Received 6 November 2023, accepted 26 November 2023, date of publication 5 December 2023,

date of current version 14 December 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3339561

A Comprehensive Review of Deep

Learning-Based PCB Defect Detection
XING CHEN1 , YONGLEI WU2 , XINGYOU HE2 , AND WUYI MING 3
1 Schoolof Artificial Intelligence and Software Engineering, Nanyang Normal University, Nanyang 473061, China
2 Henan Key Laboratory of Intelligent Manufacturing of Mechanical Equipment, Zhengzhou University of Light Industry, Zhengzhou 450002, China
3 Guangdong Provincial Key Laboratory of Digital Manufacturing Equipment, Guangdong HUST Industrial Technology Research Institute, Dongguan 523808,

China
Corresponding author: Wuyi Ming ([email protected])
This work was supported in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515140066, in part
by the Henan Provincial Key Scientific and Technological Project 222102220011, and in part by the Guangdong Provincial Key
Laboratory of Manufacturing Equipment Digitization under Grant 2023B1212060012.

ABSTRACT A printed circuit board (PCB) functions as a substrate essential for interconnecting and
securing electronic components. Its widespread integration is evident in modern electronic devices, spanning
computers, cell phones, televisions, digital cameras, and diverse apparatus. Ensuring product quality man-
dates meticulous defect inspection, a task exacerbated by the heightened precision of contemporary circuit
boards, intensifying the challenge of defect detection. Conventional algorithms, hampered by inefficiency
and limited accuracy, fall short of usage benchmarks. In contrast, PCB defect detection algorithms rooted in
deep learning hold promise for achieving heightened accuracy and efficiency, bolstered by their adeptness at
discerning novel defect types. This review presents a comprehensive analysis of machine vision-based PCB
defect detection algorithms, traversing the realms of machine learning and deep learning. It commences by
contextualizing and elucidating the significance of such algorithms, followed by an extensive exploration of
their evolution within the machine vision framework, encompassing classification, comparison, and analysis
of algorithmic principles, strengths, and weaknesses. Moreover, the introduction of widely used PCB defect
detection datasets and assessment indices enhances the evaluation of algorithmic performance. Currently,
the detection accuracy can exceed 95% at an Intersection over Union (IoU) of 0.5. Lastly, potential future
research directions are identified to address the existing issues in the current algorithm. These directions
include utilizing Transformers as a foundational framework for creating new algorithms and employing
techniques like Generative Adversarial Networks (GANs) and reinforcement learning to enhance PCB defect
detection performance.

INDEX TERMS Deep learning, defect detection, machine vision, neural networks, PCB.

I. INTRODUCTION underpinnings and pivotal components within electronic

In recent years, electronic products have witnessed products [4], PCBs must demonstrate robust stability,
widespread utilization and popularity, driven by the rapid pronounced resistance to interference, and excel in charac-
expansion of the global economy and the swift evolution of teristics encompassing high-speed transmission, heightened
information technology [1], [2]. With continuous upgrades integration levels, and compact dimensions [5]. Furthermore,
and enhanced functionalities of electronic products, the req- PCB layout encompasses the strategic arrangement of PCBs,
uisites for performance and quality in printed circuit boards constituting a critical stage in electronic device manufactur-
(PCBs) are on the rise [3]. Functioning as foundational ing. Key concepts in the PCB layout process include:
1) Layout: This involves planning and organizing the
The associate editor coordinating the review of this manuscript and placement of electronic components, connection lines, holes,
approving it for publication was Chuan Li. and other specific elements on a PCB. The quality of the
2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME 11, 2023 139017
 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

layout directly influences circuit performance, electromag- Although offering commendable reliability and accuracy, this
netic interference (EMI), thermal performance, and mainte- approach necessitates specialized equipment and intricate
nance challenges. testing procedures, resulting in a laborious and time-intensive
2) Component placement: An essential phase in the layout process. Visual inspection technology primarily relies on arti-
process, which includes the positioning of electronic compo- ficial intelligence algorithms, utilizing image processing and
nents (such as chips, resistors, capacitors, connectors, etc.) on pattern recognition techniques for swift and automated identi-
the PCB. Proper component placement is integral to ensuring fication of PCB defects [12]. However, its effectiveness relies
optimal circuit performance and signal transmission. on ongoing optimization and enhancement of the detection
3) Routing: Upon completion of the layout, the wiring algorithm. The instrument on-line detection method serves
routes must be established to interconnect the various compo- primarily to oversee PCB defects via various instruments,
nents and form the circuit. Routing requires consideration of including the high-voltage detection and insulation hydro-
factors like signal integrity, power supply, ground, and signal carbon detection methods. This approach offers the benefits
transmission lines. of convenient operation and rapid detection but occasionally
4) Power distribution: The layout process must include a faces challenges of misjudgment. The artificial visual subjec-
rational plan for the distribution of power and ground lines tive judgment method predominantly hinges on experiential
to ensure that electronic components receive a stable power and perceptual assessment, rendering it susceptible to human
supply and to minimize noise and interference within the influence; nevertheless, it finds applicability in certain spe-
circuit. cialized inspection scenarios.
Therefore, PCB layout is a pivotal step in the design and Despite the array of PCB defect detection methods, each
manufacture of electronic products. It demands that engi- approach presents inherent limitations and deficiencies. The
neers consider multiple factors concurrently—performance, selection of the suitable method must align with specific
heat dissipation, and maintainability—to ensure the ratio- requirements and circumstances. In recent years, propelled
nal layout and connection of electronic components. This by the continual progress and integration of technology, deep
careful planning ensures that the PCB meets the design learning-based algorithms have progressively gained promi-
requirements and facilitates smooth manufacturing. Hence, nence across diverse industries. Within the industrial sector,
ensuring defect detection and implementing quality con- these algorithms find utility in processing and positioning
trol measures for PCBs assume paramount significance tasks [13], [14], while also capable of identifying defects in
in the domain of PCB manufacturing and production 3C products [15], [16], [17]. Similarly, within the medical
processes [6]. field, deep learning-based algorithms aid doctors in analyzing
Nevertheless, due to the diminishing size of components, medical images for diagnostic support [18], [19]. In agri-
amplified component density, and intricate and diverse nature culture, these algorithms contribute to crop monitoring [20],
of PCB manufacturing, PCBs become susceptible to a range [21], [22], [23]. Additionally, in the military domain, deep
of factors, including mechanical friction, electrostatic inter- learning-based algorithms analyze remote sensing images to
ference, and chemical corrosion, during production. These facilitate swift positioning [24], [25], [26]. Simultaneously,
factors can readily give rise to a spectrum of defects, such PCB defect detection algorithms based on deep learning are
as missing holes, mouse bites, open circuits, shorts, spurs, poised to emerge as pivotal tools within the electronic man-
spurious copper, and broken holes [1], [7]. Figure 1 illustrates ufacturing industry, offering a broader scope of application
a variety of PCB defect types. These defects substantially and developmental opportunities.
undermine the quality and performance of PCBs. Failure to A standard PCB typically comprises several distinct layers:
detect and address these issues promptly can directly impede 1) Bottom copper layer: This foundational metal layer
the regular operation and service life of electronic devices, provides ground connections and layouts for various circuit
potentially resulting in severe accidents and incidents. Con- components. It frequently includes a soldermask layer to
sequently, swiftly and accurately detecting and locating PCB safeguard the circuit traces and a copper layer.
defects during the manufacturing process holds paramount 2) Top copper layer: This uppermost metal layer hosts the
significance. primary circuit components, wires, and signal paths, often
Presently, the detection of PCB defects stands as a crucial including pads for connecting electronic elements.
concern within the electronics manufacturing sector. Various 3) Inner layers: Beyond the bottom and top copper layers,
inspection methods have been researched and developed to a PCB may feature one or more inner layers composed of
address diverse PCB defects and align with the demands glass fiber and copper foil, which facilitate signal or power
of industrial production. These include the functional test transmission.
method, visual inspection technology, instrument on-line 4) Signal layers: These layers transmit signals between
inspection method, and manual visual subjective determi- electronic components, encompassing data, control signals,
nation method [9], [10], [11]. Among these, the functional and clock signals.
test method involves utilizing a fault simulator to assess the 5) Power layers (power planes): These layers supply
circuit board’s functionality and identify potential defects. power connections, including power and ground, typically

139018 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 1. Variety of PCB defect types: (a) Missing hole [8]; (b) Mouse bite [8]; (c) Open circuit [8]; (d) Short [8]; (e) Spur [8]; (f) Spurious copper [8].

distributing power and ground connections to ensure board methods and algorithms employing deep learning techniques
power stability. to enhance the efficacy of PCB defect detection. The efficacy
6) Ground layer (ground planes): This dedicated layer and impact of these algorithms in practical applications are
provides a stable ground connection to minimize signal inter- expounded in detail. Additionally, a thorough elucidation of
ference and enhance circuit performance. the fundamental principles underpinning deep learning and
7) Pad layer: Positioned on top of the top and bottom a concise overview of the Transformer model is furnished to
copper layers, the solder mask layer protects the circuit traces ensure readers attain a comprehensive grasp of the founda-
and copper layer to prevent short circuits, typically featuring tional aspects of the domain. Moreover, this paper introduces
pad openings to allow electronic component soldering. widely employed datasets and evaluation metrics utilized for
8) Silkscreen layer: This layer commonly includes labels, PCB defect detection. Ultimately, drawing upon the extant
component pin markings, and additional component informa- literature, prevailing algorithms are scrutinized and deliber-
tion to aid assembly and maintenance personnel in correctly ated upon, while potential developmental pathways within
identifying and handling electronic components. this realm are postulated.
These layers’ configurations can differ according to
PCB design requirements. As this review focuses on deep II. FUNDAMENTALS OF DEEP LEARNING
learning-based PCB defect detection—predominantly based A. BASIC KNOWLEDGE
on machine vision applications—defect detection usually Coined in 1986, the term deep learning (DL) initially found its
pertains to surface layers, including: place within the realm of machine learning, later extending to
1) Top copper layer: Here, the deep learning model can the domain of artificial neural networks in 2000. Deep neural
identify pad defects, such as false soldering, short circuits, networks are structured with numerous hidden layers that pro-
open circuits, offsets, and other issues. gressively abstract data features [27], [28]. This architecture
2) Bottom copper layer: This layer also accommodates empowers computers to autonomously acquire higher-level
the layout of pads and electronic components, with defect abstract features, thus addressing a gamut of tasks such as
detection similar to the top copper layer. classification, regression, clustering, and generation [29],
3) Pad layer: In this layer, the deep learning model can [30]. In contrast to conventional machine learning methods,
detect defects within the pad opening area. deep learning dispenses with the necessity for pre-designed
4) Silkscreen layer: This layer, containing information manual features, automating the process of feature extraction
about component markings and pin markings, allows deep and learning through extensive data training. An inherent
learning models to detect issues such as broken and shifted advantage of deep learning lies in its efficiency in processing
lettering. voluminous datasets, culminating in remarkable outcomes
Moreover, some PCBs may incorporate specialized layers when presented with ample data. Nonetheless, challenges
like soldermasks, inner layers, etc., which could be used for persist, encompassing the requisition for substantial labeled
detecting specific defects. data, the expenditure of costly computational resources, and
This review delivers a methodical analysis and thorough the interpretability of models. Consequently, the training
synthesis of the research literature concerning PCB defect process assumes paramount significance, especially in appli-
detection within the last decade. Its primary focus rests on cations like PCB defect detection. This phase necessitates

VOLUME 11, 2023 139019

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

data preprocessing, the construction of a problem-relevant classification challenges. In 2014, Simonyan et al. [36]
network model, and the formulation of the corresponding introduced VGG, a novel network architecture. In contrast
loss function. Lately, the strides achieved in the realm of to AlexNet, VGG boasts a deeper network structure and
deep learning predominantly owe their origins to pioneer- employs diminutive convolutional kernels to curtail param-
ing advancements in network models, loss functions, and eter counts. This prowess was affirmed by its attainment
activation functions. As technology continues its relentless of second place in the 2014 ILSVRC competition. Simi-
evolution, deep learning is poised to embrace a wider scope larly, in that year, Szegedy introduced a novel deep learning
of developmental opportunities in the future, positioning architecture named GoogLeNet [37]. While prior networks
itself as an integral component of the artificial intelligence predominantly pursued increased depth for enhanced training
landscape. Subsequently, attention will be directed towards outcomes, this approach presented challenges like overfit-
convolutional neural networks (CNNs) and the recently ting, gradient vanishing, and gradient explosion. GoogLeNet
surpassing Transformer architecture. These network struc- introduced the inception module, which not only optimizes
tures have demonstrated remarkable accomplishments across computational resource utilization but also enhances feature
varied domains, thereby unveiling novel avenues for the pro- extraction within the confines of the same computation load,
liferation of deep learning applications. thereby amplifying its effectiveness. However, as the num-
ber of network layers increases, the model becomes more
challenging to train and is susceptible to issues like gradient
B. CONVOLUTIONAL NEURAL NETWORK vanishing. In 2015, He et al. introduced residual networks
CNN [31] is a distinct neural network architecture that finds (ResNet) [38], accompanied by the design of a residual mod-
extensive application in processing various forms of data, ule tailored to mitigate the gradient vanishing issue arising
including images, speech, and natural language. Key con- from heightened network depth in deep neural networks. This
stituents within CNN encompass the convolutional layer, architectural enhancement facilitates the training of deeper
pooling layer, and fully connected layer. Skillful integration networks with greater ease. In 2016, Cai et al. [39] initially
of these layers enables the processing of data and extraction proposed a CNN leveraging cascade learning, which yielded
of features. CNNs have found widespread application within superior detection results. This further underscores the sig-
the realm of computer vision (CV), undertaking diverse nificant potential of CNNs in PCB defect detection.
tasks like image classification, target detection, semantic seg- In summary, propelled by the swift advancement of
mentation, image caricature, style migration, and generative deep learning, CNNs have achieved significant milestones
adversarial networks (GANs) [32]. These applications have within the realm of CV, fundamentally transforming image
yielded outstanding outcomes. Simultaneously, the network recognition and classification endeavors. This review cen-
architecture of CNN has undergone continual refinement ters on the canonical architectures of CNNs, encompassing
and optimization, leading to the emergence of several exem- LeNet, AlexNet, VGG, GoogLeNet, and ResNet. These
plary structures, including AlexNet, VGG, GoogLeNet, and architectures have garnered significant academic attention
ResNet, among others. These established networks exhibit and concurrently exhibited exceptional prowess in industrial
commendable performance across diverse CV tasks, serv- and practical contexts. Through a comprehensive compar-
ing as pivotal benchmarks in both research and application ison and analysis of these network architectures, a more
domains. profound comprehension of the influence exerted by diverse
In 1994, Lecun et al. introduced LeNet, one of the pio- design strategies on CNN performance can be attained. This
neering CNNs for handwriting font recognition. This seminal understanding will be instrumental in guiding future research
work played a pivotal role in shaping the subsequent evolu- endeavors within the CNN domain. Table 1 provides an over-
tion of CNNs. Krizhevsky introduced AlexNet in 2012 [33]. arching depiction of the strengths (including key attributes
Building upon the foundations laid by LeNet, AlexNet effec- and innovations) and limitations characterizing the aforemen-
tively adopted the Rectified Linear Unit (ReLU) [34] as the tioned CNNs. This review aims to furnish readers with a lucid
activation function for CNN. This choice proves superior to comprehension of both the strengths and limitations of these
Sigmoid [35] for networks with increased depth, addressing networks.
challenges like the vanishing gradient problem. Concurrently, The CNNs discussed above have witnessed a progression
CUDA (Compute Unified Device Architecture) technology from modest layer counts to substantial depths, reaching
facilitated the acceleration of deep convolutional network dozens or even hundreds of layers. Simultaneously, the
training, leveraging the robust parallel computing prowess refinement of these networks has matured, evolving in
of GPUs (Graphic Processing Units) to manage extensive tandem with advancements in activation functions, the inte-
operations during training. The remarkable achievement of gration of maximum pooling layers, and other techniques.
winning the 2012 ImageNet competition attests to the effi- Fig.2 (a) illustrates the schematic structure of the AlexNet
cacy of this approach, particularly notable for conventional network. Notably, AlexNet rescales all input images to
machine learning classification algorithms. The success of dimensions of 32 × 32 before subjecting them to succes-
AlexNet stands as empirical evidence showcasing the supe- sive convolutional operations. Subsequently, the outcomes
rior performance of CNNs in addressing large-scale image undergo compression via a fully connected layer, yielding a

139020 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

TABLE 1. A comparative analysis of strength and boundedness among classical algorithms employed in CNNs.

1 × 10 vector that encodes character weights. In Fig.2 (b), to implement the attention mechanism in image recogni-
the extended AlexNet architecture is depicted. This itera- tion, confirming its viability. This marked the inaugural
tion involves deeper network structures and integrates the integration of the attention mechanism into deep learning.
Max Pooling technique, inspired by LeNet. Furthermore, Subsequently, this concept was extended to the domain of nat-
GPU acceleration is employed during training, culminating ural language processing (NLP) [41], [42]. In 2017, Google
in the model’s remarkable success in the ImageNet compe- Research [43] introduced the Transformer model, founded
tition. Fig.2 (c) and 2(d) showcase the Inception module, on the concept of attention, which achieved remarkable
a hallmark of both VGG and GoogLeNet, co-introduced results and disrupted the NLP field. Similarly, the realm of
in a single year. VGG employs a block-oriented approach CV witnessed the emergence of numerous novel attention
to progressively extend the network, a straightforward yet models. Hu et al. [44] introduced the SENet (Squeeze-and-
efficient strategy. In contrast, GoogLeNet pursues network Excitation Network), integrating attention mechanisms into
depth expansion through width enhancement, introducing the feature channels. SENet autonomously learns the signifi-
Inception module. Over time, this module has undergone iter- cance of various channels, using this acquired information to
ative enhancements, resulting in impressive accuracy gains. amplify valuable features and attenuate less relevant ones for
Despite these advancements, the prevalent issue of excessive the current task. SE attention mechanism module as shown
network depth during that era, contributing to suboptimal in figure Fig.3 (a). Nonetheless, SENet overlooks the fea-
outcomes, remained unaddressed. In Fig.2 (e), the schematic ture space aspect. Consequently, Woo et al. [45] introduced
representation of the residual module, a significant innovation the convolutional block attention module (CBAM), which
by ResNet, is showcased. Within this module, the output merges the attention mechanism in both feature channels and
encapsulates not only the network’s result but also the sum feature space dimensions, enhancing network performance
of the original input, fostering a residual architecture. This without substantially increasing the parameter count. This
structural paradigm recurs in subsequent network iterations, module’s schematic depiction can be found in Fig.3 (b).
ultimately establishing itself as a fundamental element in the However, the prior attention mechanisms exhibited increased
landscape of CNNs. complexity, leading to heightened model intricacy. Therefore,
However, none of the aforementioned approaches consid- Wang et al. [46] introduced efficient channel attention for
ers the inherent human tendency to selectively concentrate deep convolutional neural networks (ECA-Net), character-
on specific information while disregarding other visible data ized by a reduced parameter count and evident performance
during the process of visual perception. The introduction of improvements. The schematic diagram of the ECA module
the attention mechanism, in turn, presents a more effective is presented in Fig.3 (c). While earlier attention mechanisms
solution. mainly addressed inter-channel information, disregarding
spatial location details or lacking long-range relationship
C. ATTENTION MECHANISM capabilities, Hou and colleagues [47] proposed coordi-
The concept of the Attention Mechanism originated from nate attention (CA). This method captures inter-channel
observations in human vision. In 2014, Mnih et al. [40] noted information and considers orientation-based positional infor-
that despite using multiple GPUs and priors to process image mation, enhancing the model’s ability to accurately local-
data, CNNs still required extended training periods. This ize and identify targets. Additionally, Coordinate Attention
concept was later merged with human observation images boasts a lightweight and flexible nature, permitting seamless

VOLUME 11, 2023 139021

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 2. Diagrams illustrating the network structures of classical CNNs: (a) LeNet; (b) AlexNet; (c) VGG block; (d) GoogleNet
Inception module; (e) ResNet residual structure.

integration into networks while delivering substantial perfor- width and height dimensions. Subsequently, [52] advocated
mance enhancements. The CA attention module is displayed for replacing the convolution module in ResNet entirely with
in Fig.3 (d). Attention-based relative position embedding to achieve a
In PCB defect detection, the majority of areas are typ- fully Attention-powered image model. Fig.5 (b) illustrates
ically defect-free, with only a minor proportion exhibiting the local attention layer at k = 3. Progressing from CNN
defects. The attention mechanism, by directing more focus to Transformer modules, Carion et al. [53] introduced the
toward these defect-laden areas, facilitates more efficient End-to-End Object Detection with Transformers (DETR)
defect detection. Consequently, incorporating an attention model, marking the inaugural replacement of CNN with
mechanism into the PCB defect detection model proves Transformer. DETR exhibited promising results in target
advantageous. detection, segmentation, and classification, though it lagged
behind then-state-of-the-art CNN in certain aspects. DETR’s
model structure is outlined in Fig.5 (c). With the surge of
D. TRANSFORMER
researchers investigating the Transformer’s role in CV tasks,
The Transformer architecture, introduced by Google Transformer’s advancement in the CV domain has outpaced
Research [43], represents a significant breakthrough in CNN’s performance in large-scale models [50], [54], [55].
the realm of natural language processing [48]. Traditional In PCB defect detection, defects are often minuscule,
sequential models, including recurrent neural networks rendering traditional methods ineffective for their identifica-
(RNNs) and CNNs, often encounter issues such as gradient tion. Recently, the Transformer model, with its heightened
vanishing and explosion when handling lengthy sequences. sensitivity to small targets and complex defects in diverse
In contrast, Transformer overcomes these challenges through environments, has been more suitable for PCB defect detec-
the incorporation of a self-attention mechanism, enhanc- tion. Thus, employing the Transformer model for PCB defect
ing its capability to capture extensive dependencies within detection could potentially enhance performance.
sequences. Consequently, Transformer has emerged as the
fundamental architecture within the NLP domain, and its core
structure is depicted in Fig.4 (a). E. SUMMARY
As the Transformer gained prominence in the NLP field, This section delves into the fundamental concepts of deep
scholars increasingly extended its application to the CV learning, encompassing CNNs, attention mechanisms, and
domain, exemplified by ViT (Vision Transformer) [49], [50]. the transformer model. CNNs, originating in the last century,
Attention-based approach to complement CNN’s limitations only truly began to thrive a little over a decade ago with
in ultra-long sequence modeling. This method transforms the the surge in computational power and accessibility to ample
input image dimensions [H, W, F] into two dimensions To labeled data. This led to ground-breaking results in numer-
Transformer’s spatial location information deficiency, Bello ous domains and they have now become integral to PCB
extended the concept from [51] and bolstered Attention’s defect detection and machine vision. The attention mecha-
prowess by incorporating relative position coding along the nism, a versatile module, can be efficiently utilized in both

139022 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 3. Diagram illustrating attention mechanisms: (a) SE attention mechanism module; (b) CBAM attention mechanism module; (c) ECA attention
mechanism module; (d) CA attention mechanism module.

FIGURE 4. Transformer models: (a) The transformer model introduced by Google Research for NLP applications; (b) Fundamental architecture of
ViT; (c) Transformer encoder configuration within ViT.

CNNs and transformers. Its role is to facilitate the model in algorithms and single-stage algorithms [57], [58]. Apart from
prioritizing important features while disregarding irrelevant the aforementioned CNN-based methods, a category of PCB
background information, thereby enhancing the detection of defect detection algorithms based on Transformers has also
PCB defects. The transformer model is a deep learning con- emerged. Notable two-stage algorithms encompass Region
struct built upon the self-attention mechanism. Despite its CNN (R-CNN) [59], Fast Region CNN (Fast R-CNN) [60],
late introduction, the transformer model has witnessed rapid Faster Region CNN (Faster R-CNN) [61], and Mask Region
growth due to the extensive research conducted by a myriad CNN (Mask R-CNN) [62], among others. These methods
of scholars. It has found widespread application in numerous partition PCB defect detection into two phases: firstly, region
tasks within NLP and CV, delivering outstanding results. The proposal (RP), which involves generating pre-selected boxes
potential for further development of the transformer model is potentially containing the objects to be detected; subse-
undeniable. quently, sample classification is conducted using CNNs.
The R-CNN algorithm starts by generating a sequence of
III. PCB DEFECT DETECTION METHOD BASED ON DEEP candidate regions (Region Proposals) within the input image.
LEARNING It then performs CNN feature extraction on these regions and
In the domain of PCB defect detection, detection algo- forwards the extracted features to classifiers and bounding
rithms are categorized into two primary groups: two-stage box regressors for target detection. While possessing high

VOLUME 11, 2023 139023

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 5. Transformer models in CV: (a) Schematic of the proposed attention-augmented convolutional architecture; (b) Substitution of attention
for convolution in ResNet with kernel size k=3; (c) Pioneering alternative to CNN for the DETR model.

accuracy, its speed is impeded by the need for separate CNN dimensions. This design enables SSD to excel in detecting
feature extraction for each candidate region. The algorithm’s small and multi-scale targets. Its structure is illustrated in
structure is depicted in Fig. 6 (a). Faster R-CNN, an upgraded Fig. 6 (d). In comparison to YOLO, SSD exhibits slightly
version of R-CNN, introduces the region proposal network slower performance, although it may offer heightened accu-
(RPN), a learnable network designed for rapid candidate racy in specific contexts.
region generation. Faster R-CNN seamlessly integrates the Comparing the two-stage and single-stage algorithms,
RPN with subsequent classifiers and bounding box regres- the two-stage algorithm has higher accuracy but is time-
sors, resulting in an end-to-end target detection network. This consuming and not suitable for real-time detection tasks.
integration enables candidate region generation and feature Whereas, the single-stage algorithm is faster but relatively
extraction within the same network, significantly enhancing less accurate. Transformer-based PCB target detection algo-
detection speed. The configuration of Faster R-CNN is shown rithms have also emerged, and unlike traditional CNNs, these
in Fig.6 (b). methods are based on Transformer and show excellent detec-
The other category encompasses single-stage algorithms, tion accuracy, but the detection speed is slower, the training
prominently represented by single shot multiBox detector requires higher computational resources, and it requires a
(SSD) [63] and You Only Look Once (YOLO) [64], [65], large amount of data support [70].
[66], [67], [68], [69]. These methods redefine PCB defect
detection as a regression problem. The YOLO algorithm A. BASED ONE-STAGE ALGORITHM
introduces a distinctive perspective, treating target detection Addressing the issues of poor stability and low accuracy
as a regression challenge. It accomplishes this by partition- in PCB defect detection models, Xin et al. [71] introduced
ing the input image into a grid and concurrently predicting an enhanced YOLOv4 model. This approach incorporates a
class and bounding box attributes for multiple targets within mosaic data augmentation strategy during input processing
each grid cell. YOLO achieves target detection through and replaces the leaky rectified linear unit (Leaky-ReLU)
single-shot forward propagation, rendering it exceptionally activation function in the network backbone with the Mish
rapid. However, its performance in detecting small targets activation function [72]. Additionally, the detection images
might be comparatively less effective. The structure of YOLO are automatically segmented based on the average size of
is depicted in Fig. 6 (c). The SSD algorithm also falls labeled boxes, thereby increasing the likelihood of includ-
under the single-stage target detection approach. Similar ing the target in the anchor frame. Similarly, to address the
to YOLO, it simultaneously forecasts target categories and challenges of detecting small defects against complex back-
bounding boxes on feature maps at varying scales, adapting grounds in PCBs, Zhang et al. [73] introduced a lightweight
to diverse target scales via multiple anchor frames of differing single-stage defect detection network. This network leverages

139024 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 6. Traditional target detection algorithms: (a) Flowchart of R-CNN; (b) Fundamental structure of faster R-CNN; (c) Core process of the YOLO
algorithm; (d) Network architecture schematic of the SSD algorithm.

the dual attention mechanism and a path-aggregation fea- evaluated its performance on YOLOv5 using both labeled
ture pyramid network (PAFPN) to enhance the detection of and unlabeled samples. This approach reduces the reliance
small defects. MobileNetV2, a lightweight backbone neural on labeled samples, as it utilizes both labeled and unlabeled
network, replaces ResNet101, considerably reducing model data. Moreover, a data-expanding strategy is proposed to
parameters. A dual attention mechanism is integrated to mitigate the impact of unlabeled samples. This enhancement
ensure effective feature extraction. The feature extraction is particularly evident with small data volumes; however, the
capability is further enhanced by substituting PAFPN [74] effectiveness diminishes as data volume increases. In another
for feature pyramid network (FPN) [75] in the neck. This study, Wu et al. [82] presented GSC YOLOv5, a deep learn-
improved model reduces inference time, and parameter count, ing detection approach that integrates a lightweight network
and enhances detection accuracy. Jiang et al. [76] pro- with a dual-attention mechanism. This modified algorithm
posed enhancements to the SSD network model, introducing employs lightweight Ghost Conv and Ghost Bottleneck struc-
coordinated attention in the shallow network to better han- tures to significantly reduce the model’s parameter count and
dle positional information, particularly for smaller targets. floating-point operations. Furthermore, SE and CBAM mod-
Li et al. [77] created a dataset for PCB assembly scene object ules are incorporated into the network, resulting in enhanced
detection, addressing anchor frame size-related detection accuracy and improved detection speed. Addressing con-
issues. They performed a detailed analysis of effective recep- cerns related to detection efficiency, memory consumption,
tive fields (ERF) [78] across the output layers, defining ERF and sensitivity to small defects, Xuan et al [83] adopted
ranges and introducing ERF-based anchor frame assignment a novel cross stage partial network darknet (CSPDarkNet)
rules to address anchor frame size challenges. Furthermore, as the YOLOX backbone network. This revised backbone
they designed an improved atrous spatial pyramid pooling comprises multiple inverted residual blocks and incorpo-
(ASPP) [79], [80], incorporated a channel attention module, rates coordinated attention into the network architecture,
and added contextual information to address challenges posed significantly improving the model’s capacity to detect small
by small and hard-to-detect defects. PCB defects. Notably, this modified model is lighter and
Due to the limited number of labeled PCB defect sam- more suitable for deployment on embedded systems. Zhao,
ples, the training process is influenced by unlabeled samples. Y. et al. [84] extended YOLOv5 by integrating adaptively
To address this, Wan et al. [81] introduced a defect detec- spatial feature fusion (ASFF) [85] for feature fusion. This
tion method with a data-expanding strategy (DE-SSD) and integration enables the adaptive fusion of varying levels

VOLUME 11, 2023 139025

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

of feature information across different spaces. Additionally,

they introduced the global attention mechanism (GAM) [86]
to enhance the model’s information extraction capabilities.
Zheng and colleagues [87] introduced an enhanced fully
CNN by integrating successive convolutional modules into
the MobileNetV2 architecture. This augmentation, along-
side an improved skip connection, contributes to heightened
detection speed and accuracy in contrast to VGG-16 and
ResNet-50 models. Lim and colleagues [88] developed a
novel multi-scale feature pyramid network using YOLOv5,
addressing the detection of minuscule PCB defects by lever-
aging contextual insights. The network also incorporates
the CIoU loss function to precisely determine the spa-
tial parameters, effectively capturing the exact positions of
these imperfections. Yu and co-authors [89] introduced a
lightweight, efficient network tailored for detecting minute
PCB defects. Within the backbone network, they intro-
duced diagonal feature pyramid (DFP), a mechanism for
low-cost fusion of expansive feature maps, enhancing the
detection of these subtle flaws. Additionally, they devised a
FIGURE 7. Results of PCB defect detection using four different
multi-scale necking network to accommodate various scales algorithms: (a) (b) Printed circuit boards defect detection method based
of defects. Moreover, they introduced an adaptive localization on improved fully convolutional networks [87]; (c) (d) Printed circuit
loss function, enhancing the model’s ability to discern these board quality detection method integrating lightweight network and dual
attention mechanism [82].
small-scale imperfections.
In summary, the aforementioned research has made signifi-
cant advancements in the field of PCB defect detection using to design appropriate anchor frames, enhances inter-feature
one-stage algorithms. These advancements encompass vari- map relationships, and utilizes online hard example mining
ous novel approaches, including attention mechanisms, data (OHEM) [91] to refine region of interest (ROI) prediction.
enhancement techniques, and innovative backbone networks. Addressing the limitations of template-dependent and com-
These methods have found widespread application in PCB putationally demanding traditional defect detection methods,
defect detection from images, yielding impressive outcomes Hu et al. [92] introduced an algorithm founded on Faster
across multiple dimensions such as enhanced accuracy and RCNN and Feature Pyramid FPN (Feature Pyramid Net-
detection speed. A comparative analysis of these one-stage work). This algorithm initially employs ResNet50 with
algorithms with specific alternative methods, along with their feature pyramids as its backbone network and subsequently
benefits and limitations, is presented in Table 2. It is crucial integrates generative adversarial region proposal networks
to acknowledge, however, that the one-stage algorithms are (GARPN) [93] to enhance anchor frame prediction accu-
not exempt from limitations. These limitations encompass the racy. In the same year, Li et al. [94] introduced a feature
potential for diminished performance when confronted with pyramid-based network. They incorporated an SE module
intricate and varied defective scenarios, the high reliance on into ResNet-101 to enhance network expressive power, intro-
limited sample data, and the imperative for further enhance- duced a top-down structure to elevate overall feature levels,
ments in coping with diverse dimensions, angles, and lighting and employed ROI Align in lieu of ROI Pooling to mitigate
conditions. Fig. 7 presents the results of PCB defect detection the impact of dislocations on small object defect detection.
using four different algorithms. In summary, when considering studies based on two-
stage algorithms, it’s evident that there are fewer scholarly
works compared to those focused on one-stage algorithms.
B. BASED TWO-STAGE ALGORITHM However, there’s no doubt that two-stage-based algorithms
While the single-stage algorithm offers faster performance, offer significant advantages in terms of detection accuracy.
the two-stage algorithm notably outperforms it in terms of Table 3 provides a comparison of two-stage-based algo-
detection accuracy. To achieve PCB defect detection via rithms, detailing the specific methods employed, along with
machine vision, Li et al. [90] introduced a faster-RCNN their respective advantages and limitations. Recently, one-
algorithm founded on VGG16. This algorithm incorporates stage algorithms have flourished to meet real-time detection
data expansion and RGB data enhancement. Addressing the speed requirements. However, the difference in detection
challenge of detecting tiny defects, which are inherently accuracy, compared to two-stage algorithms, is not signif-
challenging to generate and detect in real-world scenarios, icant. As a result, there are relatively fewer researchers
Ding et al. [8] introduced tiny defect detection network focusing on this field. Nonetheless, utilizing two-stage
(TDD-net). This network employs a K-means algorithm algorithms remains a valuable choice for specific tasks

139026 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

TABLE 2. A comparison of strength and boundedness in one-stage CNN algorithms.

TABLE 3. A comparison of strength and boundedness in two-stage CNN algorithms.

TABLE 4. A comparison of strength and boundedness in transformer algorithms.

demanding high detection accuracy and tailored datasets, the results of PCB defect detection using four different
without necessitating rapid detection speed. Fig. 8 presents algorithms.

VOLUME 11, 2023 139027

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

FIGURE 8. Results of PCB defect detection using four different

algorithms: (a) (b) Detection of PCB surface defects with improved
faster-RCNN and feature pyramid network [92].

C. BASED TRANSFORMER ALGORITHM

While Transformer has demonstrated proficiency in CV and
FIGURE 9. Results of PCB defect detection using four different
NLP, it does exhibit limitations when employed in visual algorithms: (a) (b) LPViT [95]; (c) (d) Transformer-YOLO [96].
inspection tasks, notably involving time constraints and
demanding equipment prerequisites. These limitations have
resulted in a scarcity of literature concerning the implemen- convolutional layer. This addition serves to amplify the depth
tation of Transformer for PCB defect detection. Nevertheless, of SA and bolster the flexibility of the attention mechanism.
notwithstanding these limitations, several studies have com- Fig. 9 presents the results of PCB defect detection using four
menced efforts to address these challenges, aiming to unlock different algorithms.
the full potential of Transformer in the realm of PCB defect Despite transformers being infrequently used in PCB
detection. By surmounting the limitations of the transformer defect detection, the evolution of computer vision and compu-
and effectively employing it in PCB defect inspection tasks, tational capabilities has catalyzed the proposal of numerous
it is anticipated to offer a more efficient and accurate solution transformer models. In 2021, Liu et al [55] trained a swin
for industrial production. transformer v2 model, enriched with three billion parameters,
An et al. [95] introduced label robust and patch corre- and introduced a post-normalization technique and scaled
lation enhanced ViT (LPViT). In their work, a novel ViT cosine attention method. This approach achieved state-of-the-
model is presented, which is founded upon the principles of art (SOTA) results across multiple visual tasks. As of 2023,
LPViT. This model prioritizes robustness while fully lever- the swin transformer v2 backbone network continues to be a
aging distinct regions of the PCB image across relationships. subject of extensive research, demonstrating exceptional.
Additionally, certain blocks are randomly masked or sub- Performance and further stimulating the evolution of large
stituted to enhance mutual understanding between different visual models. Today’s ViTs have not explicitly leveraged
image regions. Ultimately, the model undergoes training via features at different scales, which are notably crucial for
a label smoothing strategy, elevating its robustness. Chen [96] visual inputs. To address this, Wang et al [98] proposed the
employs an enhanced clustering algorithm to generate appro- crossformer in 2021, introducing a cross-scale embedding
priate anchor frames tailored to the PCB defective dataset layer (CEL) and a long-short distance attention (LSDA).
in this study. Next, CNN was abandoned in favor of shifted In 2023, they [99] further enhanced the crossformer by
window transformer (Swin-Transformer) for network feature proposing the progressive group size (PGS) and amplitude
extraction. Subsequently, the channel order in the feature map cooling layer (ACL) to mitigate challenges associated with
was adjusted to enable the network to efficiently prioritize enlarging self-attention maps and amplitude explosion. Given
more valuable information. Additionally, both convolutional the lack of a priori image information, ViT underperforms
and attentional mechanisms were integrated to enhance the in dense prediction tasks. To remedy this, Chen et al [100]
network’s feature extraction capacity. Yang et al. [97] intro- proposed the ViT-adapter in 2022, an ancillary network not
duced an enhanced YOLOv7 model. They achieved this requiring pre-training, enabling the basic ViT model to adapt
enhancement by formulating the SwinV2_TDD module, to downstream dense prediction tasks without any archi-
which facilitates the extraction of local PCB informa- tectural modifications. This significantly improved model
tion through the incorporation of an added convolutional performance for dense prediction tasks.
layer. Next, the study introduces the MFSA mechanism, These novel transformer models, designed for visual
which augments each shuffle attention (SA) branch with a macro-models, harness features at different scales and excel

139028 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

in dense prediction tasks. These tasks are also inherent in

transformer-based PCB defect detection. We posit that these
innovative transformer models and their methodologies can
be applied to PCB defect detection tasks, enhancing the per-
formance of existing models.

D. SUMMARY
In this section, we elucidate the one-stage, two-stage, and
transformer-based algorithms. One-stage algorithms, char- FIGURE 10. IoU schematic diagram: (a) Red represents the intersection
acterized by high detection accuracy and speed, cater to region [8]; (b) Yellow represents union region [8].

enterprise needs for real-time PCB defect detection. How-

ever, their efficiency may diminish in more challenging defect
scenarios. In contrast, the two-stage algorithm, despite being
slower, excels in detecting intricate defects due to its superior
accuracy, leading to its widespread adoption in factories for
PCB defect detection. The transformer-based algorithm, dis-
tinct in structure from its counterparts, has also demonstrated
commendable results in PCB defect detection and various
industrial tasks. Despite its large parameter count, it effec-
FIGURE 11. IoU limitations (while the IoU values are equal for the three
tively fulfills factory requirements for PCB defect detection. images, it’s noticeable that image (a) [8] and image (b) [8] exhibit lower
Concurrently, the transformer model exhibits impressive per- accuracy. Image(c) demonstrates superior accuracy [8].)
formance in other domains, indicating substantial potential
for further development.
Specifically, IoU’s highest value, 1, signifies complete over-
IV. ASSESSMENT METRICS, PCB DEFECT DATASETS, AND lap between the actual object area and the inferred area,
COMPARATIVE RESULTS whereas its lowest value, 0, indicates no overlapping seg-
A. ASSESSMENT METRICS ment between the actual object area and the inferred area.
PCB defect detection tasks often require assessing algorith- However, there are instances where IoU may not precisely
mic localization accuracy, which measures the discrepancy capture positional accuracy. For instance, in the three images
between the target bounding box identified by the algorithm shown in Fig.11, although the calculated IoUs for these
and the actual target bounding box [101]. For assessing algo- images (Fig.11 (c)) are equal, it’s evident that the third
rithmic positional accuracy, a widely adopted metric is the image is superior. Consequently, several enhanced iterations
intersection over union (IoU) [102], [103], defined as the of IoU have emerged, including GIoU [102], DIoU [105],
ratio of the intersection area between the target frame and CIoU [105], EIoU [106], αIoU [107], and SIoU [108]. GIoU
the actual frame to the union area. Specifically, in the context resolves the issue of the unavailability of gradient back-
of the detection outcomes and real annotations depicted in propagation for IoU’s two frames without their intersection.
Fig. 10 below, the intersection encompasses the overlapping Meanwhile, DIoU considers the prediction frame and real
portion of the two frames, while the union encompasses frame, building on the GIoU distance involving the centroid
the combined area of both frames. The extent of overlap and the distance between the minimum enclosing frames.
between detection outcomes and actual labeling is computed It also accounts for aspect ratio relationships based on DIoU.
through intersection and union ratios, facilitating algorithmic However, it does not address the gap in actual distance.
positional accuracy evaluation as presented in formula (1). Consequently, EIoU replaces DIoU’s aspect ratio with the
In practical detection scenarios, the intersection ratio thresh- actual difference between width and height, along with their
old is commonly set to 0.5. If the intersection ratio between respective confidence levels. In 2021, αIoU was introduced,
the targeted detection frame and the actual labeled frame which involves only a single parameter α in comparison to
exceeds 0.5, the algorithm is deemed to have successfully other IoUs, yet it yields superior outcomes. The presence
anticipated the target’s location [104]. Modifying the inter- of multiple angles at equidistant points impacts the actual
section ratio threshold enables adjustments to the algorithm’s loss. Therefore, in 2022, SIoU was introduced to address the
positional accuracy and miss rate. Notably, distinct thresholds influence of angles. However, at present, a modified version
may necessitate establishment for varying tasks. of IoU is solely employed in the loss function, whereas IoU
remains the standard for evaluation metrics.
A∩B Categorizing detection results in target detection tasks
IoU = (1)
A∪B involves four distinct groups [109]: predicted values aligned
Equation (1) reveals that a greater IoU signifies the prox- with positive examples are labeled as P (Positive), those
imity of the estimated object region to the actual region, aligned with negative examples are labeled as N (Negative),
thus leading to higher accuracy in the detection results. values aligning with true values are T (True), and values

VOLUME 11, 2023 139029

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

2) PRECISION
Precision refers to the ratio of correctly predicted positive
samples to the total predicted positive samples by the model.
The formula is depicted in Equation (3).
TP
Precision = (3)
TP + FP
3) RECALL
Recall, also known as the true positive rate, represents the
ratio of correctly predicted positive samples to the actual total
positive samples. The formula is illustrated in Equation (4).
TP
Recall = (4)
TP + FN
4) F1 SCORE
The F1 score provides a balanced mean between precision
and recall, effectively harmonizing the model’s accuracy and
recall. The formula is depicted in Equation (5).
2TP
F1 = (5)
2TP + FP + FN
5) AVERAGE PRECISION
Among the frequently employed evaluation metrics, Preci-
FIGURE 12. Schematic diagram of the recognition result. sion and Recall stand as crucial indicators for gauging model
performance. However, a notable observation arises from the
opposing true values are F (False). Upon classification, the equation above. Precision and Recall exhibit a contradictory
data can be organized into a confusion matrix to yield four relationship: an increase in Precision often corresponds to a
distinct combination types. decrease in Recall, and vice versa. To address this challenge,
Notably, As shown in Fig.12, TP represents the count of the Average Precision (AP) metric is introduced to compre-
correctly detected targets, encompassing instances where the hensively evaluate the model’s performance.
predicted positive sample aligns with the true positive sample; The term ‘‘average precision’’ specifically denotes the
TN signifies the count of accurately identified background mean value resulting from the integration of accuracy rates
instances, where the predicted negative sample aligns with the across distinct thresholds, spanning a recall range from 0 to 1.
true negative sample; FP designates the count of erroneously For each category, a check-accuracy-recall curve is plot-
detected targets, represented by instances where the predicted ted, enabling the calculation of the area beneath this curve,
positive sample mismatches the true negative sample; and FN referred to as the AP value. Consequently, a model’s aver-
accounts for the count of improperly missed targets, signify- age precision corresponds to the mean AP values across all
ing cases where the predicted negative sample mismatches categories, constituting the mean average precision (mAP).
the true positive sample. Notably, the mAP metric stands as one of the most extensively
Evaluating algorithm performance becomes more precise employed performance measures in target detection, quanti-
through a comprehensive tally and comparison of pre- fying a model’s proficiency in identifying multiple categories
dictions based on distinct types. Examples of assessment of targets.
metrics include Accuracy (Acc), Precision [110], Recall The computation methods for mAP metrics vary. Specif-
[111], F1 Score [26], Average Precision [112], mAP [104], ically, AP0.5 represents the average accuracy when the
mAPSmall , mAPmedium , mAPlarge [113], and FPS (Frames Per intersection-over-union (IoU) threshold exceeds 50%, and
Second) [67]. AP0.5:0.95 corresponds to the average accuracy when the
IoU threshold ranges from 50% to 95% in 5% increments.
1) ACCURACY In real-world target detection tasks, models often need to
The accuracy rate represents the proportion of correctly clas- detect targets from multiple categories, necessitating the cal-
sified samples among the total samples, thus indicating the culation of AP values for each category and subsequent
model’s precision in classifying input data. The formula is averaging to derive the mAP metric.
depicted in Equation (2). In conclusion, the mean average precision (mAP) stands
as a crucial metric in assessing target detection model perfor-
TP + TN mance. By amalgamating model accuracy and recall, mAP
Acc = (2)
TP + FN + FP + FN offers insight into a model’s ability to detect a multitude of

139030 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

diverse target categories. Consequently, this review employs C. RESULT COMPARING

mAP0.5:0.95 and [email protected] (AP0.5) as the primary evalua- This subsection will primarily focus on an in-depth analysis
tion criterion. The formula is depicted in Equation (6). and meticulous comparison of the previously cited litera-
ture. By systematically summarizing and comprehensively
Z 1 N
1 X evaluating the research outcomes from these references,
AP = P(R)dr, mAP = APi (6)
0 N we can attain a more comprehensive understanding of the
(i=1)
latest advancements in the relevant field and unveil both
commonalities and disparities therein. The results compar-
6) mAPSmall , mAPmedium , AND mAPlarge
ison encompasses various metrics, including mean average
These three metrics correspond to average accuracies for
precision (mAP), frames per second (FPS), parameter quan-
objects of different sizes. Specifically, mAPSmall pertains to
tity, and recall rate. These metrics collectively provide a
objects with an area less than 32 × 32 pixels, mAPmedium
comprehensive assessment of the performance of different
corresponds to objects with an area more than 32 × 32 pixels
algorithms or models in experimental settings. Furthermore,
and less than 96 × 96 pixels, and mAPlarge relates to objects
within the comparative experiments, detailed specifics of the
with an area greater than 96 × 96 pixels. These metrics
utilized datasets will also be elaborated upon.
provide a more detailed insight into the algorithm’s detection
Various performance metrics from the aforementioned
accuracy across small, medium, and large targets.
literature are presented in Fig.13. Fig.13 (a) illustrates a com-
parison of FPS across different algorithms, indicating that
7) FPS many of these algorithms are primarily single-stage models,
FPS serves as a metric for assessing inference speed, rep- achieving real-time detection goals, with some even reaching
resenting the quantity of images that can be processed per up to 90 FPS. Achieving a balance between inference speed
second on specific hardware. FPS holds significance as a and required accuracy is crucial in practical development
key metric for gauging model performance and its appli- to meet industrial demands. In Fig. 13 (b), a comparison
cability in real-time scenarios. Through computation of the of mAP reveals significant disparities, possibly arising from
model’s processing capacity for images in a single second, dataset diversity and difficulty levels. Nevertheless, all these
the model’s real-time performance can be assessed. Elevated algorithms yield commendable results. Fig. 13 (c) displays
FPS values signify the model’s capacity for swifter image a comparison of parameters, where lower values generally
processing, thereby enhancing real-time inference efficiency. correspond to reduced device demands and shorter inference
This attribute is crucial for numerous applications demand- times. Models with fewer parameters are suitable for deploy-
ing rapid response times, including real-time video analysis ment on mobile devices to fulfill specific requirements. The
[64], autonomous driving systems [114], and real-time object two studies with the lowest parameters often accomplished
recognition [64]. Consequently, researchers and developers this through network optimization and pruning, effectively
strive to enhance a model’s FPS value. reducing parameters without notable accuracy compromise.
Fig. 13 (d) presents [email protected] comparisons, with the major-
B. PCB DEFECT DATASETS ity of algorithms achieving around 90% accuracy, meeting
In the literature related to PCB defect detection mentioned industrial practicality. In Fig. 13 (e), recall rates across dif-
earlier, many studies utilize proprietary datasets. There- ferent algorithms are shown, with Transformer-based models
fore, this subsection focuses on introducing several publicly exhibiting higher recall rates, underscoring the significant
available datasets, which possess certain evident advan- developmental potential of Transformers in the future.
tages compared to proprietary datasets. Public datasets not
only offer greater persuasiveness but also provide more
accurate baselines. Additionally, they facilitate experiment D. SUMMARY
reproducibility among other researchers. These publicly This section articulates evaluation metrics pertinent to PCB
available datasets offer researchers more reliable standards defect detection, explores publicly accessible datasets, and
and benchmarks, fostering performance comparison and compares these metrics to the algorithms discussed in
method enhancement. Simultaneously, they provide a broader section II. Evaluation metrics afford an assessment of a
platform for progress and collaboration within the research model’s strengths and weaknesses, with a higher FPS imply-
community. Currently available datasets include PCB Defect, ing a more rapid model inference, a greater mAP and
PCB Defect-Augmented [8], DEEP PCB [115], HRIPCB [email protected] signifying superior accuracy, and a lower parame-
[116], and Micro-PCB, among others. These datasets exhibit ters value reflecting a reduced model parameter count. Within
varying characteristics such as different defect types, image the realm of PCB defect detection datasets, we concentrate
quantities, and environmental factors. Models trained on on several public collections that range in size from over a
different datasets yield varying accuracies and are suited for thousand to more than 10,000 images. These public datasets
different scenarios. Detailed information regarding currently facilitate the establishment of a baseline, simplify perfor-
available publicly accessible PCB defect detection datasets is mance comparison across models, and simultaneously, the
presented in Table 5. higher quality datasets foster advancements in PCB defect

VOLUME 11, 2023 139031

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

TABLE 5. Overview of PCB datasets.

FIGURE 13. Comparison of performance indicators in various references: (a) FPS, (b) mAP, (c) Parameters, (d) [email protected], (d) Recall.

detection. In the comparison of results, we enumerate the a clear visualization of each model’s strengths and weak-
backbone, FPS, mAP, [email protected], parameters, recall, and nesses. However, due to the use of varying datasets, these
the dataset used in each paper for every model. Utilizing comparative results do not directly adjudicate the models’
this data, we produce a suite of statistical charts, providing performance.

139032 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

TABLE 6. Multiple result comparison of the above reference algorithm used in PCB defect detection.

V. DISCUSSION circuit breaks and wire corrosion. Furthermore, PCB images

In delving into the domain of PCB defect detection, ample often suffer from noise and other interference factors due
room for development in this field becomes evident. Despite to the complexity of production environments and potential
the existence of numerous exceptional algorithms, PCB random factors during image acquisition. These interferences
images exhibit distinct dissimilarities compared to natural include noise, shadows, and variations in lighting conditions.
images, presenting unique challenges in defect detection. Lastly, efficient and rapid PCB defect detection is impera-
Moreover, PCB defects display prominent characteristics. tive. In practical production settings, the urgency to fulfill
Firstly, PCB images possess intricate structures and intricate production line requirements mandates the timely completion
details, stemming from the complex nature of circuit boards. of PCB defect detection [89].
These images contain a multitude of elements such as lines, In the face of these challenges, we acknowledge that
solder points, and components, with intricate interconnec- traditional methods have gradually demonstrated limited
tions, demanding algorithms capable of comprehending and capabilities in addressing these issues. The steps are different
analyzing complex structures. Secondly, PCB defects fre- between traditional machine vision methods and deep learn-
quently exhibit diversity and variations, with different defect ing based methods. Conventional machine vision methods
types and locations manifesting distinctive features and forms necessitate an initial pre-processing of the image to enhance
in PCB images [117]. For instance, solder point defects can its quality. Subsequently, the PCB board is isolated via image
encompass scenarios like open soldering, short circuits, and segmentation, followed by the extraction of features, such as
displacement, while line defects may involve issues such as shape, texture, color, and others, indicative of PCB defects

VOLUME 11, 2023 139033

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

within the image. The location of the defect in the PCB image in the field of CV, they are accompanied by considerable
is then identified and classified using machine learning algo- computational demands and sample requirements, limiting
rithms, and the detection results undergo post-processing. real-time detection advantages. With the continuous evolu-
In contrast, PCB defect detection grounded in deep learning tion of Transformer technology, models with comparable
can learn from substantial or even massive volumes of data. speed and accuracy to CNN have emerged. Given that
This method automatically discerns various levels of image Transformer-based models do not require non-maximum sup-
features through CNNs or transformers, eliminating the need pression (NMS) during inference, we hold confidence in their
for manual image feature design and extraction. Deep learn- substantial potential for development.
ing employs an end-to-end training approach, treating feature Through an in-depth exploration of the PCB defect detec-
extraction and classification or regression as an integrated tion domain, we have not only recognized its vast potential for
process. This eliminates the requirement for a complex development but also gained a profound appreciation for the
machine vision processing workflow. However, it is precisely remarkable performance of deep learning methods within it.
behind these challenges that deep learning-based approaches Among various approaches, CNNs, transfer learning, GANs,
begin to manifest remarkable superior performance. Relative and Transformers exhibit substantial prospects. Particularly
to traditional methods, deep learning methods exhibit multi- in the realm of PCB defect detection, deep learning methods
ple advantages stemming from their exceptional capabilities stand out with their exceptional performance, addressing the
in image processing and pattern recognition. Firstly, deep limitations of traditional methods and demonstrating note-
learning methods possess the capability of automated fea- worthy advantages in automated feature extraction, adapt-
ture extraction, eliminating the cumbersome manual feature ability, accuracy, and complex relationship handling. Despite
engineering process. In contrast, traditional methods often these significant advancements, the field of PCB defect detec-
require significant time and effort to design manual feature tion still confronts a spectrum of difficulties and challenges.
extractors to capture key information in images. Secondly, As we look towards future development, it is evident that a
deep learning methods demonstrate enhanced adaptability mixture of challenges and opportunities lies ahead.
[4]. Given the diversity in PCB defect types and shapes, tra- In addition to visible light detection, infrared imaging
ditional methods struggle to encompass all defect variations. detection is also widely used in industry. Understanding
Furthermore, deep learning methods exhibit higher accuracy. the heat transfer mechanism is crucial for infrared detec-
Through training on extensive datasets, deep learning models tion. By deeply understanding the heat transfer model, the
learn from numerous samples, thereby enhancing detection advantages and disadvantages of infrared imaging in various
accuracy. Conversely, traditional methods may be constrained defect detection and classification applications can be ana-
by specific rules and limitations, preventing them from reach- lyzed more thoroughly [118]. At the same time, revealing
ing the accuracy levels of deep learning models [77]. Lastly, the heat transfer mechanism helps to further improve the
deep learning methods can handle complex relationships. accuracy and applicability of infrared imaging. For example,
In PCB images, multiple defects may coexist with intricate the study of various heat transfer models, including electrical
interdependencies and interactions. Deep learning models discharge machining (EDM) heat transfer [119], [120], laser
adeptly capture these intricate relationships, resulting in more heat transfer [121], [122], and wire-EDM heat transfer [123],
accurate detection of multiple defects. [124], helps to have a more comprehensive understanding of
However, despite the significant achievements of deep the mechanism of infrared imaging. Furthermore, consider-
learning in PCB defect detection, certain methods still exhibit ing the differences in thermal gradients may be helpful for
more remarkable potential for future development. Firstly, detecting defects in infrared imaging.
methods based on CNNs offer substantial advantages in
detection speed and accuracy. Employing CNN for feature VI. OUTLOOK
extraction and classification of PCB images facilitates the In recent years, technologies based on deep learning
effective detection of common defect types. have demonstrated remarkable capabilities across various
Furthermore, by enhancing network structures and aug- domains, particularly in the realm of image processing. This
menting sample data, continuous improvements in detec- comprehensive review aims to investigate and analyze the
tion accuracy can be achieved [92]. Secondly, transfer application of deep learning techniques to the detection of
learning methods can leverage pre-trained models like defects in PCBs, exploring their potential, methodologies,
ImageNet for fine-tuning, and adapting to the unique challenges, and future directions.
features of PCB images. This approach reduces the (1) Enormous Potential of Deep Learning in PCB Defect
demand for training samples, enhancing both training Detection: Deep learning models, such as CNNs and
speed and detection effectiveness. Additionally, harness- Transformer models, have made significant strides in PCB
ing data augmentation through GANs can provide diverse defect detection owing to their robust feature learning and
PCB defect images for deep learning algorithms, thus bol- representation capabilities. CNNs employ stacked convolu-
stering model robustness. Lastly, PCB defect detection tion and pooling operations to learn local and global features
algorithms based on Transformers show promise. Although from raw images, progressively abstracting and compre-
Transformer-based models exhibit exceptional performance hending image content. While Transformer models have

139034 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

achieved success in natural language processing, they have precision and streamlines quality control procedures. Addi-
also demonstrated applicability in image processing tasks. tionally, increasing model interpretability and explainability,
These models leverage self-attention mechanisms to capture enhancing model decision transparency, and incorporating
global information while retaining sensitivity to local details, techniques such as reinforcement learning, transfer learning,
providing substantial support for PCB defect detection. and generative adversarial networks can further elevate the
(2) Diverse Detection Approaches: Deep learning-based performance of PCB defect detection.
PCB defect detection methods are categorized into one-stage In conclusion, deep learning-based PCB defect detection
and two-stage algorithms. One-stage algorithms directly pre- holds vast developmental prospects and application potential
dict defect locations and categories in raw images, offering in the future. Technological advancements coupled with
real-time advantages. Two-stage algorithms first extract can- in-depth research will drive improvements in quality control
didate regions from images and then classify these regions, and production efficiency within the PCB manufacturing
achieving more precise defect localization. Additionally, industry.
Transformer-based methods have shown promising results
in PCB defect detection by transforming image data into REFERENCES
sequential data for processing, capturing features at different [1] Q. Ling and N. A. M. Isa, ‘‘Printed circuit board defect detection
scales and levels. methods based on image processing, machine learning and deep learn-
(3) Significance of Evaluation: In PCB defect detection ing: A survey,’’ IEEE Access, vol. 11, pp. 15921–15944, 2023, doi:
10.1109/ACCESS.2023.3245093.
tasks, the choice of appropriate evaluation metrics is crucial [2] A. F. M. Hani, A. S. Malik, R. Kamil, and C.-M. Thong, ‘‘A review
for assessing model accuracy and stability. Common eval- of SMD-PCB defects and detection algorithms,’’ Proc. SPIE, vol. 8350,
uation metrics include precision, recall, accuracy, and F1 pp. 373–379, Jan. 2012, doi: 10.1117/12.920531.
[3] J.-H. Park, Y.-S. Kim, H. Seo, and Y.-J. Cho, ‘‘Analysis of training
score. Moreover, selecting suitable PCB defect datasets for deep learning models for PCB defect detection,’’ Sensors, vol. 23, no. 5,
training and testing is vital to ensuring model generaliza- p. 2766, Mar. 2023, doi: 10.3390/s23052766.
tion. Thoughtful dataset selection better reflects real-world [4] Q. Zhang and H. Liu, ‘‘Multi-scale defect detection of printed circuit
board based on feature pyramid network,’’ in Proc. IEEE Int. Conf.
application scenarios and more accurately evaluates model
Artif. Intell. Comput. Appl. (ICAICA), Jun. 2021, pp. 911–914, doi:
performance. 10.1109/ICAICA52286.2021.9498174.
(4) Overcoming Challenges and Issues: Despite the notable [5] D. Makwana and S. Mittal, ‘‘PCBSegClassNet—A light-weight network
progress of deep learning in PCB defect detection, challenges for segmentation and classification of PCB component,’’ Exp. Syst. Appl.,
vol. 225, Sep. 2023, Art. no. 120029, doi: 10.1016/j.eswa.2023.120029.
persist. Firstly, PCB defects exhibit diverse complexity, pos- [6] I. A. Soomro, A. Ahmad, and R. H. Raza, ‘‘Printed circuit board identifi-
ing challenges for designing robust deep learning models to cation using deep convolutional neural networks to facilitate recycling,’’
detect various defects. Different defect types may possess Resour., Conservation Recycling, vol. 177, Feb. 2022, Art. no. 105963,
doi: 10.1016/j.resconrec.2021.105963.
distinct features and morphologies, making enhancing gen- [7] M. Liukkonen, E. Havia, and Y. Hiltunen, ‘‘Computational intelligence
eralization a formidable task. Secondly, data imbalance and in mass soldering of electronics—A survey,’’ Exp. Syst. Appl., vol. 39,
scarcity of defect samples impact model training and gen- no. 10, pp. 9928–9937, Aug. 2012, doi: 10.1016/j.eswa.2012.02.100.
eralization, necessitating effective strategies for mitigation. [8] R. Ding, L. Dai, G. Li, and H. Liu, ‘‘TDD-Net: A tiny defect detection
network for printed circuit boards,’’ CAAI Trans. Intell. Technol., vol. 4,
Lastly, enhancing model interpretability and explainability is no. 2, pp. 110–116, Jun. 2019, doi: 10.1049/trit.2019.0019.
a pressing research direction in PCB defect detection, partic- [9] X. Wu, Y. Ge, Q. Zhang, and D. Zhang, ‘‘PCB defect detection using
ularly for applications demanding rigorous model decisions. deep learning methods,’’ in Proc. IEEE 24th Int. Conf. Comput. Sup-
ported Cooperat. Work Design (CSCWD), May 2021, pp. 873–876, doi:
(5) Future Development Prospects: Deep learning-based 10.1109/CSCWD49262.2021.9437846.
methods retain advantages over traditional methods. Deep [10] F. Raihan and W. Ce, ‘‘PCB defect detection USING OPENCV
learning models autonomously learn features from data, with image subtraction method,’’ in Proc. Int. Conf. Inf.
Manage. Technol. (ICIMTech), Nov. 2017, pp. 204–209, doi:
reducing the reliance on manual feature engineering. With 10.1109/ICIMTech.2017.8273538.
hardware advancements and the accumulation of large-scale [11] A. Raj and A. Sajeena, ‘‘Defects detection in PCB using image pro-
data, the potential of deep learning models becomes more cessing for industrial applications,’’ in Proc. 2nd Int. Conf. Inventive
pronounced. Hence, future research can focus on improv- Commun. Comput. Technol. (ICICCT), Apr. 2018, pp. 1077–1079, doi:
10.1109/ICICCT.2018.8473285.
ing the robustness and generalization capabilities of deep [12] M. Moganti, F. Ercal, C. H. Dagli, and S. Tsunekawa, ‘‘Automatic
learning models, exploring more effective data augmen- PCB inspection algorithms: A survey,’’ Comput. Vis. Image Understand.,
tation and defect sample generation methods to address vol. 63, no. 2, pp. 287–313, Mar. 1996, doi: 10.1006/cviu.1996.0020.
[13] Y. Tang, M. Chen, C. Wang, L. Luo, J. Li, G. Lian, and X. Zou,
challenges related to diversity and data scarcity. Meanwhile, ‘‘Recognition and localization methods for vision-based fruit picking
employing crack analysis in tandem with deep learning robots: A review,’’ Frontiers Plant Sci., vol. 11, p. 510, May 2020, doi:
is crucial for detecting PCB defects. Specifically, the uti- 10.3389/fpls.2020.00510.
[14] W. He, Z. Jiang, W. Ming, G. Zhang, J. Yuan, and L. Yin,
lization of neural networks in model training accurately
‘‘A critical review for machining positioning based on computer
identifies minuscule cracks, often imperceptible to the human vision,’’ Measurement, vol. 184, Nov. 2021, Art. no. 109973, doi:
eye, on PCB surfaces. This application is invaluable for 10.1016/j.measurement.2021.109973.
early detection of potential issues, contributing significantly [15] W. Ming, F. Shen, X. Li, Z. Zhang, J. Du, Z. Chen, and Y. Cao,
‘‘A comprehensive review of defect detection in 3C glass com-
to ensuring product quality during production. Integration ponents,’’ Measurement, vol. 158, Jul. 2020, Art. no. 107722, doi:
of deep learning with crack analysis enhances detection 10.1016/j.measurement.2020.107722.

VOLUME 11, 2023 139035

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

[16] W. Ming, F. Shen, H. Zhang, X. Li, J. Ma, J. Du, and Y. Lu, [36] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for
‘‘Defect detection of LGP based on combined classifier with dynamic large-scale image recognition,’’ 2014, arXiv:1409.1556.
weights,’’ Measurement, vol. 143, pp. 211–225, Sep. 2019, doi: [37] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
10.1016/j.measurement.2019.04.087. V. Vanhoucke, and A. Rabinovich, ‘‘Going deeper with convolutions,’’
[17] W. Ming, C. Cao, G. Zhang, H. Zhang, F. Zhang, Z. Jiang, and J. Yuan, 2014, arXiv:1409.4842.
‘‘Review: Application of convolutional neural network in defect detection [38] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
of 3C products,’’ IEEE Access, vol. 9, pp. 135657–135674, 2021, doi: recognition,’’ 2015, arXiv:1512.03385.
10.1109/ACCESS.2021.3116131. [39] N. Cai, Q. Ye, G. Liu, H. Wang, and Z. Yang, ‘‘IC solder joint inspection
[18] W. He, Z. Shi, Y. Liu, T. Liu, J. Du, J. Ma, Y. Cao, and W. Ming, based on the Gaussian mixture model,’’ Soldering Surf. Mount Technol.,
‘‘Feature fusion classifier with dynamic weights for abnormality detec- vol. 28, no. 4, pp. 207–214, Sep. 2016, doi: 10.1108/ssmt-03-2016-0005.
tion of amniotic fluid cell chromosome,’’ IEEE Access, vol. 11, [40] V. Mnih, N. Heess, A. Graves, and K. Kavukcuoglu, ‘‘Recurrent models
pp. 31755–31766, 2023, doi: 10.1109/ACCESS.2023.3257045. of visual attention,’’ 2014, arXiv:1406.6247.
[19] W. He, Y. Han, W. Ming, J. Du, Y. Liu, Y. Yang, L. Wang, Y. Wang, [41] D. Bahdanau, K. Cho, and Y. Bengio, ‘‘Neural machine translation by
Z. Jiang, C. Cao, and J. Yuan, ‘‘Progress of machine vision in the jointly learning to align and translate,’’ 2014, arXiv:1409.0473.
detection of cancer cells in histopathology,’’ IEEE Access, vol. 10, [42] M.-T. Luong, H. Pham, and C. D. Manning, ‘‘Effective approaches to
pp. 46753–46771, 2022, doi: 10.1109/ACCESS.2022.3161575. attention-based neural machine translation,’’ 2015, arXiv:1508.04025.
[20] A. Kamilaris and F. X. Prenafeta-Boldu, ‘‘Deep learning in agriculture: [43] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
A survey,’’ Comput. Electron. Agricult., vol. 147, pp. 70–90, Apr. 2018, L. Kaiser, and I. Polosukhin, ‘‘Attention is all you need,’’ 2017,
doi: 10.1016/j.compag.2018.02.016. arXiv:1706.03762.
[21] D. Shen, S. Zhang, W. Ming, W. He, G. Zhang, and Z. Xie, ‘‘Development [44] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu, ‘‘Squeeze-and-excitation
of a new machine vision algorithm to estimate potato’s shape and size networks,’’ 2017, arXiv:1709.01507.
based on support vector machine,’’ J. Food Process Eng., vol. 45, no. 3,
[45] S. Woo, J. Park, J.-Y. Lee, and I. So Kweon, ‘‘CBAM: Convolutional
Mar. 2022, Art. no. e13974, doi: 10.1111/jfpe.13974.
block attention module,’’ 2018, arXiv:1807.06521.
[22] Y.-Y. Zheng, J.-L. Kong, X.-B. Jin, X.-Y. Wang, and M. Zuo, ‘‘CropDeep:
[46] Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, ‘‘ECA-Net: Effi-
The crop vision dataset for deep-learning-based classification and detec-
cient channel attention for deep convolutional neural networks,’’ 2019,
tion in precision agriculture,’’ Sensors, vol. 19, no. 5, p. 1058, Mar. 2019,
arXiv:1910.03151.
doi: 10.3390/s19051058.
[47] Q. Hou, D. Zhou, and J. Feng, ‘‘Coordinate attention for efficient mobile
[23] V. S. Dhaka, S. V. Meena, G. Rani, D. Sinwar, K. Kavita, M. F. Ijaz, and
network design,’’ 2021, arXiv:2103.02907.
M. Wozniak, ‘‘A survey of deep convolutional neural networks applied
for prediction of plant leaf diseases,’’ Sensors, vol. 21, no. 14, p. 4749, [48] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, ‘‘BERT: Pre-training
Jul. 2021, doi: 10.3390/s21144749. of deep bidirectional transformers for language understanding,’’ 2018,
arXiv:1810.04805.
[24] L. Zhang, L. Zhang, and B. Du, ‘‘Deep learning for remote sens-
ing data: A technical tutorial on the state of the art,’’ IEEE Geosci. [49] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou,
Remote Sens. Mag., vol. 4, no. 2, pp. 22–40, Jun. 2016, doi: ‘‘Training data-efficient image transformers & distillation through atten-
10.1109/MGRS.2016.2540798. tion,’’ 2020, arXiv:2012.12877.
[25] W. Pei, Z. Shi, and K. Gong, ‘‘Small target detection with remote sensing [50] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai,
images based on an improved YOLOv5 algorithm,’’ Frontiers Neuro- T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly,
robotics, vol. 16, Feb. 2023, Art. no. 1074862. J. Uszkoreit, and N. Houlsby, ‘‘An image is worth 16×16 words: Trans-
[26] H. Tian, X. Fang, Y. Lan, C. Ma, H. Huang, X. Lu, D. Zhao, H. Liu, and formers for image recognition at scale,’’ 2020, arXiv:2010.11929.
Y. Zhang, ‘‘Extraction of citrus trees from UAV remote sensing imagery [51] P. Shaw, J. Uszkoreit, and A. Vaswani, ‘‘Self-attention with relative
using YOLOv5s and coordinate transformation,’’ Remote Sens., vol. 14, position representations,’’ 2018, arXiv:1803.02155.
no. 17, p. 4208, Aug. 2022, doi: 10.3390/rs14174208. [52] P. Ramachandran, N. Parmar, A. Vaswani, I. Bello, A. Levskaya,
[27] W. Ming, P. Sun, Z. Zhang, W. Qiu, J. Du, X. Li, Y. Zhang, G. Zhang, and J. Shlens, ‘‘Stand-alone self-attention in vision models,’’
K. Liu, Y. Wang, and X. Guo, ‘‘A systematic review of machine learning in Proc. Adv. Neural Inf. Process. Syst. Curran Associates,
methods applied to fuel cells in performance evaluation, durability pre- 2019., pp. 1–13. [Online]. Available: https://proceedings.neurips.
diction, and application monitoring,’’ Int. J. Hydrogen Energy, vol. 48, cc/paper/2019/hash/3416a75f4cea9109507cacd8e2f2aefc-Abstract.html
no. 13, pp. 5197–5228, Feb. 2023, doi: 10.1016/j.ijhydene.2022.10.261. [53] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and
[28] W. He, Z. Li, T. Liu, Z. Liu, X. Guo, J. Du, X. Li, P. Sun, and W. Ming, S. Zagoruyko, ‘‘End-to-end object detection with transformers,’’
‘‘Research progress and application of deep learning in remaining useful May 2020, arXiv:2005.12872.
life, state of health and battery thermal management of lithium bat- [54] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, ‘‘Deformable
teries,’’ J. Energy Storage, vol. 70, Oct. 2023, Art. no. 107868, doi: DETR: Deformable transformers for end-to-end object detection,’’ in
10.1016/j.est.2023.107868. Proc. Int. Conf. Learn. Represent., 2021, pp. 1–16. [Online]. Available:
[29] J. Schmidhuber, ‘‘Deep learning in neural networks: https://xueshu.baidu.com/usercenter/paper/show?paperid=1p7m0060630
An overview,’’ Neural Netw., vol. 61, pp. 85–117, Jan. 2015, doi: 70v10ua310e20pd632729&site=xueshu_se
10.1016/j.neunet.2014.09.003. [55] Z. Liu, H. Hu, Y. Lin, Z. Yao, Z. Xie, Y. Wei, J. Ning, Y. Cao, Z. Zhang,
[30] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521, L. Dong, F. Wei, and B. Guo, ‘‘Swin transformer v2: Scaling up capacity
no. 7553, pp. 436–444, May 2015, doi: 10.1038/nature14539. and resolution,’’ 2021, arXiv:2111.09883.
[31] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, ‘‘Gradient-based learn- [56] I. Bello, B. Zoph, A. Vaswani, J. Shlens, and Q. V. Le, ‘‘Attention
ing applied to document recognition,’’ Proc. IEEE, vol. 86, no. 11, augmented convolutional networks,’’ 2019, arXiv:1904.09925.
pp. 2278–2324, Nov. 1998, doi: 10.1109/5.726791. [57] L. Liu, W. Ouyang, X. Wang, P. Fieguth, J. Chen, X. Liu, and
[32] A. Radford, L. Metz, and S. Chintala, ‘‘Unsupervised representation M. Pietikäinen, ‘‘Deep learning for generic object detection: A survey,’’
learning with deep convolutional generative adversarial networks,’’ 2015, Int. J. Comput. Vis., vol. 128, no. 2, pp. 261–318, Feb. 2020, doi:
arXiv:1511.06434. 10.1007/s11263-019-01247-4.
[33] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification [58] Z.-Q. Zhao, P. Zheng, S.-T. Xu, and X. Wu, ‘‘Object detection with deep
with deep convolutional neural networks,’’ in Adv. Neural Inf. learning: A review,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 30,
Process. Syst. Curran Associates, 2012, pp. 1–9. [Online]. Available: no. 11, pp. 3212–3232, Nov. 2019, doi: 10.1109/TNNLS.2018.2876865.
https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c [59] R. Girshick, J. Donahue, T. Darrell, and J. Malik, ‘‘Rich feature hier-
8436e924a68c45b-Abstract.html archies for accurate object detection and semantic segmentation,’’ in
[34] X. Glorot, A. Bordes, and Y. Bengio, ‘‘Deep sparse rectifier neural Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Columbus, OH, USA,
networks,’’ in Proc. IEEE IWAENC, vol. 15, Jun. 2011, pp. 315–323. Jun. 2014, pp. 580–587, doi: 10.1109/CVPR.2014.81.
[35] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, ‘‘Learning rep- [60] R. Girshick, ‘‘Fast R-CNN,’’ in Proc. IEEE Int. Conf. Comput.
resentations by back-propagating errors,’’ Nature, vol. 323, no. 6088, Vis. (ICCV), Santiago, Chile, Dec. 2015, pp. 1440–1448, doi:
pp. 533–536, Oct. 1986, doi: 10.1038/323533a0. 10.1109/ICCV.2015.169.

139036 VOLUME 11, 2023

X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

[61] S. Ren, K. He, R. Girshick, and J. Sun, ‘‘Faster R-CNN: Towards [82] L. Wu, L. Zhang, and Q. Zhou, ‘‘Printed circuit board quality
real-time object detection with region proposal networks,’’ IEEE Trans. detection method integrating lightweight network and dual attention
Pattern Anal. Mach. Intell., vol. 39, no. 6, pp. 1137–1149, Jun. 2017, doi: mechanism,’’ IEEE Access, vol. 10, pp. 87617–87629, 2022, doi:
10.1109/TPAMI.2016.2577031. 10.1109/ACCESS.2022.3198994.
[62] K. He, G. Gkioxari, P. Dollár, and R. Girshick, ‘‘Mask R-CNN,’’ in Proc. [83] W. Xuan, G. Jian-She, H. Bo-Jie, W. Zong-Shan, D. Hong-Wei,
IEEE Int. Conf. Comput. Vis. (ICCV), Oct. 2017, pp. 2980–2988, doi: and W. Jie, ‘‘A lightweight modified YOLOX network using coordi-
10.1109/ICCV.2017.322. nate attention mechanism for PCB surface defect detection,’’ IEEE
[63] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and Sensors J., vol. 22, no. 21, pp. 20910–20920, Nov. 2022, doi:
A. C. Berg, ‘‘SSD: Single shot MultiBox detector,’’ in Computer Vision— 10.1109/JSEN.2022.3208580.
ECCV (Lecture Notes in Computer Science), vol. 9905, B. Leibe, [84] Y. Zhao, H. Yang, and H. Feng, ‘‘An improved YOLOv5 PCB defect
J. Matas, N. Sebe, and M. Welling, Eds. Cham, Switzerland: Springer, detection,’’ Proc. SPIE, vol. 12351, pp. 380–387, Nov. 2022, doi:
2016, pp. 21–37, doi: 10.1007/978-3-319-46448-0_2. 10.1117/12.2652341.
[64] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, ‘‘You only look once: [85] S. Liu, D. Huang, and Y. Wang, ‘‘Learning spatial fusion for single-shot
Unified, real-time object detection,’’ in Proc. IEEE Conf. Comput. Vis. object detection,’’ 2019, arXiv:1911.09516.
Pattern Recognit. (CVPR) Las Vegas, NV, USA, Jun. 2016, pp. 779–788, [86] Y. Liu, Z. Shao, and N. Hoffmann, ‘‘Global attention mechanism:
doi: 10.1109/CVPR.2016.91. Retain information to enhance channel-spatial interactions,’’ 2021,
[65] J. Redmon and A. Farhadi, ‘‘YOLO9000: Better, faster, stronger,’’ in arXiv:2112.05561.
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Honolulu, HI, [87] J. Zheng, X. Sun, H. Zhou, C. Tian, and H. Qiang, ‘‘Printed circuit
USA, Jul. 2017, pp. 6517–6525, doi: 10.1109/CVPR.2017.690. boards defect detection method based on improved fully convolu-
[66] J. Redmon and A. Farhadi, ‘‘YOLOv3: An incremental improvement,’’ tional networks,’’ IEEE Access, vol. 10, pp. 109908–109918, 2022, doi:
2018, arXiv:1804.02767. 10.1109/ACCESS.2022.3214306.
[67] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, ‘‘YOLOv4: Optimal [88] J. Lim, J. Lim, V. M. Baskaran, and X. Wang, ‘‘A deep context
speed and accuracy of object detection,’’ 2020, arXiv.2004.10934. learning based PCB defect detection model with anomalous trend alarm-
[68] C. Li, L. Li, H. Jiang, K. Weng, Y. Geng, L. Li, Z. Ke, Q. Li, M. Cheng, ing system,’’ Results Eng., vol. 17, Mar. 2023, Art. no. 100968, doi:
W. Nie, Y. Li, B. Zhang, Y. Liang, L. Zhou, X. Xu, X. Chu, X. Wei, 10.1016/j.rineng.2023.100968.
and X. Wei, ‘‘YOLOv6: A single-stage object detection framework for [89] Z. Yu, Y. Wu, B. Wei, Z. Ding, and F. Luo, ‘‘A lightweight and
industrial applications,’’ 2022, arXiv:2209.02976. efficient model for surface tiny defect detection,’’ Int. J. Speech
[69] C.-Y. Wang, A. Bochkovskiy, and H.-Y. Mark Liao, ‘‘YOLOv7: Trainable Technol., vol. 53, no. 6, pp. 6344–6353, Mar. 2023, doi: 10.1007/
bag-of-freebies sets new state-of-the-art for real-time object detectors,’’ s10489-022-03633-x.
2022, arXiv:2207.02696. [90] Y. T. Li and J. I. Guo, ‘‘A VGG-16 based faster RCNN model for
[70] Z. Zou, K. Chen, Z. Shi, Y. Guo, and J. Ye, ‘‘Object detection in 20 years: PCB error inspection in industrial AOI applications,’’ in Proc. IEEE Int.
A survey,’’ Proc. IEEE, vol. 111, no. 3, pp. 257–276, Mar. 2023, doi: Conf. Consum. Electron.-Taiwan (ICCE-TW), May 2018, pp. 1–2, doi:
10.1109/JPROC.2023.3238524. 10.1109/ICCE-China.2018.8448674.
[71] H. Xin, Z. Chen, and B. Wang, ‘‘PCB electronic component defect [91] A. Shrivastava, A. Gupta, and R. Girshick, ‘‘Training region-based object
detection method based on improved YOLOv4 algorithm,’’ J. Phys., detectors with online hard example mining,’’ in Proc. IEEE Conf. Com-
Conf., vol. 1827, no. 1, Mar. 2021, Art. no. 012167, doi: 10.1088/1742- put. Vis. Pattern Recognit. (CVPR), Las Vegas, NV, USA, Jun. 2016,
6596/1827/1/012167. pp. 761–769, doi: 10.1109/CVPR.2016.89.
[72] D. Misra, ‘‘Mish: A self regularized non-monotonic activation function,’’ [92] B. Hu and J. Wang, ‘‘Detection of PCB surface defects with improved
2019, arXiv:1908.08681. faster-RCNN and feature pyramid network,’’ IEEE Access, vol. 8,
[73] Y. Zhang, F. Xie, L. Huang, J. Shi, J. Yang, and Z. Li, ‘‘A lightweight pp. 108335–108345, 2020, doi: 10.1109/ACCESS.2020.3001349.
one-stage defect detection network for small object based on dual atten- [93] J. Wang, K. Chen, S. Yang, C. C. Loy, and D. Lin, ‘‘Region proposal
tion mechanism and PAFPN,’’ in Proc. AIP. Conf., vol. 9, Oct. 2021, by guided anchoring,’’ in Proc. IEEE/CVF Conf. Comput. Vis. Pattern
Art. no. 708097, doi: 10.3389/fphy.2021.708097. Recognit. (CVPR), Long Beach, CA, USA, Jun. 2019, pp. 2960–2969,
[74] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, ‘‘Focal loss for doi: 10.1109/CVPR.2019.00308.
dense object detection,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 42, [94] D. Li, S. Fu, Q. Zhang, Y. Mo, L. Liu, and C. Xu, ‘‘An improved PCB
no. 2, pp. 318–327, Feb. 2020, doi: 10.1109/TPAMI.2018.2858826. defect detector based on feature pyramid networks,’’ in Proc. 4th Int.
[75] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, Conf. Comput. Sci. Artif. Intell., Zhuhai China, Dec. 2020, pp. 233–239,
‘‘Feature pyramid networks for object detection,’’ in Proc. IEEE Conf. doi: 10.1145/3445815.3445853.
Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 936–944, doi: [95] K. An and Y. Zhang, ‘‘LPViT: A transformer based model for
10.1109/CVPR.2017.106. PCB image classification and defect detection,’’ IEEE Access,
[76] W. Jiang, S. Zhang, and W. Chen, ‘‘Multi-target detection of PCB defects vol. 10, pp. 42542–42553, 2022, doi: 10.1109/ACCESS.2022.
based on improved SSD,’’ Int. Core J. Eng., vol. 8, no. 6, pp. 319–325, 3168861.
2022, doi: 10.6919/ICJE.202206_8(6).0045. [96] W. Chen, Z. Huang, Q. Mu, and Y. Sun, ‘‘PCB defect detec-
[77] J. Li, W. Li, Y. Chen, and J. Gu, ‘‘Research on object detection of tion method based on transformer-YOLO,’’ IEEE Access, vol. 10,
PCB assembly scene based on effective receptive field anchor alloca- pp. 129480–129489, 2022, doi: 10.1109/ACCESS.2022.3228206.
tion,’’ Comput. Intell. Neurosci., vol. 2022, pp. 1–32, Feb. 2022, doi: [97] Y. Yang and H. Kang, ‘‘An enhanced detection method of PCB defect
10.1155/2022/7536711. based on improved YOLOv7,’’ Electronics, vol. 12, no. 9, p. 2120,
[78] W. Luo, Y. Li, R. Urtasun, and R. Zemel, ‘‘Understanding the May 2023, doi: 10.3390/electronics12092120.
effective receptive field in deep convolutional neural networks,’’ [98] W. Wang, L. Yao, L. Chen, B. Lin, D. Cai, X. He, and W. Liu, ‘‘Cross-
in Proc. Adv. Neural Inf. Process. Syst. Curran Associates, Former: A versatile vision transformer hinging on cross-scale attention,’’
2016, pp. 1–9. [Online]. Available: https://proceedings.neurips. 2021, arXiv:2108.00154.
cc/paper/2016/hash/c8067ad1937f728f51288b3eb986afaa- [99] W. Wang, W. Chen, Q. Qiu, L. Chen, B. Wu, B. Lin, X. He, and W. Liu,
Abstract.html ‘‘CrossFormer++: A versatile vision transformer hinging on cross-scale
[79] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, attention,’’ 2023, arXiv:2303.06908.
‘‘DeepLab: Semantic image segmentation with deep convolutional nets, [100] Z. Chen, Y. Duan, W. Wang, J. He, T. Lu, J. Dai, and Y. Qiao, ‘‘Vision
atrous convolution, and fully connected CRFs,’’ IEEE Trans. Pat- transformer adapter for dense predictions,’’ 2022, arXiv:2205.08534.
tern Anal. Mach. Intell., vol. 40, no. 4, pp. 834–848, Apr. 2018, doi: [101] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and
10.1109/TPAMI.2017.2699184. A. Zisserman, ‘‘The PASCAL visual object classes (VOC) challenge,’’
[80] L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, ‘‘Rethinking atrous Int. J. Comput. Vis., vol. 88, no. 2, pp. 303–338, Jun. 2010, doi:
convolution for semantic image segmentation,’’ 2017, arXiv:1706.05587. 10.1007/s11263-009-0275-4.
[81] Y. Wan, L. Gao, X. Li, and Y. Gao, ‘‘Semi-supervised defect detection [102] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese,
method with data-expanding strategy for PCB quality inspection,’’ Sen- ‘‘Generalized intersection over union: A metric and a loss for bounding
sors, vol. 22, no. 20, p. 7971, Oct. 2022, doi: 10.3390/s22207971. box regression,’’ 2019, arXiv:1902.09630.

VOLUME 11, 2023 139037

 X. Chen et al.: Comprehensive Review of Deep Learning-Based PCB Defect Detection

[103] P. Dollar, C. Wojek, B. Schiele, and P. Perona, ‘‘Pedestrian detec- [123] Y. Chen, S. Hu, A. Li, Y. Cao, Y. Zhao, and W. Ming, ‘‘Parameters
tion: An evaluation of the state of the art,’’ IEEE Trans. Pattern optimization of electrical discharge machining process using swarm
Anal. Mach. Intell., vol. 34, no. 4, pp. 743–761, Apr. 2012, doi: intelligence: A review,’’ Metals, vol. 13, no. 5, p. 839, Apr. 2023, doi:
10.1109/TPAMI.2011.155. 10.3390/met13050839.
[104] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, [124] G. Zhang, Z. Chen, Z. Zhang, Y. Huang, W. Ming, and H. Li, ‘‘A macro-
P. Dollár, and C. L. Zitnick, ‘‘Microsoft COCO: Common objects in scopic mechanical model of wire electrode deflection considering tem-
context,’’ in Computer Vision—ECCV (Lecture Notes in Computer Sci- perature increment in MS-WEDM process,’’ Int. J. Mach. Tools Manuf.,
ence), vol. 8693, D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds. vol. 78, pp. 41–53, Mar. 2014, doi: 10.1016/j.ijmachtools.2014.01.004.
Cham, Switzerland: Springer, 2014, pp. 740–755, doi: 10.1007/978-3-
319-10602-1_48.
[105] Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, ‘‘Distance-IoU
loss: Faster and better learning for bounding box regression,’’ 2019, XING CHEN was born in Hubei, in 1979.
arXiv:1911.08287. He received the master’s degree from the
[106] Y.-F. Zhang, W. Ren, Z. Zhang, Z. Jia, L. Wang, and T. Tan, ‘‘Focal Huazhong University of Science and Technology,
and efficient IOU loss for accurate bounding box regression,’’ 2021, in 2008. He is currently a Teacher with Nanyang
arXiv:2101.08158. Normal University. He has published more than
[107] J. He, S. Erfani, X. Ma, J. Bailey, Y. Chi, and X.-S. Hua, ‘‘Alpha- ten papers in academic journals, three papers in EI
IoU: A family of power intersection over union losses for bounding box search, presided over two projects of the Ministry
regression,’’ 2021, arXiv:2110.13675.
of Education, and guided students to win the first
[108] Z. Gevorgyan, ‘‘SIoU loss: More powerful learning for bounding box
prize of Henan Province in their graduation papers.
regression,’’ 2022, arXiv:2205.12740.
His research interests include intelligent control
[109] Z. Liu, Y. Gao, Q. Du, M. Chen, and W. Lv, ‘‘YOLO-extract:
and security.
Improved YOLOv5 for aircraft object detection in remote sens-
ing images,’’ IEEE Access, vol. 11, pp. 1742–1751, 2023, doi:
10.1109/ACCESS.2023.3233964.
[110] G. Cheng and J. Han, ‘‘A survey on object detection in optical YONGLEI WU was born in Henan, China, in 2001.
remote sensing images,’’ ISPRS J. Photogramm. Remote Sens., vol. 117,
He is currently pursuing the master’s degree with
pp. 11–28, Jul. 2016, doi: 10.1016/j.isprsjprs.2016.03.014.
the College of Mechanical and Electrical Engi-
[111] J. Pang, C. Li, J. Shi, Z. Xu, and H. Feng, ‘‘R2-CNN: Fast tiny
neering, Zhengzhou University of Light Industry.
object detection in large-scale remote sensing images,’’ IEEE Trans.
Geosci. Remote Sens., vol. 57, no. 8, pp. 5512–5524, Aug. 2019, doi: His research interests include machine learning
10.1109/TGRS.2019.2899955. and deep learning. He has won the silver medal in
[112] J. Zhang, J. Lei, W. Xie, Z. Fang, Y. Li, and Q. Du, ‘‘SuperY- the CCPC Henan Division, the second prize in the
OLO: Super resolution assisted object detection in multimodal remote Blue Bridge Cup National Finals, and the second
sensing imagery,’’ IEEE Trans. Geosci. Remote Sens., vol. 61, 2023, prize in the Henan Division of the National College
Art. no. 5605415, doi: 10.1109/TGRS.2023.3258666. Mathematics Competition.
[113] K. Fu, Z. Chang, Y. Zhang, G. Xu, K. Zhang, and X. Sun, ‘‘Rotation-
aware and multi-scale convolutional neural network for object detection
in remote sensing images,’’ ISPRS J. Photogramm. Remote Sens.,
vol. 161, pp. 294–308, Mar. 2020, doi: 10.1016/j.isprsjprs.2020.01.025. XINGYOU HE was born in Henan, China, in 1999.
[114] M. Siam, M. Gamal, M. Abdel-Razek, S. Yogamani, M. Jagersand, and He is currently pursuing the master’s degree with
H. Zhang, ‘‘A comparative study of real-time semantic segmentation the College of Mechanical and Electrical Engi-
for autonomous driving,’’ in Proc. IEEE/CVF Conf. Comput. Vis. Pat- neering, Zhengzhou University of Light Industry.
tern Recognit. Workshops (CVPRW), Jun. 2018, pp. 700–70010, doi: He has won the first prize in the final of the
10.1109/CVPRW.2018.00101. 2021 iCAN National Undergraduate Innovation
[115] S. Tang, F. He, X. Huang, and J. Yang, ‘‘Online PCB defect detector on a and Entrepreneurship Competition and the second
new PCB defect dataset,’’ 2019, arXiv:1902.06197. prize in the 2021 ‘‘Internet Plus’’ Undergradu-
[116] W. Huang, P. Wei, M. Zhang, and H. Liu, ‘‘HRIPCB: A challenging ate Innovation and Entrepreneurship Competition
dataset for PCB defects detection and classification,’’ J. Eng., vol. 2020, and the Seventh China National Standard ‘‘Inter-
no. 13, pp. 303–309, Jul. 2020, doi: 10.1049/joe.2019.1183. net Plus’’ Undergraduate Innovation and Entrepreneurship Competition in
[117] V. U. Sankar, G. Lakshmi, and Y. S. Sankar, ‘‘A review of various defects Henan.
in PCB,’’ J. Electron. Test., vol. 38, no. 5, pp. 481–491, Oct. 2022, doi:
10.1007/s10836-022-06026-7.
[118] X. Zhao, Y. Zhao, S. Hu, H. Wang, Y. Zhang, and W. Ming, ‘‘Progress
in active infrared imaging for defect detection in the renewable and WUYI MING was born in Hubei, China, in 1981.
electronic industries,’’ Sensors, vol. 23, no. 21, p. 8780, Oct. 2023, doi: He received the bachelor’s degree in engi-
10.3390/s23218780. neering from Huazhong Agricultural University,
[119] W. Ming, Z. Xie, J. Ma, J. Du, G. Zhang, C. Cao, and Y. Zhang, in 2002, the master’s degree in engineering from
‘‘Critical review on sustainable techniques in electrical discharge machin- Zhengzhou University, in 2010, and the Ph.D.
ing,’’ J. Manuf. Processes, vol. 72, pp. 375–399, Dec. 2021, doi:
degree in engineering from the Huazhong Univer-
10.1016/j.jmapro.2021.10.035.
sity of Science and Technology, in 2014.
[120] W. Ming, S. Zhang, G. Zhang, J. Du, J. Ma, W. He, C. Cao, and
He has done the theoretical and applied research
K. Liu, ‘‘Progress in modeling of electrical discharge machining pro-
cess,’’ Int. J. Heat Mass Transf., vol. 187, May 2022, Art. no. 122563,
in EDM and deep learning. He is currently with the
doi: 10.1016/j.ijheatmasstransfer.2022.122563. Guangdong Provincial Key Laboratory of Manu-
[121] Z. Zhang, W. Qiu, G. Zhang, D. Liu, and P. Wang, ‘‘Progress in facturing Equipment Digitization. He has published nearly 100 SCI indexed
applications of shockwave induced by short pulsed laser on surface articles (more than 1700 citations, H27) in journals, such as Journal of
processing,’’ Opt. Laser Technol., vol. 157, Jan. 2023, Art. no. 108760, Cleaner Production, Applied Surface Science, Journal of Materials Process-
doi: 10.1016/j.optlastec.2022.108760. ing Technology, Measurement, and IEEE ACCESS. His main research interests
[122] P. Wang, Z. Zhang, B. Hao, S. Wei, Y. Huang, and G. Zhang, ‘‘Inves- include digital technology and equipment for difficult-to-process materials.
tigation on heat transfer and ablation mechanism of CFRP by different He won the Second Prize of the Henan (China) Science and Technology
laser scanning directions,’’ Compos. B, Eng., vol. 262, Aug. 2023, Progress Award, in 2018.
Art. no. 110827, doi: 10.1016/j.compositesb.2023.110827.

139038 VOLUME 11, 2023