1999) IEEE Trans. Industrial S. (2002) Pattern Recognition, 35,
Electronics, 46, 219-226. 2973–2991.
[65] Bergsa L., Duffy N., Laecy G., and Mazo [89] Shu Y. and Tan Z. (2004) Proceedings of
M., (2000) Image and Vision Computing, the 5th World Congress on Intelligent
18, 951-957. Control and Automation, 3378-3380.
[66] F. Vilnrotter, R. Nevatia, and K. Price, [90] Han R. and Zhang L. (2009) Global
(1986) IEEE Transactions on Pattern Congress on Intelligent Systems, 184-
Analysis and Machine Intelligence, 188.
8:76–89. [91] Mak K. L. and Peng P. (2008) Robotics and
[67] Wechsler H. (1980) Signal Processing, 2, Computer-Integrated Manufacturing,
271-282. 24(3),359-369.
[68] Azencott R., Wang J. –P., and Younes L., [92] Arivazhagan S., Ganesan L. and Bama S.
(1997) IEEE Trans. Pattern Anal. Mach. (2006) Machine Vision and
Intell. 19,148-153. Applications,16 (6).
[69] Daugman J. G. (1985) Journal of the [93] Hou Z. and Parker J. M. (2005) Proc. of
Optical Society of America 2, 1160- seventh IEEE Workshop on Application
1169. of Computer.
[70] Clark M. and Bovik A. C., (1987) Pattern [94] Sari-Sarraf H. and Goddard J. S. (1999)
Recognition Letters 6, 261-267. IEEE Trans. Ind. Appl., 35, 1252-1259.
[71] Mallat S. G., (1989) IEEE Trans. Pattern [95] Bracewell R. N. (2000) The Fourier
Anal. Mach. Intell., 11, 674-693. transform and its applications, Boston:
[72] Liu S. –S., Jernigan M. E., (1990) Computer McGraw-Hill, 3rd ed.
Vision, Graphics and Image Processing [96] Scharcanski J. (2005) Pattern Recognition
49, 52-67. Letters, 26, 1701–1709.
[73] Escofet J., Millan M. S., Abril H., Torrecilla [97] Kim S., Lee M., and Woo K. (1999) In
E.,(1998) Proceedings of SPIE 3490, Proceedings of IEEE International
207-210. Symposium on Industrial Electronics, 3,
[74] Chan C. –H., Pang K. H., (2000) IEEE 1406–1409.
Trans. Industry Applications 36, 1267- [98] Latif-Amet L., Ertuzun A., and Ercil A.
1276. (2000) Image and Vision Computing,
[75] Sengottuvelan P, Wahi A and Shanmugam 18(6-7),543–553.
A, (2008) Research Journal of Applied [99] Jasper W. J., Garnier S.J., and Potapalli H.
Sciences, 3 (1), 26-31. (1996) Optical. Eng., 35, 3140-3149.
[76] Tsai D. and Heish C., (1999) Image and [100] Jasper W., Joines J. and Brenzovich J.
Vision Computing, 18(1),49–62. (2004) J. Textile Institute, 96, 43-54.
[77] Tsai D. and Huang T., (2003) Image and [101] Yang X., Pang G. and Yung N. (2004)
Vision Computing, 21, 307–323. Pattern Recognition, 37889-899.
[78] Mead D. C., Kasdan H. L., and Dorrity J. L. [102] Yang X., Pang G. and Yung N. (2005) IEE
(1978) US Patent No. 4,124,300. Proc.- Vis. Image Signal
[79] Castellini C., Francini F., Longobardi G., Process,152(6),715-723.
and Tirbilli B., (1992) Opt. Lasers Eng., [103] Serdaroglu A., Ertuzun A., and Ercil A.
24, 19-32. (2006) Pattern Recognition & Image
[80] Ribolzi S, Merckle J., and Gresser J. (1993) Analysis, 61-64.
Text. Res. J., 63, 61-71. [104] Kumar A. and Gupta S.(2000) Proc. IEEE
[81] Ciamberlini C., Francini F., Longobardi G., Conf. Industrial Technology, 2, 113-
Poggi P., Sansoni P., and Tiribili T., 118.
(1996) Proc. SPIE 2786, 9-18. [105] Truchetet F. and Laligant O. (2008) Journal
[82] Ciamberlini C, Francini F., Longobardi G., of Electronic Imaging.
Sansoni P., and Tiribili B.,(1996) Opt. [106] Guan S., Shi X., Cui H. and Song Y. (2008)
Eng., 35, 835-844. IEEE Pacific – Asia Workshop on
[83] Campbell J. G. and Murtagh F. (1998) Opt. Computational Intelligence and
Eng., 37, 2536-2542. Industrial Application, 266-370.
[84] Daugman J. (1980) Vision Research, [107] Guan S. and Shi X.(2008) International
20,847–856. Symposium on Information Science and
[85] Jain A. and Farrokhnia F., (1991) Pattern Engineering, 281-285.
Recognition, 24, 1167–1186. [108] Ngan H. Y. T., Pang G. K. H., Yung S. and
[86] Kumar A. and Pang G., (2000) Optical Michael K. Ng (2005) Pattern
Engineering, 39(12), 3176–3190. recognition, 38(4), 559-576.
[87] Kumar A .and Pang G., (2002) IEEE Trans. [109] Li S. (2001) Markov Random Filed Modeling
on Industry Applications, 38(2), 425– in Image Analysis.
440.

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 27

 Mahajan PM, Kolhe SR and Patil PM

[110] Cohen F., Fan Z., and Attali S. (1991) IEEE [118] Lee, Tin-Chi (2004) Fabric defect detection
Transactions on Pattern Analysis and by wavelet transform and neural
Machine Intelligence, 13(8) 803–809. network,” University of Hong Kong.
[111] Baykut A., Atalay A., Ercil A., and G¨uler M. [119] Randen T and Husoy J. H. (1999) IEEE.
(2000) Real- Time Imaging 6, 17–27. Trans. PAMI, 21 (4), 291-310.
[112] Ozdemir S., Baykut A., Meylani R., Ercil A., [120] Randen T., Husoy J. (1994) Optical
and Ert¨uz¨un A (1998) International Engineering, 33 (8), 2617-2625.
Conference on Pattern Recognition [121] Chien-Chang Chen, Chaur-Chin Chen
1738–1740. (1999) Pattern Recognition Letters, 20,
[113] Brzaković D.P., Bakić P. R, Vuiovic N. S., 783-790.
and Sari-Sarraf H. (1997) Proc. IEEE [122] Chetverikov D. and Henbury A. (2002)
Intl. Conf. Robotics and Automation, 1- Pattern Recognition, 35, 2165-2180.
8.
[114] Brzaković D. P., Bakić P. R., and
Liakopoulos A. (1995) Proc. SPIE 2597.
60-69.
[115] Campbell J. G., Fraley C., Murtagh F., and
Raftery A. E. (1996) Technical Report
No. 314, Dept. of Statistics, University of
Washington, Seattle, 1-15.
[116] Ozdemir S., Baykut A., Meylani R., Ercil A.,
and Ert¨uz¨un A. (1998) International
Conference on Pattern Recognition,
1738–1740.
[117] Bodnarova A., Bennamoun M., and Latham
S. J. (2002) Pattern Recognition, 35,
2973-2991.

Copyright © 2009, Bioinfo Publications, Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009
28
 A review of automatic fabric defect detection techniques

Table 1- In exhaustive list of fabric defect detection methods

Approach Method References
Statistical 1. Co-occurrence matrix [17, 19, 21, 22, 23, 24, 25]
2. Local linear [27, 28, 29, 30, 31, 32]
transforms
3. Fractal dimension [33, 34, 35, 36]
4. Edge detection [38, 39, 40, 41]
5. Cross-correlation [42, 43]
6. Bi-level thresholding [ 44, 45, 46, 47]
7. Morphological [48, 49]
operations
8. Neural networks [50, 51, 52, 53, 54, 55, 56, 57, 58, 59 ]
9. Histogram [60, 62, 63]

Spectral 1. Discrete Fourier [ 67, 68, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83 ]
transform
[69, 70, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93 ]
2. Gabor filter [71, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106,
107, 108 ]
3. Wavelet transform

Model 1. Gauss Markov [109, 110, 111, 112]

based random field
2. Poissonian model [113, 114]
3. Model-based [115]
clustering

Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009 29