Hindawi

Wireless Communications and Mobile Computing

Volume 2022, Article ID 1070405, 10 pages
https://doi.org/10.1155/2022/1070405

Research Article
Circuit Manufacturing Defect Detection Using VGG16
Convolutional Neural Networks

Sara A. Althubiti ,1 Fayadh Alenezi ,2 S. Shitharth ,3 Sangeetha K. ,3

and Chennareddy Vijay Simha Reddy4
1
Department of Computer Science, College of Computer and Information Sciences, Majmaah University,
Al-Majmaah 11952, Saudi Arabia
2
Department of Electrical Engineering, College of Engineering, Jouf University, Saudi Arabia
3
Department of Computer Science and Engineering, Kebri Dehar University, Kebri Dehar, Ethiopia
4
Department of Computer Science, Middlesex University, London, UK

Correspondence should be addressed to Sangeetha K.; [email protected]

Received 1 March 2022; Revised 20 March 2022; Accepted 24 March 2022; Published 16 April 2022

Academic Editor: Kuruva Lakshmanna

Copyright © 2022 Sara A. Althubiti et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Manufacturing, one of the most valuable industries in the world, is boundlessly automatable yet still quite stuck in traditionally
manual and slow processes. Industry 4.0 is racing to define a new era of digital manufacturing through Internet of Things-
(IoT-) connected machines and factory systems, fully comprehensive data gathering, and seamless implementation of data-
driven decision-making and action taking. Both academia and industry understand the tremendous value in modernizing
manufacturing and are pioneering bleeding-edge strides every day to optimize one of the largest industries in the world. IoT
production, functional testing, and fault detection equipment are already being used in today’s maturing smart factory
paradigm to superintend intelligent manufacturing equipment and perform automated defect detection in order to enhance
production quality and efficiency. This paper presents a powerful and precise computer vision model for automated
classification of defect product from standard product. Human operators and inspectors without digital aid must spend
inordinate amounts of time poring over visual data, especially in high volume production environments. Our model works
quickly and accurately in sparing defective product from entering doomed operations that would otherwise incur waste in the
form of wasted worker-hours, tardy disposition, and field failure. We use a convolutional neural network (CNN) with the
Visual Geometry Group with 16 layers (VGG16) architecture and train it on the Printed Circuit Board (PCB) dataset with
3175 RBG images. The resultant trained model, assisted by finely tuned optimizers and learning rates, classifies defective
product with 97.01% validating accuracy.

1. Introduction Computing (CC), sensors, and integrated systems. In Indus-

try 4.0, large-scale IoT systems and machine-to-machine
In recent years, the world has moved towards the Fourth communications (M2M) are integrated to enhance automa-
Industrial Revolution, which is also known as Industry 4.0. tion, communication, and self-monitoring without the need
Its defining characteristic is to build the boosted productivity for human intervention. The smart factory is a step beyond
and increased efficiency that can be seen in factories and traditional automated manufacturing environments to facto-
achieved by modern technologies [1] like the Internet of ries where it may consider fully automotive systems. The
Things (IoT), Artificial Intelligence (AI), robotics, Cloud machines are controlled by superior computational intelli-
2 Wireless Communications and Mobile Computing

gence and are linked to sensors and other devices through leather defect. Therefore, this research focuses on defect clas-
wired or wireless networks [2]. In a study [3], the integration sification using a deep learning model (CNN) with VGG16
of all physical components and digital technologies into one for better accuracy in classifying the model. Most researchers
system is called Cyber-Physical Systems (CPSs) and is ana- only concentrated on developing the software model and
lyzed in detail. In addition, the amount of sensors collecting implementing it on a personal computer (PC) or server
data in smart factories is increasing exponentially [4]. computer with graphics processing unit (GPU). The utiliza-
Among the various kinds of sensor systems found in facto- tion of another computer for classifying products is not suit-
ries, vision-based systems are the most popular and effective able with practical factories in terms of grade existed
when it comes to estimating and classifying product quality. systems. Moreover, this research focuses on learning rate
A comprehensive review of automated vision-based defect to generate a better-trained model by SGD optimizer which
detection approaches that look at numerous kinds of mate- even produce better accuracy in large dataset. This proposed
rials such as ceramics, textiles, and metals is introduced CNN with VGG16 assists to classify the defective PCB more
in [5]. accurately because of the better learning rate obtained by the
Smart manufacturing is a recent production paradigm in pretrained model.
which machines are fully networked, monitored by sensors, The organization of the article is listed as follows: Section
and managed by better computer intelligence to boost sys- 2 suggests that the defective classification can be done
tem efficiency, product quality, and sustainability while cut- through CNN and various defective product identification
ting costs. The IoT, CC, and CPS are all-important is discussed. Section 3 discusses the VGG16 architecture
supporting technologies for modern production [6–10]. with CNN for a better training model, and an SGD opti-
These contemporary manufacturing technologies collect mizer for a high learning rate is illustrated. Section 4 has
and analyze data from several stages of a product’s life cycle, evaluated the defective detection classification from the nor-
including raw materials, machine operations, facility logis- mal product through CNN with VGG16 with other existing
tics, and also human operators [8]. Thanks to the prolifera- CNN and CNN with ResNet through various parameters
tion of industrial data, data-driven intelligence combined like accuracy, precision, recall, and F1 score. Section 5 con-
with powerful analytics transforms massive volumes of data cludes that CNN with VGG16 has better accuracy in train-
into relevant and insightful information for smart ing model accuracy and evaluation with existing CNN
manufacturing. shows that CNN with VGG16 has a high accuracy of 97%.
An important aspect of product quality control in the
manufacturing environment is faulty product categorization
on the product line. Currently, numerous methods are 2. Related Work
deployed to tackle this task. These defective product classifi-
cation systems must match challenging requirements such as This review is intended to study the approach of defective
the need to work in real-time with high accuracy and robust classification done on smart factories to enhance their level
performance in noisy environments like those seen in real- through various deep learning methods. Those are discussed
world factories. With the development of machine learning by several researchers, and a few are discussed below. The
and sophisticated vision systems, feature-based defect classi- image processing method and computer vision are focused
fication algorithms are starting to be investigated and on by Xie, who carefully reviews current improvements in
applied to classify defective products using various classifiers surface identification. This research is mostly focused on
such as artificial neural networks (ANN), Bayesian network industrial applications, while surface defect detection based
classifiers, and support vector machine (SVM) [11, 12]. on image processing needs high real-time performance by
When applied in real-world factories, these feature-based matching the results of prior studies [20]. From the stand-
categorization methods might be susceptible to noise such point of the textile industry’s defect management, Ngan
as lighting fluctuations or shadows in photos. Moreover, et al. and Mahajan et al. investigated the use and develop-
some companies manufacture a large range of items across ment of flaw-detection systems frequently used in the
many product lines, making them inappropriate for the manufacturing of textile for the textile industry in the
application of feature-based approaches. advancement of defect detection [21, 22]. Thermal imaging
Deep learning-based algorithms have shown outstanding technology is widely employed in a variety of industries.
outcomes in many computer vision applications like object Aldave et al. [23] examined data from commercially avail-
detection, picture categorization, and object identification able nonexperimental IR techniques to provide cameras in
during the last several years. These successes have displayed the field of nondestructive defect detection with references.
their great potential for application in the manufacturing In both business and academics, defect detection technology
environment [13, 14]. To combine wire defect region identi- is a popular issue. Steel [24] and textile [25], which serve as
fication with faulty product categorization with a multitask crucial detection approaches, have yet to be classified by
CNN is devised [15, 16]. Other quality inspection tasks that Pernkopf and Zhang et al. The description shows how
use CNNs have been suggested to monitor various products defect-detection technology works, as well as current
such as PCBs [17, 18], metal surfaces [19], bottled wine, defect-detection equipment and alternative options. Further-
casting products, semiconductor fabrication, and light- more, the assessment, review, and summary of the research
emitting diode (LED) cup apertures, mobile phone screen, status with major technologies in the US and abroad have
cover glass of display panels, bearings, optical film, and yet to be finished.
Wireless Communications and Mobile Computing 3

Ren et al. have examined the electrical resistance tomog- assisted to identify the defective product classification with
raphy (ERT) images for presenting the depictive identifica- high accuracy.
tion through an evaluation approach based on color
histogram [26]. Similarly, Song et al. devised a classification 3. Research Methodology
methodology for detecting flaws on wood surfaces based on
the percentage color histogram feature and feature vector The goal of this research is to identify and classify defective
texture feature of picture blocks, which has been proved to product from the raw material of the manufacturing indus-
be effective in trials, especially for junction type defects tries. This study focused on real-time datasets collected from
[27]. Prasitmeeboon and Yau proposed a two-step technical the captured image through the network, and the images are
procedure for particleboard defect identification, involving preprocessed by region of interest (RoI) extraction. The pur-
the use of SVM and color histogram characteristics for pose of RoI extraction is utilized for image segmentation
defect detection and smoothing and threshold technology which influences the performance of the preprocessing.
for defect localization [28, 29]. To differentiate keyboard Once the preprocessing is done by RoI extraction, then the
light leakage faults from dust, Ren and Huang suggested an dataset is progressed for the normalization which is done
approach that combines Autoencoder and FCN with a deep for splitting each pixel by 255. The entire pixel is represented
neural network. In a test set of 1632 photos, the suggested in terms of 0 and 1 assist to learn the model better, and the
technique is shown to reduce the false positive rate of light dataset gets split into train dataset and test dataset.
leaking defects from 6.27 percent to 2.37 percent [30]. Sam- In general, the normalization is the initial process of all
pedro et al. suggested a method for the automated identifica- neural network, and the CNN works based on the deep for-
tion and diagnosis of insulator strings that contained one ward neural network (FNN) that assist in learning invariant
segmentation component for insulator strings and two diag- feature hierarchies. The distinctive attributes and invariant
nostic components for missing and damaged insulator disc present in the CNN context are the CNN features that rep-
units, respectively [31]. Balzategui et al. propose a method resent the same transformation applied at various locations
for segmenting defects in solar cell electroluminescence pic- of the model. CNN model is majorly used for training and
tures. When compared to the method of continually per- data processing in grid topology or in a spatial relationship.
forming CNN sliding windows, this methodology uses The CNN model with VGG16 architecture working is per-
FCN with unique U-net architecture to generate the defect formed as the training process and get evaluated for accom-
segmentation map in a single step [12, 32]. Gao et al. used plishing the better pretrained model for identifying the exact
FCN with faster RCNN to construct a deep learning model defective product from manufactured product progressed
for tunnel defect detection based on FCN. This model can from each unit of the smart manufacturing unit as shown
accurately and quickly detect flaws like stains, leaks, and in Figure 1.
pipeline blockages [33, 34]. Baumgartl et al. have developed
a Laser Powder Bed Fusion (LPBF) device that incorporates 3.1. Data Collection. This study used Peking University’s
off-axis in situ imaging and thermography. The photos were intelligent robot laboratory and a PCB defect dataset with
used as a source of data for the CNN model, which is used to 1245 photos and four different types of defects: false copper,
identify printing problems. This model is used to detect mouse bite, missing hole, and open circuit. The various
thermographic in situ defects in LPBF processes, and it defect images have been fed for detecting and classifying
achieves a model accuracy of 96 percent. This CNN model, them as the defective product from the normal image of
on the other hand, can only identify splatter and delamina- the PCB dataset. Figure 2 shows a categorized picture dataset
tion faults in metal LPBF. As a result, other related flaws with various defect kinds grouped into different groups.
such as cracks, porosity, and unmelted powder, as well as Based on the measurements, the callout box average size
alternative methods such as selective laser sintering (SLS), for the defect available in the data set is about 8 × 8 pixels as
are not investigated [35, 36]. Jayanthi et al. illustrate that shown in Figure 2. The images can be rotated to a specific
the application of Hough Circle Transform (HT) is to local- angle over the program for improving the model robustness
ize the targeted region of iris region efficiently through mask and confirming the detection accuracy though the image
region-based CNN during the segmentation process. The gets rotated in the ensuing detection.
proposed technique has presented with the performance of
supreme iris recognition and resulted in maximum recogni- 3.2. Data Preprocessing. The casting images are captured
tion accuracy of 99.14% which is comparatively higher than using special arrangements to ensure that the image is cap-
existing UniNet V2, AlexNet, Inception, ResNet, DenseNet, tured in stable lighting conditions. The lighting condition
and VGGNet models [37, 38]. changes over time may lead to classification errors in the
This section discusses the various approaches of defect real-world manufacturing environment, and the vision sys-
detection identification in various kinds of factories as well tem completely depends upon lighting conditions. The RoI
as various analyses from several researchers about machine extraction from the captured image is based on saturation
learning methods and CNN in defect detection of products channel details and hue saturation value (HSV) which dis-
in factories. The defect detection through image processing cusses the color HSV and that the PCB color level gets differ
can be done only through CNN, which shows better accu- from the background. Initially, the red, blue, and green
racy, and several researchers also justify that. Thus, the (RBG) image is converted to HSV color space based on a
review support in focusing the CNN with VGG16 has general equation. Let us consider R, B, G ∈ ½0, 1, the
4 Wireless Communications and Mobile Computing

Data collection of images captures from

the PCB whereas this RoI is made to be located and cropped
each product of all units from original images.

3.3. Classification of Defective Product by CNN with VGG16.

This portion of the research is devoted to determining which
RoI extraction by converting to HSV class the retrieved ROI picture belonged to. The defective
image and threshold with S-channel image
product classification training method undergoes two major
steps as follows:

Finding of RoI after morphology and (i) Step 1. Training CNN with VGG16
image localization as well as resizing.
(ii) Step 2. Optimize and retrain the network

3.3.1. Step 1: Training CNN with VGG16. To obtain better

Normalization of dataset performance, CNN methods have been deployed in embed-
ded systems. However, CNNs trained from scratch require a
dataset with several labeled entries, more time consumption,
and high power consumption in computation. Transfer
learning (TL) is introduced to tackle these issues; TL is a
technique for retraining a CNN model on a novel, small
Training images to dataset through reprocessing its feature extractor portion.
CNN with VGG16 as Testing images
training model It gets trained using an existing large dataset, in this case
the ImageNet dataset, and retrained only on the functional-
ity of classification for saving training time. Based on TL, the
pretrained model weights are fine-tuned for classifying a
Evaluation of training model Pre-trained model novel dataset. Moreover, this knowledge transfer can be sig-
nificantly improved by learning performance without
requiring efforts from work-intensive data labeling. The TL
approach has accomplished advanced results on many data-
Classification of the defective product and normal product sets instead of a specified certain amount of training time.
with product ID as output
This proposed model VGG16 is familiar and mainly consid-
ered as one step up from AlexNet because of its deeper
Figure 1: Workflow of proposed classification method for a defect architecture and smaller kernel sizes.
product.
VGG16 has acquired more information from the Alex-
Net because it is a deep learning model represented in terms
maximum region as MAX ≕ max ðR, B, GÞ and the mini- of layer numbers. The proposed VGG16 with CNN model
mum region as MIN ≕ MINðR, B, GÞ. Thus, the equation of intended on earlier model inadequacy and focused on better
H, S, and V is expressed as follows: accuracy with high efficiency. According to this research,
  there are 6 variants with similar architecture that have been
° B−G implemented and have layer numbers ranging from 11 to 19.
H : f0, if R = B = G 60 × 0 + , In this proposed VGG16 architecture, the layers involved are
MAX − MIN
16 layers which include 13 convolutional layers along with
 
° G−R max-pooling layer and 3 fully connected layers that are
if MAX = R 60 × 2 + , shown in Figure 3. The main advantage present in VGG16
MAX − MIN
architecture while compared to AlexNet is usage of smaller
  filters placed on top of each other rather than one single
R−B
if MAX = B 60° × 4 + , large filter. The concept behind this approach is closely asso-
MAX − MIN
ciated with sensitive fields of convolutional filters. However,
if MAX = G, ð1Þ each pixel receives information from 49 pixels of the previ-
ous layer through the kernel of 7 × 7. Hence, the higher sen-
MAX − MIN sitive fields can able to capture more motifs in a large area
S : f0, if R = B = G , else, ð2Þ during missing of details whereas lesser receptive fields have
MAX failed to capture the usual patterns. Thus, the VGG16 archi-
V ≔ MAX: ð3Þ tecture contains convolutional layers, max pooling layers,
and fully connected layers. Even though there are further
Eventually, the selection of the HSV threshold is based dissimilarities, the model which is implemented in this stake
on pixel range values of the HSV color spacing concerning involves 3 × 3 and 2 × 2 convolutional filters. 1 × 1 fully con-
each histogram channel. Morphology is applied subse- nected convolutions only represented the linear combina-
quently for removing the noise from the threshold images. tion of a pixel location over the layers. Indeed, the
Hence, the RoI extraction is considered to be the region of utilization of nonlinear activation and ReLU has been
Wireless Communications and Mobile Computing 5

c
vo Labellmg

c
Labellmg

c
Labellmg?

vo
c

vo
vo Labellmg

Missing_hole Spurious_copperl Open_circuit Mouse_bitel

(a) (b) (c) (d)

Figure 2: Input images with dissimilar defects on PCB: (a) missing hole; (b) spurious copper; (c) open circuit; (d) mouse bite.

224 x 224 x 64
112 x 112 x 128

56 x 56 x 256

28 x 28 x 512

14 x 14 x 512

7 x 7 x 512 1 x 1 x 4096 1 x 1 x 1000

224 x 224 x 3

Convolution + ReLU Fully connected + ReLU

Max pooling Softmax

Figure 3: VGG16 architecture with CNN.

Table 1: Dataset distributions for training, validation, and testing. 3.3.2. Step 2: Optimize and Retrain the Network. TensorRT-
Product Product Image Training images Testing based applications have performed 40 times faster than
dataset class status count Training Validating images CPU-only platforms during inference. Based on the Ten-
sorRT, the researchers were only required to focus on gener-
PCB Normal 1937 1317 232 387
ating a new application with AI-powered instead of
dataset Defective 1238 842 149 248
performance tuning for inference deployment. Then the
pre-trained models are converted to the format of Open
preferred in the AlexNet. Moreover, training of AlexNet on Neural Network Exchange (ONNX). These processes are
MNIST is not an issue in the earlier stack but VGG16 is a completed on the server whereas the TensorRT model in
quite larger model and needs more data for highly trained. the ONNX format can then be downloaded from the cloud
The MNIST dataset does not provide the preferred statistical and loaded onto the training dataset for the prediction
variety to the model but the pretrained models are available phase. The experimental findings show that the deployed
in Keras and various deep learning frameworks and libraries. model can execute in real-time for all deep learning func-
The adjustable weights and pretrained frameworks are the tions that run on the same objective function f ðxÞ for min-
general practice to acquire certain layers or all layers from imizing error and establishing fresh data input. This
those networks have performed as the backbone of the fea- research focuses on the sample with a minibatch drawn con-
ture extractor module. This method is said to be TL, and it sistently from the training set. The size of the minibatch is
speeds up deep learning models by eliminating the require- generally selected to be a comparatively less instance num-
ment of training huge models from scratch. ber that can be generated from one to a few hundred.
6 Wireless Communications and Mobile Computing

Model accuracy
1.0

0.8

Accuracy 0.6

0.4

0.2

0 10 20 30 40 50

Epoch
Train
Validation

Figure 4: Model accuracy for train and validation.

Model loss
1.6

1.4

1.2

1.0
Loss

0.8

0.6

0.4

0.2

0.0

0 10 20 30 40 50

Epoch
Train
Validation

Figure 5: Model loss for train and validation.

However, the SGD performs as the machine-learning algo- classifier CNN with VGG16 model has accomplished with
rithm that executed better once it gets trained, though it better accuracy in identifying and classifying the defective
reaches the local minimum in a reasonable amount of time. product from normal along with the manufacturing shop
Hence, the significant parameter for SGD is the learning rate ID.
which is essential for decreasing the learning rate over time. The experimental findings show that the deployed model
The learning rate is represented with iteration k. can be execute in real-time applications for all deep learning
Moreover, the pretrained model supports classifying the functions that run on the same objective function f ðxÞ for
image through its loss functions and learning rates and assist minimizing error and establishing a fresh data input. This
in performing the classifier model. Therefore, the performed paper focuses on the sample with a minibatch, which draws
Wireless Communications and Mobile Computing 7

Table 2: Evaluation of testing image by confusion matrix for various algorithm.

Confusion matrix values

S. no Algorithm
True positive (TP) True negative (TN) False positive (FP) False negative (FN)
1 CNN 337 249 10 39
2 CNN–VGG16 391 225 6 13
3 CNN-ResNet 358 241 8 28

Performance comparison for accuracy

100.00 97.01 94.33
92.28
90.00
80.00
Accuracy value (%)

70.00
60.00
50.00
40.00
30.00
20.00
10.00
0.00
CNN CNN-VGG16 CNN-ResNet
Machine learning models
Accuracy

Figure 6: Comparison of accuracy performance for various deep learning models.

consistently from the training set. The size of minibatch is at 0.01 and the training epoch at 50 for optimal iteration.
generally selected to be a comparatively less instance num- The casting dataset, a publicly accessible dataset relevant to
ber that can be generated from one to few hundreds. product categorization, is used to assess and compare the
In a neural network, the loss of gradient over each performance of the suggested technique [37]. The PCB
parameter can be computed using back propagation net- product data one contains 3175 augmented, 300 × 300 pixel
works. This will assist in initiating the model from the model RBG images for training and testing the system. The 3175
loss and measured its derivative based on the weightage of images consist of a normal product as well as defective prod-
last model layers. The derivatives have been selected in uct images shown in Table 1.
accordance with each layer that executed backward from last The original electrical wire and cropped electrical wire
layer of previous model to first layer of previous model con- datasets are used to examine the effect of applying the ROI
sidered. However, the SGD performs as the DL algorithms extraction step. The final number of cropped images is
that executed better once it gets trained, though it reaches smaller than the number of original electrical wire images
the local minimum in a reasonable amount of time. Hence, due to not including images when the ROI extraction pro-
the significant parameter for SGD is the learning rate which cess failed. For all datasets, we randomly choose 80% for
is essential for decreasing the learning rate over time. The training (70% for training and 10% for validation) and leave
learning rate has the ability in controlling the sizes of the the remaining 20% for testing. The details of each dataset
updated steps along with gradient that is represented with and the distributions used for training, validation, and test-
iteration k. Thus, selecting the suitable learning rate is signif- ing are provided.
icant to train any neural network model whereas the better Figure 4 illustrates that the accuracy of the train model
learning rate relies on every individual model and issues in begins at 35% approximately at 2 epochs and improves until
the model. 94.8% at 10 epochs. A slight fluctuation occurs from 10 to 30
epochs and becomes steady with an accuracy of 96.2% after
4. Result and Discussion 50 epochs. In the case of the validation model, the accuracy
is lower at the start and fluctuates until 10 epochs, but it
This experimental research utilized a high performance steadies from 10 to 50 epochs with accuracy of 97.4%. This
server along with the configuration of an Intel Core i7 shows that the learning rate and the optimizer have reduced
DMI2 CPU, 16GB RAM, and a Quadro K600 GPU. The the loss function and improved the TL with better knowl-
operating system used is Ubuntu 16.04.6 LST for executing edge transformation through the pretrained model.
the abovementioned GPU while training the dataset images. Figure 5 illustrates that the loss of the train model begins
The optimizer and loss function utilized in this proposed at 1.56 at 2 epochs and decreases as we approach 40 epochs.
model are SGD and cross entropy. The learning rate is set There is slight fluctuation occurs from 40 to 50 epochs
8 Wireless Communications and Mobile Computing

Performance of precision, recall and F1-score

1.00 0.97 0.98 0.97 0.98 0.98
0.93 0.93 0.95
0.90
0.90
0.80
Confusion matrix scores 0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
CNN CNN-VGG16 CNN-ResNet

Deep learning models

Precision
Recall
F1-score

Figure 7: Comparison of precision, recall, and F1 score performance for various deep learning models.

which stabilizes at a loss of less than 0.16 at 50 epochs. In the cross entropy. The positive is high in actual as well as predic-
case of the validation model, the loss begins at 1.52 and fluc- tion which assists in performing precision and recall that
tuates until 30 epochs, but it stabilizes from 30 to 50 epochs make the better weightage in F1 score. When comparing
with a loss of less than 0.1. This shows that the learning rate the performance of the proposed model CNN-VGG16, the
and the optimizer have reduced the loss function and CNN and CNN-ResNet have performed less in precision,
improved the TL with better knowledge transformation recall, and F1 score.
through the pretrained model. Moreover, the proposed CNN-VGG16 has better accu-
In this study, accuracy, precision, recall, and F1 scores racy in classifying the defective product from the normal
were used as assessment measures for the suggested CNN product in the smart factories through image processing.
with VGG16 for product categorization. Table 2 shows This initiation makes this Industry 4.0 technology reduces
how these four assessment metrics were computed using the waste as well as minimizes the cost of production.
true positive (TP), true negative (TN), false positive (FP),
and false negative (FN) outcomes from a confusion matrix 5. Conclusion
between the prediction and the reality. The ratio of correct
predictions (TP + TN) to the total number of predictions As the manufacturing industry continues to propel into the
(TP + TN + FP + FN) is known as accuracy. The number of modern age, intelligent data-driven systems are becoming
accurately predicted samples (TP) divided by the number ubiquitous in optimizing, replacing, and augmenting human
of positive samples anticipated (TP + FP) yields precision. labor. Improvements to the models behind these systems are
The number of properly predicted (TP) samples divided by becoming increasingly necessary. Our proposed model pro-
all actual positive samples (TP + FN) yields recall. The vides one such advance, in providing accurate classification
weighted average of accuracy and recall is the F1 score. for product defect in production. The model’s CNN-
The problem of defect classification in the product is binary. VGG16 architecture emphasizes layers that assist in captur-
Therefore, F1 score more reliable statistical rate evaluation ing large pixels, and the SGD optimizer generates an optimal
metric that considers both positive elements and negative learning rate for learning PCB data. This pretraining model
elements is adopted for this study. assists in implementing a better TL model to retrain CCN-
Figure 6 illustrates that the accuracy of CNN with VGG16 for PCB product dataset ROI extraction based on
VGG16 is high while compared to the existing CNN and S-channel data. Each pretrained model is deployed for deter-
CNN-ResNet. The performance of the proposed CNN- mining the defective PCB product with an accuracy of 97.3%
VGG16 is 97%, and it also determined that the TL is even in training. Moreover, the proposed CNN-VGG16 is evalu-
better for large datasets than CNN and CNN-ResNet. ated using various deep learning models such as CNN and
Figure 7 illustrates the performance of precision, recall, CNN-ResNet. Hence, the evaluation result through testing
and F1 score that describes the proposed model CNN- the image dataset of the proposed CNN-VGG16 has an
VGG16 that has high scores in all precision, recall, and F1 accuracy of 97%. Thus, the proposed CNN-VGG16 can
score that are 0.98, 0.97, and 0.98, correspondingly. The excel at determining defected PCBs before they reach further
scores of CNN-VGG16 discuss that because of a better pre- production stages. This will reduce scrap, increase yield, and
trained model using SGD optimizer and loss function as minimize production cost across the board. As the Industry
Wireless Communications and Mobile Computing 9

4.0 paradigm encompasses production around the world, [14] F. S. Alenezi and S. Ganesan, “Geometric-pixel guided single-
advancements such as ours will provide value for countless pass convolution neural network with graph cut for image
companies. dehazing,” IEEE Access, vol. 9, pp. 29380–29391, 2021.
[15] X. Tao, Z. H. Wang, Z. T. Zhang et al., “Wire defect recogni-
tion of spring-wire socket using multitask convolutional neu-
Data Availability ral networks,” IEEE Transactions on Components, Packaging
and Manufacturing Technology, vol. 8, pp. 689–698, 2018.
The data that support the findings of the study are available
[16] F. Alenezi, “Image dehazing based on pixel guided CNN with
from the corresponding author, upon reasonable request. PAM via graph cut,” Computers, Materials & Continua,
vol. 71, no. 2, pp. 3425–3443, 2022.
Conflicts of Interest [17] V. A. Adibhatla, H.-C. Chih, C.-C. Hsu, J. Cheng, M. F. Abbod,
and J.-S. Shieh, “Defect detection in printed circuit boards
The authors declare that there are no conflicts of interest. using you-only-look-once convolutional neural networks,”
Electronics, vol. 9, no. 9, p. 1547, 2020.
[18] E. Zhang, B. Li, P. Li, and Y. Chen, “A deep learning based
References printing defect classification method with imbalanced sam-
ples,” Symmetry, vol. 11, no. 12, p. 1440, 2019.
[1] G. Büchi, M. Cugno, and R. Castagnoli, “Smart factory perfor-
mance and Industry 4.0,” Technological Forecasting and Social [19] J. P. Yun, W. C. Shin, G. Koo, M. S. Kim, C. Lee, and S. J. Lee,
Change, vol. 150, article 119790, 2020. “Automated defect inspection system for metal surfaces based
on deep learning and data augmentation,” Journal of
[2] V. Alcácerac and V. Cruz-Machadoab, “Scanning the Industry
Manufacturing Systems, vol. 55, pp. 317–324, 2020.
4.0: a literature review on technologies for manufacturing sys-
tems,” Engineering Science and Technology, An International [20] X. Xie, “A review of recent advances in surface defect detection
Journal, vol. 22, pp. 899–919, 2019. using texture analysis techniques,” ELCVIA: Electronic Letters
on Computer Vision and Image Analysis, vol. 7, pp. 1–22, 2008.
[3] D. Sinha and R. Roy, “Reviewing cyber-physical system as a
part of smart factory in Industry 4.0,” IEEE Engineering Man- [21] H. Y. Ngan, G. K. Pang, and N. H. Yung, “Automated fabric
agement Review, vol. 48, pp. 103–117, 2020. defect detection—a review,” Image and Vision Computing,
[4] T. Kalsoom, N. Ramzan, S. Ahmed, and M. Ur-Rehman, vol. 29, pp. 442–458, 2011.
“Advances in sensor technologies in the era of smart factory [22] P. M. Mahajan, S. R. Kolhe, and P. M. Patil, “A review of auto-
and Industry 4.0,” Sensors, vol. 20, p. 6783, 2020. matic fabric defect detection techniques,” Advances in Compu-
[5] T. Czimmermann, G. Ciuti, M. Milazzo et al., “Visual-based tational Research, vol. 1, pp. 18–29, 2009.
defect detection and classification approaches for industrial [23] I. J. Aldave, P. V. Bosom, L. V. González et al., “Review of ther-
applications—a survey,” Sensors, vol. 20, no. 5, p. 1459, 2020. mal imaging systems in composite defect detection,” Infrared
[6] L. Wang, M. Törngren, and M. Onori, “Current status and Physics & Technology, vol. 61, pp. 167–175, 2013.
advancement of cyber-physical systems in manufacturing,” [24] F. Pernkopf, “Detection of surface defects on raw steel blocks
Journal of Manufacturing Systems, vol. 37, pp. 517–527, 2015. using Bayesian network classifiers,” Pattern Analysis and
[7] P. Wang, R. X. Gao, and Z. Fan, “Cloud computing for cloud Applications, vol. 7, pp. 333–342, 2004.
manufacturing: benefits and limitations,” Journal of [25] W. Zhang, C. Ye, K. Zheng et al., “Tensan silk-inspired hierar-
Manufacturing Science and Engineering, vol. 137, no. 4, chical fibers for smart textile applications,” ACS Nano, vol. 12,
pp. 1–10, 2015. no. 7, pp. 6968–6977, 2018.
[8] Y. Lu, X. Xu, and J. Xu, “Development of a hybrid manufactur- [26] H. Ren, K. Tian, S. Hong, B. Dong, F. Xing, and L. Qin, “Visu-
ing cloud,” Journal of Manufacturing Systems, vol. 33, no. 4, alized investigation of defect in cementitious materials with
pp. 551–566, 2014. electrical resistance tomography,” Construction and Building
[9] D. Wu, D. W. Rosen, and D. Schaefer, “Cloud-based design Materials, vol. 196, pp. 428–436, 2019.
and manufacturing: a new paradigm in digital manufacturing [27] W. Song, T. Chen, Z. Gu, W. Gai, W. Huang, and B. Wang,
and design innovation,” Computer-Aided Design, vol. 59, “Wood materials defects detection using image block percen-
pp. 1–14, 2015. tile color histogram and eigenvector texture feature,” in Pro-
[10] F. Alenezi and E. Salari, “A fuzzy-based medical image fusion ceedings of the First International Conference on Information
using a combination of maximum selection and gabor filters,” Sciences, Machinery, Materials and Energy, pp. 2022–2026,
International Journal of Engineering Science, vol. 9, pp. 118– Chongqing, China, 2015.
129, 2018. [28] P. Prasitmeeboon and H. Yau, “Defect detection of particle-
[11] M. P. Samy, S. Foong, G. S. Soh, and K. S. Yeo, “Automatic boards by visual analysis and machine learning,” in 2019 5th
optical & laser-based defect detection and classification in International Conference on Engineering, Applied Sciences
brick masonry walls,” in 2016 IEEE Region 10 Conference and Technology (ICEAST), pp. 1–4, Luang Prabang, Laos,
(TENCON), pp. 3521–3524, Singapore, 2016. 2019.
[12] F. Alenezi, A. Armghan, S. N. Mohanty, R. H. Jhaveri, and [29] F. Alenezi and K. C. Santosh, “Geometric regularized Hopfield
P. Tiwari, “Block-Greedy and CNN based underwater image neural network for medical image enhancement,” Interna-
dehazing for novel depth estimation and optimal ambient tional Journal of Biomedical Imaging, vol. 2021, Article ID
light,” Water, vol. 13, no. 23, p. 3470, 2021. 6664569, 2021.
[13] J. Wang, Y. Ma, L. Zhang, R. X. Gao, and D. Wu, “Deep learn- [30] J. Ren and X. Huang, “Defect detection using combined deep
ing for smart manufacturing: methods and applications,” Jour- autoencoder and classifier for small sample size,” in Proceed-
nal of Manufacturing Systems, vol. 48, pp. 144–156, 2018. ings of the 2020 IEEE 6th International Conference on Control
10 Wireless Communications and Mobile Computing

Science and Systems Engineering (ICCSSE), pp. 32–35, Beijing,

China, 2020.
[31] C. Sampedro, J. Rodriguez-Vazquez, A. Rodriguez-Ramos,
A. Carrio, and P. Campoy, “Deep learning-based system for
automatic recognition and diagnosis of electrical insulator
strings,” IEEE Access, vol. 7, pp. 101283–101308, 2019.
[32] J. Balzategui, L. Eciolaza, and N. Arana-Arexolaleiba, “Defect
detection on polycrystalline solar cells using electrolumines-
cence and fully convolutional neural networks,” in Proceedings
of the 2020 IEEE/SICE International Symposium on System
Integration (SII), pp. 949–953, Honolulu, HI, USA, 2020.
[33] X. Gao, M. Jian, M. Hu, M. Tanniru, and S. Li, “Faster multi-
defect detection system in shield tunnel using combination of
FCN and faster RCNN,” Advances in Structural Engineering,
vol. 22, pp. 2907–2921, 2019.
[34] P. K. R. Maddikunta, Q. V. Pham, B. Prabadevi et al., “Industry
5.0: a survey on enabling technologies and potential applica-
tions,” Journal of Industrial Information Integration, vol. 26,
article 100257, 2021.
[35] H. Baumgartl, J. Tomas, R. Buettner, and M. Merkel, “A deep
learning-based model for defect detection in laser-powder bed
fusion using in-situ thermographic monitoring,” Progress in
Additive Manufacturing, vol. 5, no. 3, pp. 277–285, 2020.
[36] M. Alazab, K. Lakshmanna, T. R. G, Q.-V. Pham, and P. K.
Reddy Maddikunta, “Multi-objective cluster head selection
using fitness averaged rider optimization algorithm for IOT
networks in smart cities,” Sustainable Energy Technologies
and Assessments, vol. 43, article ???, 2021.
[37] J. Jayanthi, E. L. Lydia, N. Krishnaraj, T. Jayasankar, R. L.
Babu, and R. A. Suji, “An effective deep learning features based
integrated framework for iris detection and recognition,” Jour-
nal of Ambient Intelligence and Humanized Computing,
vol. 12, no. 3, pp. 3271–3281, 2021.
[38] C. Iwendi, P. K. R. Maddikunta, T. R. Gadekallu,
K. Lakshmanna, A. K. Bashir, and M. J. Piran, “A metaheuris-
tic optimization approach for energy efficiency in the IOT net-
works,” Software: Practice and Experience, vol. 51, no. 12,
pp. 2558–2571, 2020.