The Journal of The Textile Institute

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tjti20

Fabric defect detection and classification using

modified VGG network

R. S. Sabeenian, Eldho Paul & C. Prakash

To cite this article: R. S. Sabeenian, Eldho Paul & C. Prakash (2023) Fabric defect detection
and classification using modified VGG network, The Journal of The Textile Institute, 114:7,
1032-1040, DOI: 10.1080/00405000.2022.2105112

To link to this article: https://doi.org/10.1080/00405000.2022.2105112

Published online: 28 Jul 2022.

Submit your article to this journal

Article views: 303

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tjti20
THE JOURNAL OF THE TEXTILE INSTITUTE
2023, VOL. 114, NO. 7, 1032–1040
https://doi.org/10.1080/00405000.2022.2105112

ARTICLE

Fabric defect detection and classification using modified VGG network

R. S. Sabeeniana, Eldho Paula and C. Prakashb
a
Department of Electronics and Communication Engineering, Sona College of Technology, Salem, India; bDepartment of Handloom and
Textile Technology, Indian Institute of Handloom Technology, Fulia, Ministry of Textile, Govt. of India, Shantipur, India

ABSTRACT ARTICLE HISTORY

Fabrics and garments are some of the most important utilitarian items for human beings. The textile Received 11 October 2021
industry is rapidly growing with various fabric models and different attractive designs. To hold up the Accepted 22 June 2022
development of the textile industry, rigid measures have to be taken to check the quality of fabric
KEYWORDS
during manufacturing. Thus, skilled workers are required to screen the quality of the fabric manually.
Artificial intelligence; fabric
This paper focuses on designing a deep learning framework to detect various fabric types and classify defect detection; computer
the defects using artificial intelligence. The proposed work has two phases; in phase 1, the input vision fabric inspection
image is preprocessed with a novel Pseudo–Convolutional Neural Network (P-CNN) having zero tun- system; convolutional
able parameters. In phase 2, a modified Convolutional Neural Network (CNN) is applied to the prepro- neural network;
cessed image to detect and classify major fabric defects. The Convolutional Neural Network (CNN) Gaussian noise
uses appropriate hidden layers to acquire an undeniable degree of accuracy for defect classification
using images. The dataset consists of five different fabric defects such as broken pick defects, pattern
defects, weft yarn deformity, soiled fabrics, and plain fabric defects that are considered for training
and validation of the proposed architecture. The performance of the proposed network is measured
using metric parameters such as sensitivity, specificity, and accuracy. The proposed technique has high
accuracy for different fabric types used for testing the creation network.

1. Introduction defects, patterns defects, soiled fabric, yarn defects, and plain
fabric defects. Some of the defects may occur due to malfunc-
Fabrics and garments are important commodities of our day
tioning or breaking down of machines. The product’s imper-
today life regardless of caste, creed, and geographic location.
fections influence the selling price of the product at the last
Varieties of fabrics are available on the market with different
stage of production, leading to a loss in the business.
designs and quality. Primarily, fabrics are produced using On a generalized basis, we can classify deformities into
strands of cotton, fleece, and many more. Modern clothing two sorts: major defects and minor defects. Some common
has engineered filaments of synthetic materials like rayon, major defects in fabrics are weft yarn defects like bars in the
polyester, etc. Defects in fabrics that occurred during produc- weft, burling, stains, knots, missing ends, and stripes in the
tion or due to improper handling at the sale point reduce the warp, as shown in Figure 1. Minor imperfections are mild
selling price up to a maximum of 50%. Quality control in the deformities that can be modified by further processing tech-
fabric industry is a very important functional unit of the gar- niques, such as fabric with holes, stains, broken pick defects,
ment industry to ensure that the quality product is up to the soiled fabric and plain fabric defectswhich are considered
required level in the market. Unfortunately, in most develop- minor defects and can be further modified with some proc-
ing countries, as a matter of cost reduction, human labour is essing techniques. The other common defects in fabrics are
involved in the defect detection process. Manual inspection of shown in Figure 2. Yarn defects are caused by defective
the quality check requires a large number of skilled labourers. yarn. Major imperfections lead to scrapping of the entire
Coincidently, the approach of finding the deformities physic- fabric material. These deformities are caused by defective
ally is a time-consuming process. Human fatigue causes errors samples being processed in weaving and designing.
during quality checks. Thus, an automated defect recognition Efficient image classification and defect identification is a
system is required for classifying imperfections using intelli- challenging as well as hot topic in the field of computer
gent image analytics. There are various computer vision-aided vision. Image classification using deep learning involves data
automated imperfection detection systems. This system helps pre-processing, feature extraction, and a training stage. The
in finding manufacturing defects in fabrics that are missed by Convolutional Neural Network (CNN) is used in various
the naked eyes of a human inspector. computer vision applications such as classification, pattern
Various researches have categorized more than seventy recognition, and object detection as it preserves spatial and
different types of fabric defects. For our work, we reviewed temporal features of the images. There are various computer
and examined five common defects, such as broken pick vision-based algorithms in the literature, such as the

CONTACT R. S. Sabeenian [email protected]

ß 2022 The Textile Institute
 THE JOURNAL OF THE TEXTILE INSTITUTE 1033

Figure 1. Major fabric defects.

MRMRFM & MRCSF method for silk fabric defect detection database has a total of 1,800 leaf images captured from 32
and identification based on computer vision in handloom varieties of plants. The dimensionality reduction technique
silk fabrics (Sabeenian et al., 2011)], the modified Alex net PCA (Principal Component Analysis) produces twelve fea-
architecture for classification (Shanthi & Sabeenian, 2019), tures, which are given as input to the PNN. This method
and discrete curve let transform based fabric defect detec- achieves 90% accuracy. Kang and Park (2009) have provided
tion (Anandan & Sabeenian, 2018), which are used for a strategy for classification using fusion neural network clas-
image classification and fault identification in fabrics. The sifiers and MPEG descriptors such as shading layout fea-
articles (Eldho Paul et al., 2021; Hanbay et al. 2016; Hu tures and histogram. The method classifies the input image
et al., 2020; Jing et al., 2019; Li et al., 2019; Mei et al., 2018; into six distinct classes. This literature has utilized a data-
Sabeenian et al., 2019; Shi et al., 2021; Wang et al., 2021; base of around 2000 images for developing the model. This
Zhang et al., 2019; Zhang & Tang, 2019) also describe framework has an accuracy of 78.4%.
recent developments and new methodologies in the field of Dodge et al. (2018) proposed an algorithm for ear data rec-
defect identification and classification. ognition. This network is trained with two different data sets:
Giacinto and Roli (2001) have proposed a neural network AWE (Annotated Web Ears) and CVLE (Computer Vision
for image classification. The framework has been trained Laboratory Ear Dataset). Firstly, image features are extracted,
and tested with a dataset that includes remote sensing data and the classifier is chosen. Secondly, the training of the net-
samples that were taken from horticultural land. The classi- work uses transfer learning of the DNN (deep neural network)
fication is done based on pixel values and classified into five classifier. The network gives high accuracy for the augmented
categories. Three distinct sorts of ensembles were made, dataset. Wang et al. (2018) successfully classified six
with a classification accuracy of around 89%. Park et al. fabric defects using the Deutsche Arbeitsgemeinschaft f€ ur
(2004) propose an automated methodology for content- Mustererkennung (DAGM) dataset. This literature uses a multi-
based image recovery by designing a deep neural network. layered deep CNN architecture for grouping or classifying
The technique proposed in the literature has three stages of more than 1000 images with defective and non-defective
operation, such as preprocessing, followed by feature extrac- images. The architecture accomplishes accuracy of around 98%.
tion, and classification. In the preprocessing stage, image The novelty and contributions of our work are summar-
resizing operations, grayscale conversions, and noise cancel- ized as:
lation are performed. This method utilizes extracted GLCM
features, which are used for training the classifier. Over
We propose a defect classification model using P-CNN
3000 images (training) and 300 images (testing) were used and modified CNN for removing noisy artifacts and
for implementing the network. This method gives a classifi- defect detection in textile fabrics respectively.
cation accuracy of 81.7%. Wu et al. (2007) developed a net-
The novel Pseudo-CNN with Weight Initialized Adaptive
work classifying various leaf images using a probabilistic Window (WIAW) filters is used in the preprocessing
neural network (PNN). The input layer, radial bias layer, stage improves the performance of defect classification
and competitive layer are the different layers of PNN. The by retrieving salient image features.
1034 R. S. SABEENIAN ET AL.

Figure 2. (A) Fabric with floats; (B) fabric with oil stain; (C) fabric broken pick defect; (D) soiled fabric; and (E) fabric with hole.

The computational complexity of proposed model is classification from nonlinear noisy defective datasets, which
reduced compared to the VGG net without compromis- is not addressed in the contemporary existing methods. (iii)
ing the accuracy. The conventional VGG Net has 138 The novel PCNN is capable of filtering nonlinear noises in
million parameters, whereas the proposed modified CNN the dataset, which eventually provides good flexibility to the
has a reduced number of parameters, i.e., 59 million. defect detection CNN model. (iv) We have customized

The metric parameters like specificity and selectivity are VGG Net for fabric defect detection with reduced computa-
comparable to various state-of-art techniques. tional complexity. The computational complexity was
reduced by 57% compared to VGG Net. The VGG net
The advantages of the proposed method compared with is very effective towards classifying 1000-class problems
the existing methods are summarized as: (i) The proposed efficiently. However, our method is proposed for industry-
method will have a very high impact on automatizing the specific problems like defect detection in plain fabric data-
textile industry for defect identification and classification. sets, which does not require complex deep neural network
(ii) The proposed method uses a novel preprocessing PCNN models. (v) The metric parameters like accuracy, specificity
with Weight Initialized Adaptive Window Filters (WIAWF) and selectivity of the proposed method are comparable to
for removing nonlinear noises, especially Gaussian Noise deeper convolutional neural network like Google Net, Dense
(GN) and Salt and Pepper Noise(SPN) in the dataset. Net, ResNet, DNN etc. The rest of the paper is organized as
Hence, we have addressed the problem of defect detection a proposed method followed by results and discussions,
with data preprocessing and defect classification using including metric parameters such as accuracy, precision, and
modified CNN. This enables good quantitative results for recall. The proposed work attempts to classify five different
 THE JOURNAL OF THE TEXTILE INSTITUTE 1035

defective samples from the fabric dataset. Finally, concluding weight of 0 and 1. As the window glides through the
remarks are given for the defect classification system. noisy imageðM x M x cÞ, the coefficients of the filter
WFxFxc ði þ a, j þ b) are updated. Convoluting a noisy image
with filter coefficients ðWFxFxc ) yields a volume feature
2. Proposed method mapðM x M x kÞ: The first layer is followed by a max -
The proposed method consists of a preprocessing stage, i.e. pooling layer (Layer 2). The (2,2) pooling layer reduces the
Pseudo CNN (PCNN) and a deep neural classifier using volume feature ðM x M x kÞ to½ðM  1Þx ðM  1Þx k,
convolutional neural network for classifying various defects suppressing noise. Layer 3 follows the pooling layer, and it
in plain fabrics. just has a single convolutional filter of sizeð1  1 x kÞ to
generate a pre-processed image. The loss function is not
estimated in P-CNN because there is only one forward pass
2.1. Pre-processingfilter (PCNN) and no back propagation with no tunable parameters.
While analyzing the fabric dataset, we observed certain Hence, we named this network as Pseudo-CNN. To validate
abrupt pixel transitions in different locations and other ran- the denoising efficiency of PCNN, we have taken some real
dom, noisy artifacts in the data samples. The common non- time images from the BSD dataset. We have manually cor-
linear noises in textile fabrics are Gaussian Noise (GN) and rupted images by adding Gaussian Noise (GN) and Salt and
Salt and Pepper Noise (SPN). In order to filter these nonlin- Pepper Noise (SPN) respectively. The preprocessing filter
ear noises, we have designed a novel Pseudo–Convolutional (PCNN filtering) results of some test images and real times
Neural Network (P-CNN) to remove the nonlinear pixel per- images from the BSD dataset with SPN noise level S ¼ 90is
turbations in the dataset. The preprocessing noise filtering shown in Figure 3.
stage offers more accuracy and flexibility to the defect classifi-
cation in the proposed method. The noise probabilistic equa- 2.2. Modified Convolutional neural network
tion of nonlinear mixed noise is given by equation 1
8 9 The emergence of various deep learning architectures finds
< dmin with probabilty p=2 = its applications in various fields such as image classification,
X ði, jÞ ¼ dmax with probabilty p=2 (1)
: ; object detection, image segmentation, defect classification,
X ði, jÞ þ vði, jÞ with probability 1p and practical industry-specific problems. The challenges in
P-CNN is a three-layered network with Weight Initialized developing deep networks for defect detection in the textile
Adaptive Window Filters (WIAWF) that serves as a prepro- industry are the lack of availability of public datasets and
cessing filter. The first layer has k adaptive ðWFxFxc ) of the computational complexity for very deep models. The
size FXF ¼ ð2d þ 1ÞXð2d þ 1Þ, where depth of the kernel, textile industry required the implementation of high-speed
d ¼ 1, 2, 3, . . . k: The number of color channels (c ¼ 1 & c ¼ 3 and real-time data processing for analyzing the quality of
for gray and RGB images) is represented by c. i.e. P-CNN the fabrics. Most of the deep learning approaches have
extracts k distinct feature maps from the input image using k either single-stage or two-stage detectors for classifying
½l defects in fabrics. Single-stage detectors are fast, but detec-
feature detectors. The lth layer output Yi, j is given by equation
½l ½l tion accuracy is lower, whereas double-stage detectors have
2, where Xi, j , Wi, j or WFxFxc ) represents noisy input image
and the weight initialized adaptive window filter, respectively. high accuracy at the expense of computational timing.
" X ½ l  ½ l !# The proposed method is compared with faster RCNN
½ 1
Yi,lj ¼ X W (2) [Wei et al., 2018), Cascade CNN (Li & Li, 2021), and
FF i, j2F i, j i, j ResNet (He et al., 2016), DNN (Shanthi et al., 2022), Google
Net (Szegedy et al., 2015), and DenseNet-201 (Huang et al.,
The filter ðWFxFxc ) has a receptive field of F X F over 2017). The proposed model is derived from the famous
the image. The depth of the feature map (k) is set to ten VGG network with some modifications (Simonyan &
since the number of weighted adaptive window filters in Zisserman, 2014). The number of output classes changes
layer 1 is set to ten. Extraction of relevant information is from 1000 to 5 (since we are interested in classifying five
attributed to improving the size of the receptive field at the major defects in fabric data). To train our proposed network
expense of computational complexity. The filter coefficients faster and have quick error convergence, we have used
of layer 1 and layer 3 in P-CNN ðWFxFxc Þ are initialized to transfer learning (Zhang et al., 2017) with pretrained
1,0 and 1, respectively using equation 3. weights in the untrained model. The hyperparameters such
Y
F as learning rate, batch size, dropout and activation functions
fWFxFxc ði þ a, j þ bÞg remain unchanged. By valid padding and eliminating pool-
a, b, c¼1 ing layers, the dimension in hidden space is converted to an
8 9
< 1 dmin  X ði þ a, j þ bÞ  dmax >
> (3) input image. For training the model, the loss function is
=
¼ 0 X ði þ a, j þ bÞ ¼ dmax combined with the residual strategy and SGD weights with
>
: >
; weight decay of 0.0001, momentum of 0.9, and a mini-batch
1 X ði þ a, j þ bÞ ¼ dmin
size of 128. We used 30 epochs to train the model, with the
Uncorrupted pixels were given a maximum weight of learning rate decaying from 1e-1 to 1e-4. The test was run
þ1, while corrupted pixels (dmax & dmin) were given a in a python environment on a PC with a core i5 processor
1036 R. S. SABEENIAN ET AL.

Figure 3. P-CNN filtering results of SPN noise level S ¼ 90.

Figure 4. Block diagram of proposed CNN Architecture.

running at 3.3 GHz, 8GB RAM, and a python environment. Table 1. Different layers and feature map of the proposed CNN architecture.
While designing the deep model, if the parameters of the Layer (type) Output Shape Parameters(P)
network are too small, it will be difficult to implement com- Input_1 Input Layer) (None, 224, 224, 3) 0
plex data textures, and if the network parameters are too block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
high, the redundancy and computational complexity will be block1_ (MaxPooling2D) (None, 112, 112, 64) 0
high. After a detailed investigation into the features learned block2_conv3 (Conv2D) (None, 112, 112, 128) 73856
by various layers, we have tried to fine-tune VGG Net for block2_conv4 (Conv2D) (None, 112, 112, 128) 147584
block2_poo(MaxPooling2D) (None, 56, 56, 128) 0
specific problems like fabric defect detection. Our architec- block3_conv5 (Conv2D) (None, 56, 56, 256) 295168
ture has 11 weight layers, which includes nine convolutional block3_conv6 (Conv2D) (None, 56, 56, 256) 590080
layers and two fully connected networks. Most of the defect block3_conv7 (Conv2D) (None, 56, 56, 256) 590080
block3_(MaxPooling2D) (None, 28, 28, 256) 0
features can be distinguished before the ninth layer of the block4_conv8 (Conv2D) (None, 28, 28, 512) 2359808
VGG Net with a receptive field of 3  3 filters. However, block4_ (MaxPooling2D) (None, 14, 14, 512) 0
VGG Net is an efficient network made for classifying 1000- block4_conv9 (Conv2D) (None, 14, 14, 512) 2359808
block4_(MaxPooling2D) ( None, 7, 7, 512) 0
class problems. Industry-specific problems such as fabric flatten (Flatten) (None, 25088) 0
defect detection and multi class classification (e.g. five defect Fc1 (Dense) (None, 2048) 51382272
classes) can be solved by removing some layers with a Fc2 (Dense) (None, 1024) 2098176
Predictions (Dense) (None, 5) 5125
trade-off between accuracy and redundancy of data. As a Total parameters:62,033,728
result, the computational complexity can be reduced without
a compromise in the accuracy of detection. The proposed
CNN model uses four convolution layers, each with its own
activation or regularization, followed by a max pool and fully the input image is filtered by the first convolutional layer,
connected layers. The architecture of the proposed modified and the size of the output volume is defined based on the
convolutional neural network is shown in Figure 4. Firstly, mathematical upgrade (4).
 THE JOURNAL OF THE TEXTILE INSTITUTE 1037

nf þ 2p compared to classical algorithms. Generally, machine learn-

þ1 (4)
s ing algorithms require hand-crafted features for designing
classifiers, whereas deep networks learn hidden features
Where n, f, p, and s represent the input image size, filter from the data. The success of deep learning models in the
size, valid padding, and stride, respectively. The parameters fields of image restoration and computer vision has tremen-
of the different layers of the modified convolutional neural dous impact on these fields. The conventional model-based
network are given in Table 1. The network parameters in a methods are not strong enough to characterize salient image
convolutional layer can be calculated by equation (5). features/complex image priors.

P ¼ ½ðM x N x DÞ þ 1Þ K (5)
3. Result and discussion
where M, N, D, K represents the kernel width, kernel
height, number of filters in the previous layer, number of The literature survey reveals different types of fabric defects
filters in the current layer respectively. and defect detection techniques associated with the textile
The final fully connected layer has five class labels using industry. We have performed experiments to detect and clas-
the Soft Max classifier. Deep learning algorithms can predict sify five different fabric defects, such as broken pick defects,
decision boundaries and have easier design methodologies patterns defects, soiled fabric, yarn defects, and plain fabric
defects. In phase 1, the input images are preprocessed by a
Table 2. PSNR values of PCNN filtering at different noise levels. novel Pseudo CNN filter (preprocessing filter). Figure 3 and
SPN Noise Density S ¼ 30% Table 2 shows the denoising results and PSNR values of the
Image r ROR[29] CAI[30] WESNER[31] CNN[32] PCNN PCNN in real-time images corrupted by nonlinear mixed
Lena 10 29.38 31.25 32.42 31.50 32.86 noises. With increasing the Gaussian noise level (r), one can
20 26.29 29.34 30.75 29.55 28.18 observe that the PSNR values of PCNN are reduced com-
30 25.22 28.77 29.55 28.35 26.14
Plane 10 28.52 30.82 31.25 32.33 32.89
pared to the state-of-the-art nonlinear noise filters like ROR
20 26.03 28.54 29.56 29.88 28.65 (Xiong & Yin, 2012), CAI (Cai et al., 2010), WESNER (Jiang
30 24.84 27.33 28.55 28.36 26.44 et al., 2014) and CNN (Islam et al., 2018). However, our
Peppers 10 28.94 31.73 32.34 32.19 32.54
20 26.54 29.87 30.30 30.81 29.41 novel PCNN is very effective in removing low-density nonlin-
30 25.20 28.22 29.29 29.42 27.31 ear mixed noises in the fabric datasets. Due to the unique
pattern and profile, denoising of the plain fabrics is not
Table 3. Comparison of different CNN techniques. adequate to validate the performance of PCNN. Hence, we
Method Accuracy % Specificity Selectivity have chosen some real-time images from the BSD dataset
Modified Alex Net 89.52 84.58 81.38 and test images for comparing the metric parameters of
Google Net 85.29 81.65 82.35 PCNN with various standard denoising techniques.
ResNet-50 91.32 86.24 91.58 In phase 2, we train our proposed CNN model with an
Faster RCNN 92.56 88.78 90.25
Cascade RCNN 90.58 89.51 91.54 80% data set (preprocessed images) for training and keep
Densenet-201 85.6 87.22 83.33 20% for validation. The network was trained for 25 epochs,
DNN 92.98 91.54 91.88
Proposed Method 93.92 92.51 92.12
or iterations, and obtained an accuracy of 93.92%. It is
noticed that the validation accuracy and losses are optimum

Figure 5. Comparisons of Accuracy, Specificity, and Selectivity of Various CNN Methods.

1038 R. S. SABEENIAN ET AL.

between the 10th and 15th epochs, respectively. The pro- computational complexity is reduced by 57% compared to
posed CNN is compared with various state-of-the-art tech- the original VGG-16 network. It should also be noted that
niques such as modified Alex Net (Shanti & Sabeenian the conventional VGG-16 network is capable of classifying
et al., 2019), faster RCNN (Wei et al., 2018), Cascade CNN data into 1000 classes, whereas the proposed model is
(Li & Li, 2021), and ResNet (He et al., 2016), DNN (Shanthi problem-specific and application-oriented with five out-
et al., 2022), Google Net (Szegedy et al., 2015), and put classes.
DenseNet-201 (Huang et al., 2017). Table 3 and Figure 5
show the comparisons of various convolutional neural net-
works with respect to accuracy, specificity, and selectivity. 3.1. Data-set description
In the modified convolutional neural network, around The proposed model has been trained and tested with a data-
10.4025% of the entire parameters are associated with convo- set of fabrics with five major defects. All the images were
lutional layers, and 89.5913% of the parameters are linked captured with a smart camera, the N1772C (1.6 GHz Atom
with fully connected layers. The conventional VGG-16 net- processor, VGA, color, 65fps). The practice of increasing the
work consists of 138 million parameters. The convolutional amount of data needed to train a model is known as data
layers and fully connected layers have around 18 million and augmentation. Deep learning models frequently demand a
120 million parameters, respectively, whereas the proposed large amount of training data, which is not always available.
fine-tuned VGG Net has around 59 million parameters, Hence, we have used position augmentation techniques such
including 6 million convolutional layer parameters and 53 as cropping, flipping, padding, and rotation to improve data
million fully connected layer parameters. Hence, the volume. Around 6749 images were taken for defect classifica-
tion (i.e. 5399 and 1350 images were used for training and
testing the performance of our network, respectively). All the
images are resized to 224x224x3 to make them compatible
with the VGG network. Figure 6 shows the distribution of
datasets with different defect classes.
To validate the effectiveness of the proposed algorithm,
metric parameter such as the confusion matrix is calculated
using test data. Table 4 shows the confusion matrix for test-
ing samples.
Precision and recall are performance matrix parameters
extracted from the confusion matrix. Figure 7 shows the
Figure 6. Data set with defect classes. results of the precision and recall parameters for the testing

Table 4. Confusion matrix for testing samples.

Broken Pick Pattern Soiled Weft Yarn Plain Fabric Total Number Per Class
Data Samples Defect Defects Fabric Defects Defects of Images Accuracy (%)
Broken pick defect 207 4 3 9 7 230 90.00
Pattern defects 0 244 4 2 6 256 95.31
Soiled fabrics 6 4 291 6 3 310 93.87
weft yarn defects 1 1 4 260 4 270 96.29
Plain fabric Defects 4 5 3 6 266 284 93.66

Figure 7. Precision and Recall of testing samples.

 THE JOURNAL OF THE TEXTILE INSTITUTE 1039

ratio of the sum of true positive and false negative samples

to the total number of samples (93.925%).

4. Conclusion
This paper has focused on designing a novel preprocessing
filter and deep convolutional neural network architecture
for fabric defect classification. For better classification per-
formance, the layers of the convolutional neural network are
customized for defect detection. The proposed algorithm is
also capable of filtering nonlinear mixed noise in the fabric
datasets using PCNN. The modified convolutional neural
network classifies various fabric defects from a database
consisting of both defected and non-defected samples. Our
work is simulated considering five common defects in the
fabric industry. The proposed algorithm has an impact on
automating the textile industry for quality control. The pro-
posed method is more computationally efficient in terms of
the reduced number of computationally complex operations
than VGG Net. The quantitative and qualitative results of
the proposed architecture are competent with various exist-
ing methodologies. However, for better reliability, the pro-
posed system should periodically train with new occurrences
of defects in the textile industry. In the future, we will
investigate the scope of the proposed network for defect
classification in striped fabrics and real-time noisy plain fab-
ric data sets.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
C. Prakash http://orcid.org/0000-0003-2472-6765

References
Anandan, P., & Sabeenian, R. S. (2018). Fabric defect detection using
discrete curvelet transform. Procedia Computer Science, 133,
Figure 8. (a). CNN accuracy vs Epochs. (b). CNN Loss vs Epochs.
1056–1065. https://doi.org/10.1016/j.procs.2018.07.058
Cai, J. F., Chan, R. H., & Nikolova, M. (2010). Fast two-phase image
data set. Recall is the ratio of the number of true positives deblurring under impulse noise. Journal of Mathematical Imaging
to the sum of the numbers of true positives and false nega- and Vision, 36(1), 46–53. https://doi.org/10.1007/s10851-009-0169-7
tives. True positives are data items that the model classifies Dodge, S., Mounsef, J., & Karam, L. (2018). Unconstrained ear recog-
nition using deep neural networks. IET Biometrics, 7(3), 207–214.
as positive but are actual positive (i.e. correct), while false
https://doi.org/10.1049/iet-bmt.2017.0208
negatives are data points that the model classifies as negative Eldho Paul, R. S., Paramasivam, M. E., Valliappan, M., & Kavin
but are actual positive (incorrect). Whereas precision is Kumar, K. A. (2021). A comparative study of fabric detection and
defined as the ratio of true positives to the sum of the num- classification using efficient hybrid learning algorithm. Turkish
bers of true positives and false positives, False positives are Journal of Physiotherapy and Rehabilitation, 32(2), 942–947.
Giacinto, G., & Roli, F. (2001). Design of effective neural network ensem-
occurrences in which a model labels something as positive
bles for image classification purposes. Image and Vision Computing,
when it is actual negative. 19(9–10), 699–707. https://doi.org/10.1016/S0262-8856(01)00045-2
Figures 8a and b shows the accuracy and loss vs. number € uven, O.
Hanbay, K., Talu, M. F., & Ozg€ € F. (2016). Fabric defect detec-
of iterations, respectively. The training accuracy improves tion systems and methods—A systematic literature review. Optik,
with an increasing number of epochs. The computational 127(24), 11960–11973. https://doi.org/10.1016/j.ijleo.2016.09.110
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for
time also increases with the number of epochs. Based on
image recognition [Paper presentation]. Proceedings of the IEEE
our trial-and-error analysis, we have estimated that 25 Conference on Computer Vision and Pattern Recognition (pp. 770–778).
epochs are sufficient for better classification accuracy. The Hu, G., Huang, J., Wang, Q., Li, J., Xu, Z., & Huang, X. (2020).
overall accuracy of the proposed network is given by the Unsupervised fabric defect detection based on a deep convolutional
1040 R. S. SABEENIAN ET AL.

generative adversarial network. Textile Research Journal, 90(3-4), Electrical Engineering, 76, 56–64. https://doi.org/10.1016/j.compele-
247–270. https://doi.org/10.1177/0040517519862880 ceng.2019.03.004
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Shi, B., Liang, J., Di, L., Chen, C., & Hou, Z. (2021). Fabric defect
Densely connected convolutional networks [Paper presentation]. detection via low-rank decomposition with gradient information
Proceedings of the IEEE Conference on Computer Vision and and structured graph algorithm. Information Sciences, 546, 608–626.
Pattern Recognition (pp. 4700–4708). https://doi.org/10.1016/j.ins.2020.08.100
Islam, M. T., Rahman, S. M., Ahmad, M. O., & Swamy, M. N. S. (2018). Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks
Mixed Gaussian-impulse noise reduction from images using convolutional for large-scale image recognition [Paper presentation]. Proceedings of
neural network. Signal Processing: Image Communication, 68, 26–41. the International Conference Learn Represent (pp. 1–14).
Jiang, J., Zhang, L., & Yang, J. (2014). Mixed noise removal by Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D.,
weighted encoding with sparse nonlocal regularization. IEEE Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015). Going deeper
Transactions on Image Processing : A Publication of the IEEE Signal
with convolutions [Paper presentation]. Proceedings of the IEEE
Processing Society, 23(6), 2651–2662.
Conference on Computer Vision and Pattern Recognition (pp. 1–9).
Jing, J. F., Ma, H., & Zhang, H. H. (2019). Automatic fabric defect
Wang, T., Chen, Y., Qiao, M., & Snoussi, H. (2018). A fast and robust
detection using a deep convolutional neural network. Coloration
Technology, 135(3), 213–223. https://doi.org/10.1111/cote.12394 convolutional neural network-based defect detection model in prod-
Kang, S., & Park, S. (2009). A fusion neural network classifier for uct quality control. The International Journal of Advanced
image classification. Pattern Recognition Letters, 30(9), 789–793. Manufacturing Technology, 94(9–12), 3465–3471. https://doi.org/10.
https://doi.org/10.1016/j.patrec.2008.06.009 1007/s00170-017-0882-0
Li, F., & Li, F. (2021). Bag of tricks for fabric defect detection based on Wang, H., Duan, F., & Zhou, W. (2021). Fabric defect detection under
Cascade R-CNN. Textile Research Journal, 91(5–6), 599–612. https:// complex illumination based on an improved recurrent attention
doi.org/10.1177/0040517520955229 model. The Journal of the Textile Institute, 112(8), 1273–1279.
Li, Y., Luo, H., Yu, M., Jiang, G., & Cong, H. (2019). Fabric defect https://doi.org/10.1080/00405000.2020.1809918
detection algorithm using RDPSO-based optimal Gabor filter. The Wei, B., Hao, K., Tang, X. S., & Ren, L. (2018, June). Fabric defect
Journal of the Textile Institute, 110(4), 487–495. https://doi.org/10. detection based on faster RCNN [Paper presentation]. International
1080/00405000.2018.1489951 Conference on Artificial Intelligence on Textile and Apparel) (pp.
Mei, S., Yang, H., & Yin, Z. (2018). An unsupervised-learning-based 45–51). Springer.
approach for automated defect inspection on textured surfaces. Wu, S. G., Bao, F. S., Xu, E. Y., Wang, Y. X., Chang, Y. F., & Xiang, Q. L.
IEEE Transactions on Instrumentation and Measurement, 67(6), (2007, December). A leaf recognition algorithm for plant classification using
1266–1277. https://doi.org/10.1109/TIM.2018.2795178 probabilistic neural network [Paper presentation]. 2007 IEEE International
Park, S. B., Lee, J. W., & Kim, S. K. (2004). Content-based image clas- Symposium on Signal Processing and Information Technology (pp.
sification using a neural network. Pattern Recognition Letters, 25(3), 11–16). IEEE. https://doi.org/10.1109/ISSPIT.2007.4458016
287–300. https://doi.org/10.1016/j.patrec.2003.10.015 Xiong, B., & Yin, Z. (2012). A universal denoising framework with a
Sabeenian, R. S., Paramasivam, M. E., Anand, R., & Dinesh, P. M. (2019). new impulse detector and nonlocal means. IEEE Transactions on
Palm-leaf manuscript character recognition and classification using
Image Processing : A Publication of the IEEE Signal Processing
convolutional neural networks. Computing and Network Sustainability,
Society, 21(4), 1663–1675.
75, 397–404. https://doi.org/10.1007/978-981-13-7150-9_42
Zhang, H., Liu, S., Ge, Z., & Li, P. (2019). Defect detection of yarn-
Sabeenian, R. S., Paramasivam, M. E., & Dinesh, P. M. (2011).
Detection and location of defects in handloom cottage silk fabrics dyed shirt cut piece based on denoising convolutional autoencoder.
using MRMRFM & MRCSF. International Journal of Technology Journal of Basic Science of Textile Universities, 32(2), 119–125.
and Engineering System, 2(2), 172–176. Zhang, B., & Tang, C. (2019). A method for defect detection of yarn-
Shanthi, T., Paramasivam, M. E., Prakash, C., Manju, K., Paul, E., dyed fabric based on frequency domain filtering and similarity
Anand, R., Dinesh, P. M., Sabeenian, R. S., & Raja, D. (2022). Robust measurement. Autex Research Journal, 19(3), 257–262. https://doi.
fabric defects inspection system using deep learning architecture. org/10.1515/aut-2018-0040
Journal of Testing and Evaluation, 50(1), 20200778–20200132. https:// Zhang, K., Zuo, W., Chen, Y., Meng, D., & Zhang, L. (2017). Beyond a
doi.org/10.1520/JTE20200778 gaussian denoiser: Residual learning of deep cnn for image denois-
Shanthi, T., & Sabeenian, R. S. (2019). Modified Alexnet architecture ing. IEEE Transactions on Image Processing : A Publication of the
for classification of diabetic retinopathy images. Computers & IEEE Signal Processing Society, 26(7), 3142–3155.