Research Reports on Computer Science

https://ojs.wiserpub.com/index.php/RRCS

Research Article

Deep Learning Based Fabric Defect Detection

Syeda Rabia Arshad* and Muhammad Khuram Shahzad

Faculty of Computing, National University of Sciences and Technology, Islamabad, Pakistan

Email: sarshad.msds20seecs@seecs.edu.pk

Received: 25 December 2023; Revised: 25 February 2024; Accepted: 5 March 2024

Abstract: Ensuring quality standards is a crucial stage within the textile sector. Automated classification of the
fabric defects is a vital step during the fabric manufacturing process in order to prevent any faulted fabric from
being supplied to the market. The defects on the surface of the fabric were manually identified by the
individuals but this poses problems in terms of human-error and is also time-consuming. Efforts have been made
to achieve better precision in defect detection through image processing studies, leading to the development of
automated systems. In this study, some high-performing deep learning models are applied including ResNet and
VGG-16 and illustrated how these algorithms can be used in the domain of textile manufacturing for fabric
defect detection. A combination of images are used ranging from patterned and textured to plain for better
defects recognition on any given fabric. The algorithm VGG-16 has displayed 73.91% accuracy while the
ResNet algorithm has shown 67.59% accuracy.

Keywords: deep learning, computer vision, ResNet, VGG-16, fabric defect detection

1. Introduction
Fabric defects are flaws in the material that diminish their quality, reducing their value and utility. Once
these defects are in the market, they cannot be rectified or reversed. Hence, ensuring a flawless product is
essential to deliver high-quality goods [1]. Quality control plays a pivotal role in guaranteeing customer
satisfaction and reducing production costs.
Defects are caused in the fabric during the manufacturing process due to various reasons like weaving,
knitting process, raw material, spinning and dying. As a result, faulted fabrics are created which may lower the
customer satisfaction. Substandard products are mostly returned to the textile mill for repair or replacement
which increases the overall cost apart from the production charges. Hence, it is necessary to identify the defects
on the fabric (if any) because only then the manufacturers will be able to figure out how the defects are being
caused and then the necessary steps could be taken to rectify this issue. Once we know how the defects are being
manifested, the process of manufacturing could be improved so that the defects do not occur in the first place,
instead of fixing the defects whenever it occurs. For creating high-quality products to be delivered into the
market, it is crucial that the fabric manufacturing process is improved from scratch [2, 3].
Another challenge in fabric defect detection is that detecting surface defects on fabric is a manual process
that necessitates the employment of an individual to physically inspect each piece of fabric, thoroughly
examining it from all angles in search for potential flaws. This is a time-consuming task, as a single person
cannot perform it as rapidly as needed, resulting in increased labour costs since one person alone cannot
efficiently handle this labour-intensive duty. Furthermore, this manual inspection process is prone to errors,
given that fabric is produced rapidly by machines, making it challenging to identify defects with the naked eye.
Copyright ©2024 Syeda Rabia Arshad, et al.
DOI: https://doi.org/10.37256/rrcs.3120244156
This is an open-access article distributed under a CC BY license
(Creative Commons Attribution 4.0 International License)
https://creativecommons.org/licenses/by/4.0/

Volume 3 Issue 1 |2024| 1 Research Reports on Computer Science

Some flaws may go unnoticed, potentially leading to the production of defective fabric. Hence, there is a
compelling need for the automation of this process. Automated defect detection, without the need for human
intervention, will eliminate the manual labour traditionally associated with this task. Being automated, it will
exhibit a reduced susceptibility to errors, scrutinise details thoroughly, and, not least, prove to be cost-effective.
Along with finding out any defects on the fabric, it is important to classify the defects into their respective types.
The defects must be occurring in a specific pattern or might show some similar behaviour. If we figure out some
similar types of defects, categorization can get much easier. One of the many benefits of categorization is that
the manufacturing process can be improved by focusing on those particular steps that are causing some specific
types of defects. With the count of each class of defect, we can know which step in the process is the most
problematic.
Some other researches done in this field are those that employ the methodology of grey-level co-
occurrence matrix [4, 5], Gabor filter [6, 7], CZI-Net [8] and others but these mentioned here are either too
computationally heavy or are not easily scalable as the number of samples increases. In this study, an automated
solution for fabric defect detection is created using deep learning algorithms like ResNet and VGG-16. This
research aims to provide an efficient solution to overcome the limitations arising from a lack of data. Within this
study, an automated resolution is created for the issue being discussed. Along with the detection of defects, this
system also classifies the detectors into their respective categories which are vertical defects, horizontal defects,
and holes. The system is trained on plain and patterned fabrics which means it is able to detect defects on any
type of fabric, be it plain, textured or patterned. The system inputs the images of the fabric and it detects defects
on the surface along with its respective categories. This research has produced remarkable results, characterised
by a substantial enhancement in performance. Moreover, it has achieved equitable outcomes across all three
categories, effectively mitigating the considerable data imbalance.

2. Literature Review
Talking about some of the studies related to this field, a method known as Gabor filter which utilises one-
class classification (OCC) alongside Gabor filters. In OCC, a skilled classifier can define a classification
boundary based solely on the information from the positive class. Consequently, OCC simplifies the problem by
focusing exclusively on positive fabrics. This eliminates the need to collect and handle various fabric defects, as
well as worries about imbalanced datasets or the lack of negative samples. The researchers have developed a
Gabor filter bank with a range of orientations and bandwidths to analyse different fabric textures. This method
facilitates the creation of customised Gabor filter designs for each fabric texture [9]. The Gabor filter acts as a
strong representation of the simple cells found in the human visual cortex, effectively capturing textures. It is
formed by multiplying a sinusoidal waveform with a Gaussian function. The researchers focused on assessing
the model's effectiveness with plain, patterned, and rotated fabric images.
Another research done in this field is the grey-level co-occurrence matrix, also referred to as the spatial
grey-level dependence method, which aggregates the probability of certain pairs of grey levels co-occurring
under two criteria: the distance between pixels and their relative orientation [10]. The probability assessment is
determined by tallying the occurrences of various grey levels within a defined region, while taking into account
a specified displacement value. It's imperative for the statistics to remain unchanged despite shifts in grey levels,
ensuring that classification is not influenced by tone variations [11]. To generate the co-occurrence data, various
parameters must be defined, such as the window size, pixel distances (d), orientation (theta), and the number of
quantified grey levels (G). A moving window systematically moves across the entire image, computing the
GLCM at each position it covers. Subsequently, the energy feature is extracted from the GLCM. Then these
energy values are compared to a reference, and if any deviation exceeding a specific threshold is detected, it is
identified as a defect. The smallest window size must be designed to encompass a minimum of one periodic
pattern within the pattern or the texture.
After analysing these related works in the field of fabric defect detection, it can be concluded that most of
these studies use a window or a filter-based approach. The Gabor filter uses a linear filter for texture analysis. It
checks for any specific frequency content in the image around the point of analysis. Due to the linear nature of
the filter, it takes too much time to perform the defect detection process. It captures the features and stores them
into a feature vector and the dimensions of the feature vector also keeps on increasing due to the linearity of the
filter. This also causes redundancy of features which reduces the recognition rate. On the other hand, the grey-

Research Reports on Computer Science 2 | Syeda Rabia Arshad, et al.

level co-occurrence is a matrix that is defined over an image to be the distribution of co-occurring pixel values
at a given offset. It converts an image into matrix of pixels and finds the combination of grey levels found
within the image. However, the limitation of this approach is that it has limited capability of capturing texture
information.
Our research caters for all of these gaps. It uses deep learning algorithms which do not use the linear filter,
rather the ResNet uses the residual connections within the network and this helps with maintaining the
information flow throughout the network. And VGG-16, which is a type of convolutional neural network (CNN),
uses very small convolutional filters, usually 3x3 and this is how the problem of linear filtering is handled.
These powerful deep learning architectures are also designed to deal with a huge dataset with hundreds or
thousands of images, so the issue of high processing time, tardiness or scalability is also resolved. Deep learning
models learn through data and since the dataset that is being used consists of a variety of fabric images including
plain, patterned, and textured fabric, that’s how these models can detect defects on any type of fabric. In this
way, the drawback of restriction in capturing detailed texture information is also covered in this research.

3. Materials and Methods

In order to achieve the best results from a deep learning model, it is important that we feed good and
relevant samples of images to our model. Then carefully selecting the best suited combination of
hyperparameters of your model also plays a vital role for the training of the model. We tweaked some of the
hyperparameters of our model like learning rate, number of epochs to check where our model performs its best.
Then testing your model after training on unseen data is necessary to see how your model performs in a real-life
situation where the data is similar to the training data but is not exactly the same.

3.1 Dataset
Having a good dataset is the key to achieving the optimal results in a deep learning model. In our dataset,
we included a variety of sample images including varied patterns, textured and plain fabric images. The dataset
is divided into three classes of defects which are vertical defects, horizontal defects and holes. Vertical and
horizontal defects are those in which the fabric has a stretch mark in form of a line on top or bottom and in the
left or right (respectively) of the fabric while holes are those defects in which the fabric has a circle or oval
shaped flaw anywhere on the fabric.
All the images including the training and testing images have been converted to grey-scale. The benefit of
working with monochrome images is that it reduces the number of colour channels which in return reduces the
computation power required to pre-process the data. Less number of channels means lesser number of
parameters our model needs to take into account for classification. During pre-processing, all the images were
resized to 224 x 224 pixels. The dataset comprises 3630 images in total which are split into training and testing
samples in a 80:20 ratio, where the 80% of our dataset is reserved for training purposes and the rest 20% is used
for testing our model.
The dataset is curated from three different sources, first is One-Shot Dataset for Fabric Detection shown in
figure (Figure 1), second is Fabric Defect Detection dataset from Kaggle [12] shown in figure (Figure 2), and
the third is Tilda Textile Texture Database [13] which is shown in figure (Figure 3). Some data samples from all
of these datasets are given below:

One-shot dataset:

Figure 1. One-shot Fabric Defect Dataset

Volume 3 Issue 1|2024| 3 Research Reports on Computer Science

Fabric defect dataset:

Figure 2. Kaggle Fabric Defect Dataset

Tilda:

Figure 3. Tilda Fabric Defect Dataset

The types of images present in the dataset are plain, textured and patterned images. Plain fabric is a
material with minimal patterning and usually solid or tonal in colour. Patterned fabrics are the ones that have a
repeating of an element or motif used to create a unique decoration on fabric. Texture refers to the feel and
appearance of fabric, encompassing qualities such as roughness or smoothness, coarseness or fineness, softness
or stiffness etc. and these textures can be different on each fabric. The image samples of these fabric types are
displayed in the image below (Figure 4).

Figure 4. Plain, Patterned and Textured images

3.2 ResNet
A Residual Neural Network (ResNet) is a type of deep learning model where the weight layers are trained
to learn residual functions with respect to the input of each layer. ResNets effectively address the well-known
issue of vanishing gradients. This problem occurs in deep neural networks when, during back propagation,
gradients become extremely small after multiple applications of the chain rule, leading to weight values not
updating and hindering learning [14]. ResNets counter this by allowing gradients to flow directly through skip
connections, which enables them to propagate backward from later layers to earlier filters, preventing the

Research Reports on Computer Science 4 | Syeda Rabia Arshad, et al.

vanishing gradient problem and facilitating learning [15]. A pre-trained ResNet model is used by TensorFlow.
In the below figure (Figure 5), the summary of layers in ResNet model is shown:

Figure 5. ResNet Layers Summary

Before training our model we first pre-processed our dataset. All images underwent resizing to dimensions
of 224 x 224 pixels, and a batch size of 10 was chosen. Given that there were multiple classes to predict, the
class mode is configured as “sparse.” Our dataset pre-processing began with the utilisation of Keras
ImageDataGenerator [16]. This tool was employed to transform the data based on the specified parameters. All
images underwent resizing to dimensions of 224 x 224 pixels, and a batch size of 10 was chosen.
The total parameters of our model are 266,411,910 out of which 207,834,118 are the trainable parameters
and 58,577,792 are non-trainable parameters. The ResNet architecture comprises an initial convolution and
pooling step, succeeded by four consecutive layers that exhibit similar behaviour. Each of these layers adheres

Volume 3 Issue 1|2024| 5 Research Reports on Computer Science

to an identical pattern. They carry out a 3x3 convolution with a consistent feature map dimension of [64, 128,
256, 512] respectively. Additionally, they skip the input every two convolution operations. Notably, the width
and height remain unchanged throughout the layer notably, the width and height remain unchanged throughout
the layer. The optimiser used in this model is ’Adam’ optimizer and the activation function used is ’ReLU’.
The ’learning rate’ chosen for this model is 0.001, the number of epochs is 50 and the output size is 3. In the
following image (Figure 6), ResNet architecture can be seen.

Figure 6. ResNet Architecture Diagram

3.3 VGG-16
Visual Geometry Group (VGG) is 16 layers in depth. This model is also based on the Convolutional Neural
Network (CNN). A convolutional neural network consists of an input layer, an output layer, and several hidden
layers in between. This model comprises a total 13 convolutional layers, 5 max pooling layers, and 3 dense
layers, adding up to 21 layers in terms of architecture. However, it has only 16 weight layers, meaning there are
16 layers with learnable parameters. VGG-16 expects an input tensor with dimensions of 224 pixels in height,
224 pixels in width, and 3 channels for RGB colour information. One of the distinctive features of VGG16 is its
emphasis on simplicity and uniformity in architectural design. Instead of employing a multitude of
hyperparameters, VGG-16 predominantly utilises 3x3 convolution layers with a stride of 1, consistently
applying the same padding technique. Additionally, it employs max pooling layers with 2x2 filters and a stride
of 2. This consistent and straightforward design is a notable characteristic of VGG16. This model’s design
adheres to a uniform arrangement of convolutional and max- pooling layers across its entire architecture.
Following the stack of convolutional layers in VGG16, there are three Fully-Connected (FC) layers. The final
layer is the softmax layer.
We pre-processed the data by resizing it into 224 x 224 pixels. A little scaling is applied between the range
of 0.8 and 1.0. Then image normalisation was applied. Image normalisation is a crucial process that ensures
accurate comparisons across various data acquisition methods and different instances of textures. Normalising
pixel values is particularly advised for imaging modalities that don’t directly correspond to absolute physical
quantities. After that, we have used a pre-trained model by PyTorch [17] which had 135,310,918 total
parameters out of which 1,050,374 are trainable while 134,260,544 are non-trainable parameters. The batch size
was set to 128, Learning rate to 0.005, dropout rate to 0.4 [18] and the output size is 3. The summary of our
VGG-16 model can be seen in the figure (Figure 7) and the architecture diagram of the VGG-16 model is
displayed in (Figure 8).

Research Reports on Computer Science 6 | Syeda Rabia Arshad, et al.

 Figure 7. VGG-16 Layers Summary

Figure 8. VGG-16 architecture

Volume 3 Issue 1|2024| 7 Research Reports on Computer Science

4. Results and Discussion
ResNet is a very complex deep learning model. It requires high computational resources and has a large
number of parameters to train. That’s why it requires a large number of data samples categorised in a large
number of classes. If the dataset is large and varied enough, then the ResNet does a great job in identifying
patterns and classifying the images. The accuracy achieved from the ResNet model is 67.59%. The model was
executed for a total of 50 epochs. The graph of loss recorded during training and testing of the model can be
seen in figure (Figure 9). Likewise, the training and testing accuracies over the training and testing period of the
ResNet model can be seen in figure (Figure 10).

Figure 9. ResNet Training and Testing Loss

Figure 10. ResNet Training and Testing Accuracy

Research Reports on Computer Science 8 | Syeda Rabia Arshad, et al.

 The graph of loss recorded during the training and testing phase of the VGG-16 model is depicted in figure
(Figure 11). Similarly, the training and testing accuracies over the training and testing period of the VGG-16
model can be seen in figure (Figure 12).

Figure 11. VGG-16 Training and Testing Loss

Figure 12. VGG-16 Training and Testing Accuracy

VGG-16 has more parameters than ResNet and therefore is a deeper architecture and it has shown better
performance than ResNet with an overall accuracy of 73.91% achieved at 13th epoch. The model was executed
for a total of 30 epochs. This transfer learning model re-used the knowledge of pre-trained weights and therefore,
lesser parameters were trained due to which, the accuracy achieved was higher than ResNet. Transfer learning
made the model size substantially smaller, because it used global average pooling instead of the fully connected
layers. Since the weights are not updated and the knowledge is transferred throughout the model, the use of
global average pooling is a smarter option than fully connected layers. This gives the model an opportunity to
effortlessly flow the knowledge through the model.

Volume 3 Issue 1|2024| 9 Research Reports on Computer Science

 The comparison of accuracy, precision, recall, and F1 Score of both the models can be seen in the below
table (Table 1).

Table 1. Result comparison of models

Model ResNet VGG-16

Accuracy 67.59 % 73.91 %

Precision 0.26 0.64

Recall 0.27 0.65

F1 Score 0.26 0.64

5. Conclusion
Artificial Intelligence finds applications in a wide range of industries, and it has demonstrated impressive
achievements in the textile sector. Computer Vision (CV) is harnessed to spot imperfections on fabric surfaces
by recognizing patterns and singling out any unusual designs, deeming them as defects. These defects may take
the form of horizontal or vertical flaws, or even holes. This research showcases the advantages of employing
various deep learning methods for the identification of defects on fabric surfaces. Additionally, the study
categorises these defects into different types, such as horizontal, vertical, and holes. The accuracy obtained from
ResNet was 67.59% while the accuracy obtained from VGG-16 was 73.91% which was higher than ResNet.

6. Limitations and Future Work

In future, we would like to employ better deep learning techniques like MobileNet, VGG-19 etc. so that
better performance results can be achieved. Right now this research is mainly on image based dataset, in future
we would like to extend the work to videos as well so that real-time fabric defect detection can be made possible.
This study is also limited to one defect per image, further enhancements can be made which detect multiple
types of defects in an image if there are any.

Acknowledgments
This acknowledgment serves as a tribute to everyone who contributed to the meaningfulness of my
research journey.

Conflict of Interest
There is no conflict of interest for this study.

References
[1] Karayiannis, Y.A.; Stojanovic, R.; Mitropoulos, P.; Koulamas, C.; Stouraitis, T.; Koubias, S.;
Papadopoulos, G. Defect detection and classification on web textile fabric using multiresolution
decomposition and neural networks. In Proceedings of the IEEE International Conference on Electronics,
Circuits and Systems, Pafos, Cyprus, 5–8 September 1999, https://doi.org/10.1109/ICECS.1999.813221.
[2] Chan, C.-H.; Pang, G. Fabric defect detection by Fourier analysis. IEEE Trans. Ind. Appl. 2000, 36, 1267–
1276, https://doi.org/10.1109/28.871274.
[3] Jain, A.K.; Tuceryan, M. The Handbook of Pattern Recognition and Computer Vision. Chapter 11, Texture
analysis. World Scientific Publishing: Singapore, 1992.

Research Reports on Computer Science 10 | Syeda Rabia Arshad, et al.

[4] Chellappa, R.; Manjunath, B.S. Texture classification and segmentation: tribulations, triumphs and tributes.
In Foundations of Image Understanding, pp. 219–240, Springer: Boston, MA, USA, 2001. https://doi.org/
10.1007/978-1-4615-1529-6_8.
[5] Zhou, H.; Chen, Y.; Troendle, D.; Jang, B. One-Class Model for Fabric Defect Detection. arXiv preprint,
2022, arXiv:2204.09648, https://doi.org/10.48550/arXiv.2204.09648.
[6] Tong, L.; Wong, W.; Kwong, C. Differential evolution-based optimal Gabor filter model for fabric
inspection. Neurocomputing 2016, 173, 1386–1401, https://doi.org/10.1016/j.neucom.2015.09.011.
[7] Mak, K.-L.; Peng, P.; Yiu, K.-F.C. Fabric defect detection using multi-level tuned-matched Gabor filters. J.
Ind. Manag. Optim. 2012, 8, 325–341, https://doi.org/10.3934/jimo.2012.8.325.
[8] Lowe, D.G. Distinctive Image Features from Scale-Invariant Keypoints. Int. J. Comput. Vis. 2004, 60, 91–
110, https://doi.org/10.1023/b:visi.0000029664.99615.94.
[9] Raheja, J.L.; Ajay, B.; Chaudhary, A. Real time fabric defect detection system on an embedded DSP
platform. Optik 2013, 124, 5280–5284, https://doi.org/10.1016/j.ijleo.2013.03.038.
[10] Kula, J.; Tunak, M.; Linka, A. Real-time quality control of fabric based on multivariate control charts. In
Proceedings of the 7th international conference TEXSCI 2010, Liberec, Czech Republic, 6–8 September
2010.
[11] Ranathunga, S. Fabric Defect Dataset. Kaggle. Available online: https://www.kaggle.com/datasets/rmshas
hi/fabric-defect-dataset (accessed on 24 December 2023).
[12] Ben Salem, Y.; Abdelkrim, M.N. Texture classification of fabric defects using machine learning. Int. J.
Electr. Comput. Eng. (IJECE) 2020, 10, 4390–4399, https://doi.org/10.11591/ijece.v10i4.pp4390-4399.
[13] Targ, S.; Almeida, D.; Lyman, K. Resnet in resnet: Generalizing residual architectures. arXiv preprint,
2016, arXiv:1603.08029, https://doi.org/10.48550/arXiv.1603.08029.
[14] He, F.; Liu, T.; Tao, D. Why ResNet Works? Residuals Generalize. IEEE Trans. Neural Networks Learn.
Syst. 2020, 31, 5349–5362, https://doi.org/10.1109/tnnls.2020.2966319.
[15] Gulli, A.; Pal, S. Deep learning with Keras. Packt Publishing Ltd: Birmingham, UK, 2017.
[16] Tammina, S. Transfer learning using VGG-16 with Deep Convolutional Neural Network for Classifying
Images. Int. J. Sci. Res. Publ. (IJSRP) 2019, 9, 143–150, https://doi.org/10.29322/ijsrp.9.10.2019.p9420.
[17] Arshad, S.R.; Obaid, I.; Gull, R.; Shahzad, M.K. Steel Defect Classification Using Machine Learning. In P
roceedings of 2022 16th International Conference on Ubiquitous Information Management and Communic
ation (IMCOM), Seoul, South Korea, 3–5 January 2022, https://doi.org/10.1109/IMCOM53663.2022.972
1728.
[18] Israr, H.; Khan, S.A.; Tahir, M.A.; Shahzad, M.K.; Ahmad, M.; Zain, J.M. Neural Machine Translation
Models with Attention-Based Dropout Layer. Comput. Mater. Contin. 2023, 75, 2981–3009,
https://doi.org/10.32604/cmc.2023.035814.

Volume 3 Issue 1|2024| 11 Research Reports on Computer Science