FABRIC DEFECT DETECTION BASED ON IMPROVED FASTER RCNN

International Journal of Artificial Intelligence & Applications (IJAIA), Vol.12, No.4, July 2021
DOI: 10.5121/ijaia.2021.12402 23
FABRIC DEFECT DETECTION BASED ON
IMPROVED FASTER RCNN
Yuan He, Han-Dong Zhang, Xin-Yue Huang and Francis Eng Hock Tay
Department of Mechanical Engineering, National University of Singapore, Singapore
ABSTRACT
In the production process of fabric, defect detection plays an important role in the control of product
quality. Consider that traditional manual fabric defect detection method are time-consuming and
inaccuracy, utilizing computer vision technology to automatically detect fabric defects can better fulfill the
manufacture requirement. In this project, we improved Faster RCNN with convolutional block attention
module (CBAM) to detect fabric defects. Attention module is introduced from graph neural network, it can
infer the attention map from the intermediate feature map and multiply the attention map to adaptively
refine the feature. This method improve the performance of classification and detection without increase
the computation-consuming. The experiment results show that Faster RCNN with attention module can
efficient improve the classification accuracy.
KEYWORDS
Fabric defects detection, Faster RCNN, Convolutional block attention module, Deep learning
1. INTRODUCTION
In order to produce high quality garments, it is an important step to apply a defect detection link
in the process of fabric manufacturing to ensure the quality. Defect detecting is the process to
find out and locate defects on the surface of fabric. Finding out defects on fabric also improves
the efficiency of manufacturing process by abandoning unqualified intermediate products.
Traditionally, manual inspection which carried out on wooden board is the only method to assure
the quality of textile. Sometimes workers also do fine defects detection with the help of
equipment like magnifiers and microscopes. Manual defect detection can do prompt correction of
small defects. However, error may occur due to fatigue, and small defects are usually undetected
[1].
Since fabric defect detection has a great effect on the quality control of textile manufacture and
the conventional manual inspection method does not suit the requirement of developed automated
manufacture, automatic fabric defects detection becomes a natural way to improve fabric quality
and lower labor cost. Fortunately, with the development of deep learning technology and the
progress of computer vision technology, a new automated fabric detection method which can
replace manual inspection appears. By applying computer vision and machine learning
technology, automated visual inspection is widely used to detect the surface defects of machined
parts and components. According to the research of Rajalingappaa Shanmugamani [2] published
in 2015, visual inspection method can provide rapid quantitative assessment and improve quality
and productivity.
Two defects detection algorithms are compared in this project, the Faster RCNN and Faster
RCNN with convolutional block attention module (CBAM). The difference between these two

24
algorithms is that the backbone net is different. The backbone net for Faster RCNN is Resnet-
50[21], which is a 50 layers deep neural network and used for feature extraction of defects,
classification and regression. CBAM will combine with Resnet-50 to improve the performance.
Both algorithm detector is Faster RCNN. By comparing the result of two detect algorithms, the
role of attention module will be revealed, and the effect of the faster R-CNN is going to be
shown.
2. RELATED WORK
At present, textile defects detection approaches can be simply divided into spectral approach and
learning approach. Gabor filters provide the optimal joint position in spatial and frequency
domain [3], it becomes the most popular approach in spectral-based method. The initial
application of Gabor filter is to build a filter bank with numerous sets of filters, which is
predetermined the parameters in frequency and orientation [4]. In [5], Shu calculates the
frequency and direction data obtained from 16 Gabor filters convolution with 4 different angles
and scales to detect fabric defects. This method accuracy will be affected by the computationally
intensive and the frequency plane coverage. Bodnarova [6] utilizes optimal filters to reduce the
amount of filters, the computation time is greatly reduced resulting in an increased speed of
detection. However, the correct choice of optimal filter is difficult and crucial. Tong [7] has
developed composite differential evolution (CoDE) to optimize the parameters of Gabor filters,
and get high performance in limited samples. LI [8] integrated Gabor filter and Gaussian mixture
model to inspect simple texture defects, the classification accuracy from 360 images of 9
different defect types reach 87%.
Neural network has advantages in feature extraction, segmentation and optimization tasks of
fabric defects detection area [9, 10]. In 2001, Stojanovic [11] proposed a three-layer back-
propagation neural network for low-cost fabric real-time detection, and the accuracy achieved
86%. Kumar [12] combines forward neural network with Principal Component Analysis (PCA)
for faster detection. In [13], Kuoproposes a three-layers back-propagation neural network to
detect white fabric defects. This model a high dimensional system by non-linear regression
algorithm, achieve 91.88% recognition accuracy for 160 simple defects image. Asimilar
architecture of network is proposed in [14], the detection accuracy of holes and oil stain of twill
fabric reach 91% and 100%. However, the amount of sample is limit and the reliability is
unknown. Semnani and Vadood [15] develop an intelligent model based on artificial neural
network to estimate the appearance of knitted fabrics.
The current textile defect detection field mainly uses adjusted Faster RCNN algorithms as
backbone network for inspection. Since textile industry is an important industry in China,
researchers from China have more sufficient dataset to experiment. In the past two years, Che
[16], Cai [17] and An [18] proposed their improved Faster RCNN method to recognize and
localize the fabric defects, and got good performance. However, they chose to describe their
algorithm in Chinese, which makes it difficult to communicate with other researchers. Zhou [19]
combined Faster RCNN with Feature Pyramid Network and Deformable Convolution Network to
accurately extract the defects, and reduced the time-consuming. The dataset Zhou’ group
usedwas established in 2007 (DAGM dataset), the quality of fabric image is relatively poor, and
the types of defects also small. Therefore, it is questionable whether this algorithm can be used
for real-time detection. Peng [20] introduced an improved prior anchor network to enhance the
Faster RCNN performance. Peng’ group used amount of industrial cloth data to train the network,
and achieve 98.6% accuracy. Unfortunately, the dataset is still not released, and the result of this
algorithm is unreliable.

25
3. METHOD
3.1. Date Augmentation
In this paper, we collected 185 fabric defect images in three different types from textile factory.
They are broken hole, fly yarn and drop needle. Since the database is too small, we adapt random
real-time data augmentation to increase the amount of image, we flip the image in horizontal and
vertical direction, and distort the image.
(a) (b) (c)
(d) (e) (f)
Fig 1. The examples of fabric defects image and augmentation results. (a) is drop needle. (b) is broken
hole. (c) is fly yarn. (d) and (e) are (b) flipped in horizonal and vertical direction, (f) is (b) distorted and
resized in standard size.
The shape of image needs to be resized to 600 × 600 before training. In order to ensure no
distortion, we scale the original image to a certain proportion and add gray bars (scale on the long
side, fill 0 on the short side), and the resulting new image shape is 512 × 512. In the validation
process, just need to resize images normally to ensure the same size suitable for neural network.
While in the training process, there is warpAffine for database augmentation. We use random
scaling (from 0.25 to 1.75), random cropping, random flip (probability is 0.5) and color jittering.
Unfortunately, our dataset size is not enough for multiple augmentation method. Random
augmentation will actually lead to decrease in accuracy. After several experiment, we only use
our dataset as input. The dataset contains 500 defects image and 1500 defects free images.

26
3.2. Faster RCNN
Faster R-CNN does the job of object detection in this project. It is developed based on R-CNN
and Fast R-CNN technology and was proposed by Girshick and his team in 2015 [21]. Faster R-
CNN integrates the step of creating boundary boxes into CNN model. The overall frame of Faster
R-CNN is shown in Figure 2. Faster RCNN contains four main parts, which are convolution
layers, region proposal network, region of interest (RoI) pooling and classification.
In this paper, we use ResNet-50 [22] as the backbone net. Firstly, the input fabric defects image
feature map is extracted by convolutional layer with ReLu activation function and pooling layers.
Region proposal network (RPN) is used to generate proposals with feature matrix. We use
boundary conditions and non-maximum suppression [23] to select the appropriate anchor on
feature map. The output of regression layer indicates the coordinate position of fabric defect. RoI
pools collect region proposal and feature maps, the bounding box regression provides the final
exact target box position.
Table 1. ResNet-50 Architecture
Layer Name Output Size ResNet-50
conv1 112 × 112 × 64 7 × 7, 64, stride 2
conv2_x 56 × 156 × 64
3 × 3 max pool, stride 2
[
1 × 1, 64
3 × 3, 64
1 × 1, 256
] × 3
conv3_x 28 × 28 × 128 [
1 × 1, 128
3 × 3, 128
1 × 1, 512
] × 4
conv4_x 14 × 14 × 256 [
1 × 1, 256
3 × 3, 256
1 × 1, 1024
] × 23
conv5_x 7 × 7 × 512 [
1 × 1, 512
3 × 3, 512
1 × 1, 2048
] × 3
average pool 1 × 1 × 512
softmax 1000
CBAM (Convolutional Block Attention Module) [24] is an efficient improvement algorithm in
detection presented by Sanghyun Woo, Jongchan Park and their team in 2018. The structure of
these two modules shows in Figure 3.The function of the entire attention module can be
expressed by the following two equations. The first equation represents the function of channel
attention module and the second equation shows the function of spatial attention module.

27
Fig 2. An illustration of Faster R-CNN model
Fig 3. Diagram of each attention sub-modules
𝐹′
= 𝑀𝑐(𝐹) ⊕ 𝐹 (1)
𝐹′′
= 𝑀𝑠(𝐹′) ⊕ 𝐹′
⊕ represents element-wise multiplication. 𝑀𝑐 represents the operation of attention extraction
on the channel dimension, and𝑀𝑠represents the operation of attention extraction on the spatial
dimension. 𝐹 ∈ ℝ𝐻×𝑊×𝐶
is the intermediate feature map, 𝐹′
is the product of feature map
after channel attention module process, 𝐹′′
represents the final output feature map after
channel attention and spatial attention. 𝐻 means the height of input feature map, 𝑊 is the
width of input feature map. C is the channel number of input feature map.
Resize
600 × 600 × 3
Backbone (Resnet50)
Feature map
38 × 38 × 1024
3 × 3 conv
1 × 1 conv
1 × 1 conv
Proposal
Softmax
RoI Pooling
RPN
FC layers
FC layers
FC layers
Softmax
Bbox reg
Bbox reg
Class
Position

28
Fig 4. CBAM integrated with a ResBlock in ResNet
As shown in Figure 3, the input feature map suffers global average-pooling 𝐹𝑎𝑣𝑔
𝑐
and globalmaxi-
pooling𝐹𝑚𝑎𝑥
𝑐
in channel attention module. It is worth to note that the global average pooling and
global maximum pooling are used in parallel, which can minimize the information loss during the
pooling process. 𝐹𝑎𝑣𝑔
𝑐
and 𝐹𝑚𝑎𝑥
𝑐
forward to a shared network which composed of multi-layer
perceptron with one-hidden layer, to produce channel attention map 𝑀𝑐. It can be represented as
following equation (2):
𝑀𝑐(𝐹) = 𝜎 (𝑀𝐿𝑃(𝐴𝑣𝑔𝑃𝑜𝑜𝑙(𝐹)) + 𝑀𝐿𝑃(𝑀𝑎𝑥𝑃𝑜𝑜𝑙(𝐹)))
= 𝜎(𝑊1(𝑊0(𝐹𝑎𝑣𝑔
𝑐
)) + 𝑊1(𝑊0(𝐹𝑚𝑎𝑥
𝑐
))) (2)
Where 𝜎 presents sigmoid function, 𝑊0 ∈ ℝ𝐶/𝑟×𝐶
,𝑊1 ∈ ℝ𝐶×𝐶/𝑟
is the MLP layers wight, 𝑟 is the
reduction rate. First layer has ReLu activation function.
Different from channel attention module, spatial attention module applies a convolutional layer to
generate a spatial attention map to concatenate max-pooling and average-pooling. It can be
represented as following equation (2):
𝑀𝑐(𝐹) = 𝜎 (𝑓7×7(𝐴𝑣𝑔𝑃𝑜𝑜𝑙(𝐹));𝑀𝐿𝑃(𝑀𝑎𝑥𝑃𝑜𝑜𝑙(𝐹)))
= 𝜎(𝑓7×7|𝐹𝑎𝑣𝑔
𝑠
;𝐹𝑚𝑎𝑥
𝑠 |) (3)
Where 𝜎 presents sigmoid function, 𝑓7×7
represents a convolution operation with 7 × 7 size
filter. In this paper, we use sequential arrangement attention modules.
4. EXPERIMENT AND RESULT
In this paper, we used 90 percent samples in the database as training dataset. For one experiment,
50 epochs are run with a base learning rate of 0.001 and 0.0001 with another 50 epochs. In both
of the two iteration stages, the learning rate will decay to 92% of the original in each iteration.
Batch size is 1 and we carry 2 experiments for each method separately. Totally, we have taken
40,000 iterations on each method. The test environment is a HP desktop with an Inter(R)
Core(TM) i5-4200 3.3 GHZ CPU, the simulation software is python2.7.
(a) (b)

29
(c) (d)
(e)
(f)
Fig 3. Results display and comparison between the different backbones: Resnet50 (left figure) and
Resnet50 + CBAM (right figure). (a), (b) drop needle, (c), (d), (e)broken hole, (f) fly yarn.
There are 185 images in the database, 10% of which is test dataset. Limited by the sample size,
only 3 types of defects are involved in training and testing. All of the three types of defects can
be detected under both the Faster R-CNN with a modified backbone net and the normal one as
shown in the Fig 3. we just show some examples of the detection results. It can be seen that both
these two methods can detect the fabric defects accurately. For the two figures in each group, the
left one is the result of the normal Faster R-CNN using Resnet 50 as backbone; the right one is
the result of the Faster R-CNN with a modified backbone using Resnet50 and CBAM. Both of
the two experiments use Faster R-CNN as the detector. It is clear that using Faster R-CNN with
CBAM leads a higher confidence of the detected defect object compared with the normal one.
After adapting CBAM, the confidence is 2%-3% higher than before. Therefore, the detection
model can be regarded more reliable.
Table 2. Object detection AP (%) of each class and mAP (%) on the test dataset.
Backbone Detector AP of broken
hole
AP of drop
needle
AP of fly
yarn
mAP
ResNet-50 Faster R-CNN 93.26 61.03 55.50 69.93
ResNet-50+CBAM Faster R-CNN 95.31 61.89 55.48 70.89
Global recognizable ratio is the percentage of samples that can be recognized in all test samples.
Correspondingly, we can describe the percentage of samples that can be recognized in a certain
class in all samples of that class as class recognizable ratio. Broken hole is the most easily to be
detected since its structure feature is obvious and simple. As for drop needle, it is not very easy to
be detected since it has many kinds of different structures in the database. Fly yarn is most
difficult to be detected since its sample size is small and it does not have special structures.

30
Table 3. Global recognizable ratio (%) and class recognizable ratio (%) on the test dataset.
Backbone Detector Global ratio Broken hole Drop needle Fly
yarn
ResNet-50 Faster R-CNN 46 86 50 36
ResNet-50+CBAM Faster R-CNN 62 100 56 67
In addition, to determine whether every recognized sample is classified correctly, the ratio of
samples with correct classification in all samples that can be recognized is accuracy. It can also
be divided into global accuracy and category accuracy, like the recognizable ratio. Adapting
CBAM makes the global accuracy becomes 1% higher than before. For each class, the accuracy
of drop needle is the highest, although it is not easy to be detected out, its accuracy is high. With
CBAM, broken hole can reach 100% accuracy and accuracy of fly yarn is also improved.
Table 4. Global accuracy (%) and class accuracy (%) of each class on the test dataset.
Backbone Detector Global accuracy Broken hole Drop needle Fly
yarn
ResNet-50 Faster R-CNN 87 83 100 71
ResNet-
50+CBAM
Faster R-CNN 88 100 100 83
In the experiment, we use RGB images with shape (600,600,3) to simulate the real manufacturing
environment for training and testing. The detection speed is 1.10 images per second, which
proved this model is not suitable for the real-time detection. If the video captured by industrial
camera is input, this model will cause a significant delay. The detection speed shows in [19] and
[20] is close to 19 images per second. However, they use the grey level image with shape
(128,128). Theoretically, the design of RPN requiring high computation-consuming, which leads
to the slow detection speed. Application of CBAM in our experiment used to improve the
detection accuracy, but how to improve the detection speed is still a research gap which need to
be solved in the future.
5. CONCLUSIONS
In this paper, we introduce a kind of attention mechanism called CBAM into Faster RCNN and
compare the performance of normal Faster RCNN and the improved Faster RCNN with CBAM
in fabric defects detection area. For Faster RCNN part, we use Resnet50 as the backbone to
extract the feature map, feature map is used in the RoI pooling directly and RPN for proposal
extraction meanwhile. Soft-NMS is involved in the RPN and classifier to remove the extra boxes
and improve the detection accuracy. CBAM is a simple and effective attention module for feed
forward convolutional neural networks.
Given an intermediate feature map, the CBAM module sequentially infers the attention map
along two independent dimensions (channel and space), and then multiplies the attention map
with the input feature map for adaptive feature optimization. Since CBAM is a lightweight
general-purpose module, the overhead of this module can be ignored and it can be seamlessly
integrated into any CNN architecture, and it can be trained end-to-end together with the basic
CNN. The Channel Attention Module compresses the feature map in the spatial dimension to
obtain a one-dimensional vector before performing the operation. When compressing in the
spatial dimension, not only Average Pooling but also Max Pooling is considered. Average
Pooling and Max Pooling can be used to aggregate the spatial information of the feature map,
send it to a shared network, compress the spatial dimension of the input feature map, and sum and

31
merge element by element to generate a channel attention map. As far as a picture is concerned,
channel attention focuses on which content on the picture is important. Average pooling has
feedback for each pixel on the feature map, while maximum pooling is used for gradient back
propagation calculation, only the place with the largest response in the feature map has gradient
feedback.
In this paper experiment is carried out with the normal Faster RCNN with and without CBAM to
verify the positive effect of the CBAM attention module. The experiment result proved CBAM
can get better accuracy and recognizable ratio in our fabric defect database. Dueto the difference
in image quality and pattern complexity, some defects cannot be successfully detected. Collecting
fruitfulness high-quality fabric defect and defect-free image, updating existing network model
will be the focus of future research. The different training time between CBAM and Faster
RCNN is negligible. However, compared with the traditional approaches, the training time is
more time-consuming. Reduce the required time for training the model, realize real-time
detection is crucial to whether this technology can be applied in industrial production, which is
also the focus of future research.
REFERENCES
[1] Ngan, H. Y., Pang, G. K., & Yung, N. H. (2011). Automated fabric defect detection—a review.
Image and vision computing, 29(7), 442-458.
[2] Shanmugamani, R., Sadique, M., &Ramamoorthy, B. (2015). Detection and classification of surface
defects of gun barrels using computer vision and machine learning. Measurement, 60, 222-230.
[3] Kumar, A., & Pang, G. K. (2002). Defect detection in textured materials using Gabor filters. IEEE
Transactions on industry applications, 38(2), 425-440.
[4] Mak, K. L., & Peng, P. (2008). An automated inspection system for textile fabrics based on Gabor
filters. Robotics and Computer-Integrated Manufacturing, 24(3), 359-369.
[5] Shu, Y., & Tan, Z. (2004). Fabric defects automatic detection using Gabor filters. In Fifth World
Congress on Intelligent Control and Automation (IEEE Cat. No. 04EX788)(Vol. 4, pp. 3378-3380).
IEEE.
[6] Bodnarova, A., Bennamoun, M., & Latham, S. (2002). Optimal Gabor filters for textile flaw
detection. Pattern recognition, 35(12), 2973-2991.
[7] Tong, L., Wong, W. K., &Kwong, C. K. (2016). Differential evolution-based optimal Gabor filter
model for fabric inspection. Neurocomputing, 173, 1386-1401.
[8] Zhang, Y., Lu, Z., & Li, J. (2009). Fabric defect detection and classification using gabor filters and
gaussian mixture model. In Asian conference on computer vision (pp. 635-644). Springer, Berlin,
Heidelberg.
[9] Jain, A. K., Duin, R. P. W., & Mao, J. (2000). Statistical pattern recognition: A review. IEEE
Transactions on pattern analysis and machine intelligence, 22(1), 4-37.
[10] Egmont-Petersen, M., de Ridder, D., &Handels, H. (2002). Image processing with neural networks—
a review. Pattern recognition, 35(10), 2279-2301.
[11] Stojanovic, R., Mitropulos, P., Koulamas, C., Karayiannis, Y., Koubias, S., & Papadopoulos, G.
(2001). Real-time vision-based system for textile fabric inspection. Real-Time Imaging, 7(6), 507-
518.
[12] Kumar, A. (2003). Neural network based detection of local textile defects. Pattern Recognition, 36(7),
1645-1659.
[13] Kuo, C. F. J., Lee, C. J., & Tsai, C. C. (2003). Using a neural network to identify fabric defects in
dynamic cloth inspection. Textile Research Journal, 73(3), 238-244.
[14] Yin, Y., Zhang, K., & Lu, W. (2009). Textile flaw classification by wavelet reconstruction and BP
neural network. In International Symposium on Neural Networks (pp. 694-701). Springer, Berlin,
Heidelberg.
[15] Semnani, D., &Vadood, M. (2010). Improvement of intelligent methods for evaluating the apparent
quality of knitted fabrics. Engineering Applications of Artificial Intelligence, 23(2), 217-221.

32
[16] Che, Xiang-jiu(2019). Fabric defect recognition algorithm based on improved Fast RCNN. Jilin
DaxueXuebao. Gongxue Ban = Journal of Jilin University. Engineering and Technology Edition,
1(1), 2038.
[17] Cai, Zhao-Xin. (2021). Fabric Defect Recognition System Based on Faster RCNN.
JisuanjiXitongYingyong, 2, 83.
[18] An, Meng. (2021). Fabric Defect Detection Method Based on Improved Faster R-CNN
.XiaoXxnWeixinJisuanjiXitong, 42(5), 1029.
[19] Zhou, H., Jang, B., Chen, Y., &Troendle, D. (2020, September). Exploring Faster RCNN for Fabric
Defect Detection. In 2020 Third International Conference on Artificial Intelligence for Industries
(AI4I) (pp. 52-55). IEEE.
[20] Peng, P., Wang, Y., Hao, C., Zhu, Z., Liu, T., & Zhou, W. (2020). Automatic Fabric Defect Detection
Method Using PRAN-Net. Applied Sciences, 10(23), 8434.
[21] Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with
region proposal networks. arXiv preprint arXiv:1506.01497.
[22] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
[23] Neubeck, A., & Van Gool, L. (2006, August). Efficient non-maximum suppression. In 18th
International Conference on Pattern Recognition (ICPR'06) (Vol. 3, pp. 850-855). IEEE
[24] Woo, S., Park, J., Lee, J. Y., &Kweon, I. S. (2018). Cbam: Convolutional block attention module. In
Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).
AUTHORS
Yuan He is a Ph.d. Candidate in National University of Singapore, Singapore. He
received the B.Sc. degress in Northeastern University, P.R. China. His research interests
include defects detection, computer vision and image processing.
Dr. Francis TAY is currently an Associate Professor with the Department of
Mechanical Engineering, Faculty of Engineering, National University of Singapore. Dr.
Tay is the Deputy Director (Industry) for the Centre of Intelligent Products and
Manufacturing Systems, where he takes charge of research projects involving industry
and the Centre.
Zhang Handong is a master candidate in National University of Singapore, Singapore.
He received the B.S. in Harbin Institute of Technology, P.R. China. His research
interests include defects detection and mechanical design.
Xin-Yue Huang is a Master Candidate in University of Singapore, Singapore. She
received the B.Sc. degree in Sichuang University, P.R. China. Her research interests
include computer vision and image processing.

FABRIC DEFECT DETECTION BASED ON IMPROVED FASTER RCNN

More Related Content

What's hot (20)

Similar to FABRIC DEFECT DETECTION BASED ON IMPROVED FASTER RCNN (20)

Recently uploaded (20)

FABRIC DEFECT DETECTION BASED ON IMPROVED FASTER RCNN