REVIEW ARTICLE

Published Online: September 14, 2024

https://doi.org/10.31893/multirev.2025011

A deep learning approach for automated PCB defect

detection: A comprehensive review
Tay Sui Jiana | Muhammad Hafidz Fazli Md Fauadia | Saifudin Hafiz Yahayaa |
Ahamad Zaki Mohamed Noorb | Adi Saptaric
aFakulti
Teknologi dan Kejuruteraan Industri dan Pembuatan, Universiti Teknikal Malaysia Melaka, 76100 Durian Tunggal, Malaysia.
bManufacturing Section, Universiti Kuala Lumpur, Malaysian Spanish Institute Kulim Hi-Tech Park, 09000 Kulim, Kedah, Malaysia.
cDepartment of Industrial Engineering, President University, Kota Jababeka, Cikarang Baru, Bekasi 17550, Indonesia.

Abstract The ever-expanding market for electronic devices has significantly heightened the demand for high-quality printed
circuit boards (PCBs). Even minor defects in PCBs can pose substantial safety risks for end-users. This article provides a
comprehensive review on deep learning-based approaches for PCB defect detection. Our exploration covers various critical
aspects, including the classification of PCB defects, automated vision inspection (AVI) techniques, object detection
methodologies, and the widespread adoption of deep learning models. Specifically, we focus on the state-of-the-art
approach known as region-based Fully Convolutional Network with Feature Pyramid Networks (FPN-RFCN). Additionally,
we discuss effective data augmentation techniques and commonly used evaluation metrics in this domain. This review
provides valuable insights for researchers, practitioners, and industry professionals engaged in PCB quality assurance.
Keywords: deep learning, feature pyramid network, PCB, defect detection, image processing, smart manufacturing

1. Introduction
The widespread adoption of consumer electronics, including smartphones, laptops, smart televisions, and tablets, has
driven the demand for high-quality printed circuit boards (PCBs) produced in large quantities. However, maintaining high PCB
quality at such scales presents substantial challenges (Chung et al., 2023; Fung et al., 2024). As fundamental components in
electronic devices, PCBs must exhibit robust stability, high resistance to interference, and advanced features such as enhanced
integration, compactness, and high-speed transmission capabilities (Chen et al., 2023).
Effective PCB quality control is crucial. Even minor defects can lead to significant issues, including reduced manufacturing
quality due to undetected flaws, increased production costs from reworking or scrapping faulty products, delays in time to
market while allocating extra time for quality assurance, and potential safety hazards such as short circuits, component failure,
or overheating that pose risks to users (Fung et al., 2024; Łyczek & Skarka, 2024; Reichenstein et al., 2022; Rosli et al., 2018).
The advent of Industry 4.0 (IR 4.0) and the rapid growth of AI have prompted the integration of deep learning techniques
into automated optical inspection (AOI) systems across various industries. These advancements have significantly improved
the accuracy and speed of detecting various PCB defects (Tai et al., 2021). However, achieving zero-defect production remains
a formidable challenge for modern electronics manufacturers (Saadallah et al., 2022; Rosli et al., 2016).
This study focuses on identifying the most critical defects in PCBs. Additionally, this article reviews the deep learning
approaches used for defect detection in printed circuit boards (PCBs). Consequently, a generic deep learning-based approach
for detecting critical defects is proposed.
2. Types of PCB Deficits
Any anomalies or irregularities in the PCB structure or the assembled components are considered PCB defects. These
imperfections can cause electronic devices to malfunction, shorten their lifespan, or reduce their reliability. Schubeck et al.
(2021) classified PCB defects into four types: trace, component, solder, and via and pad. Each type of defect is further classified
into subclasses, with each subclass consisting of several related defects. In addition, a study by Sankar et al. (2022) examined
various defects found in PCBs, collecting data from the PCB fabrication industry from April 2017 until July 2020.
Currently, PCB defects can be classified as bare PCB defects or component-mounted PCB defects, which include through-
hole defects and surface mount device (SMD) defects. A bare PCB is a circuit board without any electronic components
mounted on its surface, whereas a through-hole PCB is full of holes where components are inserted and soldered. The SMD
refers to components mounted on the surface of a PCB. In their papers, both Schubeck et al. (2021) and Sankar et al. (2022)
focused only on component-mounted PCB defects. On the basis of an exhaustive review, the types of PCB defects classified are
comprehensively illustrated in Figure 1.

Multidiscip. Rev. (2025) 8:e2025011 Received: July 8, 2024 | Accepted: August 11, 2024
Jian et al. (2025) 2

Figure 1 Taxonomy for PCB defect classification.

3. Defect Inspection Via the Vision System

3.1. Automated vision inspection
Automated vision inspection (AVI), also known as automatic optical inspection, represents a significant advancement
over traditional human vision inspection methods (Silva et al., 2019). Techniques that are utilized for AVI vary depending on
the specific requirements and characteristics of the inspection task. These include:

• Object detection: Identifying all objects in an image and pinpointing their locations and scales through bounding boxes
(Ge et al., 2020).
• Image classification: The core of image classification research always revolves around feature extraction, the process of
categorizing entire images into predefined classes or categories on the basis of their visual characteristics. This process
proves valuable in scenarios where the goal is to determine whether an entire image meets certain criteria or belongs
to a particular class (Chen et al., 2021).
• Image segmentation: A fundamental and pivotal step. This technique divides an image into distinct regions on the basis
of specific characteristics and extracts the target of interest (Wang et al., 2020).
• Feature extraction: This method simplifies data complexity by creating a straightforward representation for each
variable within a feature space, which is often achieved through linear combinations of the original input variables.
Feature extraction typically involves transforming the original data into features with enhanced pattern recognition
capabilities, in contrast to the original data, which may possess weaker recognition capabilities as features (Arif et al.,
2022).

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 3

• Template matching: Scans the image across a dense array of subwindows and compares each one with a template
representing the object (Mercier et al., 2021). This method proves valuable for detecting objects or patterns that closely
resemble the template.
• Optical character recognition (OCR): Additionally, text detection and recognition involve extracting information from
input images and translating it into digital text. This process involves transmitting information from the image domain
to the text domain (Wang et al., 2021).
• Anomaly detection: Additionally, outlier detection or novel detection involves identifying data instances that deviate
significantly from the majority of data points (Pang et al., 2021).
• Motion analysis: Relies on data from motion capture systems. There are three types of motion capture systems: marker-
based systems, vision-based systems, and volumetric capture systems. Marker-based systems can generate highly
accurate motion tracking data, but attaching markers to human bodies is inconvenient and time-consuming. Vision-
based motion capture systems use computer vision to identify motion from images or videos. Volumetric motion
capture systems go a step further than vision-based systems by generating textured and animated 3D character models
instead of skeleton models (Zhu et al., 2022).
• 3D vision involves capturing and analysing three-dimensional information from images or depth sensors. These
techniques are beneficial for tasks such as measuring object dimensions, detecting surface defects, or performing
volumetric inspections (Linton et al., 2023).
3.2. Object detection technique
A core task in computer vision involves simultaneously classifying and pinpointing potential objects within an image (Li
et al., 2023). The utilization of object detection techniques by the public has shown a consistent upwards trend from 1998--
2021, as illustrated in Figure 2 (Zou et al., 2023).

(a) (b)
Figure 2 Number of publications in (a) object detection from 2010--2023. (Data from Google scholar advanced search: allintitle: “object
detection”); and (b) defect detection in PCB from 2010--2023. (Data from Google scholar advanced search: allintitle: “defect detection PCB”).

Several methods have been employed for object detection:

• Traditional computer vision approach: Two fundamental techniques within this domain are edge detection and the
Hough transform. Edge detection plays a critical role in image processing by identifying discontinuities in intensity, which
often correspond to object boundaries. In edge detection, the input image is passed through a filter to detect lines or
boundaries of shapes and then extract and highlight the edge lines within the image (Dong et al., 2022). The Hough
transform (HT) is a representative technique for line detection in digital images. Owing to its simplicity and efficiency,
the HT can be extended to identify other regular shapes, such as circles and rectangles. The main principle of the HT
involves voting for evidence from the image domain to the parameter domain, followed by the detection of shapes in
the parametric domain by identifying local-maximal responses. This technique is sensitive to lighting variations and
occlusions, resulting in noisy outputs containing irrelevant lines (Zhao et al., 2022).
• Deep learning-based approaches: There are two powerful deep learning approaches for detecting objects in images:
convolutional neural networks (CNNs) and transformer models, as illustrated in Figure 3 (Sun et al., 2024).
• Traditional machine learning approaches: Support vector machines (SVMs), decision trees, random forests, and gradient
boosting machines (GBMs) are several machine learning algorithms that can be used for object detection. SVM is a
powerful machine learning technique used for classification tasks. It is designed to identify the optimal hyperplane
within the feature space that maximizes the separation between two sets of training samples (Mohan et al., 2020). A
random forest classifier consists of numerous trees, each grown via a form of randomization. The terminal nodes, which

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 4

are the leaf nodes, are assigned labels representing estimations of the posterior probability distribution across the image
classes. Every internal node within the decision tree structure incorporates a test function. This test function serves to
optimally partition the data space on the basis of the feature characteristics of the data points. To classify an image, it
performs classification for each tree, and the resulting leaf distributions are aggregated (Boateng et al., 2020). The
gradient boosting machine (GBM) constitutes a powerful technique in machine learning. It works by combining multiple
weaker learners to create a powerful ensemble with superior predictive ability. GBMs are successful in various
prediction tasks, including spam filtering, online advertising, fraud detection, anomaly detection, and computational
physics (Lu & Mazumder, 2020).
• Feature-based approach: A traditional approach in computer vision that relies on extracting informative features from
images to identify and locate objects. Two common feature-based object detection techniques include the histogram
of oriented gradients (HOG) and the Haar cascade. The core principle of the HOG algorithm lies in its computation of
gradients as local descriptors within an image. These gradients are subsequently normalized locally, resulting in features
that are invariant to the object location and robust to illumination changes within the image (Ghaffari et al., 2020). The
core principle of Haar cascade revolves around the utilization of Haar-like features, which are simple contrast-based
descriptors employed for object identification within an image. These features are subsequently leveraged to train a
cascaded structure of classifiers. The Haar cascade approach involves training a cascade of classifiers, which are
progressively more complex. This cascaded structure allows for the efficient discarding of nonrelevant image regions
during the detection process. While this approach offers significant advantages in terms of computational efficiency, it
can be limited in its ability to handle variations in object appearance, such as orientation, scale, and lighting conditions
(Georgiev et al., 2023).
• Hybrid approach: This approach involves combining multiple techniques to leverage their respective strengths. These
approaches are characterized by the strategic integration of diverse methodologies, encompassing both established CV
techniques and deep learning methods. This combination offers a distinct advantage in capitalizing on the
heterogeneous computing capabilities available on edge devices. Heterogeneous computing refers to the ability to
utilize a variety of processing units, such as CPUs and GPUs, within a single device (O’Mahony et al., 2020).

Figure 3 Classification of deep learning-based object detection. Source: Sun et al. (2024).

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 5

3.3. Application of object detection in PCBs

Zhou et al. (2021) proposed a multiview template matching method for PCB component detection, as opposed to
manual inspection. The experiment compared multiview and single-view approaches and demonstrated a significant
improvement in performance with the multiview approach. Bahrebar et al. (2022) proposed the use of machine learning to
predict PCB corrosion conditions. Predictive models are employed to forecast the levels of electrochemical migration (ECM)
and leakage current (LC) within electronic systems subjected to corrosive environments. Supervised machine learning
approaches are utilized, with predictions from the ECM and LC models guiding the training process. Commonly used algorithms
in this domain include k-nearest neighbors (k-NNs), decision trees (DTs), random forests (RFs), and support vector machines
(SVMs). As a result, the SVM and RF algorithms achieve the highest scores in terms of the F1 score, AUC, accuracy, sensitivity,
and precision for PCB corrosion prediction under diverse critical factor combinations.
Piliposyan and Khursheed (2022) proposed a hybrid object detection method that combines traditional computer vision
with the Siamese neural network (SNN) for detecting hardware Trojan components on PCBs. An SNN is a combination of a CNN
and a feedforward fully connected neural network. A hardware Trojan is an extra component implanted on a PCB. A private
PCB dataset was used for training the model. The dataset was categorized into three types: small (surface area within 4 to 9
mm²), medium (surface area within 15 to 50 mm²), and large (surface area 280 mm² and above). The proposed method
achieved an overall accuracy of 95.6%.
Lou et al. (2023) proposed a modification of the Vision Transformer (ViT) for multiscale feature extraction. The method
uses a multistage ViT architecture, with each stage focusing on capturing features at a specific scale. Subsequently, 1x1
convolution layers are employed between adjacent stages. These convolution layers act as connectors, effectively increasing
the dimensionality of the extracted features by combining information from different scales. For the final object detection
algorithm, the detection head of YOLOx was adopted. The model was tested on three datasets and achieved a mAP of 69.3%
for the MSCOCO 2017 dataset, 86.0% for the PASCAL VOC 2007 dataset, and 99.9% for a private PCB ultrasonic image dataset.
Zhang et al. (2024) proposed LDD-Net, a lightweight network designed for detecting defects on PCBs. LDD-Net comprises
three components: lightweight feature extraction (LFE), a multiscale aggregation network (MAN), and LD-Head. The LFE is used
for extracting features from images, the MAN is used for fusing multiscale feature maps, and the LD-Head acts as the detector,
performing object classification and regression for defect detection. The dataset used is the Test PCB dataset, which is an
enhanced version of the public synthesis PCB dataset. The model achieved a mAP of 95.9%.
Fung et al. (2024) proposed the SF-PSPyramid for the Faster R-CNN for PCB defect detection. The structure includes a
bottom-up pathway, composed of C1 to C5, and a top-down pathway, composed of P1 to P6. The model achieved 98.6% AP50
for the Deep PCB dataset, whereas for the public synthesis PCB dataset, the model achieved 99.3% AP50. The proposed
method, object detection approach, and contribution of the study are summarized in Table 1.
Table 1 Summary of object detection applications in PCB defect detection.
Authors Proposed Method Object Detection Approach Contribution
Has great detection on PCB
Zhou et al. (2021) Multiview template matching Template matching approach
component
SVM and RF algorithms achieved
Bahrebar et al. (2022) Machine Learning Machine learning approach highest score for corrosive condition
prediction
High accuracy of 95.6% for detecting
Piliposyan & Khursheed. Combination of traditional CV
Hybrid approach three type of surface area for
(2022) and SNN
hardware trojan
Vision Transformer (ViT) for A high mAP of 99.9% for private PCB
Lou et al. (2023) Deep learning approach
multiscale ultrasonic image dataset
Perform well in detecting PCB defect
Zhang et al. (2024) LDD-Net Deep learning approach
which achieved 95.9% of mAP
Perform well on both public synthesis
Fung et al. (2024) SF-PSPyramid Deep learning approach dataset (99.3% AP@50) and Deep PCB
dataset (98.6% AP @50).

3.4. Deep learning technique for PCB vision inspection

Deep learning is a technique rooted in neural networks and inspired by the structure of the human brain. A neural
network typically comprises three layers: an input layer, a hidden layer, and an output layer. In the input layer, artificial
neurons, also known as nodes or units, function as feature detectors, representing specific data. This information is then
transmitted to the hidden layer where computations occur. Each neuron in the hidden layer receives weighted inputs from the

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 6

previous layer and applies an activation function to compute an output. The final layer, the output layer, generates the desired
output. If the output does not match expectations, the weights are adjusted iteratively until the desired output is achieved
(Robert, 2024).
Deep learning involves two primary stages: training and testing. During the training stage, a substantial amount of data
undergoes a labelling process, which defines identifiable features. Through a comparison of these features, the model learns
to identify patterns, enabling accurate predictions and decisions when encountering similar data in the future. In the testing
stage, the model applies its acquired knowledge to analyse new data and make predictions (Patel and Thakkar, 2020).
A convolutional neural network (CNN) is a specialized type of neural network designed for analysing visual data,
specifically images. Compared with traditional neural networks, CNNs are more intricate and incorporate filters, kernels, and
pooling layers. In the CNN architecture, the input is an image, and the second layer involves a convolutional layer where filters
and kernels are applied to extract important features from the input image. The convolutional layer produces outputs known
as feature maps, which are then passed to the pooling layer. In the pooling layer, feature maps undergo downsampling to
reduce computational complexity and extract crucial features. Typically, a CNN includes multiple convolutional layers and a
pooling layer. Fully connected layers serve as activation layers for prediction and generating outputs, as illustrated in Figure 4.
The adoption of CNNs in PCB vision inspection has significantly enhanced performance, offering benefits such as reduced
computational time, robust handling of variable and complex visual data, and improved detection accuracy (Robert, 2024).

Figure 4 Convolutional neural network architecture. Source: Li et al. (2022)

Ding et al. (2019) proposed a tiny defect detection network (TDD-Net) specifically designed to identify tiny defects on
PCB surfaces. While it builds upon the Faster R-CNN architecture, TDD-Net incorporates three significant modifications to
Faster R-CNN by implementing a data augmentation technique, designing reasonable anchors, and a multiscale feature fusion
strategy for the model. The experimental results indicate a 98.80% mAP. Hu & Wang proposed a combination of a feature
pyramid network (FPN) with Faster R-CNN and a guided anchor-based region proposal network (GARPN). FPN with Faster R-
CNN as the backbone for defect detection. GARPN uses a multiscale RPN rather than a standard RPN. The experimental results
indicate a 95.6% mAP.
Xia et al. (2021) proposed a combination of parallel high-definition feature extraction (PHFE), focal loss (FL) and a region-
based fully convolutional network (RFCN) for inspecting PCB minor defects. The experimental results indicate a 97.3% mAP,
which is 12.3 higher and 6.7 higher than those of YOLOv3 and Faster R-CNN, respectively. Lan et al. proposed an improvement
to YOLOv3. There are three improvements to YOLOv3. First, a batch normalization (BN) layer is combined with a convolutional
layer. Second, the GIoU loss function is implemented rather than the standard loss function, which is the mean square error
(MSE). Finally, the K-means++ clustering algorithm is used. The experimental results indicate a precision of 92.98% mAP
compared with SSD (84.67% MAP) but slightly less precision compared with Faster R-CNN (94.72% mAP).
Li et al. (2022) proposed an extended FPN model for hard sample PCB detection. Moreover, the focal loss function is
implemented in the model to reduce the loss of samples and increase the weight of hard samples. This approach achieved a
96.2% mAP. Liu et al. proposed Gaussian intersection of union (GsIoU) loss with YOLOv4. The GSIoU addresses the problem of
having multiple bounding boxes predicted for a single object. It addresses this by using a Gaussian function to correlate these
predictions. Boxes predicted at identical positions are treated collectively. The experimental results revealed an 86.9% AP.
Moreover, Zeng et al. (2022) proposed ABFPN for defect detection. An atrous spatial pyramid pooling (ASPP)-balanced
FPN (ABFPN), a novel approach for detecting small objects, combines the strengths of three different techniques for small

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 7

object detection. ABFPN is flexible and can be utilized in existing object detection systems. Two main parts are used to combine
image features: the skip-ASPP module and the balanced module. The proposed improved PCB defect detection (IPDD)
framework consists of four parts: an input layer, an enhanced ResNeXt-152 as the backbone, ABFPN as the neck part and a
cascaded RCNN as the detection head. The experimental results indicate an 85.59% mAP compared with Faster R-CNN (73.2%
mAP) and a slight improvement compared with RefineDet (83.8% mAP).
Moreover, Yu et al. (2022) proposed ES-Net for tiny defect detection. ES-Net consists of three parts: the backbone, neck,
and head. The system leverages the CSPDarknet53 network as its backbone for feature extraction, while the aggregated feature
guidance module (AFGM) is the neck. The DSH (dynamic scale-aware head) serves as the final module of the network. It utilizes
the fused features from the ESP module, which incorporates information from all scales. This allows DSH to perform multiscale
object prediction, considering details from both coarse- and fine-grained feature maps. DSH acts as the prediction head of the
network. According to the experimental results on the PCB defect dataset, ES-Net with CSPDarknet53 achieves a 97.5% mAP.
Chen et al. (2023) proposed a simple CNN model to classify normal and component defects. The model mainly adopts
the method of ensemble learning. This model combines the strengths of multiple mini models. Instead of using one model, it
leverages the diverse features extracted by two submodules. The probabilities are used for these initial outputs to generate
the fourth output. The experimental results for the proposed CNN model achieve an accuracy of 96.96%.
Ling et al. (2023) proposed an enhancement to the YOLOv8 object detection model. The primary improvements focus
on the backbone architecture, which incorporates ghost convolution and C2 focal modules. C2 focal modules address the
challenge of detecting densely packed objects across various scales. Traditionally, this is achieved by generating and merging
feature maps of different sizes. The proposed approach implements FocalNeXt blocks as a substitute for the original C2f module
within YOLOv8. By introducing the C2Focal module, the model achieves competitive performance while maintaining minimal
computational costs. In addition, a new Sig-IoU loss function for bounding box regression has been introduced to replace the
original loss function for YOLOv8. The experimental results indicate that 87.7% mAP is achieved. The deep learning technique
that has been proposed for detecting PCB defects is summarized in Table 2.
Table 2 Summary of deep learning techniques for PCB defect detection.
Authors Proposed Approach Contribution Limitations
Achieved high mAP of 98.8% on • Higher training time
Ding et al. (2019) TDD-Net
detecting six types of defects • Expensive computational cost
• Unable have a great performance when
The GARPN has perform well
Faster R-CNN with FPN detecting different PCB dataset, even
Hu & Wang (2020) compared to standard RPN for Faster
and GARPN after adding the training dataset to
R-CNN
model
Success to detect and achieve high
Lan et al. (2021) Improvement YOLOv3 precision and improve in detection • Lower mAP compared to Faster R-CNN
rate for one stage object detection
• The model It is not clear how the model
The data enhancement has improved
arrives at its defect classifications, which
the accuracy of the detection and
Xia et al. (2021) PHFE and FL-RFCN could be a concern in safety-critical
focal loss function has perform well
applications.
in deal with imbalance dataset
• No real-world applications
• Even the mAP achieved high accuracy of
Has successful detect for tiny defect
Yu et al. (2022) ES-Net 76.2% but still not suitable for real
detection
industry application
• Not suitable for real-world industrial
Has a better performance compared
environments
Li et al. (2022) Extended FPN to original FPN and achieved high
• Model performance may affect by light
precision for hard samples
and noise
Has the ability to solve the problem
Liu et al. (2022) Gaussian-IoU loss deal with multiple bounding boxes • Not real-time yet
for single object condition
• Model exhibits low performance on
Has achieved high 98.8% AP at 50 for
Zeng et al., 2022 ABFPN COCO 2017 dataset, and VisDrone 2019
PCB defect detection
dataset
CNN model based on Success in achieved high accuracy
• Only detected small region rather than
Chen et al. (2023) depth wise separable detection and able to replace the
whole image
convolution traditional CNN
High performance for 87.7% of mAP
at threshold 0.5 and high inference • Unable to deal with imbalanced PCB
Ling et al. (2023) Improvement YOLOv8
speed has the possibility for real components
application

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 8

4. FPN-RFCN: A State-of-Art Approach

FPN generates feature pyramids with robust semantic information at various scales to obtain useful details on small
objects without increasing the amount of calculation and occupied memory. It uses a CNN to generate a set of hierarchical
features that encode semantic information at different scales in the pyramid. The different levels of features in the hierarchical
pyramid represent the objects in the image and their contextual information from different perspectives (Hu and Wang, 2020).
The region-based fully convolutional network (R-FCN) was proposed by Dai et al. (2016) and is composed of 4 main
parts, including the CNN backbone, region proposal network (RPN), classification, and regression, as illustrated in Figure 5. The
main contribution of this network is position-sensitive score maps. The R-FCN architecture employs a modified version of the
ResNet-101 architecture as its backbone. Any CNN can be used as the backbone for the RFCN. This modification entails the
removal of the average pooling layer and fully connected layer and uses only the convolutional layer of ResNet101 to generate
the feature maps. Since the last convolutional block in ResNet101 is 2048-dimensional, a randomly initialized 1x1 convolutional
layer with 1024 dimensions was attached to reduce the dimension. To capture spatial information within each region of interest
(ROI), the RFCN architecture divides each ROI into a k × k grid, creating k² individual bins. Each bin in this grid corresponds to a
specific position-sensitive score map generated by the final convolutional layer. This layer uses a k²(C+1)-channel convolution
to account for both the object category (C) and background (+1) (Du & Wang, 2021).

Figure 5 Architecture of the RFCN. Source: Dai et al. (2016).

The region proposal network (RPN) is a fully convolutional network that operates on a feature map derived from a CNN.
It employs a set of predefined boxes called anchors to generate k anchor boxes of varying sizes and aspect ratios positioned at
different locations across the feature map. These anchor boxes are then processed through two fully connected layers, a
classification layer and a box regression layer, as depicted in Figure 6 (Zhang et al., 2023). In the RFCN architecture, the RPN
identifies candidate regions of interest (RoIs) that potentially correspond to object locations. This efficiency stems from the
use of shared feature representations extracted from the image by both the RPN and R-FCN. The R-FCN architecture
subsequently utilizes these proposed RoIs in conjunction with the shared features to generate position-sensitive score maps
for classification.

Figure 6 RPN representation. Source: Zhang et al. (2023)

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 9

4.1. Data augmentation techniques

Recent advancements in artificial intelligence research have led to significant developments in novel network
architectures and the exploitation of increasingly powerful graphics processing unit (GPU) computational capabilities.
However, the exploration of data augmentation techniques remains an area that warrants more attention. Data augmentation
techniques enhance the diversity and adequacy of training datasets by generating synthetic data. This augmented data can be
seen as being sampled from a distribution closely resembling real-world data, thereby enabling the dataset to capture a more
comprehensive range of characteristics. These methods are versatile and applicable across various computer vision tasks,
including semantic segmentation, object detection, and image classification. Image data augmentation broadly encompasses
two approaches: basic and advanced techniques. The methodologies for both basic and advanced approaches are illustrated
in Figure 7, as outlined by Yang et al. (2022).

Figure 7 Taxonomy of data augmentation methods. Source: Yang et al. (2022)

Furthermore, deep transfer learning (DTL) presents a solution to the problem of limited data, which is a significant
hurdle for many deep learning applications. DTL uses knowledge gained from source tasks and datasets to accelerate learning
on different but related target tasks. Unlike standard deep learning models initialized with random weights, DTL utilizes
pretrained weights from models trained on the source task. This pretraining serves as a foundation, enabling the model to
learn the target task more efficiently and achieve superior performance. DTL allows source and target data to have different
distributions. On the other hand, DTL involves the transfer of preacquired knowledge from the source task to the target task,
and these tasks are not necessarily interrelated or learned simultaneously (Iman et al., 2023). A recent study by Zhuang et al.
(2021) proposed four main categories for DTL approaches:
i. Innovation-based approach: This approach focuses on feature selection within source data instances. It applies
differential weighting to these features, optimizing the relevance for use with the target data.
ii. Feature-based/mapping-based: This approach aims to make the source and target data more similar by transforming
the features into a more homogeneous format. The study has further divided this category into two subcategories:
a. Asymmetric: This method focuses on aligning the source features with the target features.
b. Symmetric: This method focuses on identifying a fundamental set of features that are relevant to both the source
and target data and transforms both sets of features into new spaces.
iii. Parameter-based/Model-based: This approach prioritizes the reuse of knowledge encoded within a pretrained model.
This is achieved by implementing various combinations of freezing, fine-tuning, or adding new layers to the pretrained
network architecture.
iv. Relational-based/adversarial-based: This approach focuses on the extraction of transferable features that are applicable
to both the source and target data. This is achieved through two primary methods: leveraging the established logical
relationships or rules gleaned from the source domain and employing techniques inspired by the architecture and
functionality of generative adversarial networks (GANs).
Deep transfer learning (DTL) primarily uses model-based approaches to address the challenge of domain adaptation.
Since source and target data often have different distributions, these approaches adjust the network architecture, also known
as the model, to overcome this disparity and achieve better performance. In simpler terms, DTL focuses primarily on leveraging
model-based techniques for knowledge transfer. There are many ways to adapt models for deep transfer learning (DTL). These

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 10

techniques often involve a combination of pretraining, freezing, fine-tuning, and/or adding new layers. A deep learning network
trained on a source dataset is referred to as a pretrained model, which consists of pretrained layers that encapsulate learned
weights. Freezing and fine-tuning are techniques that utilize some or all layers of pretrained models to train the model on
target data. Freezing refers to a technique where specific layers within a pretrained model have their own weights or
parameters locked, whereas fine-tuning refers to the parameters or weights that utilize pretrained values initialized with
pretrained values instead of random initialization for the whole network or specific chosen layers. Another recent DTL
technique is based on freezing a pretrained model and adding a new layer to that model for training on target data.
A loss function and a performance metric are both instrumental in evaluating the effectiveness of a deep learning model.
However, both serve distinct purposes within the training process. The loss function continuously evaluates the discrepancy
between the model's predictions and the actual targets. It plays a crucial role during the training phase. A table summarized
by Terven et al. (2023) provides a comparative overview of four loss functions commonly utilized in object detection tasks:
smooth L1, IoU loss, focal loss, and YOLO loss. This research focuses on the implementation of a loss function specifically
designed for imbalanced datasets. Focal loss (FL) is effective for handling highly imbalanced class problems because it addresses
the imbalance between difficult and easy-to-classify examples. Even easy samples contribute minimal loss to the overall
training objective, but a larger volume of easy examples will lead to a cumulative loss that may outweigh the loss from hard
examples. Thus, FL focuses on learning more on hard examples by downweighting the loss contribution from correctly classified
or confidently predicted examples (Lin et al., 2021). The mathematical formula for FL is defined as follows:

FL(pt ) = -αt (1 - pt )γ log(pt ) (1)

where:

• pt refer to the predicted probability

• αt refers to the weighting factor
• γ refers to the focusing parameter
• where pt represents the model confidence in predicting the correct class for a specific sample.

A higher value close to 1 indicates greater confidence. αt is the hyperparameter controlling the importance of each
sample within the loss calculation. It is often set to the inverse class frequency to balance the loss across all classes. γ is another
hyperparameter that determines the degree to which easy examples are downweighted. Usually, the value is set between 2
and 4. A higher γ value leads to a stronger focus on hard-to-classify examples during training.
In addition to focal loss, another approach for handling imbalanced datasets is focal Tversky loss (FTL). This loss function
forces the model to focus on harder training examples, which include minority class segments (Sinha & Senapati, 2023). The
formula is defined as follows (Das and Zhang, 2020):
∑𝑁
𝑖=1 𝑝0𝑖 𝑔0𝑖
𝑇(𝛼, 𝛽) = ∑𝑖 𝑝0𝑖 𝑔0𝑖 +𝛼 ∑𝑖 𝑝0𝑖 𝑔0𝑖 +𝛽 ∑𝑖 𝑝0𝑖 𝑔0𝑖
(2)
𝛾
𝐹𝑇𝐿(𝛼, 𝛽, 𝛾) = (1 − 𝑇(𝛼, 𝛽)) (3)

where:

• T (𝛼, 𝛽) refers to the Tversky index (TI).

• P is the predicted label.
• G is the ground truth label.
p0i indicates the probability that detected object i belongs to an actual object, whereas p1i indicates the probability that
detected object i does not belong to an actual object. g0i is 1 for an actual object and 0 for a nonobject; similarly, g1i is 1 for a
nonobject and 0 for an actual object. Additionally, 𝛼 𝑎𝑛𝑑 𝛽 are the hyperparameters used to determine the relative importance
of penalizing false positives and false negatives. The 𝛼 and 𝛽 in the FTL equation act as the balancing factors, whereas the γ
controls the difficulty level assigned to the misclassified ROIs.
On the other hand, performance metrics are a critical component of model evaluation following the training process.
These metrics assess the model's ability to generalize novel data and generate accurate predictions. Performance metrics
enable the comparative analysis of different models or configurations to choose the optimal model (Terven et al., 2023). The
standard evaluation criteria for performance metrics include accuracy (Acc), precision (Precision), recall (Recall), the F1 score,
and the mean average precision (mAP). The four mathematical metrics were defined as follows and can be computed using
various software (Awang et al., 2015):
TP+TN
Acc = (4)
TP+TN+FP+FN

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 11

TP
Precision = (5)
TP+FP
TP
Recall = (6)
TP+FN
precision x recall
F1 = 2 x (7)
precision + recall

where:

• True positive (TP) represents the number of correct predictions for positive samples.
• True negative (TN) indicates the number of correct predictions for negative samples.
• False negative (FN) refers to the number of incorrect predictions for positive samples.
• False positive (FP) represents the number of incorrect predictions for negative samples.
4.2. Dataset
Zhang and Liu (2021) introduced a Faster R-CNN with an FPN for PCB defect detection. The experiment uses two publicly
available datasets, the public synthesis PCB dataset and the deep PCB dataset, and one private dataset, the PCB dataset with
components. In this research, two publicly available PCB datasets are combined into one dataset and utilized for the model.
The public synthetic PCB dataset contains single-defect images in the RGB color space, whereas the deep PCB dataset offers
images of PCBs with multiple defects in grayscale format, as illustrated in Figure 8 and Figure 9. These combinations of datasets
were able to enhance model robustness and generalizability. The model is expected to learn spatial patterns from the grayscale
images in the Deep PCB dataset and leverage the rich color information from the Public Synthetic PCB dataset. This combined
training strategy aims to improve the model's ability to detect a wider variety of defects, including those that may not exhibit
distinct color variations. The datasets are divided into three sets: a training set, a validation set, and a test set. Seventy percent
of the dataset is used for training the model. The 20% of the dataset is allocated as the validation dataset, which is used for
fine-tuning the model, and the remaining 10% of the dataset serves as the test set, which acts as new data to evaluate the
model performance.

Figure 8 Public synthesis PCB defect image. Source: Zhang and Liu (2021) Figure 9 Deep PCB defect image. Source: Zhang and Liu (2021)

4.3. Proposed model for PCB defect detection

The architecture for the proposed method is the combination of an FPN with an RFCN. The PCB is defected as the input
image and fed into the CNN. The CNN performs feature extraction and generates feature maps as outputs. The feature maps
are fed into both the RPN and FPN in parallel. The RPN generates the objectness score and bounding box regression. The
bounding box regression is used to refine the bounding box. The refined bounding box acts as a reference for the position-
sensitive score map stage. For the FPN stage, the FPN generates multiple levels of feature maps whose low level contains higher
resolution to capture fine details, and higher levels contain lower resolution but capture a larger field of view. These multiple
levels of feature maps are passed to the position-sensitive score map stage. In the position-sensitive score map stage, the
refined bounding boxes help select the most appropriate level of detail from the multiscale feature maps extracted by the FPN.
This chosen feature map is then used to generate position-sensitive score maps. Each element in the score maps represents a
location within the candidate region and predicts the likelihood of a defect being present. A threshold is applied to these scores,
and regions exceeding it are considered potential defects. These position-sensitive score maps are used to vote by using
nonmax suppression (NMS), which removes redundant bounding boxes, and the remaining ones with the highest scores are
classified to identify defects. These classified defects are then displayed visually on the image. Figure 10 illustrates the proposed
model architecture.

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 12

Figure 10 Proposed model for PCB defect detection based on an improved FPN-RFCN method.

5. Final Considerations
In conclusion, the integration of the FPN-RFCN in automated PCB defect detection through deep learning represents a
significant improvement in ensuring the reliability and quality of electronic devices. This review highlights the evolution from
traditional methods to sophisticated deep learning architectures, emphasizing their ability to detect diverse PCB defects with
high precision and efficiency. While challenges such as dataset diversity and model optimization persist, ongoing advancements
in deep transfer learning and loss function adaptation promise further enhancements in robustness and real-world
applicability. As Industry 4.0 continues to evolve, leveraging these technologies not only enhances manufacturing efficiency
but also fosters safer and more reliable consumer electronics, paving the way for a future where zero-defect production
becomes a tangible reality.
Acknowledgment
The authors would like to express our gratitude and appreciation to Fakulti Teknologi dan Kejuruteraan Industri dan
Pembuatan, Universiti Teknikal Malaysia Melaka, for the support provided in enabling the implementation of this project.
Ethical Considerations
Not applicable.
Conflict of Interest
The authors declare no conflict of interests.
Funding
This research did not receive any financial support.
References
Arif, Z. H., Mahmoud, M. A., Abdulkareem, K. H., Mohammed, M. A., Al-Mhiqani, M. N., Mutlag, A. A., & Damaševičius, R. (2022).
Comprehensive Review of Machine Learning (ML) in Image Defogging: Taxonomy of Concepts, Scenes, Feature Extraction, and Classification
techniques. IET Image Processing, 16(2), 289–310. https://doi.org/10.1049/IPR2.12365
Awang, N., Fauadi, M. H. F. M., & Rosli, N. S. (2015). Image Processing of Product Surface Defect Using Scilab. Applied Mechanics and Materials
vols. 789–790, pp. 1223–1226. https://doi.org/10.4028/www.scientific.net/amm.789-790.1223
Bahrebar, S., Homayoun, S., & Ambat, R. (2022). Using machine learning algorithms to predict failure on the PCB surface under corrosive
conditions. Corrosion Science, 206, 110500. https://doi.org/10.1016/J.CORSCI.2022.110500
Boateng, E. Y., Otoo, J., Abaye, D. A., Boateng, E. Y., Otoo, J., & Abaye, D. A. (2020). Basic Tenets of Classification Algorithms K-Nearest-
Neighbor, Support Vector Machine, Random Forest and Neural Network: A Review. Journal of Data Analysis and Information Processing, 8(4),
341–357. https://doi.org/10.4236/JDAIP.2020.84020
Chen, I. C., Hwang, R. C., & Huang, H. C. (2023). PCB Defect Detection Based on Deep Learning Algorithm. Processes, Vol. 11, Page 775, 11(3),
775. https://doi.org/10.3390/PR11030775

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 13

Chen, L., Li, S., Bai, Q., Yang, J., Jiang, S., & Miao, Y. (2021). Review of Image Classification Algorithms Based on Convolutional Neural Networks.
Remote Sensing, Vol. 13, Page 4712, 13(22), 4712. https://doi.org/10.3390/RS13224712
Dai, J., Li, Y., He, K., & Sun, J. (2016). R-FCN: Object Detection via Region-based Fully Convolutional Networks. Advances in Neural Information
Processing Systems, 379–387. https://arxiv.org/abs/1605.06409v2
Das, K., & Zhang, Q. (2020). Convolutional Recurrent Residual U-Net Embedded with Attention Mechanism and Focal Tversky Loss Function
for Cancerous Nuclei Detection. Queen Mary University of London, United Kingdom. https://arxiv.org/abs/2010.04416v1
Ding, R., Dai, L., Li, G., & Liu, H. (2019). TDD-net: a tiny defect detection network for printed circuit boards. CAAI Transactions on Intelligence
Technology, 4(2), 110–116. https://doi.org/10.1049/TRIT.2019.0019
Dong, Y., Wang, Z., & Guo, W. (2022). Overview of edge detection algorithms based on mathematical morphology. In Proceedings of Electronic
and Automation Control Conference (IAEAC), 2022-October, 1321–1326. https://doi.org/10.1109/IAEAC54830.2022.9930043
Du, Y., & Wang, Q. (2021). Multi-Angle Face Detection Based on Improved RFCN Algorithm Using Multi-Scale Training. In Proceedings of 2021
IEEE 6th International Conference on Intelligent Computing and Signal Processing, ICSP 2021, 319–322.
https://doi.org/10.1109/ICSP51882.2021.9408676
Fung, K. C., Xue, K. W., Lai, C. M., Lin, K. H., & Lam, K. M. (2024). Improving PCB defect detection using selective feature attention and pixel
shuffle pyramid. Results in Engineering, 21, 101992. https://doi.org/10.1016/J.RINENG.2024.101992
Ge, C., Wang, J., Wang, J., Qi, Q., Sun, H., & Liao, J. (2020). Towards automatic visual inspection: A weakly supervised learning method for
industrial applicable object detection. Computers in Industry, 121, 103232. https://doi.org/10.1016/J.COMPIND.2020.103232
Georgiev, A. G., Hristov, G., Plamen Zahariev, A., & Diyana Kinaneva, A. (2023). Exploring Object Detection Algorithms: A Comprehensive
Overview and Comparative Study 15. Proceedings of University of Ruse-2023, 62.
Ghaffari, S., Soleimani, P., Li, K. F., & Capson, D. W. (2020). Analysis and Comparison of FPGA-Based Histogram of Oriented Gradients
Implementations. IEEE Access, 8, 79920–79934. https://doi.org/10.1109/ACCESS.2020.2989267
Hu, B., & Wang, J. (2020a). Detection of PCB Surface Defects with Improved Faster-RCNN and Feature Pyramid Network. IEEE Access, 8,
108335–108345. https://doi.org/10.1109/ACCESS.2020.3001349
Iman, M., Arabnia, H. R., & Rasheed, K. (2023). A Review of Deep Transfer Learning and Recent Advancements. Technologies, Vol. 11, Page
40, 11(2), 40. https://doi.org/10.3390/TECHNOLOGIES11020040
Lan, Z., Hong, Y., & Li, Y. (2021). An improved YOLOv3 method for PCB surface defect detection. In Proceedings of 2021 IEEE International
Conference on Power Electronics, Computer Applications, ICPECA 2021, 1009–1015. https://doi.org/10.1109/ICPECA51329.2021.9362675
Li, C. J., Qu, Z., Wang, S. Y., Bao, K. H., & Wang, S. Y. (2022). A Method of Defect Detection for Focal Hard Samples PCB Based on Extended
FPN Model. IEEE Transactions on Components, Packaging and Manufacturing Technology, 12(2), 217–227.
https://doi.org/10.1109/TCPMT.2021.3136823
Li, Y., Miao, N., Ma, L., Shuang, F., & Huang, X. (2023). Transformer for object detection: Review and benchmark. Engineering Applications of
Artificial Intelligence, 126, 107021. https://doi.org/10.1016/J.ENGAPPAI.2023.107021
Li, Z., Liu, F., Yang, W., Peng, S., & Zhou, J. (2022). A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. IEEE
Transactions on Neural Networks and Learning Systems, 33(12), 6999–7019. https://doi.org/10.1109/TNNLS.2021.3084827
Lin, W., Wu, P., & Xiao, X. (2021). Boundary Focal Loss for Class Imbalanced Learning. In Proceedings of 14th International Congress on Image
and Signal Processing, BioMedical Engineering and Informatics. https://doi.org/10.1109/CISP-BMEI53629.2021.9624339
Ling, Q., Isa, N. A. M., & Asaari, M. S. M. (2023). Precise Detection for Dense PCB Components Based on Modified YOLOv8. IEEE Access, 11,
116545–116560. https://doi.org/10.1109/ACCESS.2023.3325885
Linton, P., Morgan, M. J., Read, J. C. A., Vishwanath, D., Creem-Regehr, S. H., & Domini, F. (2023). New Approaches to 3D Vision. Philosophical
Transactions of the Royal Society B, 378(1869). https://doi.org/10.1098/RSTB.2021.0443
Liu, X., Hu, J., Wang, H., Zhang, Z., Lu, X., Sheng, C., Song, S., & Nie, J. (2022). Gaussian-IoU loss: Better learning for bounding box regression
on PCB component detection. Expert Systems with Applications, 190, 116178. https://doi.org/10.1016/J.ESWA.2021.116178
Lou, L., Lu, K., & Xue, J. (2023). Multi-Scale Vision Transformer for Defect Object Detection. Procedia Computer Science, 222, 397–406.
https://doi.org/10.1016/J.PROCS.2023.08.178
LU, H., & MAZUMDER, R. (2020). Randomized Gradient Boosting Machine. SIAM Journal on Optimization, 30(4), 2780–2808.
https://doi.org/10.1137/18M1223277
Mercier, J. P., Garon, M., Giguere, P., & Lalonde, J. F. (2021). Deep template-based object instance detection. In Proceedings of 2021 IEEE
Winter Conference on Applications of Computer Vision, WACV 2021, 1506–1515. https://doi.org/10.1109/WACV48630.2021.00155
Mohan, L., Pant, J., Suyal, P., & Kumar, A. (2020). Support Vector Machine Accuracy Improvement with Classification. In Proceedings of 12th
International Conference on Computational Intelligence and Communication Networks, 477–481.
https://doi.org/10.1109/CICN49253.2020.9242572
O’Mahony, N., Campbell, S., Carvalho, A., Harapanahalli, S., Hernandez, G. V., Krpalkova, L., Riordan, D., & Walsh, J. (2020). Deep Learning
vs. Traditional Computer Vision. Advances in Intelligent Systems and Computing, 943, 128–144. https://doi.org/10.1007/978-3-030-17795-
9_10
Pang, G., Shen, C., Cao, L., & Hengel, A. Van Den. (2021). Deep Learning for Anomaly Detection. ACM Computing Surveys (CSUR), 54(2).
https://doi.org/10.1145/3439950

https://www.malque.pub/ojs/index.php/mr
Jian et al. (2025) 14

Patel, P., & Thakkar, A. (2020). The upsurge of deep learning for computer vision applications. International Journal of Electrical and Computer
Engineering (IJECE), 10(1), 538–548. https://doi.org/10.11591/ijece.v10i1.pp538-548
Piliposyan, G., & Khursheed, S. (2022). Computer Vision for Hardware Trojan Detection on a PCB Using Siamese Neural Network. 2022 IEEE
Physical Assurance and Inspection of Electronics, https://doi.org/10.1109/PAINE56030.2022.10014967
Rosli, N.S., Fauadi, M., Awang, N.F., Noor, A.Z.M. (2018). Vision-based defects detection for glass production based on improved image
processing method. Journal of Advanced Manufacturing Technology, 12(1(1)), 203-212.
Rosli, N.S., Fauadi, M., Awang, N. (2016). Some technique for an Image of defect in inspection process based on image processing. Journal of
Image and Graphics, 4(1), 55-58.
Schubeck, J., Koblah, D., Botero, U. J., & Forte, D. (2021). A Comprehensive Taxonomy of PCB Defects. University of Florida, Gainesville, USA.
Silva, L. H. D. S., Azevedo, G. O. D. A., Fernandes, B. J. T., Bezerra, B. L. D., Lima, E. B., & Oliveira, S. C. (2019). Automatic Optical Inspection
for Defective PCB Detection Using Transfer Learning. 2019 IEEE Latin American Conference on Computational Intelligence.
https://doi.org/10.1109/LA-CCI47412.2019.9037036
Sinha, A., & Senapati, K. K. (2023). Analysing Satellite Images using Segmentation with U-net and Focal Tversky Loss. In Proceedings of 2023
6th International Conference on Information Systems and Computer Networks. https://doi.org/10.1109/ISCON57294.2023.10112138
Sun, Y., Sun, Z., & Chen, W. (2024). The evolution of object detection methods. Engineering Applications of Artificial Intelligence, 133, 108458.
https://doi.org/10.1016/J.ENGAPPAI.2024.108458
Terven, J., Cordova-Esparza, D. M., Ramirez-Pedraza, A., & Chavez-Urbiola, E. A. (2023). Loss Functions and Metrics in Deep Learning.
https://arxiv.org/abs/2307.02694v2
Udaya Sankar, V., Lakshmi, · Gayathri, & Siva Sankar, · Y. (2022). A Review of Various Defects in PCB. Journal of Electronic Testing 1, 3.
https://doi.org/10.1007/s10836-022-06026-7
Wang, H., Pan, C., Guo, X., Ji, C., & Deng, K. (2021). From object detection to text detection and recognition: A brief evolution history of
optical character recognition. Wiley Interdisciplinary Reviews: Computational Statistics, 13(5), e1547. https://doi.org/10.1002/WICS.1547
Wang, Z., Wang, E., & Zhu, Y. (2020). Image segmentation evaluation: a survey of methods. Artificial Intelligence Review, 53(8), 5637–5674.
https://doi.org/10.1007/S10462-020-09830-9/METRICS
Xia, S., Wang, F., Xie, F., Huang, L., Wang, Q., & Ling, X. (2021). An Efficient and Robust Target Detection Algorithm for Identifying Minor
Defects of Printed Circuit Board Based on PHFE and FL-RFCN. Frontiers in Physics, 9, 661091.
https://doi.org/10.3389/FPHY.2021.661091/BIBTEX
Yang, S., Xiao, W., Zhang, M., Guo, S., Zhao, J., & Shen, F. (2022). Image Data Augmentation for Deep Learning: A Survey.
https://arxiv.org/abs/2204.08610v2
Yu, X., Lyu, W., Zhou, D., Wang, C., & Xu, W. (2022). ES-Net: Efficient Scale-Aware Network for Tiny Defect Detection. IEEE Transactions on
Instrumentation and Measurement, 71. https://doi.org/10.1109/TIM.2022.3168897
Zeng, N., Wu, P., Wang, Z., Li, H., Liu, W., & Liu, X. (2022). A Small-Sized Object Detection Oriented Multi-Scale Feature Fusion Approach with
Application to Defect Detection. IEEE Transactions on Instrumentation and Measurement, 71. https://doi.org/10.1109/TIM.2022.3153997
Zhang, L., Chen, J., Chen, J., Wen, Z., & Zhou, X. (2024). LDD-Net: Lightweight printed circuit board defect detection network fusing multi-
scale features. Engineering Applications of Artificial Intelligence, 129, 107628. https://doi.org/10.1016/J.ENGAPPAI.2023.107628
Zhang, Q., & Liu, H. (2021). Multi-scale defect detection of printed circuit board based on feature pyramid network. In Proceedings of 2021
IEEE International Conference on Artificial Intelligence and Computer Applications, 911–914.
https://doi.org/10.1109/ICAICA52286.2021.9498174
Zhang, W., Zhu, Q., Li, Y., & Li, H. (2023). MAM Faster R-CNN: Improved Faster R-CNN based on Malformed Attention Module for object
detection on X-ray security inspection. Digital Signal Processing, 139, 104072. https://doi.org/10.1016/J.DSP.2023.104072
Zhao, K., Han, Q., Zhang, C. Bin, Xu, J., & Cheng, M. M. (2022). Deep Hough Transform for Semantic Line Detection. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 44(9), 4793–4806. https://doi.org/10.1109/TPAMI.2021.3077129
Zhou, J., Liu, Y., Zhang, X., & Yang, Z. (2021). Multi-view based template matching method for surface defect detection of circuit board.
Journal of Physics: Conference Series, 1983(1), 012063. https://doi.org/10.1088/1742-6596/1983/1/012063
Zhu, Y., Detig, C., Kane, S., & Lourie, G. (2022). Kinematic Motion Analysis with Volumetric Motion Capture. In Proceedings of International
Conference on Information Visualisation, 2022-July, 61–66. https://doi.org/10.1109/IV56949.2022.00019
Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., & He, Q. (2021). A Comprehensive Survey on Transfer Learning. In Proceedings
of the IEEE, 109(1), 43–76. https://doi.org/10.1109/JPROC.2020.3004555
Zou, Z., Chen, K., Shi, Z., Guo, Y., & Ye, J. (2023). Object Detection in 20 Years: A Survey. In Proceedings of the IEEE, 111(3), 257–276.
https://doi.org/10.1109/JPROC.2023.3238524

https://www.malque.pub/ojs/index.php/mr