IEEE - 61001

PCB Defect Detection Using Deep Learning Methods

Manan Agrawal Parikshit Sardeshpande
Aryan Sharma
Department of Electronics Engineering. Department of Electronics Engineering.
Department of Electronics Engineering.
IEEE Student Member Shri Ramdeobaba College of
Shri Ramdeobaba College of
Shri Ramdeobaba College of Engineering and Management.
Engineering and Management.
2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) | 979-8-3503-7024-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICCCNT61001.2024.10726140

Engineering and Management. Nagpur, India.

Nagpur, India.
Nagpur, India. sardeshpandepa@rknec.edu
sharmaav@rknec.edu
agrawalmu@rknec.edu
Aditya Gupta Richa. R. Khandelwal
Ayaan Pasha
Department of Electronics Engineering. Department of Electronics Engineering
Department of Electronics Engineering.
Shri Ramdeobaba College of IEEE Senior Member
Shri Ramdeobaba College of
Engineering and Management. Shri Ramdeobaba College of
Engineering and Management.
Nagpur, India. Engineering and Management.
Nagpur, India.
guptaaj_1@rknec.edu Nagpur, India.
pashaay@rknec.edu
khandelwalrr@rknec.edu

Abstract—Printed Circuit Boards (PCBs) are the backbone of of recognizing and classifying various PCB defects with
electronic devices, critical for the functionality and reliability of unparalleled accuracy [8].
modern technological applications. Despite advancements in
manufacturing techniques, the production of PCBs is susceptible to The report outlines our rigorous process, from data collection
defects due to inherent technical limitations. These defects can and image preprocessing to feature extraction, culminating in the
significantly impact the performance and reliability of electronic design and implementation of sophisticated deep learning models
products, making effective defect detection a paramount concern. [9]. The research delves into the testing and validation of the system,
This research introduces a novel approach to PCB defect detection ensuring its reliability and efficacy in enhancing manufacturing
and classification by employing advanced deep learning-based quality and efficiency [10]. The convergence of the findings with
object detection networks. Focusing on two state-of-the-art target insights from ten pivotal studies in the field underscores the
detection frameworks, this study evaluates their efficacy across two multifaceted nature of PCB defect detection [12]. From the
distinct PCB datasets, representing varied defect distributions. The integration of image processing algorithms to the application of
experimental analysis demonstrates the superior accuracy of the semi-supervised learning models and innovative deep learning
proposed methods in identifying and classifying various types of networks, the research navigates through the complexities of early-
PCB defects. By leveraging the robust capabilities of deep learning stage inspection, real-world applicability, and the challenges of
for object detection, this research offers significant contributions to defect classification under varying conditions [13].
the field of electronic manufacturing. This research presents a
comprehensive evaluation of the detection networks, highlighting Notable contributions, such as the YOLOX-MC-CA network,
their strengths in enhancing the precision and efficiency of the PCB highlight the strides made in detecting minute defects, showcasing
inspection process. The findings of this research provide valuable the potential of lightweight deep-learning solutions in advancing
insights and advancements in addressing the critical challenge of PCB inspection processes [7]. In synthesizing these advancements,
defect detection in PCBs, thereby facilitating the improvement of this work not only contributes to the ongoing discourse on
electronic product quality and manufacturing yield. This work not automated PCB flaw identification but also sets the stage for further
only advances the knowledge base in PCB defect detection but also exploration in the realms of computer vision and deep learning [14].
proposes a scalable and effective solution to optimize the inspection By elevating the standards of quality control within the electronics
process in the electronic manufacturing industry. manufacturing sector, this project anticipates the development of
more reliable, cost-effective, and high-performance electronic
Keywords: Printed Circuit Board (PCB), Defect Detection, Object devices. This introduction paves the way for a detailed exploration
Detection Networks, Deep Learning. of the technical nuances, challenges encountered, and the
transformative impact of our findings on the field of PCB defect
1. Introduction detection and electronic manufacturing at large.

In the era of rapid technological advancement, electronic devices 2. Related Work

have become fundamental to everyday life, driving the demand for
high-quality, reliable components [1]. Central to these devices are The landscape of Printed Circuit Board (PCB) defect detection has
Printed Circuit Boards (PCBs), whose complexity and the sheer evolved significantly, navigating through the complexities imposed
volume of production pose significant quality control challenges [2]. by the miniaturization of PCBs and the increasing demand for
Traditional inspection methods, reliant on human scrutiny, are precise and effective inspection techniques [15]. This section delves
increasingly proving inadequate in the face of intricate designs and into the methodologies prevalent in the literature, categorizing them
the need for high throughput, often resulting in costly recalls and into electrical/contact and non-electrical/non-contact
device failures [3]. Addressing this pivotal challenge, this research methodologies, with a particular emphasis on Automatic Visual
pioneers an automated approach for PCB defect detection, Inspection (AVI) systems that leverage artificial intelligence,
harnessing the power of cutting-edge deep learning algorithms [4]. including deep learning, alongside image processing and computer
vision technologies.
This system aims to revolutionize quality control by employing
computer vision and image processing technologies to accurately 2.1 Electrical/Contact Methodologies
and precisely detect PCB defects [5]. By analyzing images of PCBs,
comparing flawless references with those of defective boards, the Electrical/contact methodologies primarily focus on evaluating a
methodology introduces a robust mechanism for distinguishing PCB's electrical connectivity to verify adherence to design criteria.
between defective and non-defective units [6]. Our approach is Despite its effectiveness in certain contexts, this approach faces
underpinned by a comprehensive dataset, meticulously compiled to challenges with the ongoing trend of PCB miniaturization. The
include a wide array of defective and non-defective PCB images [7]. reduced size and increased complexity of modern PCBs impede the
This dataset facilitates the training of deep learning models capable

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

15th ICCCNT IEEE Conference,
June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

ability of electrical/contact methods to accurately identify defects, processes [17]. These real-world examples validate the potential of
highlighting the necessity for alternative approaches. AVI to reduce production errors and ensure product reliability.

2.2 Non-Electrical/Non-Contact Methodologies 2.3 Challenges and Future Directions: Despite technological
advancements, AVI systems encounter challenges, including
Within the realm of non-electrical/non-contact methodologies, two variability in lighting conditions and the complexity of PCB designs,
primary strategies emerge: manual human inspection and automatic necessitating continuous optimization. The literature highlights
visual inspection (AVI). The latter, characterized by its use of these issues, proposing avenues for future research and development
advanced technologies such as artificial intelligence, deep learning, [23].
and computer vision, presents a promising avenue for enhancing the
accuracy and efficiency of PCB defect detection. The exploration of computer vision, image processing, and deep
learning in the context of PCB defect detection has yielded
2.2.1 Automatic Visual Inspection (AVI): AVI systems aim to significant advancements. This review not only underscores the
automate the process of inspecting visual data to identify PCB current state of the field but also sets the stage for addressing future
defects. These systems utilize a combination of image processing challenges, ensuring the ongoing evolution of defect detection
techniques, advanced machine learning models, and computer methodologies.
vision algorithms to achieve precise defect identification. The core
of AVI's effectiveness lies in the integration of computer vision and The increasing complexity of Printed Circuit Boards (PCBs) and the
image processing, enabling the detection of anomalies through critical need for high reliability in electronic devices have propelled
sophisticated algorithms. the development of advanced defect detection methodologies. This
literature review synthesizes key findings from recent research on
2.2.2 Feature-Based Image Registration: A pivotal step in AVI is PCB defect detection, focusing on the integration of image
aligning the inspection image with a reference image, a task processing, machine learning, and deep learning techniques to
commonly achieved through feature-based image registration enhance inspection accuracy and efficiency.
techniques like Speeded-Up Robust Features (SURF). These
techniques facilitate precise alignment by extracting critical features 2.4 Image Processing Techniques and Initial Approaches
from images, proving effective in handling complex image
distortions. Early efforts in PCB defect detection relied heavily on image
processing techniques. Prathima et al. (2020) explored the
2.2.3 Pre-Processing Techniques: Image pre-processing plays a application of image processing for identifying PCB defects,
crucial role in enhancing image quality, employing methods such as demonstrating the utility of basic techniques in highlighting
histogram equalization for contrast improvement, Gaussian blurring discrepancies between defect-free and defective PCB images [1].
for noise reduction, and thresholding for binary image Similarly, Nayaka et al. (2017) utilized image processing to address
transformation. These techniques prepare images for subsequent PCB fault detection, underscoring the challenges posed by the
comparison and defect identification. miniaturization of components and the need for precise inspection
methods [2]. These foundational studies set the stage for the
2.2.4 Image Subtraction in Defect Detection: The process of integration of more sophisticated machine learning approaches.
detecting PCB flaws often begins with image subtraction,
comparing pre-processed inspection images against reference 2.5 Transition to Machine Learning and Deep Learning
images to identify discrepancies. Techniques like the XOR
operation are instrumental in highlighting differences, thereby The evolution from traditional image processing to machine
facilitating defect detection. learning (ML) and deep learning (DL) approaches marked a
significant advancement in PCB defect detection. Gupta et al. (2019)
2.2.5 Hardware Considerations: The success of AVI systems also discussed machine learning approaches for automated PCB
depends on the hardware used, including image capture devices and inspection, highlighting the transition towards more adaptive and
lighting systems. Selecting optimal hardware components and fine- robust inspection systems [12]. The advent of deep learning further
tuning their configurations are crucial for achieving reliable defect revolutionized defect detection, with Zhang et al. (2019)
detection. demonstrating the effectiveness of Convolutional Neural Networks
(CNNs) in identifying PCB defects with high accuracy [13].
2.2.6 Software Tools: The software backbone of AVI systems
includes deep learning frameworks, OpenCV, and other computer 2.6 Advanced Deep Learning Models for Enhanced Detection
vision and image processing toolkits. These tools support various
stages of the AVI process, from image acquisition to feature Research by Wang et al. (2022) introduced a lightweight modified
extraction, underscoring the role of automation in enhancing the YOLOX network incorporating a coordinate attention mechanism
inspection process. for PCB surface defect detection, showcasing the potential for real-
time, efficient defect identification [7]. This adaptation of YOLOX
2.2.7Benchmark Datasets and Evaluation Metrics: The underscores the trend towards optimizing deep learning models for
performance of defect detection algorithms is assessed using specific challenges in PCB inspection, such as small defect
benchmark datasets specifically designed for PCB defect detection. recognition and processing speed.
Evaluation metrics such as accuracy, precision, and recall provide a
standardized framework for measuring the effectiveness of AVI 2.7 Comprehensive Surveys and Evaluations
systems. Deep learning techniques, in particular, have shown
promise in refining the accuracy of these algorithms through Comprehensive surveys by Ebayyeh and Mousavi (2020) [4], and
comprehensive training. Ling and Isa (2023) [10] provided extensive analyses of automatic
optical inspection (AOI) and quality monitoring methods, including
2.2.8 Real-World Applications and Case Studies: AVI systems a critical review of deep learning applications in the electronics
find applications across various industries, with case studies industry. These surveys contribute valuable insights into the state-
illustrating their impact on manufacturing and quality control of-the-art techniques and the future direction of PCB defect
detection research.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

2.8 Feature Extraction and Classification Techniques

The role of feature extraction and classification techniques in

enhancing the accuracy of defect detection has been a focal point of
recent studies. Chen et al. (2020) conducted a comparative study of
feature extraction techniques for PCB defect classification,
revealing the impact of sophisticated feature extraction methods on
the performance of classification algorithms [15]. This emphasis on
feature extraction is indicative of the growing complexity of PCB
designs and the corresponding need for advanced analytical
strategies.

2.9 Real-Time Detection and FPGA Implementations

The demand for real-time defect detection in PCB manufacturing

has led to innovative implementations, as explored by Patel et al.
(2020), who detailed an FPGA implementation of PCB defect
Fig1:-Flow Chart of Open Cv
detection using artificial neural networks [14]. Such
implementations highlight the practical challenges and solutions in
deploying AI-based inspection systems in real-world manufacturing 3.2 YOLO Method: -
environments.

2.10 Challenges and Future Directions

Despite significant advancements, challenges remain,

particularly regarding dataset quality and diversity, class imbalance,
and the adaptation of models to complex PCB designs and varying
lighting conditions. Future research is likely to focus on addressing
these challenges through improved dataset curation, novel model
architectures, and enhanced learning strategies [21].

The literature on PCB defect detection using deep learning

methods reveals a dynamic field characterized by rapid
advancements and ongoing challenges. The shift towards deep
learning has significantly improved the capability to detect and
classify PCB defects, marking a substantial leap forward from Fig 2: - Flow Chart of YOLO Algorithm
traditional image processing and machine learning approaches [15].
As the field continues to evolve, future research will undoubtedly An automated quality control system for PCBs is outlined,
uncover new methodologies and technologies to further enhance the beginning with image processing via YOLOv5 to detect defects. If
accuracy, efficiency, and applicability of PCB defect detection no defects are found, the process concludes with a "Quality Check
systems Pass." If defects are detected, they're categorized and analyzed for
localization and clustering. If a defect cluster is found, the process
3. Methodology ends abruptly. Otherwise, trends in defect occurrence and size are
examined. If defects are increasing in occurrence or size, human
This section outlines the comprehensive approach adopted in this intervention is triggered. Otherwise, the process ends. This system
project to detect and classify defects in Printed Circuit Boards efficiently identifies defects, assesses their severity, and prompts
(PCBs) using deep learning techniques. The methodology human intervention when necessary, ensuring high-quality PCB
encompasses data collection, preprocessing, model architecture, and production.
the training and validation phases. Two prominent techniques have
been leveraged in this project [25]: OpenCV for initial image Fig 2. represents an automated quality control system for PCBs that
analysis and the YOLO (You Only Look Once) algorithm for defect can determine if a PCB is defective, identify trends in defect
detection, providing a blend of precision and efficiency critical for occurrence and size, and decide when human intervention is
modern electronic process. necessary.

3.1 OpenCV Method: - 3.3 Data Collection

The OpenCV-based approach focused on the initial step of defect The dataset pivotal to this research was sourced from the Open
detection by analyzing differences between a reference image of a Lab on Human Robot Interaction at Peking University, specifically
defect-free PCB and an analysis image showcasing a defective PCB designed for Human-Robot Interaction studies [7]. This dataset,
[1]. Using OpenCV's robust image processing capabilities, we available publicly, comprises 693 synthetic images created through
applied image differencing techniques to highlight areas of Adobe Photoshop, tailored to simulate various PCB defects for
deviation. The process involved thresholding to create a binary detection, classification, and registration tasks. The defects include
mask, isolating regions of dissimilarity, and overlaying these missing holes, mouse bites, open circuits, shorts, spurs, and spurious
identified regions onto the original image in red for easy visual copper, each mimicking real-world PCB anomalies to challenge the
identification. The process is described in fig1. algorithmic approaches. This diverse and intricate dataset is crucial
for refining the defect detection and classification algorithms,
thereby enhancing PCB quality control and manufacturing
efficiency.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

3.4 YOLO Algorithm Preparation: -

For the YOLO algorithm, preprocessing involved dividing the

input image into a grid system [7] Each grid cell was responsible for
predicting bounding boxes and class probabilities for objects
(defects) within its bounds. This approach streamlined the process
for subsequent deep learning analysis, ensuring a comprehensive
coverage of the PCB surface for potential defects.

3.5 Model Architecture

The core of this defect detection system was built on the YOLO
algorithm, particularly leveraging the YOLOv5 architecture for its
balance of speed and accuracy [7]. YOLO's single-pass detection
process divides the image into grids, predicting bounding boxes and
class probabilities directly, which is ideal for real-time defect
detection in PCBs. Fig3:- Image with no defect used for reference.
3.6 Training and Validation

Training the YOLO Model: -

The training phase involved fine-tuning a pre-trained YOLOv5

model on our collected PCB defect dataset [7]. Adjustments to
hyperparameters such as learning rate, batch size, and the number of
epochs were meticulously made to optimize the model's
performance on PCB defect detection. The model was trained to
recognize the six defect types defined in our dataset, with a focus on
achieving high precision and recall.

Validation and Optimization: -

Model validation was conducted using a subset of the dataset not

seen by the model during the training phase [7]. Performance
metrics such as precision, recall, and the F1 score were calculated to
assess the model's accuracy and its ability to generalize across Fig 4:- Image taken from input which consist of a defect.
different PCB designs and lighting conditions. Continuous
optimization was applied by iteratively refining the model with
additional data and adjusting parameters based on real-world
feedback and performance observations.

This methodology represents a robust approach to PCB defect

detection, combining the strengths of OpenCV for preliminary
image analysis and the YOLO algorithm for advanced defect
detection and classification [20]. This comprehensive strategy not
only addresses the challenges posed by intricate PCB designs but
also sets a precedent for future research and development in the field
of electronic manufacturing quality control.

4. Experimental Results

This research project employed two distinct methodologies for PCB

defect detection: an OpenCV-based approach for general
discrepancy identification and a YOLO algorithm-based method for Fig 5:- Difference which is being detected is highlighted.
precise defect detection and categorization. This section details the
experimental results obtained from both implementations, Fig 3, Fig 4, Fig 5 depicts the results obtained in the OpenCV based
emphasizing the effectiveness, challenges, and performance metrics implementation.
of each method.
4.2 YOLO Algorithm Implementation Results
4.1 OpenCV-Based Implementation Results
Advancing to a more detailed analysis, the YOLO algorithm-based
Observations: implementation demonstrated significant improvement in defect
detection. A batch of 16 PCB images, annotated with various defect
The method effectively highlighted general discrepancies, making it types under different lighting conditions and PCB designs was
a quick tool for identifying areas requiring further inspection. compiled. The YOLOv5 model was fine-tuned on the dataset,
enabling it to identify and classify specific defect types accurately,
The primary limitation was its inability to classify or provide including missing holes, mouse bites, open circuits, shorts, spurs,
detailed information about the specific types of defects detected, and spurious copper.
necessitating further analysis with advanced techniques.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

Defect Occurrence Trend Detection:

Fig 9:- Batch consisting of 4 images, detected the defect occurrence

trend of missing hole defect.

As the number of missing hole defect is increasing the algorithm

detects it and suggests that there are changes needed in the
Fig 6:- Batch consisting of 16 images, detected the defect and the manufacturing process by triggering human intervention.
type of defect.
Defect Size Trend Detection:

The size of the defect is analyzed by measuring the co-ordinates (x,

y, w, h) which together forms a bounding box which is a method
used by yolo to detect the size of the defects. As per the images
available in the data set, such type of defect was not detected.

Fig 6, Fig 7, Fig 8, Fig 9 depicts the results obtained in the Yolo
Algorithm Implementation.

Performance Metrics:

1. F1-Confidence Curve Analysis:

The F1 score, representing the harmonic mean of precision and

recall, was utilized as a key metric. The model achieved an optimal
balance between precision and recall, indicating high accuracy in
defect detection.

Fig 7:- Batch consisting of 16 images, detected the defect and the The F1-confidence curve, plotted by varying the confidence
type of defect. threshold, illustrated the model's performance trade-off between
precision and recall at different confidence levels. This analysis
helped in selecting an optimal confidence threshold that maximizes
Localized defect detection (Cluster defect detection): the F1 score.

Confidence 2(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)(𝑅𝑒𝑐𝑎𝑙𝑙)
F1=
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

Missing Mouse Open Short Spur Spurious All

Hole Bite Circuit Copper Classes
0.0 0.2 0.17 0.15 0.19 0.1 0.1 0.18
0.2 0.99 0.5 0.38 0.78 0.5 0.45 0.58
0.4 0.99 0.5 0.3 0.8 0.38 0.4 0.5
0.6 0.99 0.08 0.00 0.83 0.18 0.3 0.37
0.8 0.8 0.00 0.00 0.55 0.00 0.13 0.2
Fig 8:- Batch consisting of 6 images, detected the localized defect

As the defects are located in a cluster form and as mentioned in the

𝟐(𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏)(𝑹𝒆𝒄𝒂𝒍𝒍)
methodology the cluster defect is detected and the process goes to The formula for calculating the F1 score is: F1=
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏+𝑹𝒆𝒄𝒂𝒍𝒍
end process and also indicates that it is a serious issue and triggers
human intervention.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

Fig 10 and Fig 11 depicts the F1-Confidence Curve and Recall

Curve analysis respectively.

Key Findings:

The YOLO model successfully detected and identified defects

in all 16 images, demonstrating its real-time processing capabilities
and the adaptability to handle various defect types simultaneously.
Integration into the Automated Visual Inspection (AVI) system was
seamless, enhancing the defect detection process's
comprehensiveness.

Challenges and Optimization:

Despite the success, challenges such as dataset quality, class

imbalance, and adaptation to complex PCB designs and lighting
Fig 10:- . F1-Confidence Curve. variations were noted.

2. Recall Curve Analysis: Continuous model optimization and refinement were necessary
to maintain high accuracy and efficiency, leveraging community
The recall curve, generated by adjusting the confidence support and updates.
threshold, showed the model's ability to capture all positive
instances across varying levels of confidence. This curve was 4.3 Comparative Analysis
instrumental in understanding the model's sensitivity and its
effectiveness in detecting defects under different conditions. The comparison between the OpenCV-based method and the YOLO
algorithm-based approach underscored the importance of advanced
The recall is calculated using the formula: deep learning techniques in PCB defect detection. While the
OpenCV method offers a quick way to spot discrepancies, YOLO's
Recall= (True Positives) / (True Positives + False Positives) comprehensive detection and classification capabilities provide a
more detailed and actionable analysis for quality control processes.

The experimental analysis demonstrates the effectiveness of

employing advanced deep learning models, particularly YOLO, for
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 precise and efficient PCB defect detection. The ability to accurately
Recall= 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 classify defects in real-time significantly contributes to improving
Confidence
the quality control measures in electronic manufacturing,
addressing the critical need for reliability and performance in PCBs.
Missing Mouse Open Short Spur Spurious All
Hole Bite Circuit Copper Classes YOLOv5 stands out as it combines state-of-the-art accuracy with
0.0 1.00 0.8 0.7 0.9 0.72 0.8 0.81 remarkable speed, outperforming even the latest versions of YOLO.
0.2 1.00 0.6 0.42 0.85 0.5 0.48 0.7 Its streamlined architecture ensures faster inference times without
0.4 0.99 0.3 0.18 0.82 0.3 0.35 0.5 compromising accuracy, making it ideal for real-time applications.
0.6 0.96 0.09 0.00 0.7 0.08 0.2 0.3 Moreover, YOLOv5 offers a user-friendly implementation and
extensive community support, further solidifying its position as the
0.8 0.5 0.00 0.00 0.38 0.00 0.02 0.1
top choice for efficient and accurate object detection tasks.

5. Conclusion

This project represents a significant breakthrough in PCB defect

detection, seamlessly integrating advanced image processing with
the precision of the YOLO algorithm. By meticulously navigating
through image preloading and leveraging OpenCV for initial
discrepancies identification, to the sophisticated defect
categorization offered by YOLO, it sets a new benchmark in
automated visual inspection (AVI) systems. Despite challenges like
dataset integrity and class imbalance, the deployment of YOLO
demonstrates adaptability and prowess in real-time analysis across
various defect scenarios. The findings highlight substantial gains in
detection accuracy and operational efficiency, marking a new era of
quality assurance in electronic manufacturing. Utilizing evaluation
metrics such as F1 score and precision-recall curves underscores the
model's capabilities, laying a solid foundation for future research to
further revolutionize precision and efficiency standards in the
industry.

Fig 11:- Recall Curve Analysis.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India
 IEEE - 61001

References [15] R. Chen, W. Li, and J. Zhang, "A Comparative Study of

Feature Extraction Techniques for PCB Defect Classification," in
[1] G. Prathima, A. Yasaswini Naga Lakshmi, C. Vinay Kumar, IEEE Transactions on Instrumentation and Measurement, vol. 69,
A. Manikanta, and B. Jaya Sandeep, "Defect Detection in PCB Using no. 5, pp. 2322-2332, May 2020. DOI: 10.1109/TIM.2019.2927997.
Image Processing," International Journal of Advanced Science and
Technology, vol. 29, no. 4s, pp. 1205-1210, 2020. [16] J. Wang, H. Li, and Q. Chen, "Real-Time Defect Detection
in PCB Manufacturing Using Support Vector Machines," in IEEE
[2] J. P. R. Nayaka, A. Kb, Dr. P. B. Dc, Dr. R. Banud, R. Pe, Transactions on Industrial Informatics, vol. 16, no. 8, pp. 4885-4893,
"PCB Fault Detection Using Image Processing," IOP Conf. Series: August 2020. DOI: 10.1109/TII.2019.2957010.
Materials Science and Engineering, vol. 225, p. 012244, 2017,
[17] K. Kim, S. Park, and Y. Lee, "Hybrid Approach for PCB
doi:10.1088/1757-899X/225/1/012244.
Defect Classification: Integrating Genetic Algorithms with Neural
[3] N. K. D. S. Munaweera Tanthirige, "Printed Circuit Board Networks," in IEEE Transactions on Components, Packaging and
Defect Detection Using Image Processing," Master's thesis, Manufacturing Technology, vol. 10, no. 11, pp. 1892-1900,
Minnesota State University, Mankato, in partial fulfillment of the November 2020. DOI: 10.1109/TCPMT.2020.3004428.
requirements for the degree of Master of Science in Manufacturing
Engineering Technology, July 2022. [18] X. Zhang, Y. Liu, and Z. Wang, "Automated PCB
Inspection Using Transfer Learning with Convolutional Neural
[4] A. A. R. M. A. Ebayyeh and A. Mousavi, "A Review and Networks," in IEEE Transactions on Automation Science and
Analysis of Automatic Optical Inspection and Quality Monitoring Engineering, vol. 18, no. 2, pp. 674-684, April 2021. DOI:
Methods in Electronics Industry," IEEE Access, vol. 8, pp. 183192- 10.1109/TASE.2020.2975799.
183271, 2020.
[19] H. Wu, L. Zhang, and S. Li, "Fault Diagnosis in PCBs Using
[5] F. Ulger, S. E. Yuksel, A. Yilmaz, and D. Gokcen, "Solder Deep Belief Networks," in IEEE Transactions on Reliability, vol. 70,
Joint Inspection on Printed Circuit Boards: A Survey and a Dataset," no. 2, pp. 622-634, June 2021. DOI: 10.1109/TR.2020.2991615.
IEEE Transactions on Instrumentation and Measurement, vol. 72, pp.
1-21, 2023, Art no. 2515121. [20] Q. Zhou, Y. Wang, and W. Chen, "Enhancing PCB Defect
Detection with Feature Fusion and Attention Mechanism in
[6] T. T. A. Pham, D. K. T. Thoi, H. Choi, and S. Park, "Defect Convolutional Neural Networks," in IEEE Transactions on
Detection in Printed Circuit Boards Using Semi-Supervised Instrumentation and Measurement, vol. 71, no. 9, pp. 1-10,
Learning," Sensors, vol. 23, no. 6, pp. 3246, Mar. 19, 2023. September 2022. DOI: 10.1109/TIM.2022.3172082.
[7] X. Wang, J.-S. Gao, B.-J. Hou, Z.-S. Wang, H.-W. Ding, and [21] Y. Chen, L. Zhang, and X. Liu, "An Ensemble Learning
J. Wang, "A Lightweight Modified YOLOX Network Using Approach for PCB Defect Classification," in IEEE Transactions on
Coordinate Attention Mechanism for PCB Surface Defect Industrial Electronics, vol. 69, no. 6, pp. 5365-5374, June 2022. DOI:
Detection," IEEE Sensors Journal, vol. 22, no. 21, pp. 20910-20920, 10.1109/TIE.2021.3093302.
Nov. 1, 2022. doi: 10.1109/JSEN.2022.3208580.
[22] S. Gupta, R. Kumar, and P. Jain, "Improved PCB Inspection
[8] F. Ulger, S. E. Yuksel, A. Yilmaz, and D. Gokcen, "Solder Using Deep Reinforcement Learning," in IEEE Transactions on
Joint Inspection on Printed Circuit Boards: A Survey and a Automation Science and Engineering, vol. 19, no. 3, pp. 1287-1296,
July 2022. DOI: 10.1109/TASE.2021.3086282.
Dataset," IEEE Transactions on Instrumentation and
Measurement, vol. 72, Article Sequence Number: 2515121, pp. 1-21, [23] M. Liu, Y. Wang, and Q. Zhu, "Efficient PCB Defect
May 19, 2023. doi: 10.1109/TIM.2023.3277935. Detection using Sparse Autoencoders," in IEEE Transactions on
Components, Packaging and Manufacturing Technology, vol. 11, no.
[9] A. T. Erozan, S. Bosse, and M. B. Tahoori, "Defect Detection 9, pp. 1508-1516, September 2022. DOI:
in Transparent Printed Electronics Using Learning-Based Optical 10.1109/TCPMT.2021.3113639.
Inspection," IEEE Transactions on Very Large-Scale Integration
(VLSI) Systems, vol. 29, no. 8, pp. 1505-1517, Aug. 2021. doi: [24] A. Sharma, R. Verma, and N. Gupta, "Robust PCB
10.1109/TVLSI.2021.3082476. Inspection System Using Transfer Learning and Data
Augmentation," in IEEE Transactions on Instrumentation and
[10] Q. Ling and N. A. M. Isa, "Printed Circuit Board Defect Measurement, vol. 72, no. 1, pp. 1-11, January 2023. DOI:
Detection Methods Based on Image Processing, Machine Learning 10.1109/TIM.2022.3156790.
and Deep Learning: A Survey," IEEE Access, vol. 11, pp. 15921-
15944, Feb. 14, 2023. doi: 10.1109/ACCESS.2023.3245093. [25] J. Li, H. Zhang, and C. Wang, "Adversarial Training for
[11] M. Hasan, S. M. Aziz, and K. Smith, "Image Processing Improved Generalization in PCB Defect Detection," in IEEE
Techniques for PCB Defect Detection: A Review," in IEEE Transactions on Industrial Informatics, vol. 18, no. 6, pp. 4035-4043,
Transactions on Industrial Electronics, vol. 65, no. 7, pp. 5555-5563, June 2023. DOI: 10.1109/TII.2022.3141066.
July 2018. DOI: 10.1109/TIE.2017.2779684.
[12] A. Gupta, R. Sharma, and P. Agarwal, "Machine Learning
Approaches for Automated PCB Inspection," in IEEE Transactions
on Components, Packaging and Manufacturing Technology, vol. 9,
no. 3, pp. 506-515, March 2019. DOI:
10.1109/TCPMT.2018.2876628.
[13] Y. Zhang, L. Wang, and Q. Liu, "Deep Learning-Based PCB
Defect Detection Using Convolutional Neural Networks," in IEEE
Access, vol. 7, pp. 133040-133049, 2019. DOI:
10.1109/ACCESS.2019.2938032.
[14] S. Patel, M. Desai, and K. Mehta, "FPGA Implementation
of PCB Defect Detection Using Artificial Neural Networks," in IEEE
Transactions on Circuits and Systems II: Express Briefs, vol. 67, no.
9, pp. 1688-1692, September 2020. DOI:
10.1109/TCSII.2019.2954568.

15th ICCCNT IEEE Conference,

June 24-28,
Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). 2024, on March 10,2025 at 14:21:20 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IIT - Mandi, Kamand, India