2024 International Conference on Power, Energy, Control and Transmission Systems (ICPECTS)

HNNLS: Robust Design to Identify Textile Fabric

Defects Using Hybrid Neural Network Assisted
Learning Strategy
2024 International Conference on Power, Energy, Control and Transmission Systems (ICPECTS) | 979-8-3315-0884-5/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICPECTS62210.2024.10780104

Madasamy Raja. G. B.Deepa Rajesh. S

Department of IT Department of IT Department of CSE
Paavai Engineering College Paavai Engineering College Paavai Engineering College
Tamilnadu, India Tamilnadu, India Tamilnadu, India
[email protected] [email protected] [email protected]

Kurakula Sai Satish D. Baburao Sreedhar Burada,

Department of ECE Department of CSE (AI & ML) Department of CSE
Amrita Vishwa Vidyapeetham CMR Technical Campus, Koneru Lakshmaiah Education
Amritapuri, India Telangana, India. Foundation,
[email protected] [email protected] Andhra Pradesh, India.
[email protected]

Abstract— The rising cost of labor combined with increased required, this procedure may be exceedingly slow and
automation in the textile sector has made studying methods for laborious, wearing workers down to the point where they
detecting flaws in fabrics an area of intense interest in recent make mistakes. As a result, even though they are much slower
years. In this study, we present a machine-learning technique than modern systems, traditional ones can only reach an
for automatic fabric defect identification that employs deep
accuracy of 60% to 75%. That's why there's a growing need
neural network models that have already been pretrained. Pre-
processing the fabric pictures using standard image processing for automated, in-line visual inspection systems to keep
techniques yields improved data that is then used to train the product quality at a high standard. Automatic flaw detection
networks. In order to train and categorize many fabric flaws, systems have several benefits over manual inspection, the
the Deep Convolutional Neural Network (DCNN) is utilized. In most notable of which are increased productivity,
the performed simulations, the existing textile dataset allows for dependability, and uniformity.The faults result from improper
a maximum categorization accuracy of 97.34%. This classifier application of the materials and sloppiness during production
model-based detection and categorization system may help [2]. The large number and different types of flaws that are
humans locate manufacturing flaws in fabric with a high degree characterized by uncertainty make it difficult to check the
of accuracy.
actual fabric faults manually. In contrast to the potential
Keywords— Fabric Defect Classification, Artificial productivity gains from using a real-time automated system
Intelligence, Defect Classifier, Convolutional Neural Network, for defect identification, manual inspection only catches a
Deep Neural Network; small fraction of problems owing to human weariness.
The current method of detecting fabric flaws, which
I. INTRODUCTION relies on skilled inspectors, has many drawbacks. The first
The fabric business has always relied on human step is to ensure that the inspectors you send out are qualified.
visual examination to detect and find flaws in the fabric. As an added disadvantage, the precision with which flaws may
However, the human detection rate is only up to 12 meters per be spotted decreases as the amount of time spent gazing at the
minute, and the task is dull, highly repetitive, a waste of cloth increases [3]. As a result, there is a rising need for an
human resources, and expensive. The complexity of modern automated fabric fault detection system to help the industry
textile manufacturing makes it difficult for employees to detect, diagnose, and stop these flaws from happening again.
pinpoint the exact position of flaws, despite the ease with Fabric flaws such as these are very common and include rips,
which humans can recognize them. Fabric faults also cause holes, scratches, stretches, loose yarn, unclean patches, broken
fabric prices to drop, costing manufacturers anywhere from 45 points, misprints, color bleeding, and more. Fig. 1 depicts a
percent to 65 percent of revenue. As a result, the existing few of these examples.
reliance on manual labor calls for an alternative detection
approach that is both fast and accurate [1]. As the field of deep
learning and machine vision has advanced, several new
detection methods that leverage the best of both worlds—deep
learning and machine vision—have arisen to supplant older,
more labor-intensive approaches, such as manual inspection
and image processing.
During the quality control phase, it is essential that
any flaws in the fabric be uncovered. Fabrics with these flaws
often sell for 45–65 percent less. Manual personnel or
operators make up the backbone of conventional inspection
systems. A machine will move the cloth as they look for flaws
in it. Motorized unrolling devices have long been used to
provide cloth to workers without folds or variations in
thickness. Due to the high degree of visual skill and focus Fig. 1. Textile Fabric Defects

979-8-3315-0884-5/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.
 approached from many different angles, however current
Material quality, mechanical considerations, dye type, yarn technologies fall short due to the wide variety of fabric
size, and human factors are only few of the variables that textures and imperfections. Author [6] offer a GAN-based
might change the outcome of a textile manufacturing process. system for detecting flaws in textiles. Taking into account the
Defects in textiles are most commonly defined as flaws in the difficulties encountered in practical settings, the suggested
fabric's outermost layer. Fabric flaws come in a wide variety fabric defect detection system can study previously collected
of forms, and the vast majority of them are the result of faulty fabric defect samples and automatically adjust to the varying
procedures or broken machinery. If there are flaws in the textures of the fabric being used. To pinpoint, we tune a
production process, the finished product will suffer and so will multi-type defect-detecting deep semantic categorization
a lot of money and time. Manufacturing flaws from an earlier network specifically for fabrics. We also tried to train a multi-
step will trickle down to the next phase, affecting the final stage GAN to generate plausible flaws in otherwise brand-
product. As a result, quicker and earlier business losses can be new, error-free samples. The conditional distribution of flaws
mitigated if flaws in the fabric are discovered sooner rather is first investigated by training a texture-conditioned GAN.
than later. Thus, contemporary fabric producers use effective Our goal is to construct plausible damaged sections from a
fabric defect detection as a means of cost management, given new fabric. The created flaws are then fused to
product value enhancement, and core competency particular places using a GAN-based fusion network. Lastly,
development. the well-trained multistage GAN continually updates the
Automatic fabric flaw detection plays a crucial role existing fabric defect datasets and helps fine-tune the
in today's textile production facilities. Fabric defect semantic segmentation network so that defects may be
identification has traditionally relied on manual visual detected more accurately in a variety of settings. Extensive
examination, which is both inefficient and costly. As a result, studies on a variety of standard fabric samples are performed
the textile sector needs automated fabric flaw identification to to validate our suggested method's detection performance.
cut costs and boost output. The detection techniques are the Fabric fault detection is a crucial part of quality
backbone of a full-featured online system for detecting flaws control. Using solely defect-free images for training, we
in textile fabrics. Over the past few decades, a great deal of present a two-stage, unsupervised framework for fabric flaw
time and effort has been invested by academics and engineers recognition that is effective and practical in real-time. The
in this subject towards the development of secure and efficient fabric image is repaired and reconstructed using a lightweight
techniques [4]. Automatic detection techniques for fabric reconstructive system, and the location of the defect is
flaws are more accurate and efficient than human inspectors. determined using a lightweight decision system [7]. The
Detecting flaws in fabric has been an active area of study in proposed method performs admirably across three separate
computer science and mechanical engineering. fabric datasets, with an F1 score of 983.4%, a model size of
Decision-making in many areas can benefit from the 162.6K, a computational complexity of 146.48M floating-
use of artificial intelligence applications, which can be both point operations, as well as a rate of identification of 122.59
cost-effective and efficient. The impact of AI on the frames per second. We propose a two-stage method for fast
manufacturing sector is far more than anyone anticipated. and accurate fabric inspection in a real-world manufacturing
Industrial intelligence is progressing at a breakneck pace setting.
thanks to the large number of academics and engineers Several low-rank representation based approaches
working on the problem. Mechanical, data science, network, have been presented as of late for identifying flaws in textiles.
communication, and information security expertise are all The majority of them convert the low-rank breakdown
brought together in industrial artificial intelligence, with the challenge into a nuclear norm minimization (NNM) issue in
end goal of improving manufacturing through the application an effort to achieve convexity. In order to find a solution to
of AI algorithms to problems in the manufacturing sector [5]. the standard NNM problem, singular values in matrices must
The textile sector, like any other, must find its own method to be dealt with consistently. But this doesn't work in the case
modify the production process in order to participate in of identifying flaws in fabrics since the individual elements
industry 4.0. The textile business is massive and intricate. of the matrix have actual significance. Here we propose a
Spinning, weaving, dying, printing and finishing, and garment weighted double-low-rank decomposition (WDLRD) method
production are the mainstays of the textile manufacturing for classifying how to deal with singular values in matrices.
process, which consists of a sequence of intricate and well- As a result, a fabric image's defining characteristics are more
organized steps. The success of a business relies on the likely to be preserved [8]. Furthermore, WDLRD is distinct
consistency and high quality of the textile fabric manufactured from other low-rank-based methods in that it assumes a
by the whole manufacturing line. Deep learning neural defective fabric image is capable of being separated into two
networks (NNs) are crucial for speeding up and improving the low-rank matrices, a low-rank defect-free matrix as well as a
accuracy of inspections and fault identification. With the help low-rank defect matrix, due to the fact that defect-free and
of AI, the textile sector is entering a new era that promises incorrect regions are usually made up of homogeneous
revolutionary change and promising prospects. Motivated by objects with a high correlation. In addition, WDLRD makes
the recent successes in deep learning research, this paper fault diagnosis more robust in many settings by employing a
proposes a Convolutional Neural Network framework to noise term to dampen surface noise and other interference.
discover flaws in simple fabrics. WDLRD also incorporates a defect prior into the target
function to aid in locating problematic areas. The proposed
II. RELATED STUDY optimization problem may be solved rapidly and effectively
Defect identification in fabrics is a fascinating but using an iterative strategy based on modified Lagrange
difficult field of study. Fabric defect identification has been multipliers.

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.
 The majority of existing methods for detecting flaws The suggested approach is to create a principled
in fabrics are adaptations of the standard deep learning technique that guarantees full automation of flaw
network design. Few, however, tune the deep learning identification and makes it capable of effectively dealing with
network model in accordance with the properties of the a wide variety of fabrics and problems. In Fig. 2, we explain
fabrics; even fewer take into account the magnitude of the the three phases that make up our process: image pre-
model's parameters [9]. Fabric module (FM) is a lightweight processing, defect predicting, and identifying defects. In the
module proposed in this work to represent fabric qualities. first step, the picture undergoes pre-processing that draws
We start with YOLOX-Nano as our foundational network and attention to the defect. The second step involves checking if
swap out the SPP module for FM. Using Tianchi's publicly the box image has any issues. Training a deep convolutional
available data, we put the network model through its paces neural network (CNN) to forecast errors is also a part of this
and compared it to the winning entry from the contest's procedure. The third and final phase is roughly localizing the
preliminary round. The experimental findings demonstrate faults in the image through analysis of the data collected in
that our approach is superior to existing methods in terms of the second. The following sections expand on the
accuracy and speed, making it ideal for automatic fabric flaw aforementioned topics in further depth.
identification.
The ability to detect flaws in fabrics is crucial for
B. Dataset and Implementation
effective quality assurance. Many deep learning algorithms
have been presented as an alternative to the laborious and TILDA, as far as we are aware, is a dataset created for the
expensive manual detection process. The YOLO series of express purpose of finding bugs. Unfortunately, the
algorithms has seen widespread use in recent years thanks to collecting conditions and small sample size of this dataset
the rapid evolution of object detecting technologies [10]. make it inappropriate for the suggested model. We decided to
Using a Squeeze-and-Excitation (SE) module, we offer a construct a dataset for fabric defect identification analysis
YOLOv5 as a foundation for developing a robust fabric because of its potential usefulness in a wide variety of
identification system [11]. To enhance the YOLOv5 method, contexts. The majority of the photos in this collection (3000
SE-YOLOv5 incorporates the SE module into the YOLOv5 total) are from the Tianchi competition, and 2000 of them
base and replaces the Leaky Rectified Linear Unit (ReLU) have flaws. Additionally, there are three distinct types of
activation function with the Activate Or Not (ACON) flawed photographs. All pictures are 25601920 pixels in size,
activation algorithm of the YOLOv5 cross stage partial with the trimmed portion acting as a positive sample and the
(CSP). The proposed model SE-YOLOv5 beats YOLOv5 in untrimmed portion as a negative sample. Here, it's important
terms of accuracy, speculative ability, and durability, making to note that both "good" photos and "innocent" areas of
it ideal for the task of recognizing fabric faults. images with flaws contribute to the negative sample
collection. Sample data set displayed in Fig. 3.
III. METHODOLOGY
A. Detection Algorithm

Preprocessing
Fig. 3. Dataset Sample
C. Image Normalization
When an image is normalized, the distribution of its pixels
is made consistent. The transform function from the pandas
package was used to make this happen, with the variables
CNN
mean and standard deviation set to 0.5.

D.Dimension Reduction
In order to simplify the training process, we can limit the
number of channels by employing a dimensionality reduction
..
.

approach. When the original RGB image is converted to a

grayscale image, the number of channels employed is
reduced from three to one. The process has two parts:
choosing and extracting features. Principal component
Defected Fabric Non - Defected Fabric
analysis (PCA), linear discriminant analysis (LDA), and
generalized discriminant analysis (GDA) are all commonly
Fig. 2. Proposed Architecture cited in the research literature.

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.
 E. Histogram Segmentation G.Optimization Algorithm
Histogram equalization is commonly used to enhance Updating the weights for every successive iteration is
contrast by changing picture intensities. The key objective is essential for achieving the primary aim of continuous training
to ensure that the intensity is spread out evenly over the in a NN, which is to decrease the error as well as improve the
histogram, making the image more legible. Here, we employ accuracy. Optimizers are used to achieve the optimal learning
the standard equalization technique described in to guarantee rate. The adaptive moment estimation (Adam) approach,
that all of the pixels are sharp. which is similar to the standard stochastic gradient descent
(SGD) technique, was applied in our model. Adam optimizer
takes the finest features of AdaGrad and RMSProp and
F. Defect Prediction
combines them.
The settings for the specialized CNN architecture utilized
in this flaw finder may be found here. To ensure that the roll IV. RESULTS AND DISCUSSIONS
can be reviewed as rapidly as possible, a simpler architecture Both CNN models are trained using use Tensorflow, a
was utilized in place of more complicated state-of-the-art popular deep learning platform. We use a random seed to
deep learning models so that the pictures from several seed the model with initial data during training. In this
cameras may be processed in parallel. flaw categorization is example, we start with a learning rate of 1e-4 and decrease it
left until later to allow the system to focus on flaw detection by 0.04 every four epochs. In order to prevent over-fitting,
at this stage. Fig. 4 provides a high-level representation of the we used a weight-decay hyper-parameter of 1e-6 for the CNN.
CNN's layers and their relative sizes. Four convolution and To further refine the model, an ADAM optimizer is used.
max-pool layers are followed by two fully-connected layers Here, the activation function is represented by Relu. Training
in this design. Each and every activation was performed using time is reduced and the risk of a vanishing or exploding
ReLU. This CNN was trained and verified using the Fabric- gradient is mitigated by using batch normalization after
Net-Dataset, as was previously indicated. The network's activation. This trained model is used to foresee whether or
ability to distinguish the fabric defect traits was further not flaws exist in the given box-image. Then, we may extract
validated by visualizing activations at each layer. a feature map that is an accurate representation of the image's
defect distribution. The global defect detection model can tell
us if and what kinds of flaws are present in the original picture
by locating them in this feature map. Image before any
processing is shown in Fig. 5.
I nput
1D Convolution MaxPooling 1D Convolution MaxPooling

Output

Flattern MaxPooling 1D Convolution

Defect Non - Defect

Fig. 4. CNN Architecture

Fig. 5.Preprocessed Image
To reduce the input size from 224 by 224 by 1 to 28 by 28
by 192, the network's first stages comprised four The original photos of the damaged cloth are shown in Fig.
convolutional layers and two pooling layers. The convolution 6. The following graphic demonstrates the image histogram
kernel is just 3 by 3, down from the more typical 7 by 7 and prior to enhancement, revealing why the original photos look
5 by 5, to prevent representational bottlenecks. The two fuzzy when computing cost is taken into account.
Inception components are put into effect right after the
Inception portion of the network, with Module 1 of Inception
is used to make the network sparser, whereas module 2 is
used to make the grid sparser without increasing its size. This
final part is indicative of the structure of a convolutional
network. The initialization phase of a network produces an
output of 7 by 7 by 2048 elements. After that, we combine
the feature vectors and lower the feature's dimension with a
7x7 pooling layer and a 1x1 convolution layer. To avoid
overfitting, we employ a dropout layer with a 40% ratio of
discarded outputs. The last stage of the network is a classifier
Fig. 6. Original Image and Histogram Image.
made up of a linear layer with softmax loss.

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.
 A. Evaluation Metrics
In this research, we evaluate the effectiveness of several
defect detection algorithms using the following four criteria.
Successful detection rates (Acc), area under the curve (AUC),
and false alarm rates (FR) are used as assessment measures.
The suggested approach outperforms the other networks in
terms of accuracy (97.34%), making it the preferred option.
Table I and Fig. 7 show the results of several test accuracy
measurements for different architectural designs.

Table I. Evaluation of Model Accuracy.

Model Accuracy
SVM 93.44
KNN 91.76
Inception 93.99 Fig. 9. Train Accuracy and Test Accuracy
MobileNet 92.83
The following comparison between the effectiveness of the
ResNet50 94.10
suggested technique and the other five representation defect
VGG16 90.19 identification techniques serves to further confirm the
VGG19 91.71 efficacy of the suggested structure in fabric identification.
Proposed Model 97.34 Some of the most prevalent methods to detecting fabric
defects is the differential evolution-based optimal Gabor
filter model (DEP), which has shown to be more successful
on fabric imperfections inspection than certain classical
fabric recognition methods in terms of recognition precision
and computational cost. To better detect faults in linear
fabric, such as those in the weft or warp directions, the Gabor
filters employed in this model are adjusted in the horizontal
and vertical dimensions, making the DEP a powerful tool. It
is generally agreed that the supervised methodology known
as the Gabor filter bank-based model (GB) is the gold
standard for identifying textiles.
Limitations in the detection approach are due to the DEP
model's reliance on feature extraction, especially is
performed using Gabor filters. Because of its sensitivity to
texture fluctuations, the GB model has a high false alarm rate.
Moreover, the results of these learning-based methods (both
DL and suggested) are much superior to those of other
Fig. 7. Accuracy Comparison of Proposed Model. approaches on our datasets. When compared to DL, the
suggested technique performs admirably in terms of detection
B. Learning Curve accuracy, but it is substantially more time-consuming.
Training loss and validation loss are shown to reduce in However, compared to existing approaches, the false alarm
Fig. 8, while training accuracy and validation accuracy are rate of the suggested technique is significantly lower. In
shown to improve in Fig. 9. After 25 iterations of training, the addition, the superior defect-detection accuracy of the
AlexNet model achieved an accuracy of 97.34 percent during suggested approach is reflected in its superior AUC
training and 0.183 percent during validation.
V. CONCLUSSION AND FUTURE SCOPE
In this research, we provide a powerful framework based
on deep learning with the intention of finding flaws in the
material. To begin, the fabric image is improved by
employing a very effective HE. Then, we use a convolutional
neural network (CNN) based on a sparse network architecture
called Inception to forecast where flaws could be. A feature
map is constructed that is similar to the input image after the
prediction has been made. The Deep Convolutional Neural
Network (DCNN) classifier was built, and after extensive
testing and training, it achieved a maximum accuracy of
97.34 percent. The experimental findings prove that the
suggested structure is superior. The main drawbacks of deep
learning-based systems are their high computing cost and
Fig. 8. Training Loss and Validation Loss

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.
 time consumption compared to more conventional 2022 4th International Conference on Natural Language Processing
(ICNLP), Xi'an, China, 2022, pp. 20-25, doi:
approaches. Consequently, the present difficulty of this 10.1109/ICNLP55136.2022.00012.
investigation is how to reduce detection time without [6] R. Manikandan, T. Mathumathi, C. Ramesh, S. Arun, R.
sacrificing accuracy. The long-term goal of this effort is to Krishnamoorthy and S. Padmapriya, "Preservation of Higher Accuracy
make it more applicable to real-time applications and to Computing in Resource-Constrained Devices Using Deep Neural
include support for a wider variety of fault classes. Including Approach," 2022 Second International Conference on Artificial
Intelligence and Smart Energy (ICAIS), 2022, pp. 172-177, doi:
all feasible hardware parts, ideally those used in most textile 10.1109/ICAIS53314.2022.9742923.
inspection facilities. [7] D. Xia, Z. Yu and X. Deng, "A Real-time Unsupervised Two-stage
. Framework for Fabric Defect Detection," 2021 3rd International
Academic Exchange Conference on Science and Technology
REFERENCES Innovation (IAECST), Guangzhou, China, 2021, pp. 535-538, doi:
10.1109/IAECST54258.2021.9695639.
[1] S. Zhao, J. Wang, J. Zhang, J. Bao and R. Zhong, "Edge-cloud
collaborative fabric defect detection based on industrial internet [8] A. D. Gupta, K. Sathiyasekar, R. Krishnamoorthy, S. Arun, R.
architecture," 2020 IEEE 18th International Conference on Industrial Thiyagarajan and S. Padmapriya, "Proposed GA Algorithm with H-
Informatics (INDIN), Warwick, United Kingdom, 2020, pp. 483-487, Heed Protocol for Network Optimization using Machine learning in
doi: 10.1109/INDIN45582.2020.9442236. Wireless Sensor Networks," 2022 Second International Conference on
Artificial Intelligence and Smart Energy (ICAIS), 2022, pp. 1402-
[2] B. Zhang and C. Tang, "Study on Detection Method of Low Contrast 1408, doi: 10.1109/ICAIS53314.2022.9743120.
Fabric Defects Image Based on Regional Variation Saliency," 2019
IEEE 4th Advanced Information Technology, Electronic and [9] Y. Y. Yin and L. Q. Li, "A lightweight algorithm for woven fabric
Automation Control Conference (IAEAC), Chengdu, China, 2019, pp. defect detection," 2022 China Automation Congress (CAC), Xiamen,
1103-1106, doi: 10.1109/IAEAC47372.2019.8997628. China, 2022, pp. 1025-1029, doi: 10.1109/CAC57257.2022.10055203.
[3] G. Vladimir, I. Evgen and N. L. Aung, "Automatic Detection and [10] L. Zheng, X. Wang, Q. Wang, S. Wang and X. Liu, "A Fabric Defect
Classification of Weaving Fabric Defects Based on Digital Image Detection Method Based on Improved YOLOv5," 2021 7th
Processing," 2019 IEEE Conference of Russian Young Researchers in International Conference on Computer and Communications (ICCC),
Electrical and Electronic Engineering (EIConRus), Saint Petersburg Chengdu, China, 2021, pp. 620-624, doi:
and Moscow, Russia, 2019, pp. 2218-2221, doi: 10.1109/ICCC54389.2021.9674548.
10.1109/EIConRus.2019.8657318. [11] R. Krishnamoorthy, S. Arun, N. Sujitha, K. M. Vijayalakshmi, S.
[4] W. Chong, W. Jinghua, W. Jing and D. Huan, "Fabric Defect Detection Karthiga and R. Thiagarajan, "Proposal of HMAC based Protocol for
Method Based on Projection Location and Superpixel Segmentation," Message Authenication in Kerberos Authentication Protocol," 2022
2022 4th International Conference on Natural Language Processing Second International Conference on Artificial Intelligence and Smart
(ICNLP), Xi'an, China, 2022, pp. 20-25, doi: Energy (ICAIS), 2022, pp. 1443-1447, doi:
10.1109/ICNLP55136.2022.00012. 10.1109/ICAIS53314.2022.9742992....
[5] W. Chong, W. Jinghua, W. Jing and D. Huan, "Fabric Defect Detection
Method Based on Projection Location and Superpixel Segmentation,"

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on February 19,2025 at 03:42:25 UTC from IEEE Xplore. Restrictions apply.