Fabric Defect Detection Using Computer Vision Techniques A

Hindawi
Mathematical Problems in Engineering

Volume 2020, Article ID 8189403, 24 pages
https://doi.org/10.1155/2020/8189403
Review Article
Fabric Defect Detection Using Computer Vision Techniques:
A Comprehensive Review
Aqsa Rasheed,1 Bushra Zafar,2 Amina Rasheed,3 Nouman Ali ,1 Muhammad Sajid,4
Saadat Hanif Dar,1 Usman Habib ,5 Tehmina Shehryar,1
and Muhammad Tariq Mahmood 6
1
Department of Software Engineering, Mirpur University of Science & Technology (MUST), Mirpur-10250, AJK, Pakistan
2
Department of Computer Science, Government College University, Faisalabad 38000, Punjab, Pakistan
3
Department of Textile Design, University of Gujarat, Hafiz Hayat Main Campus, Gujarat-50700, Punjab, Pakistan
4
Department of Electrical Engineering, Mirpur University of Science & Technology (MUST), Mirpur-10250, AJK, Pakistan
5
Department of Computer Science, National University of Computer and Emerging Sciences, Islamabad,
Chiniot-Faisalabad Campus, Chiniot 35400, Pakistan
6
Future Convergence Engineering, Korea University of Technology and Education, 1600, Chungjeol-ro, Byeongcheon-myeon,
Cheonan 31253, Republic of Korea
Correspondence should be addressed to Muhammad Tariq Mahmood; [email protected]
Received 3 August 2020; Revised 29 September 2020; Accepted 23 October 2020; Published 16 November 2020
Academic Editor: Sajad Azizi
Copyright © 2020 Aqsa Rasheed et al. )is is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
)ere are different applications of computer vision and digital image processing in various applied domains and automated
production process. In textile industry, fabric defect detection is considered as a challenging task as the quality and the price of any
textile product are dependent on the efficiency and effectiveness of the automatic defect detection. Previously, manual human
efforts are applied in textile industry to detect the defects in the fabric production process. Lack of concentration, human fatigue,
and time consumption are the main drawbacks associated with the manual fabric defect detection process. Applications based on
computer vision and digital image processing can address the abovementioned limitations and drawbacks. Since the last two
decades, various computer vision-based applications are proposed in various research articles to address these limitations. In this
review article, we aim to present a detailed study about various computer vision-based approaches with application in textile
industry to detect fabric defects. )e proposed study presents a detailed overview of histogram-based approaches, color-based
approaches, image segmentation-based approaches, frequency domain operations, texture-based defect detection, sparse feature-
based operation, image morphology operations, and recent trends of deep learning. )e performance evaluation criteria for
automatic fabric defect detection is also presented and discussed. )e drawbacks and limitations associated with the existing
published research are discussed in detail, and possible future research directions are also mentioned. )is research study provides
comprehensive details about computer vision and digital image processing applications to detect different types of fabric defects.
1. Introduction industrial products [5–8]. Natural elements such as wool,

cotton, a composite of polyester, or nylon can be used to
Computer vision and image classification-based models are create textile fabric bib9 [8, 9]. Sophisticated machines are
used in various applied domains including industry-based used in textile industry to create this fabric, and defects are
problems [1–4]. Clothing is considered as one of the basic located through the inspection process. Traditionally, in-
requirements for human life, and the history of textile inspection process is completed by using manual human ef-
dustry is as old as human civilization. Fabric is considered as forts to ensure the quality of fabric [6, 10]. )e price of fabric
a main element for human clothing and is also used in many that is sent to the market depends on the number of co-
2629, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8189403 by Gaziantep Universitesi, Wiley Online Library on [27/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 Mathematical Problems in Engineering
occurrence of defects and price increase with the increase in images and the grayscale image processing can reduce the
the number of defects [9, 11, 12]. As a defect is detected, the computational complexity. )e colored fabric defects have
production process is stopped and the details about the their own importance and detection models for example
occurred defect are recorded with its location by the ma- misarranged warp yarns fabric defects [14–16].
chine operator. )e main drawbacks associated during Generally, a set number of yarns are used for the for-
manual inspections are as follows: (1) training of individuals mation of fabrics, and different techniques such as knitting,
is required to make them fabric inspector; (2) major defects weaving, and felting are used for producing fabric. )e type
can be detected while small defects can be ignored due to of the fabric varies based on the fibers used, the fabric
human carelessness; (3) lot of human effort is required to formation technique, machinery used for their production,
locate fabric defects; and (4) it is very difficult for fabric and the finishing techniques. )ere are three main chal-
inspectors to keep focus on the production process for a time lenges related to the fabric defect detection. (1) )ere exists a
that is more than 10 minutes and all of this can lead to a low broad range of fabrics exhibiting different characteristics.
efficiency of production. According to research [3], 60–75% Application of general algorithms to varied texture types is
human accuracy is reported to detect fabric defects and the difficult and may lead to instability in the traditional defect
wastage due to fabric defects leads to the high price of detection methods. (2) )e varied characteristics and cat-
product in market. egories of the fabric defects contribute to the difficulties. (3)
Due to these reasons, it is recommended to apply some It is extremely difficult to collect large numbers of defect
automated processes to detect the fabric production defects samples particularly of some rare types, which results in data
and this process can save labor cost. Computer simulations imbalance or even complete failure of traditional supervised
are used for this purpose and textile products are refined methods. To deal with these challenges, numerous re-
through this process and can provide higher inspection searchers have made substantial efforts. According to the
quality. Computer vision and digital image processing plays literature, the fabric defect detection methods are broadly
a vital role in textile industry for this inspection process bib9 categorized as motif-based approaches and non-motif-based
[2, 9]. )e inspection of fabric is carried out during the approaches [6, 8, 17]. More research is focused on non-
production process that is considered as a real-time appli- motif-based group; therefore, this article mainly covers non-
cation. Automated inspection in textile industry is per- motif-based methods. In this review article, we aim to
formed while using computer vision and digital image present an overview about fabric defect detection for textile
processing techniques. Figure 1 represents the automatic industry. )e proposed study presents a detailed overview
fabric defect system. Figure 1(a) shows the scheme based on about approaches for fabric defect detection. )e perfor-
manual efforts, and defect in this case will be detected by mance evaluation criteria for automatic fabric defect de-
humans. Figure 1(b) shows an automated process for fabric tection is also presented and discussed in detail. )e
defect detection. According to the published research, more drawbacks and limitations associated with the existing
than 70 different possible fabric defects are reported and published research are discussed in detail, and possible
some common defects are shown in Figure 2. )ese defects future research directions are also mentioned. )is research
are caused by different reasons and can be classified as major, study provides comprehensive details about computer vision
minor, or critical defect (depending on the severity of the and digital image processing applications to detect different
defects [13]. Some of the frequently occurring major defects types of fabric defects.
are broken pick, ends out, float, holes, stitches, knots, loose
threads, starting mark, oil stains and marks, bad selvedge, 2. Histogram-Based Approaches
double pick, snarls, cracks, and smash. Hole is classified as a
major defect which is caused by many reasons such as by A histogram is a display of statistical information computed
broken needle or due to defective machine. Oil stain is on the basis of the number of co-occurrence of gray levels in
mainly caused due to excessive oil from the machinery. an image [18, 19]. According to the literature [20], the spatial
Multiple netting are minor defects caused when multiple approaches have the advantage of being computationally
broken threads are combined together. Broken end appears simple but have weak performance in the detection of small
in the fabric when warp yarns break during weaving. )ick defects. Wavelet transform-based approaches outperform
and thin bar defects appear when variations occur in yarn. the spatial methods in terms of computational efficiency and
Broken weft appears when in weaving filling yarn is broken. performance (particularly for detecting small fabric defects).
Wrong weft defect appears when the process weaving does Ye et al. [20] applied fuzzy inductive reasoning (FIR) for the
not follow a schematic pattern during weft insertion. )ere detection of defects in fabrics. )e histogram statistics
are various categories of fabric defects and many ways to extracted from the images are used to distinguish between
classify them such as common fabric defects or classification the faulty and the faultless fabric images. Figure 3 shows the
based on fabric color. In this article, mostly grayscale fabric RGB (red, green, and blue) channels and combined histo-
defect detection algorithms and classification models are gram of defective images. Since faultless fabric exhibits a
discussed in detail. Most of the computer vision algorithms regular repetitive global pattern, FIR can help to detect fabric
are designed for grayscale images and use different feature defects by analyzing the fabric structure. To analyze the
extraction approaches to create discriminative image rep- fabric texture, the histogram projections of the images are
resentations. Other than this, the grayscale image processing considered a good choice since the grayscale variations occur
can extract descriptors more easily as compared to the color in a certain manner along a warp or weft. Four classes of
Mathematical Problems in Engineering 3
(a) (b)
Figure 1: Fabric defect detection using manual and automated inspection.
Multiple Oil Broken Wrong

Broken end Thick bar Thin bar Hole
netting stain weft weft
Figure 2: Examples of textile fabric defects.
fabric defects, i.e., wrong draw, oil warp, slubby yarn, and create a distinction between defective and defect-free re-
harness skip are examined by using this approach [20]. )e gions. )e second step involves the extraction and selection
difference between the predicted and the real class values of saliency histogram features for effective discrimination
triggers an alarm vector that indicates the presence of detect. between faulty and faultless fabric images. Lastly, classifi-
Besides defect occurrence, the proposed approach also cation is done using binary support vector machine (SVM)
identifies the location of defects. that was trained by using defective and nondefective fabric
Based on the characteristics of cord fabric, Zhang et al. samples. )e proposed approach resulted in accurate de-
[21] proposed a multiple windows gray ratio (MWGR) for tections as compared to the state-of-the-art research and can
the detection of cord fabric defects. )e main motivation be applied for defect detection in both patterned and
behind MWGR is that the normal cord fabric grayscale nonpatterned fabric images. Table 1 represents a summary
image exhibits an alternating gray and white pattern. A about various histogram-based approaches proposed for
threshold is used to split the image foreground and the fabric defect detection. Histogram-based methods are noise
background regions. )e image is partitioned into several sensitive, and in nonregular textures, it has low detection
regions and the gray ratio change is analyzed to determine rate.
the defect in window. )e proposed approach [21] achieved
satisfactory results for the detection of cord fabric defects 3. Color-Based Approaches
and enhanced the automated quality control for fabric defect
detection (FDD). Li et al. [22] proposed an algorithm based In computer vision and digital image processing, color is
on visual saliency for defect detection in both patterned and used in various applications that are using visible spectrum
nonpatterned fabrics. )e features computed from the sa- [23–25]. Color is extracted as an important visual feature in
liency maps are used for the detection of fabric defects. At various fabric defect detection-based approaches [26]. Nu-
the first step, the algorithm generates the saliency maps to merous research efforts have been made to improve the
1200
1000
800
600
400
200
0
0 50 100 150 200 250
Red channel
Green channel
Blue channel
Figure 3: RGB histogram of image.
Table 1: A summary about histogram-based approaches.

Author Dataset Performance evaluation criteria Proposed model and purpose
Faulty images with four kinds of fabric
Instantaneous error computed by Applied fuzzy inductive reasoning for
Ye et al. defects which are slubby yarn, harness
taking the difference between the detection of defects in fabrics with regular
[20] skip, oil warp, and wrong draw are used to
real and predicted class values repetitive global texture
examine this approach
Sample images of cord fabric including Proposed an approach for detection of cord
Zhang four kinds of cord fabric defects, i.e., fabric defects by combining gray ratio
Gray ratio method
et al. [21] broken end, hard size, split seam, and extracted from multiple windows with
longitude clinging together are used threshold segmentation
)is article presents a saliency-based
Li et al. Defective and defect-free fabric images are
Accuracy, precision, and recall approach for automatic detection of fabric
[22] used
defects
accuracy and efficiency of FDD with varying viewpoints; value of a pixel by an integrated measure of color dissim-
however, FDD still remains a challenging task for compli- ilarity and positional similarity. To enhance the reliability
cated textures [26]. Zhang et al. [26] claimed the visual and robustness of results, the contrast of both the defective
saliency of defective fabric image regions, by arguing that the and the nondefective regions is enhanced by applying
defective and the nondefective local regions of fabric images multiscale dissimilarity analysis. )e overview of the pro-
have visual differences in color. In addition to this, the posed method for FDD is shown in Figure 4. Experiments
defective regions are not widespread in the image; rather, are conducted using MATLAB R2014a on Intel Core pro-
they are grouped in a small region. Based on these obser- cessor with 6 GB RAM and 3.2 GHz frequency. Experimental
vations, the authors proposed to determine the defective results conducted with diverse fabric images achieved a
better performance as compared to the state-of-the-art re- algorithm. )e detection results obtained using the proposed
sults for complicated patterns. However, for box-patterned method are compared with the state-of-the-art methods
and motif-based images, the effectiveness of the proposed such as morphological operation, Gabor filter, and local
saliency method could not be established. Table 2 represents binary pattern. According to experimental results, the
a summary about color-based approaches proposed for proposed method achieved higher precision and detection
fabric defect detection. efficiency as compared with the traditional methods. )e
overview for dictionary learning phase and defect classifi-
4. Segmentation-Based Approaches cation is shown in Figure 5. Table 3 represents a summary
about dictionary learning-based approaches proposed for
Segmentation is used to divide the image into subregions, fabric defect detection.
and the success of any computer vision algorithm depends
on the effectiveness of the applied segmentation technique 6. Texture-Based Approaches
[27, 28]. )e high computational cost associated with image
segmentation is considered as one of the limitations of Texture analysis refers to the characterization of regions in
approaches that are relying on image segmentation [29, 30]. an image by their texture content and is widely used in
To detect defects in plain woven fabrics, Guan et al. [31] different computer vision application domains [34, 35].
investigated an approach based on the segmentation of Liang et al. [36] proposed an intelligent integrated system
regions of interest (ROI) from defected images. )e pro- for the evaluation of yarn surface appearance. )ree
cessing is done by using images in grayscale mode, and an methods are used for feature extraction: attention-driven
image enhancement technique is applied to highlight the method is applied using saliency maps for fault detection
regions with defects. To further enhance the accuracy and to and wavelet texture features and various statistical mea-
reduce the algorithm complexity, the noise is removed by sures are applied to formulate appearance description of
applying the low-pass filtering which highlights the defected yarn surface. )e fuzzy ARTMAP neural network (NN) is
regions. )e defected regions are then segmented with edge employed to learn features for grading and classification
detection based on first-order derivatives. )e Roberts op- of yarn surface quality. A comparative analysis is per-
erator was observed to efficiently detect the edge of defect formed among the fuzzy ARTMAP, backpropagation NN,
regions and resulted in better accuracy. )e ROIs are then and SVM to obtain the optimal classification performance
obtained by isolating the defect regions based on boundary. (Figure 6 shows working of SVM). )e experimental
)e experimental results demonstrated that segmentation of results demonstrated that the proposed approach while
defective regions was done successfully which enhanced the using fuzzy ARTMAP achieved superior results to classify
defect detection rate [31]. Table 2 represents a summary yarn surfaces.
about segmentation-based approaches proposed for fabric According to [37], fabric defects are basically abrupt
defect detection. features and the wavelet response for background and
abrupt features is totally different, and hence, the wavelet can
5. Dictionary Learning-Based Approaches accurately detect the defects. )e methods used for FDD
using wavelets are categorized as nonadaptive and adaptive;
Dictionary learning-based approaches have shown good the former methods use existing bases in combination with
results for fabric defect detection [3]. Li et al. [32] proposed a some feature extraction method for detection. However, the
learned dictionary-based visual saliency approach for FDD. existing wavelet bases have their drawbacks, and the non-
)is approach partitioned the image into blocks, and a adaptive methods do not attain satisfactory results for all
dictionary is then constructed for both normal and defective different fabric defects types. Wen et al. [37] proposed an
image samples based on sorted local binary features cor- approach for the detection of different image fabric defects
relation. )e N largest correlation coefficients are computed by applying adaptive wavelet bases. )e adaptive wavelet
between each test image block and the dictionary, and the bases were designed to deal with fabric defects and they
remaining correlation coefficients are set to zero. A saliency generally comprise a training stage and a detection stage.
map is then generated based on the sum of nonzero coef- )ey used a hybrid approach by the combination of wavelet
ficients related to the defective image samples. )e saliency bases and an optimized method to deal with different im-
maps created by the proposed approach are compared with ages. A cost function model is used to obtain a filter coef-
the state-of-the-art methods. )e proposed approach suc- ficient for the detection of fabric defects. A limitation of the
cessfully localized defective regions for complex fabric proposed approach is that the regular images are required to
images. train the model for the detection of patterned fabric defects.
In another research, Zhou et al. [33] proposed a novel Experimental results demonstrated that the proposed ap-
texture feature extraction technique for the detection and proach effectively achieved accurate detection of fabric
classification of fabric defects based on multiscale dictionary defects.
learning. A linear summation is applied to the multiscale According to Karlekar et al. [38], efficiency and real time
dictionaries for accurate representation of image features. To are important factors that should be considered for fabric
create a classification and detection model, an algorithm is defect detection systems. Karlekar et al. [38] used wavelet
proposed that can deal with the limitation of parameter analysis since it detects line features efficiently with low
setting and low accuracy in the original differential evolution complexity. )e images are first converted to grayscale,
Input
Pyramid image
Channel Channel Channel

L A B
Dissimilarity of color feature

Similarity of positional features
Dissimilarity
Similarity
Defect map
Figure 4: Example of the classifier model for textile fabric defect detection.
Table 2: A summary on color-based and segmentation-based approaches.

Performance evaluation
Author Dataset Proposed model and purpose Approach
criteria
Inspired from the human visual system,
A dataset of 130 defected fabric images Detection rate and other
Zhang they proposed a saliency metric-based
(patterned, nonpatterned, and gray parameters influencing the Color-based
et al. [26] approach for the detection of fabric
fabric textures) is used detection performance
defects
Plain woven fabric images with typical Investigated an approach for automatic
Guan Segmented regions of the Segmentation-
defects such as end out, mispick, hole, segmentation of defect regions from
[31] images were observed based
double end, and double pick are used plain woven fabric images
Image patches Patch vectors Approximated feature vector
Dictionary learning phase

Dictionary set
Defect-free images
Training ADE-RELM phase

ADE-RELM
Defect-free images
Testing phase
ADE-RELM
Testing images
Figure 5: Dictionary learning phase and defect classification overview.
Table 3: A summary on dictionary learning-based approaches.

Defect images of different types such as Comparison of saliency map and the Proposed an approach to detect
Li et al.
mispick, wrong draw, and fuzzy ball, from detection results of the proposed method fabric defects using learned
[32]
published research were used with the state-of-the-art methods dictionary-based visual saliency
Proposed a texture feature
Zhou Sample images from 6 different defect
Classification correct rate (CCR) extraction technique based on
et al. [33] categories were used
multiscale dictionary learning
histogram equalization, and median filter are applied for the significant features for defect classification. )resholding is
reduction of noise from fabric images. )e wavelet de- then applied for the segmentation of image defects, and
composition is applied at different scales to extract morphology filter is applied for removal of the remaining
fabric texture representation and for the detection of fabric

flaws.
In another recent research, Hanbay et al. [42] proposed a
Gap vision-based system for inspection of fabric defects for
circular knitting machine. )e proposed system is com-
prised of two modules, i.e., feature extraction and defect
classification. Initially, a fabric defect detection database is
constructed. For feature extraction from fabric images,
shearlet transform is applied on the normalized images
acquired continuously from one line scan. )e means and
variances computed from all subbands are concatenated to
create a high-dimensional feature vector. For defect detec-
tion, a three-layered artificial neural network (ANN) is used
and shown in Figure 7. )e performance evaluation is done
on a single jersey knitted fabrics on a circular knitting
machine in a textile factory. For quantitative evaluation, five
different indexes are used. )e proposed approach attained
better outcomes as compared to the related research on
Support vectors for C1
fabric defect detection. Table 4 represents a summary about
Support vectors for C2 texture-based approaches proposed for fabric defect de-
tection. Image acquisition is done by using line-scan CCD
Optimal hyperlane camera and Intel Core i7 (3.5 GHz) with 16 GB RAM. It is
Figure 6: SVM-based classification approach. concluded that the wavelet transformation-based ap-
proaches are mainly best suited for real-time application due
to its computational performance.
noise. )e proposed method obtained satisfactory perfor-
mance for thresholding and segmentation of fabric defects 7. Frequency Domain Operation
caused by warp knitting machine. Yapi et al. [39] introduced
a generalized learning-based approach based on the statis- For frequency domain operations, an image is converted
tical signatures to distinguish between defective and defect- from spatial domain to frequency domain [19]. Fast Fourier
free fabrics. )e proposed system is able to deal with textile transformation (FFT) can be used to convert an image from
fabrics of varying types, ranging from simple to more spatial domain to frequency domain [19]. According to
complex ones. )e system is comprised of three steps: Ismail et al. [43], the originality and the examination of
preprocessing, classifier training, and testing. )e descrip- fabric by using human eyes cannot guarantee that fabric is
tors representing the fabric are robust to noise and changes 100% original and also it is difficult to process in case of
in illumination and are invariant to translation and scale jersey fabric. )erefore, Ismail et al. [43] proposed a method
transformation. It was concluded that the proposed ap- based on FFT to investigate the quality of the jersey fabric
proach has provided better performance as compared to the (especially sport jersey fabric). For this purpose, they build a
state-of-the-art research. sample or prototype to examine the quality of the fabric and
Hamdi et al. [40] proposed an algorithm for unsuper- to analyze their method based on their structure pattern of
vised defect detection in patterned fabric images. )e the fabric. On the base of Fourier transformation spectrum,
proposed algorithm comprises three steps: preprocessing, a the authors [43] used the magnitude and phase graph to
texture filtering step, and defect classification step. )e compare its results of test samples with the original sample.
preprocessing step involves improving the contrast of the To differentiate both images, absolute error and mean error
image and the extraction of fabric periodic pattern. During are also computed. )e results for magnitude and phase
texture filtering, the image is segmented into different re- graph of two samples (original and test) are compared, and if
gions based on texture and is filtered to remove the effect of there is a difference between samples, then the test sample is
noise. For defect classification, the filtered image is parti- fake, and if there is no difference, then the sample is original.
tioned into blocks and the statistics aggregated from the For defect detection in fabric with periodic surface
median of all blocks are used with k-means to distinguish the texture, Hu et al. [44] proposed an unsupervised method
defected blocks with high difference value. Wu et al. [41] that is based on a combination of discrete Fourier transform
presented a dictionary-based method to enhance the texture (DFT) and discrete wavelet transform (DWT). )is method
fabric representation and boost the computation speed that only depends on the periodic surface texture to reduce some
can assist in the detection of fabric flaws. Discrete cosine problems between the sample under examination and used
transform is used to learn dictionary by applying the al- for comparison (reference sample), i.e., orientation, scale,
ternating least-squares method. )e proposed approach and unbalanced-illumination problems. It neither depends
provided remarkable results in terms of consistent results, on any feature extraction approach nor it relies on any
computation speed, and error rate. )e proposed approach sample for judgement [44]. As it is known that Fourier
can be applied to investigate the impact of fabric density on domain is suitable for the periodicity of an object, so this
and similarity measure has been used to recognize the de-

fective and defect-free units. For the verification of proposed
method, the analysis has been performed on the following
types of defects: warp or weft-lacking, holes, and on man-
ually (artificial) added defect, and reliable and fast result has
been achieved, but the results are not mentioned explicitly.
Table 5 represents a summary about frequency domain-
based approaches proposed for fabric defect detection.
8. GLCM-Based Defect Detection

In textile industry, the quality of products and name of the
brand are maintained by the production of defect-free
products. Detection of defects can be done by using
manual human efforts and this process can be automatic.
Raheja et al. [47] proposed a fast defect detection system
(FDDS) embedded on DSP (which is digital signal pro-
Figure 7: Architecture of artificial neural network (ANN).
cessor electronic board used for experiments). It is a
GLCM-based method and implemented on a Texas In-
defining characteristic has been used to eliminate the struments Evaluation Module. Here the sliding window
background periodic pattern by zero-masking their domi- analysis is used for the computation of textural energy of
nant frequency in each sample under examination. )e the fabric image from the GLCM. For implementation,
image obtained by Fourier-restored residual is then sub- Texas TMS320DM642 instrument evaluation module is
jected by wavelet shrinkage, and lastly thresholding is ap- used with two decoders, one encoder with three on-board
plied to get binary images which are then combined for a video ports. )e analysis is performed on every frame, and
single output. Hu et al. [44] performed experiments on the GLCM feature correlation, energy, homogeneity, and
number of real textured surfaces where all image samples contrast are extracted from defective image samples. But
were composed of 8-bit gray level and 256 × 256 pixels. For energy feature has been used to compute results and to
performance evaluation, two groups, c1r1 and c1r3, from the find out defects in texture images by taking the advantage
TILDA textile dataset have been used and false-positive rate of features that they give the most noticeable peaks in
(FPR), true-positive rate (TPR), and accuracy (Acc) have faulty area.
been measured. Arnia et al. [48] proposed a method based on gray-level
Sakhare et al. [45] proposed two domain techniques that co-occurrence matrix (GLCM) using its two features con-
are spatial and spectral for the detection and classification of trast and energy. Discrete cosine transform (DCT) is used
defects. )e first approach contains morphological and with the objective to evaluate and comply with their method
statistical operations which have been used to find out the on MPEG (Moving Picture Experts Group) and MJPEG
intensity values. )e second approach contains wavelet and (Motion Joint Photographic Experts Group) video com-
Gabor filter, FFT, and DCT (discrete cosine transform) to pression standards to reduce space and cost savings by
obtain transformed image which is further analyzed for the replacing the frame. Before this, GLCM has been used with
detection of defects. Performance evaluation is performed many other methods for the detection of defects. In the
for four types of defects: missing warp, missing weft, hole, proposed method, DCT images are used to preserve the
and tom out in terms of percentage accuracy. Sakhare et al. appearance and details of the image in low- and high-fre-
[45] detected the fabric defects and proposed a combination quency bands, respectively. Moreover, for the effectiveness
of spatial and spectral techniques and performed compar- of this approach using DCTb-Is many compression rates and
ison with various methods. for defect detection lowest possible rate in DCTb-Is have
According to the literature [3], defect detection in the been evaluated. )e results showed that the compression rate
Yarn-Dyed fabric is still limited due to its three-layered greater than 25% can only able to detect and locate the stains,
structure of patterns, fabric textures, and defects. Zhang et al. drops, holes, and drop stitches from DCTb-Is.
[46] proposed a method based on frequency domain fil- For the classification of defects in textile fabric, Zhang
tering, similarity measurement, and distance matching et al. [49] proposed a hybrid approach based on LBP
function for the detection of defects in the Yarn-Dyed fabric. (local binary patterns) and GLCM (gray-level co-oc-
)e established algorithm does not completely separate the currence matrix) to differentiate and to classify defects.
three different structures (patterns, fabric textures, and )e production process can be improved by producers if
defects), and basically, this multilayered structure is the defects in fabric are correctly differentiated [49]. Zhang
major reason for complexity in yarn-dyed fabric for finding et al. used LBP to extract the local features of defective
defects as separation of these layers is a difficult task [46]. In images and GLCM to extract the information of texture
this method [46], frequency filtering is used to separate the features. For the verification of their proposed hybrid
two layers (patterns and fabric textures) of the fabric, and approach, authors performed experiments on 6 types of
distance matching function to measure the pattern period defects where each defect includes 100 images of
Table 4: A summary about texture-based approaches.

Proposed an intelligent yarn grading
system based on integrated features,
Liang 576 yarn board images categorized into
Classification accuracy comparison which are used to train fuzzy ARTMAP
et al. [36] 5 distinct categories are used.
neural network for evaluation of yarn
surface appearance.
Proposed an adaptive wavelet-based
Wen et al. Sample images with different defects are Segmented images comparison with the approach by optimizing the cost function
[37] used. state-of-the-art methods to obtain the wavelet filter coefficient or
detection of fabric detects.
Proposed an approach based on wavelet
Karle-kar )e wavelet analysis is performed with a
— transform for the detection of fabric
et al. [38] defected image.
defects caused by warp knitting machine.
Image-level performance metrics:
Introduced a learning-based algorithm
Experiments are performed on the detection rates, false alarm rates, and
Yapi et al. that uses labeled examples to train
TILDA database which comprises detection accuracy. Local-level
[39] classifier to distinguish between defective
defect-free and defective images. performance metrics: local precision,
and defect-free fabrics.
local recall, and local accuracy
Two distinct databases are used to )is paper introduces an automatic
validate the effectiveness of the )e performance evaluation of the unsupervised algorithm for defect
Hamdi proposed algorithm. One of which is proposed algorithm is reported in terms detection in patterned fabrics, based on
et al. [40] used mainly in research publications of precision, recall, and detection success texture filtering using standard deviation
and the second one was captured by rates filtering and k-means clustering for defect
authors. classification.
Proposed a dictionary learning-based
42 defect-free and 3 defected image Time comparison, root mean square
Wu et al. approach for stable fabric texture feature
samples were used for the investigation error, and peak signal-to-noise ratio
[41] representation and improving the
of results. were used for comparison
computation speed.
Proposed a fabric defect detection system
Feature defect database was used for Five different indexes are used for based on statistical modelling of shearlet
Han-Bay simulations. For real-time evaluation, quantitative evaluation, i.e., specificity, subbands and ANN is used for defect
et al. [42] images are captured using line-scan sensitivity, accuracy, negative predictive detection. )e proposed system is
camera. value, and positive predictive value evaluated using single jersey knitted
fabric.
Table 5: A summary on the frequency domain operation.

Proposed fast Fourier transformation (FFT)
50 original and 50 fake fabric samples,
Magnitude and phase graph are to investigate the originality of the sport
Ismail total 100 fabric samples are tested which
used to compare the results of test jersey fabric and inspect Fourier
et al. [43] consist of Arsenal, Manchester United,
samples with the original sample transformation spectrum to detect the
Italia, Chelsea, and Liverpool brand
authenticity of the fabric
False-positive rate (FPR), true- Proposed an unsupervised method that
Hu et al. Two groups, c1r1 and c1r3, from the
positive rate (TPR), and accuracy combines discrete Fourier transform (DFT)
[44] TILDA textile dataset have been used
(Acc) and discrete wavelet transform (DWT)
)e database that includes four types of Proposed two domain techniques: spatial
Sakhare )e percentage accuracy has been
defects: missing warp, missing weft, hole, and spectral for the detection and
et al. [45] used
and tom out are used classification of defects
Presented a method based on frequency
Similarity measure has been used
Zhang domain filtering, similarity measurement,
Not explicitly mentioned to recognize the defective and
et al. [46] and distance matching function for the
defect-free units
detection of defects in yarn-dyed fabric
256 × 256 from the TILDA database to get results for Defect detection in yarn-dyed fabric is complex due to
classification average accuracy. )e results showed that its multilayered structure, so Zhu et al. [50] proposed an
the classification accuracy remarkably increased by the approach for defect detection in yarn-dyed fabric which is
proposed hybrid approach (LBP and GLCM), and with based on autocorrelation function and GLCM and defect
the increase in values of dimension of features, accuracy detection is done by using GLCM window and texture
can be increased. features. In their approach, Zhu et al. [50] selected a
detection window based on the periodic pattern of yarn- which are used for feature extraction, and kernel principal
dyed fabric and used Euclidean distance to differentiate component analysis (KPCA) is used for the reduction of
between two images. Defects such as oil stain, holes, high-dimension data to identify many uniform and struc-
broken weft, stretched warp, and wrong weft are con- tural detects. OTSU thresholding approach is used to in-
sidered to conduct experiments on two- and three-colored dicate defects, and this method is effective for the
yarn-dyed fabric. )e results have shown that the pro- identification of holes, thread errors, oil spot, and object on
posed method has high accuracy. Fabric defect detection surface of weaving structures. For the effectiveness of the
based on the GLCM technique is proposed by Zhang et al. proposed method, the experiments are conducted on de-
[51] where the distance between angular relation (theta) fective and defect-free fabric images taken from TILDA
and pixel pair is calculated. Experiments have been textile dataset, and the true detection rate of the proposed
conducted on defective and nondefective images gathered method is calculated. TILDA is a standardized textile texture
from different sources. Different window size of 10 × 10, database developed as a part of the research programme
20 × 20, 50 × 50, and 60 × 60, d � 1, and theta � 0° are the “Automatic Visual Inspection of Technical Objects” by the
used experimental factors. Zhang et al. [51] compared working group Texture Analysis of the DFG (Deutsche
results in terms of percentage detection rate and the re- Forschungsgemeinschaft). )e database comprised of 8
sults indicate that the proposed method is effective for the groups of representative textile kinds. Each class comprised
detection of fabric defects. Table 6 represents a summary of 50 TIF grayscale images of size 786 × 512 pixels. )e entire
about GLCM-based approaches proposed for fabric defect textile texture database contains 3200 TIF images with a total
detection. size of 1.2 Gbytes [54]. )e test samples defective and
GLCM-based methods depend on the matrix of co- nondefective with their binary and fused images of class C1
occurrence that is substantially influenced due to noise. )e from TILDA textile dataset are represented in Figure 8. Intel
limitations of GLCM-based approaches are as follows: it has (R) i5 (2400M) with 3.10 GHz, 4 GB RAM, and Windows 7
high computational cost, and in high-resolution images, it operating system is used as implementation platform. )us,
gives less performance. Other than these findings, optimal the result shows that the proposed algorithm has less
displacement factor is difficult and it depends on the scaling computational cost for real-time defect detection and has
and rotation and requires a procedure for feature selection. high detection rate.
For texture having large size primitive cannot analyze tex- According to the literature [55], presence of detects in
ture, and due to undirected distance, it may loss valuable fabric reduces its price from 45% to 65%. To improve the
information. quality of the fabric and for automatic defect detection, Li
et al. [55] proposed a method that relies on multiscale
9. Feature Fusion-Based Defect wavelet transform and Gaussian mixture model based on
Detection Techniques expectation-maximization (EM). Multiscale wavelet trans-
form is used to extract features from images, and Gaussian
According to Zhu et al. [52], the fabric is pressed in the dyeing mixture model’s main idea is to explain the feature’s
device during the heating roller calendar process to enhance probability space distribution. Six types of defect segmen-
the aesthetic of the fabric for soft glow on the fabric surface. tation results are discussed. Experiments are conducted to
Sometimes, the heating roller becomes defective due to the evaluate the results where 12 samples of texture images with
fabric seam, and then, fabric defects will also appear when the 256 grayscales are used from TILDA texture database.
defective roller will be used in the continuous calendaring According to Li et al. [56], there are so many types of
process and it is important to detect the defect in this case. defects, and common linear defects of warp knitting fabrics
Zhu et al. [52] proposed two methods to detect seam using are due to broken end of warp yarns which will become
images, textural analysis by wavelet energy method (say M1: significantly large if the warp knitting machine is not
method 1) and by feature extraction in CIELAB color space stopped immediately. So, to detect defects (warp knitted
(say M2: method 2) for texture analysis in the image. Energy fabric defects of vertical orientation) automatically, Li et al.
characteristics for M1 of the fabric image are calculated, and [56] proposed an automated vision examination system
for fabric seam detection, characteristic parameters such as based on Gabor filters and pulse coupled neural network
variation coefficient mean and standard deviation are used for (PCNN). Li et al. enhanced the image contrast using Gabor
M2 in CIELAB color space, and then, data obtained from filters and they used simplified PCNN for segmentation
these two techniques were used to determine the threshold purpose with some adaptive parameters to reduce the
value used for seam detection. )e results of both proposed computational complexity. To remove noise, morphological
methods have been compared in terms of accuracy and operations are performed and Figure 9 represents the ar-
computation time. )e system with Core i7 processor chitecture of the automated inspection system. )ey con-
(2.2 GHz frequency), MATLAB 2020b, and Windows 7 is ducted experiments for the evaluation of system and used
used to perform this research [52]. )e results have shown two images as test images which have been captured by a
that the proposed method M2 (textural analysis in CIELAB smart camera. Comparison has been performed with Gabor
color space) is better in terms of accuracy and time. and wavelet methods in terms of true-positive rate (TPR),
Jing et al. [53] proposed a new technique for real-time positive predictive value (PPV), and accuracy (Acc).
automated defect detection in textile cloth. )e proposed MATLAB R2012B is used for implementation with the
research is based on the multiple Gabor filters (MGFs), following hardware specifications: AMD Athlon 3.01 GHz,
Table 6: A summary on the GLCM-based operations.

Energy feature has been used to Proposed a GLCM-based method named FDDS
Raheja Different defected images have
compute results and to find out defects embedded on DSP using a sliding window
et al. [47] been used
in texture images technique to detect defects from the fabric image
For defect detection of holes, Proposed a GLCM-based approach for real-time
Arnia stains, and drop stitches, TILDA textile detect identification and find out the
Different coefficients of DCT are used
et al. [48] textile texture database has been lowest possible rate to detect defects such as
used holes, stains, and drop stitches
Proposed a hybrid approach based on LBP (local
Zhang Total 600 images of 256 × 256 from binary patterns) and GLCM (gray-level co-
Classification average accuracy
et al. [49] TILDA database have been used occurrence matrix) to differentiate and to
classify defects
Zhang Dataset of images gathered from Percentage detection rate has been Proposed a technique based on the GLCM
et al. [49] different sources used technique for fabric defect detection
Presented a new method for defect detection in
Euclidean distance has been used to
Zhu et al. Two- and three-colored yarn-dyed yarn-dyed fabric which is based on
differentiate between defects detected
[50] fabric has been used autocorrelation function and gray-level co-
images and the template image
occurrence matrix (GLCM)
Holes Object on the surface Oil spots Thread error

Defect
type
C1R1EABS C1R1EACP C1R1E4N1 C1R1E4N7 C1R1E2N3 C1R1EADR C1R1EAHZ
Defect-free
image
Defective
image
Binary
image
Fused
image
Figure 8: Defective images of class C1 from the TILDA textile dataset.
3.25 GB RAM. )e results of actual experiments have shown Intel Core i5-337um Hynix DDR3L, 8 GB RAM, and Seagate
that the proposed automated vision examination system is ST750LM022 HN-M7502MBB disk is used to evaluate the
effective and 98.6% is the detection accuracy. research [57].
Peng et al. [57] proposed a quick and efficient defect In manufacturing industries such as fabric or steel,
detection method for the detection of hole, stain, crease, quality control is an important factor for surface defect
hook wire, broken weft, and other defects by installing a detection. Huangpeng et al. [58] proposed an unsuper-
camera to capture a fabric image in a weaving circle system. vised approach for automatic defect detection of struc-
)eir proposed method relies on Blob, Canny, and Rotating tured or nearly structured textured surface using W-LRR
Integral algorithm which are used for feature point detec- (weighted low-rank reconstruction) method based on
tion, contour identification, and grayscale integration ap- texture prior. Texture prior map of image is calculated first
proach. Results are compared with other techniques used to and then W-LRR has been used to detect defects. )e
detect the flaws. )e false, missing, and recognition rates of limitations of the proposed method are as follows: (1) the
the proposed algorithm are 1%, 1%, and 98% which shows performance depends on the texture prior map to some
the effectiveness of the proposed approach. )e system with extent where the results for prior map are not best and the
Industry Ethernet normalized feature vectors serve as an input to SVDD. In

with POE
HMI Smart
the last step, for testing, new samples have been calculated
controller camera and labeled with classifier (SVDD). According to Zhou
et al. [62], as there is no standard dataset for experiment
and benchmark comparison, two basic types of fabric,
Smart
camera
twill fabric and plain fabric images with 256 gray levels of
Fabric
AC
winder
512 × 512 size, have been used. In these experiments, 10
contactor
different and common textures types with the aim to use
three defect types structural, intensity, and combination
of both are used. For evaluation purpose, 12 defect-free
Smart and 3 defective images from each type have been used and
Warp knitting
camera false detection rate (FDR) and correct detection rate
machine motor (CDR) have been measured. )ese experiments are per-
formed on OS Linux and MATLAB R2009b is used for the
implementation of this research [62].
Figure 9: Architecture of the automated inspection system. For the quality assurance of a textile fabric, auto in-
spection of defects is mandatory. Lizarraga et al. [63]
proposed a rough-set-based textile analysis (RTA) for
detection result will be affected and (2) the proposed defect detection in pattern-based textile fabric while using
method takes defect as a foreground. Synthetic and real local binary features. In this approach, for feature ex-
experiments have been performed for the effectiveness of traction, local binary features (LBPs) and coordinated
the proposed method. Synthetic experiments are per- clusters representation (CCR) are separately extracted
formed with different patch sizes. Computational accu- from the same image. For the evaluation of the proposed
racy has been measured with the objective to find out the system, defective and defect-free images are used and the
best patch size for further experiment and 10 synthetic process is performed using two phases: testing phase and
defective images have been tested. In experiments, training phase. )e standard UHK dataset provided by the
benchmark fabric defect dataset (which contains five types University of Hong Kong has been used and it has images
of defects) consisting of 25 defective and 25 nondefective for the most common defects in textile which are hole (H),
images from each pattern belonging to dot, star, and box is thin bar (TnB), broken end (BE), thick bar (TkB), and
used. Huangpeng et al. performed quantitative compar- multiple netting (MN)). For training [63], 500 mixed
ison for detection accuracy in terms of F-measure, true- images are used, and from each pattern, 25 defective
positive rate (TPR), accuracy (Acc), false-positive rate images are used for testing. DSR (detection success rate)
(FPR), and positive predictive value (PPV). )e result of and average success rate (ASR) for different patterned
the proposed method has been compared with that of the fabrics are calculated and the result has been compared
five state-of-the-art methods and it shows that it has better with the state-of-the-art methods. Analyzing time for one
detection accuracy and computational efficiency. )e image using WEKA/MATLAB is 60 ms on an ordinary
experiments are performed on Intel (r) Core (TM) Core i3 Intel processor.
2.00 GHz and MATLAB is used for implementation. For automatic defect detection, Feng et al. [64] presented
Figure 10 shows the architecture of PNN (probabilistic a model based on Small-Scale Over-Completed Dictionary
neural network) model and Figure 11 shows the classi- (SSOCD) using sparse coding (SC) which is basically
fication model based on PNN. Table 7 represents a hardware accelerated system. From the industrial environ-
summary about feature fusion-based approaches pro- ment, fabric images are acquired and then training SSOCD
posed for fabric defect detection. of SC of defect-free fabric images takes place (in training,
images are preprocessed using Gabor filter and patches are
10. Sparse-Based Operation extracted). )en the features are extracted by measuring
coding co-coefficients for inspection of defects. For analysis
)e use of sparse image representation has shown good and experiments, TILDA textile dataset of twill fabric has
results in various automation processes [59–61]. Due to been used and results are compared with other methods.
this benefit of sparse representation, Zhou et al. [62] Detection ratio has been used as an evaluation index. For
proposed a sparse dictionary reconstruction technique to performance evaluation, results have been tested while using
detect defects in fabric. To detect anomalies, support different number of cores and on platforms. )e result shows
vector data description (SVDD) has been used as SVDD is that the proposed method can inspect the defects with in-
more versatile in defining decision boundaries. As a first dustrial detection speed and with effective parallel efficiency.
step, the sample of the normal image is subdivided into Intel Core 2 Duo CPU E7300 is used as hardware with a
patches, and then with a defined parameter, the dictionary clock frequency of 2.27 GHz. )e proposed method is
learning process is completed. In the second step, these transplanted on TMS320C6678 with single-core and mul-
patches of defect-free images are used as training, and ticore series. Table 8 represents a summary about sparse
from these images, two features (correlation coefficient feature-based approaches proposed for fabric defect
and standard deviation) have been extracted and detection.
X1
X2
X3 Output
Xn
Feature Input Hidden Summation Output

vector layer layer layer layer
Figure 10: Simple architecture of PNN.
11. Morphology-Based Approaches further evaluated for real-time analysis on prototyped defect
inspection system (which is a low-cost system and has been
Morphological image processing is a collection of nonlinear developed) where twill weaving fabric containing many
operations related to the shape or morphology of features in an defects has been used for the experiment and 276 image
image [65]. To detect cord fabric defects from an image, Wang frames have been used and analyzed. Both frameworks are
et al. [66] presented a method based on morphological oper- installed on a Pentium four PC with 2.9 GHz and 1 GB RAM.
ations and discussed some traditional methods of edge detection. According to Jayashree et al. [68], manual and electronic
To enhance the effectiveness of the proposed research [66], the inspection of defects in gray plain fabric fails to find out
original image of cord fabric serum spot defects is used and the micronatured defects. So for defect detection in fine gray
results of all kinds of edge detection are discussed (differential, plain weave fabric, Jayashree et al. [68] proposed a hybrid
Laplace, Sobel, morphological dilation, morphological erosion, approach (HA) based on correlation approach (CA) fol-
and improved morphological erosion detection methods). )us, lowed by morphological approach (MA). Using this hybrid
result comparison showed that proposed improved shape approach, false alarm rate (FAR) has been reduced.
erosion method is fast, accurate, and good for preserving edge Jayashree et al. [68] performed experiments on GFDD for
details and noise elimination effective function. HA and CA and MA separately to see the effect of this
A novel defect detection technique using morphological approach and the result has been compared in terms of
filters (MF) has been proposed by Mak et al. [67] for the overall detection accuracy (ODA) using DSA (defect search
inspection of woven fabrics. In this proposed technique, algorithm). )e improvement of HA can be seen from the
pretrained GWN (Gabor wavelet network) has been used to comparison although this method identifies different defects
extract features of the textile fabric and then these extracted and has improved results, but the drawback of this HA
features have been used in the structuring element which is approach is the selection of the correct defect model and the
used further to perform some morphological operations to repetitive selection of element. Table 9 represents a summary
differentiate the defects and for the removal of background. about morphology-based approaches proposed for fabric
Morphological operations have been used as it is clear from defect detection. Morphological-based operations are suit-
the previous research study that false alarm rate can be able for nonperiodic fabric defect type and they are shape-
reduced by using MF [67]. Extensive off-line experiments sensitive and best suited for segmentation.
have been performed on 78 images gained from the Manual
of Standard Fabric Defects in the Textile Industry for the 12. Deep Learning-Based Approaches
evaluation of the proposed method which contain many
fabric defects and the following parameters for evaluation Deep neural networks are considered as a new framework
are measured: true detection (TD), false alarm (FA), and for modelling biological vision [69]. )e range of the vision
missed detection (MD). Moreover, this (technique) has been tasks tackled by the deep nets is expanding rapidly and they
Image
database
RGB
histogram
Feature f m1 f m2 f m3 Sm1 Sm2 Sm3 km1 km2 km3

database
Probabilistic
neural
network
Classes C1 C2 C3 C4 CN
Figure 11: Image classification-based model using PNN.
do represent a quantum leap as compared to earlier com- good results in various fields of computer vision including
puter vision systems. Deep networks can approximate textile image analysis [73–75].
functions and dynamics by learning from examples [70]. According to Mei et al. [17], traditional computer vision
Most of the deep learning models such as convolutional techniques are not robust to handle the existing require-
neural networks, recurrent neural networks, and deep belief ments of textile industry. Inspired from the recent trends
networks are dominating since the last decade of research about the use of deep learning in various computer vision-
[71]. )ese models have evolutionary paths lasting more based problems, they applied an automated unsupervised
than half a century and have a diversity of directions. For deep network approach to detect and localize the fabric
example, convolutional neural network (CNN) is evolved defects without manual intervention.
from the prior knowledge of biological vision systems [72]. Wei et al. [76] proposed a modified faster regional-based
)e initial development of neural networks fulfills the thrust convolutional neural network (faster RCNN) that is based on
to simulate human brains. )e maturity of these shallow the structure of VGG net. )e network is modified to enhance
networks (biological vision system) into deep architectures the performance while using the fabric defect benchmark. )e
significantly reduces the requirement of resources while fabric dataset is expanded to reduce the effect of input data.
retaining the representation power. )e recent focus of ReLU activation function is used with pooling and convo-
research is on the use of deep learning models and they have lution layers to extract the image feature map. Region
Table 7: A summary on fusion feature-based operations.

Four light fabric samples are used Proposed two methods to detect seam
Results on 100 test samples have been
(made up of polyester filament); the using images, textural analysis by
Zhu et al. used to compare the results of both
first three are with same color of fabric wavelet energy method and by features
[52] proposed methods in terms of accuracy
while the fourth one is with different extraction in CIELAB color space in the
and time
color of fabric image
Proposed a new technique for real-time
Defect-free and defective (fabric)
automated defect detection in textile
Jing et al. images which have been taken from the Computational cost and high detection
cloth which is based on the multiple
[53] TILDA textile texture database have rate have been used
Gabor filters and kernel principal
been used
component analysis.
Proposed a technique for quick and
Peng et al. )e false, missing, and recognition rate efficient defect detection method which
Not mentioned
[57] have been used to compare the result is based on Blob, Canny, and Rotating
Integral algorithm
Proposed method that relies on
12 samples of textures (256 gray levels)
multiscale wavelet transform and
Li et al. [55] with 23 defects from the TILDA textile No specific criteria discussed
Gaussian mixture model based on EM
texture database are used
algorithm
Results have been compared in terms of
Proposed a hybrid method for fabric
accuracy, positive predictive value, and
Li et al. [56] Only two images are used inspection based on Gabor filter and
true-positive rate with the other two
pulse coupled neural network (PCNN)
methods (Gabor and wavelet)
Synthetic and real experiments have Accuracy (Acc), F-measure, true-positive
Huang- Proposed an unsupervised approach for
been performed; for real experiment, rate (TPR), false-positive rate (FPR), and
peng et al. automatic defect detection using the W-
benchmark fabric defect dataset has positive predictive value (PPV) have been
[58] LRR method based on texture prior
been used used to quantify the detection accuracy
Table 8: A summary on sparse-based operations.

No standard dataset has been used. 12
Proposed a sparse dictionary
Zhou et al. defect-free and 3 defective images from False detection rate (FDR) and correct
reconstruction technique to detect
[62] each type of fabric: twill and plain detection rate (CDR) have been used
defects in fabric
images have been used
DSR (detection success rate) and average
Lizarraga- Proposed a rough-set-based textile
)e standard UHK dataset provided by success rate (ASR) for different patterned
Morales et al. analysis (RTA) for defect detection
the University of Hong Kong is used fabric have been calculated and result has
[63] in pattern-based textile fabric
been compared with the state of the art
Detection ratio has been used as an
Presented a model based on Small-
Feng et al. TILDA textile dataset of twill fabric has evaluation index And for performance
Scale Over-Completed Dictionary
[64] been used evaluation, results have been tested while
(SSOCD) using sparse coding (SC)
using different number of cores
Table 9: A summary about morphology-based approaches.

Performance evaluation
Author Dataset Proposed model and purpose
criteria
Presented improved morphological
Wang et al. Original image of cord fabric serum spot Edge detection with influence
erosion detection method based on
[66] defects is used of noise
morphological operations
For off-line evaluation: 78 fabric images have
been used from the Manual of Standard Fabric True detection (TD), false Proposed a novel defect detection
Mak et al.
Defects in the Textile Industry, and for real- alarm (FA), and missed technique using morphological filters
[67]
time evaluation: a long piece of twill weaving detection (MD) (MFs) for the inspection of woven fabrics
fabric has been used.
)e results have been Proposed a hybrid approach based on
Jayashree 115 benchmark normal samples and 75 warp
compared in terms of overall correlation approach followed by
et al. [68] break samples have been used
detection accuracy (DA) morphological approach
proposal network (RPN) is used for the generation of evaluated by using fabric defects (FBDFs), and it consists of
bounding box regression and foreground anchors to compute 2575 optical fabric images covered by 20 categories and 4193
proposals. In the last step, the computed proposals are fed to defect instances with a spatial resolution of 2446 × 1000
the ROI pooling layers so that they can be used by the network pixels.
of softmax layer to classify the images. To reduce the over- Wei et al. [80] proposed a method for defect classifi-
fitting, vertical and horizontal flips are used to augment the cation using CNN with compressing sensing based on small
textile dataset. )e fabric defects detected are broken pick, sample sizes. According to Wei et al. [80], CNN yet not
felter, drawback, sundries, broken end, and oil stains. )e achieved good classification results on small sample sizes;
authors have created their own dataset with defects to evaluate that is why the compressive sensing approach is being used
the proposed model. )e created dataset consists of 06 classes with CNN. Fabric defect image (FDI 500) dataset is used for
with 135 fabric defects per class. )e proposed research is the experiment and the images are acquired from industrial
compared with traditional computer vision approaches, and monitors of textile factory [80]. )e sample images contain
experimental results show the effectiveness of the proposed normal, slub, oil stain, broken end, felter, double flat, and
faster RCNN model. High training cost is one of the draw- mispick images. For the implementation of method and for
backs of the proposed faster RCNN model [76]. evaluation purpose, the operating system Ubuntu 14.04 with
Hu et al. [77] proposed unsupervised learning approach TensorFlow 0.8.0, 128 GB RAM, and 4 NVIDIA GeForce
based on deep convolutional generative adversarial network 1080 GPUs is used. Classification accuracy is calculated and
(DCGAN) to locate the surface defects for texture. )e has been compared with many other models. )e reported
standard DCGAN is extended by using a new encoder block classification accuracy of the proposed model is 97.9%.
that can reconstruct query image with no defect and normal Jeyaraj et al. [81] proposed an innovative method for the
texture. )e reconstructed images are then subtracted from detection of defects based on advance deep learning approach.
the image with defect to find the defected regions. )e defects ResNet512-based CNN is used to learn the features from
are located in the image by using probability of occurrence on images, and according to Jeyaraj et al., the proposed method
the basis of likelihood map. A fusion map is created by using also localizes the minute defects. For experiments, i7 with
likelihood and residual map. )e resultant map creates 8 GHz (frequency) 1 TB disk capacity and NVIDIA GPU is
uniform gray levels for the defect-free regions while devia- used. TILDA textile dataset is used for experiments, and many
tions over the defected area. )e proposed research is eval- quantitative values are computed. )e classification accuracy
uated by using a variety of real fabric defect samples. is used to compute and to validate the accuracy of the method
According to Liu et al. [78], the high computational cost and results have been compared with some other classifiers
and training time are main drawbacks associated with the also as SVM and Bayesian [81]. Table 10 represents a sum-
use of deep learning-based approaches. )ese drawbacks mary about deep learning-based approaches proposed for
limit the applications of deep learning-based models for fabric defect detection.
embedded and handheld devices. To reduce the computa- For minor defect detection, Wen et al. [82] presented a
tional cost, authors [78] proposed a computational efficient CNN-based method patch extractor triple matric (PETM)
CNN-based framework (named as YOLO-LFD) that is using which is self-similar estimation algorithm, and to preprocess it,
a reduction in feature dimensions. K-means dimension most likely defective patches are extracted from the original
reduction is applied for the computation of ground trues and image. )e minor fabric defect images (MFDI) that contain
initial anchor boxes. )e proposed research provides a 100 normal and 100 minor defective images dataset is built. )e
compression of YOLO-v2 model 241 MB to 8 MB with an size of patched Np � 10, and n � 4 with 40 × 40 is used. From
increase in detection speed with a multiplicative factor of 100 defective fabric images, 901 normal patches and 99 de-
2.6x. COCO dataset is used in this research to pretrain the fective patches are extracted. )e true-positive rate (TPR), true-
deep network, and fine-tuning of network is performed by negative rate (TNR), and accuracy (Acc) are used as evaluation
using a fabric defect dataset that addressed the problem of parameters, and accuracy has been compared with some other
limited training samples. Fabric image benchmarks with methods. )e reported accuracy for each element of the
3000 samples and five classes are used for the evaluation of proposed method is 79.32%.
proposed research and results are compared with SSD, Jin et al. [83] proposed Mobile-Unet method based on
YOLO-v3, and Tiny-Yolo-v3 models using the same envi- CNN for fabric defect detection. )e proposed method is
ronment and parameters. 2000 images were used for training efficient as it achieves pixel-level classification of defects and
while 1000 were selected for testing purpose. )e proposed best suited for online FDD. To improve the detection process
research has outperformed the state of the art in terms of and accuracy, median frequency balancing loss function is
input parameters and defect detection speed. On the basis of used. Intel Core i7-7700K 4.2 GHz, with 128 GB RAM,
the experimental result, it will be safer to conclude that the NVIDIA GeForce 1080Ti, and PyTorch deep learning
proposed research can be used in an embedded device to framework on Windows 10 OS is used as the platform for this
detect fabric defects in industry [78]. Figure 12 shows an research. )e benchmarks named as Fabric Images (FI) and
example of classification by VGG-16 model. Wu et al. [79] Yarn-dyed Fabric Images (YFI) are used to train and test the
proposed an automated approach for fabric defect detection proposed method. Recall, precision, F1-measure, and inter-
that can meet industrial requirement. To achieve the main section over union (IoU) are used as evaluation matrices.
objective, a stacked feature pyramid is used with deformable Learning-based approaches require intensive computation
convolution filter to sort defect patterns. Approach is for feature vectors having large size. )ey serve as black box
Input [224 × 224] × 3
Conv {3 × 3],/1
Conv [3 × 3],/1
Conv [2 × 2],/2
Conv [3 × 3],/1
Conv [3 × 3],/1
CNN layers
Conv [2 × 2],/2
Conv [3 × 3],/1
Conv [3 × 3],/1
Conv [2 × 2],/2
Conv [3 × 3],/1
Conv [3 × 3],/1
Conv [3 × 3],/1
Conv [2 × 2],/2
Conv [3 × 3],/1
Conv [3 × 3],/1
Conv [3 × 3],/1
Conv [2 × 2],/2
Fully connected layers[4069]

Fully connected layers
Fully connected layers[4069]
Fully connected layers [5/7]
C1 C2 C3 C4 CN
Convolution + ReLU Fully connected + ReLU
Max pooling Softmax
Figure 12: Example of classification by VGG-16 model.
character, with abundance of feature, and there is difficulty in measure the defect detection performance. Commonly,
copying and they are not suited for the detection of small defects. there are two such ways that are being mostly used which are
(a) sensitivity and specificity and (b) detection success rate.
13. Performance Evaluation Criteria Some other metrics used are false alarm rate, detection rate,
and positive predictive value. Many authors used these
For fabric defect detection, the performance can be evalu- terms/metrics to evaluate their performance, but the se-
ated in many ways as there is no standard way or method to lection of the “evaluation criteria” mainly depends on the
Table 10: A summary about deep learning-based approaches.

Fabrics [84] with 62 samples, KTH-TIPS
Automated unsupervised deep network
Mei et al. [85] with 128 samples, Kylberg Texture
Precision, recall, and F1-measure approach to detect and localize the fabric
[17] [86] with 132 samples, and ms-Texture
defects
[17] with 50 samples
A modified faster regional-based
Wei et al. )e self-created dataset consists of 06 Computation time to detect fabric defect convolutional network method (faster
[76] classes with 135 fabric defects per class and accuracy RCNN) that is based on the structure of
VGG net
ROC graph analysis based on the true-
positive rate (TPR, detection rate), the
false-negative rate (FNR, missed An unsupervised learning approach based
Hu et al. Ngan et al. [87, 88], the TILDA textile detection), the false-positive rate (FPR, on deep convolutional generative
[77] [89], a self-built dataset [77] false alarm rate), the positive predictive adversarial network (DCGAN) to locate
value (PPV, precision), and the accuracy the surface defects for texture
(Acc, comparison of the time, detection
performance) is used
Authors proposed a computational
Fabric image benchmarks with 3000
efficient CNN-based framework (named
samples and five classes are used for the
as YOLO-LFD) that is using a reduction
Liu et al. evaluation of proposed research and Accuracy, speed in milliseconds, model
in feature dimensions and K-means
[78] results are compared with SSD, YOLO- size in MB
dimension reduction is applied for the
v3, and Tiny-Yolo-v3 models using the
computation of ground trues and initial
same environment and parameters
anchor boxes
FBDF consists of 2575 images covered
A stacked feature pyramid is used with
Wu et al. by 20 defect categories and 4193 Accuracy and variance precision that
deformable convolution filter to sort
[79] instances with a spatial resolution of illustrate the interclass accuracy stability
defect patterns based on deep learning
2446 × 1000 pixels
Proposed a method for defect
Wei et al. Fabric defect images (FDI 500) dataset is classification using CNN with
Classification accuracy is used
[80] used compressing sensing based on small
sample sizes
Proposed innovative method for
Jeyaraj
Standard TILDA textile dataset is used Classification accuracy is used detection of defects based on advance
et al. [81]
deep learning approach
)e minor fabric defects images (MFDI) )e true-positive rate (TPR), true-
Wen Presented a CNN-based method PETM
that contain 100 normal and 100 minor negative rate (TNR), and accuracy (Acc)
et al. [82] (patches extractor triple matric)
defective images dataset is built are used as evaluation parameters
)e benchmarks named as Fabric Recall, precision, F1-measure, and
Jin et al. Proposed Mobile-Unet method based on
Images (FI) and Yarn-dyed Fabric intersection over union (IoU) are used as
[83] CNN for fabric defect detection
Images (YFI) are used evaluation matrices
nature of the approach, application domain, and environ-

ment and on other factors. Before going into depth of these Detected as
criteria, we must know some basic terms that are used to Defective Defect-free
measure these terms such as true positive (TP), true negative
(TN), false positive (FP), and false negative (FN) that are
discussed below and are shown in Figure 13.
(1) TP: defective images which are detected or identified Defective True positive False negative
as defective (TP) (FN)
Images
(2) TN: defect-free images which are detected or iden-

tified as defect-free
(3) FP: defect-free images which are detected or iden- Defect-free False positive True negative
tified as defective (FP) (TN)
(4) FN: defective images which are detected or identified
as defect-free
Figure 13: Basic explanation of true positive (TP), true negative
)us, (TN), false positive (FP), and false negative (FN).
Sensitivity: it is defined as the correct identification of

defective samples and represented as follows: Total no. of sample images detected correctly
DSR � .
Total no. of sample images
TP (3)
sensitivity � . (1)
TP + FN
It is known as detection accuracy (DA) and can be
It is also known as recall. expressed as follows:
Specificity: it is defined as correct detection of non-
defective samples and expressed as follows:
TN + TP
detection accuracy(DA) � DSR � .
TN + FP + FN + TP
TN (4)
specificity � . (2)
TN + FP
False alarm rate (FAR): it can be calculated as
Detection success rate (DSR): it can be calculated as
follows:
no. of nondefective image samples detected as defective

false alarm rate(FAR) � . (5)
no. of defect − free image samples
Detection rate (DR): it can be calculated as follows:
no. of defective image samples detected correctly

detection rate(DR) � . (6)
no. of defective image samples
Positive predictive value (PPV): it can be calculated as )e imbalanced detection is determined by average rec-
ognition precision and variance fluctuation between different
categories [79]. Precision for a single class predicts a model’s
TP ability to classify only the relevant objects. However, recall gives
positive predictive value (PPV) � . (7) a measure of the ability of the model to identify all of the
TP + FP relevant cases. )e imbalance between inter- and intraclass
fabric defects results in different values of average precision
It is also known as precision. (AP) among categories. Firstly, mean AP (mAP) is determined
Li et al. [22] used accuracy, precision, and recall and for overall performance evaluation of detectors which refers to
Zhang et al. [26] used detection rate. Hanbay et al. [42] used the mean of AP of all categories. Secondly, since the data
five different indexes for quantitative evaluation, i.e., spec- distribution varies significantly among different classes termed
ificity, sensitivity, accuracy, negative predictive value, and as interclass imbalance, ROC curve is considered a better
positive predictive value. Hamdi et al. [40] used precision, choice as compared to the PR curve as it takes into account
recall, and detection success rates. Hu et al. [44] used false- both +ve and −ve examples. To compute the imbalanced
positive rate (FPR), true-positive rate (TPR), and accuracy detection between different categories, variance precision (VP)
(Acc). Zhang et al. [49] used percentage detection rate. is used as shown in equation (8), which illustrates the interclass
Accuracy (Acc), F-measure, true-positive rate, false-positive accuracy stability:
rate, and positive predictive value have been used to quantify 􏽶��
􏽴
the detection accuracy [58]. Hu et al. [77] used receiver 1 C 2
operating characteristic (ROC) graphs analysis based on the VP � 􏽘 APk − mAP􏼁 . (8)
C k�1
true-positive rate (TPR, detection rate), the false-negative
rate (FNR, missed detection), the false-positive rate (FPR,
false alarm rate), the positive predictive value (PPV, pre- )irdly, the intraclass imbalance is computed that relies
cision), and the accuracy (Acc, comparison of the time, on size variance and the AP for large, medium, and small
detection performance). objects is divided by cross scale 962 and 2562 .
14. Limitations and Future Work region of textile fabric. )e current trends are shifted to the
use of deep learning models that require high computational
Traditionally, the quality of textile fabrics is determined by cost for training and lot of training data. )e use of feature
human-oriented analysis of textile fabric defects. However, fusion such as color and texture with deep learning can be a
the manual inspection leads to lower productivity and higher possible future direction. Real-time defect detection is still
market losses. In this study, automated fabric defect de- an open research area in this domain, and only few research
tection methods are discussed in eleven groups. Since there models are reported for real-time applications. In future,
are many different types of fabrics and defects, a single defect detection using small handheld devices with less
method that can run on all fabric types and contain these computational cost will attract textile industry. According to
defects has not been found. Each method has its strengths this review, the research models based on fabric defect
and weaknesses that are discussed in subsequent sections. In detection are evaluated by using different image benchmarks
the vast majority of the methods examined, the authors have that are created by the users according to their requirements.
created their own database under different lighting condi- One of the limitations of existing research is the number of
tions. Some studies have used the TILDA fabric database, publicly available datasets except TILDA textile dataset. It
but it is not easily accessible for all researchers as it is paid. will be good if some common image benchmarks will be
Anonymously accessible fabric databases are needed to created and shared at pubic repertoires so that researchers
develop objective and reliable methods. Since a large number can use them as standard for further research. )is will
of fabric defect detection methods exist in the literature, provide an opportunity in a way that researchers will be able
their comparison is helpful for researchers to find the op- to compare their research to sort out the best approach. All
timal method depending on fabric type and defect. However, of these are the possible future research directions for au-
it should be considered that the studies are conducted using tomated fabric defect detection.
different databases, different parameters, and varied imaging
systems, hence making the validity and reliability of methods Data Availability
far from objectivity. )is review article covers the state-of-
the-art research on textile fabric defect detection methods, All the details about data availability is mentioned within
and it is thought that comprehensive studies similar to this this manuscript.
one will contribute significantly to the textile industry. For
future research, there is a critical need for some common Conflicts of Interest
publicly available fabric image databases that will open up
new directions for the researchers in textile industry. Hybrid )e authors declare no conflicts of interest.
approaches with a combination of deep features with other
visual features are also considered as a promising direction References
for future research.
[1] A. Latif, A. Rasheed, U. Sajid et al., “Content-based image
retrieval and feature extraction: a comprehensive review,”
15. Conclusion Mathematical Problems in Engineering, vol. 2019, Article ID
9658350, 2019.
In this review article, we have presented an overview about [2] Y. Li, H. Luo, M. Yu, G. Jiang, and H. Cong, “Fabric defect
fabric defect detection for textile industry. According to the detection algorithm using RDPSO-based optimal Gabor fil-
literature, the process based on manual human inspection is ter,” ;e Journal of the Textile Institute, vol. 110, no. 4,
not sufficient and effective to meet the current need of textile pp. 487–495, 2019.
industry. To enhance the quality and to decrease the pro- [3] W. Wong and J. Jiang, “Computer vision techniques for
duction cost of the final textile product, it is the current detecting fabric defects,” in Applications of Computer Vision
in Fashion and TextilesElsevier, Amsterdam, Netherlands,
requirement that the inspection process must be done by
2018.
using some industrial automation. Computer vision and [4] T. Czimmermann, G. Ciuti, M. Milazzo et al., “Visual-based
digital image processing can provide a base that can cover defect detection and classification approaches for industrial
this industrial gap. We have discussed the techniques based applications-a survey,” Sensors, vol. 20, no. 5, p. 1459, 2020.
on eleven different subgroups with performance criteria for [5] P. Dario, S. Kolhe, and P. Patil, “A review of automatic fabric
fabric defect detection. )e basic theme of each approach is defect detection techniques,” Advances in Computational
discussed with advantages and disadvantages while the Research, vol. 1, no. 2, pp. 18–29, 2009.
tabular representation is used to summarize the research. [6] H. Y. T. Ngan, G. K. H. Pang, and N. H. C. Yung, “Automated
)e pictorial representation is used in the review article to fabric defect detection-a review,” Image and Vision Com-
provide an overview of the popular research models for puting, vol. 29, no. 7, pp. 442–458, 2011.
fabric defect detection. )e image benchmarks used in each [7] E. W. T. Ngai, S. Peng, P. Alexander, and K. K. L. Moon,
“Decision support and intelligent systems in the textile and
research model are discussed in detail.
apparel supply chain: an academic review of research articles,”
According to this review, significant research has been Expert Systems with Applications, vol. 41, no. 1, pp. 81–91,
reported to detect by using texture, frequency domain, 2014.
GLCM, feature fusion, sparse feature representation, image [8] K. Hanbay, M. F. Talu, and Ö. F. Ö F.ay/, “Fabric defect
morphology, and deep learning-based approaches. For any detection systems and methods-A systematic literature re-
defect model, the main requirement is to sort the defective view,” Optik, vol. 127, no. 24, pp. 11960–11973, 2016.
[9] J. Wen and W. Wong, “Fundamentals of common computer [25] A. Alzu’bi, A. Amira, and N. Ramzan, “Semantic content-
vision techniques for fashion textile modeling, recognition, based image retrieval: a comprehensive study,” Journal of
and retrieval,” in Applications of Computer Vision in Fashion Visual Communication and Image Representation, vol. 32,
and TextilesElsevier, Amsterdam, Netherlands, 2018. pp. 20–54, 2015.
[10] D. Siegmund, T. Samartzidis, B. Fu, A. Braun, and A. Kuijper, [26] K. Zhang, Y. Yan, P. Li, J. Jing, X. Liu, and Z. Wang, “Fabric
“Fiber defect detection of inhomogeneous voluminous tex- defect detection using salience metric for color dissimilarity
tiles,” in Mexican Conference on Pattern RecognitionSpringer, and positional aggregation,” IEEE Access, vol. 6, pp. 49170–
Berlin, Germany, 2017. 49181, 2018.
[11] M. F. Nisha, P. Vasuki, and S. M. M. Roomi, “Survey on [27] P. Jaglan, R. Dass, and M. Duhan, “A comparative analysis of
various defect detection and classification methods in fabric various image segmentation techniques,” in Proceedings of the
images,” Journal of Environmental Nanotechnology, vol. 6, 2nd International Conference on Communication, Computing
no. 2, pp. 20–29, 2017. and Networking, pp. 359–374, Chandigarh, India, 2019.
[12] A. Kumar, “Computer-vision-based fabric defect detection: a [28] S. S. Chouhan, A. Kaul, and U. P. Singh, “Image segmentation
survey,” IEEE Transactions on Industrial Electronics, vol. 55, using computational intelligence techniques: review,” Ar-
no. 1, pp. 348–363, 2008. chives of Computational Methods in Engineering, vol. 26, no. 3,
[13] S. N. Niles, S. Fernando, and W. D. G. Lanerolle, “A system for pp. 533–596, 2019.
analysis, categorisation and grading of fabric defects using [29] H. J. He, C. Zheng, and D. W. Sun, “Image segmentation
computer vision,” Research Journal of Textile and Apparel, techniques,” in Computer Vision Technology for Food Quality
vol. 19, no. 1, p. 59, 2015. EvaluationElsevier, Amsterdam, Netherlands, 2016.
[14] J. Zhang, J. Wang, R. Pan, J. Zhou, and W. Gao, “A computer [30] Y. Guo, Y. Liu, T. Georgiou, and M. S. Lew, “A review of
vision-based system for automatic detection of misarranged semantic segmentation using deep neural networks,” Inter-
warp yarns in yarn-dyed fabric: part I: continuous segmen- national Journal of Multimedia Information Retrieval, vol. 7,
tation of warp yarns,” ;e Journal of the Textile Institute, no. 2, pp. 87–93, 2018.
vol. 109, no. 5, pp. 577–584, 2018. [31] M. Guan, Z. Zhong, and Y. Rui, “Automatic defect seg-
[15] J. Zhang, J. Wang, and R. Pan, “A computer vision-based mentation for plain woven fabric images,” in Proceedings of
system for automatic detection of misarranged warp yarns in the 2019 International Conference on Communications, In-
formation System and Computer Engineering (CISCE),
yarn-dyed fabric: part II: warp region segmentation,” ;e
pp. 465–468, Haikou, China, July 2019.
Journal of the Textile Institute, vol. 110, no. 9, pp. 1359–1367,
[32] C. Li, R. Yang, Z. Liu, G. Gao, and Q. Liu, “Fabric defect
2019.
detection via learned dictionary-based visual saliency,” In-
[16] J. Wang, J. Zhang, L. Wang, R. Pan, J. Zhou, and W. Gao, “A
ternational Journal of Clothing Science and Technology, vol. 28,
computer vision-based system for automatic detection of
no. 4, pp. 530–542, 2016.
misarranged color warp yarns in yarn-dyed fabric: part III:
[33] Z. Zhou, C. Wang, X. Gao et al., Fabric Defect Detection and
yarn layout proofing,” ;e Journal of the Textile Institute,
Classifier via Multi-Scale Dictionary Learning and an Adaptive
pp. 1–9, 2020.
Differential Evolution Optimized Regularization Extreme
[17] S. Mei, Y. Wang, and G. Wen, “Automatic fabric defect
Learning Machine, Fibres & Textiles in Eastern Europe, Łódź,
detection with a multi-scale convolutional denoising Poland, 2019.
autoencoder network model,” Sensors, vol. 18, no. 4, p. 1064, [34] X. Xie, “A review of recent advances in surface defect de-
2018. tection using texture analysis techniques,” ELCVIA Electronic
[18] M. Pietikäinen, T. Mäenpää, and J. Viertola, “Color texture Letters on Computer Vision and Image Analysis, vol. 7, no. 3,
classification with color histograms and local binary patterns,” pp. 1–22, 2008.
in Proceedings of the Workshop on Texture Analysis in Ma- [35] D. Ping Tian et al., “A review on image feature extraction and
chine Vision, Florence, Italy, October 2002. representation techniques,” International Journal of Multi-
[19] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image media and Ubiquitous Engineering, vol. 8, no. 4, pp. 385–396,
Processing Using MATLAB, Pearson Education India, New 2013.
Delhi, India, 2004. [36] Z. Liang, B. Xu, Z. Chi, and D. Feng, “Intelligent charac-
[20] Y. Ye, “Fabric defect detection using fuzzy inductive rea- terization and evaluation of yarn surface appearance using
soning based on image histogram statistic variables,” in saliency map analysis, wavelet transform and fuzzy ARTMAP
Proceedings of the 2009 Sixth International Conference on neural network,” Expert Systems with Applications, vol. 39,
Fuzzy Systems and Knowledge Discovery, pp. 191–194, Tianjin, no. 4, pp. 4201–4212, 2012.
China, August 2009. [37] Z. Wen, J. Cao, X. Liu, and S. Ying, “Fabric defects detection
[21] W. Zhang, J. Zhang, Y. Hou, and S. Geng, “MWGR: a new using adaptive wavelets,” International Journal of Clothing
method for real-time detection of cord fabric defects,” in Science and Technology, vol. 26, no. 3, pp. 202–211, 2014.
Proceedings of the 2012 International Conference on Advanced [38] V. V. Karlekar, M. Biradar, and K. Bhangale, “Fabric defect
Mechatronic Systems, pp. 458–461, Tokyo, Japan, September detection using wavelet filter,” in Proceedings of the 2015
2012. International Conference on Computing Communication
[22] M. Li, S. Wan, Z. Deng, and Y. Wang, “Fabric defect detection Control and Automation, pp. 712–715, Pune, India, February
based on saliency histogram features,” Computational Intel- 2015.
ligence, 2019. [39] D. Yapi, M. S. Allili, and N. Baaziz, “Automatic fabric defect
[23] R. Datta, D. Joshi, J. Li, and J. Z. Wang, “Image retrieval,” detection using learning-based local textural distributions in
ACM Computing Surveys, vol. 40, no. 2, pp. 1–60, 2008. the contourlet domain,” IEEE Transactions on Automation
[24] A.-M. Tousch, S. Herbin, and J.-Y. Audibert, “Semantic hi- Science and Engineering, vol. 15, no. 3, pp. 1014–1026, 2017.
erarchies for image annotation: a survey,” Pattern Recogni- [40] A. A. Hamdi, M. S. Sayed, M. M. Fouad, and M. M. Hadhoud,
tion, vol. 45, no. 1, pp. 333–345, 2012. “Unsupervised patterned fabric defect detection using texture
filtering and K-means clustering,” in Proceedings of the 2018 [56] Y. Li and C. Zhang, “Automated vision system for fabric
International Conference on Innovative Trends in Computer defect inspection using Gabor filters and PCNN,” Spring-
Engineering (ITCE), pp. 130–144, Aswan, Egypt, February erPlus, vol. 5, no. 1, p. 765, 2016.
2018. [57] D. Peng, G. Zhong, Z. Rao, T. Shen, Y. Chang, and M. Wang,
[41] Y. Wu, J. Zhou, N. T. Akankwasa, K. Wang, and J. Wang, “A fast detection scheme for original fabric based on Blob,
“Fabric texture representation using the stable learned dis- canny and rotating integral algorithm,” in Proceedings of the
crete cosine transform dictionary,” Textile Research Journal, 2018 IEEE 3rd International Conference on Image, Vision and
vol. 89, no. 3, pp. 294–310, 2019. Computing (ICIVC), pp. 113–118, Chongqing, China, June
[42] K. Hanbay, M. F. Talu, Ö. F. Özgüven, and D. Öztürk, “Real- 2018.
time detection of knitting fabric defects using shearlet [58] Q. Huangpeng, H. Zhang, X. Zeng, and W. Huang, “Auto-
transform,” Tekstil Ve Konfeksiyon, vol. 29, no. 1, pp. 1–10, matic visual defect detection using texture prior and low-rank
2019. representation,” IEEE Access, vol. 6, pp. 37965–37976, 2018.
[43] N. Ismail, W. M. Syahrir, J. M. Zain, and H. Tao, “Fabric [59] D. G. Lowe, “Distinctive image features from scale-invariant
keypoints,” International Journal of Computer Vision, vol. 60,
authenticity method using fast Fourier transformation de-
no. 2, pp. 91–110, 2004.
tection,” in Proceedings of the International Conference on
[60] H. Bay, T. Tuytelaars, and L. Van Gool, “Surf: speeded up
Electrical, Control and Computer Engineering 2011 (InECCE),
robust features,” in Proceedings of the European Conference on
pp. 233–237, Kuantan, Malaysia, June 2011.
Computer Vision, Graz, Austria, May 2006.
[44] G.-H. Hu, Q.-H. Wang, and G.-H. Zhang, “Unsupervised [61] W. Eng, V. Koo, and T. Lim, “IPDDF: an improved precision
defect detection in textiles based on Fourier analysis and dense descriptor based flow estimation,” CAAI Transactions
wavelet shrinkage,” Applied Optics, vol. 54, no. 10, on Intelligence Technology, vol. 5, no. 1, pp. 49–54, 2020.
pp. 2963–2980, 2015. [62] J. Zhou, D. Semenovich, A. Sowmya, and J. Wang, “Sparse
[45] K. Sakhare, A. Kulkarni, M. Kumbhakarn, and N. Kare, dictionary reconstruction for textile defect detection,” in
“Spectral and spatial domain approach for fabric defect de- Proceedings of the 2012 11th International Conference on
tection and classification,” in Proceedings of the 2015 Inter- Machine Learning and Applications, pp. 21–26, Boca Raton,
national Conference on Industrial Instrumentation and FL, USA, December 2012.
Control (ICIC), pp. 640–644, Pune, India, May 2015. [63] R. A. Lizarraga-Morales, F. E. Correa-Tome, R. E. Sanchez-
[46] B. Zhang and C. Tang, “A method for defect detection of yarn- Yanez, and J. Cepeda-Negrete, “On the use of binary features
dyed fabric based on frequency domain filtering and similarity in a rule-based approach for defect detection on patterned
measurement,” Autex Research Journal, vol. 19, no. 3, 2018. textiles,” IEEE Access, vol. 7, pp. 18042–18049, 2019.
[47] J. L. Raheja, B. Ajay, and A. Chaudhary, “Real time fabric [64] T. Feng, L. Zou, J. Yan et al., “Real-time fabric defect detection
defect detection system on an embedded DSP platform,” using accelerated small-scale over-completed dictionary of
Optik, vol. 124, no. 21, pp. 5280–5284, 2013. sparse coding,” International Journal of Advanced Robotic
[48] F. Arnia and K. Munadi, “Real time textile defect detection Systems, vol. 13, no. 1, p. 1, 2016.
using GLCM in DCT-based compressed images,” in Pro- [65] J. Deng, Z. He, G. Weng, L. Sun, B. Zuo, and C. Wang,
ceedings of the 2015 6th International Conference on Modeling, “Detection of chemical fabric defects on the basis of mor-
Simulation, and Applied Optimization (ICMSAO), pp. 1–6, phological processing,” ;e Journal of the Textile Institute,
Istanbul, Turkey, May 2015. vol. 107, no. 2, pp. 233–241, 2016.
[49] L. Zhang, J. Jing, and H. Zhang, “Fabric defect classification [66] D. Wang and H. Liu, “Edge detection of cord fabric defects
based on LBP and GLCM,” Journal of Fiber Bioengineering image based on an improved morphological erosion detection
and Informatics, vol. 8, no. 1, pp. 81–89, 2015. methods,” in Proceedings of the 2010 Sixth International
[50] D. Zhu, R. Pan, W. Gao, and J. Zhang, “Yarn-dyed fabric Conference on Natural Computation, pp. 3943–3947, Yantai,
defect detection based on autocorrelation function and China, August 2010.
GLCM,” Autex Research Journal, vol. 15, no. 3, pp. 226–232, [67] K. L. Mak, P. Peng, and K. F. C. Yiu, “Fabric defect detection
using morphological filters,” Image and Vision Computing,
2015.
vol. 27, no. 10, pp. 1585–1592, 2009.
[51] X. Zhang and X. Fan, “Fabric defect detection based on GLCM
[68] V. Jayashree and S. Subbaramn, “Hybrid Approach using
approach,” in Proceedings of the 6th International Conference
correlation and morphological approaches for GFDD of plain
on Information Engineering for Mechanics and Materials,
weave fabric,” in Proceedings of the 2012 IEEE Control and
Huhhot, Inner Mongolia, July 2016. System Graduate Research Colloquium, pp. 197–202, Shah
[52] B. Zhu, J. Liu, R. Pan, S. Wang, and W. Gao, “Fabric seam
Alam, Selangor, Malaysia, July 2012.
detection based on wavelet transform and CIELAB color [69] N. Kriegeskorte, “Deep neural networks: a new framework for
space: a comparison,” Optik, vol. 126, no. 24, pp. 5650–5655, modeling biological vision and brain information processing,”
2015. Annual Review of Vision Science, vol. 1, no. 1, pp. 417–446,
[53] J. Jing, X. Fan, and P. Li, “Automated fabric defect detection 2015.
based on multiple Gabor filters and KPCA,” International [70] N. Kriegeskorte and T. Golan, “Neural network models and
Journal of Multimedia and Ubiquitous Engineering, vol. 11, deep learning,” Current Biology, vol. 29, no. 7, pp. R231–R236,
no. 6, pp. 93–106, 2016. 2019.
[54] Pattern Recognition and Image Processing, https://lmb. [71] R. Ding, L. Dai, G. Li, and H. Liu, “TDD-net: a tiny defect
informatik.uni-freiburg.de/resources/datasets/tilda.en.html. detection network for printed circuit boards,” CAAI Trans-
[55] P. Li, H. Zhang, J. Jing, R. Li, and J. Zhao, “Fabric defect actions on Intelligence Technology, vol. 4, no. 2, pp. 110–116,
detection based on multi-scale wavelet transform and 2019.
Gaussian mixture model method,” ;e Journal of the Textile [72] H. Wang and B. Raj, “On the origin of deep learning,” 2017,
Institute, vol. 106, no. 6, pp. 587–592, 2015. http://arxiv.org/abs//170207800.
[73] L. Deng, An Overview of Deep-Structured Learning for In-

formation Processing, 2011.
[74] Y. Tingting, W. Junqian, W. Lintai, and X. Yong, “)ree-stage
network for age estimation,” CAAI Transactions on Intelli-
gence Technology, vol. 4, no. 2, pp. 122–126, 2019.
[75] G. Hu, J. Huang, Q. Wang, J. Li, Z. Xu, and X. Huang,
“Unsupervised fabric defect detection based on a deep con-
volutional generative adversarial network,” Textile Research
Journal, vol. 90, no. 3-4, pp. 247–270, 2020.
[76] B. Wei, K. Hao, Xs Tang, and L. Ren, “Fabric defect detection
based on faster RCNN,” in Proceedings of the International
Conference on Artificial Intelligence on Textile and Apparel,
pp. 45–51, Hong Kong, July 2018.
[77] G. Hu, J. Huang, Q. Wang, J. Li, Z. Xu, and X. Huang,
“Unsupervised fabric defect detection based on a deep con-
volutional generative adversarial network,” Textile Research
Journal, vol. 90, no. 3-4, pp. 247–270, 2020.
[78] Z. Liu, J. Cui, C. Li, M. Wei, and Y. Yang, “Fabric defect
detection based on lightweight neural network,” in Proceed-
ings of the Chinese Conference on Pattern Recognition and
Computer Vision (PRCV), pp. 528–539, Xiʼan, China, No-
vember 2019.
[79] Y. Wu, X. Zhang, and F. Fang, “Automatic fabric defect
detection using cascaded mixed feature pyramid with guided
localization,” Sensors, vol. 20, no. 3, p. 871, 2020.
[80] B. Fang, K. Hao, X.-s. Tang, and Y. Ding, “A new method
using the convolutional neural network with compressive
sensing for fabric defect classification based on small sample
sizes,” Textile Research Journal, vol. 89, no. 17, pp. 3539–3555,
2019.
[81] P. R. Jeyaraj and E. R. S. Nadar, “Effective textile quality
processing and an accurate inspection system using the ad-
vanced deep learning technique,” Textile Research Journal,
vol. 90, no. 9-10, pp. 971–980, 2020.
[82] Z. Wen, Q. Zhao, and L. Tong, “CNN-based minor fabric
defects detection,” International Journal of Clothing Science
and Technology, 2020.
[83] J. Jing, Z. Wang, M. Rätsch, and H. Zhang, “Mobile-Unet: an
efficient convolutional neural network for fabric defect de-
tection,” Textile Research Journal, 2020.
[84] C. Kampouris, S. Zafeiriou, A. Ghosh, and S. Malassiotis,
“Fine-grained material classification using micro-geometry
and reflectance,” in Proceedings of the European Conference on
Computer Vision, pp. 778–792, Amsterdam, Netherlands,
October 2016.
[85] M. Fritz, E. Hayman, B. Caputo, and J. O. Eklundh, ;e kth-
Tips Database, 2004.
[86] G. Kylberg, Kylberg Texture Dataset v. 1.0, Centre for Image
Analysis, Swedish University of Agricultural Sciences Uppsala,
Sweden, 2011.
[87] M. K. Ng, H. Y. T. Ngan, X. Yuan, and W. Zhang, “Patterned
fabric inspection and visualization by the method of image
decomposition,” IEEE Transactions on Automation Science
and Engineering, vol. 11, no. 3, pp. 943–947, 2014.
[88] H. Y. T. Ngan and G. K. H. Pang, “Regularity analysis for
patterned texture inspection,” IEEE Transactions on Auto-
mation Science and Engineering, vol. 6, no. 1, pp. 131–144,
2008.
[89] D. Germany, Tilda Textile Texture Database, 1996.

Fabric Defect Detection Using Computer Vision Techniques A

Uploaded by

Copyright:

Available Formats

Fabric Defect Detection Using Computer Vision Techniques A

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fabric Defect Detection Using Computer Vision Techniques A

Uploaded by

Copyright:

Available Formats

Hindawi

Mathematical Problems in Engineering

Correspondence should be addressed to Muhammad Tariq Mahmood; [email protected]

Academic Editor: Sajad Azizi

1. Introduction industrial products [5–8]. Natural elements such as wool,

Figure 1: Fabric defect detection using manual and automated inspection.

Multiple Oil Broken Wrong

Figure 2: Examples of textile fabric defects.

0 50 100 150 200 250

Table 1: A summary about histogram-based approaches.

Channel Channel Channel

Dissimilarity of color feature

Table 2: A summary on color-based and segmentation-based approaches.

Image patches Patch vectors Approximated feature vector

Dictionary learning phase

Training ADE-RELM phase

Figure 5: Dictionary learning phase and defect classiﬁcation overview.

Table 3: A summary on dictionary learning-based approaches.

fabric texture representation and for the detection of fabric

and similarity measure has been used to recognize the de-

8. GLCM-Based Defect Detection

Table 4: A summary about texture-based approaches.

Table 5: A summary on the frequency domain operation.

Table 6: A summary on the GLCM-based operations.

Holes Object on the surface Oil spots Thread error

Figure 8: Defective images of class C1 from the TILDA textile dataset.

Industry Ethernet normalized feature vectors serve as an input to SVDD. In

Feature Input Hidden Summation Output

Figure 10: Simple architecture of PNN.

Feature f m1 f m2 f m3 Sm1 Sm2 Sm3 km1 km2 km3

Figure 11: Image classiﬁcation-based model using PNN.

Table 7: A summary on fusion feature-based operations.

Table 8: A summary on sparse-based operations.

Table 9: A summary about morphology-based approaches.

Input [224 × 224] × 3

Fully connected layers[4069]

Fully connected layers[4069]

Fully connected layers [5/7]

Convolution + ReLU Fully connected + ReLU

Max pooling Softmax

Figure 12: Example of classiﬁcation by VGG-16 model.

Table 10: A summary about deep learning-based approaches.

nature of the approach, application domain, and environ-

(2) TN: defect-free images which are detected or iden-

Sensitivity: it is deﬁned as the correct identiﬁcation of

no. of nondefective image samples detected as defective

Detection rate (DR): it can be calculated as follows:

no. of defective image samples detected correctly

[73] L. Deng, An Overview of Deep-Structured Learning for In-

You might also like