2023 International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE) | 978-1-6654-9260-7/23/$31.

00 ©2023 IEEE | DOI: 10.1109/IITCEE57236.2023.10090861

A Pre-trained YOLO-v5 model and an Image

Subtraction Approach for Printed Circuit Board
Defect Detection
Anagha Lailesh K Jolinson Richi A Preethi N
Department. of Data Science Department. of Data Science Department. of Data Science
Christ University Christ University Christ University
Pune, Lavasa, India Pune, Lavasa, India Pune, Lavasa, India
[email protected] [email protected] [email protected]

Abstract— Almost every electronic product used regularly computer vision in terms of effectiveness, reliability, and
contains printed circuit boards, which in addition to being used quality. Numerous studies on finding PCB faults have been
for business purposes are also used for security applications. conducted in the last several years. However, the studies have
Manual visual inspection of anomalies and faults in circuit yet to be extremely successful and, regrettably, cannot find
boards during manufacture and usage is extremely challenging.
Due to a shortage of training data and the uncertainty of new
microscopic problems. Despite using modern image
abnormalities, identifying undiscovered flaws continues to be processing techniques like image subtraction, the PCB
complicated. The YOLO-v5 technique on a customized PCB manufacturing sector still needs to catch up with the quality
dataset is used in the study to incorporate computer vision to inspection process [1]. With the growing popularity of
detect six potential PCB defects. The algorithm is designed to be consumer electronics, accurate PCB fabrication has become
feasible, deliver precise findings, and operate at a considerable even more important. To suit the requirements of complicated
pace to be effective. A technique of image subtraction is also systems, the designs of PCBs are becoming increasingly
implemented to detect flaws in printed circuit boards. The sophisticated as the electronics industry develops, containing
structural similarity index, a perception-based method, gauges many layers, concealed vias, and dormant components [19].
how similar non-defective and defective PCB images are to one
another.
The intricacy of the system provides a great chance for
potential attackers to deliberately change the design and
Keywords—printed circuit board, defect detection, YOLO-v5, reveal vulnerabilities in the PCB distribution network. There
image subtraction, structural similarity index, computer vision, are not many instances of PCB fault detection techniques in
electronic manufacturing, automatic optical inspection the literature. Convolutional neural networks (CNNs) have
achieved notable advancements in various applications,
I. INTRODUCTION including object and image identification. Specifically, object
The fundamental element of several electronic devices, recognition techniques and CNNs based on regions are used
including motherboards and graphics cards, is a printed to perform object detection.
circuit board. PCBs build electronic channels and offer
physical stability such that they enable sophisticated AOI and AVI devices are utilized in conjunction with
communications that are required to take place continuously these techniques. PCB businesses must educate and employ
for electrical gadgets to function. To meet expectations and a sizable number of people for quality inspection following
needs while maintaining quality standards, PCBs—the core the conventional inspection procedure, in addition to utilizing
unit of every electronic product—must be developed and such a costly machine. In addition to investing in such an
manufactured with extreme accuracy. PCBs are constructed expensive machine, the PCB sector makes significant
of fiberglass, composite epoxy, and laminated materials that investments in the training and upkeep of a sizable staff
hold chips, transistors, capacitors, and other electrical dedicated to quality assurance using conventional inspection
components. Many PCBs created during the manufacturing instruments [2]. To overcome these industrial difficulties,
process include flaws because of improper handling or several computer vision algorithms have been created by
technical issues with the manufacturer. The flawed or researchers to address the problems and obstacles
defective PCBs should be identified and separated during the encountered in earlier studies. When the background is
quality inspection. distinct from the flaw and the fault patterns on the surfaces
are consistent, conventional image processing techniques can
There have historically been several conventional produce the desired results [16]. In such scenarios, flaw
approaches to identifying flaws, but the techniques have been identification is frequently aided by methods like corner
ineffective or too quick. High-quality pictures required for detection, thresholding in gray-scale pictures, and image
visual examination require superior optical illumination segmentation.
systems and sufficient image-capturing hardware [20].
Through optoelectronic components and sensors, machine Numerous you-only-look-once (YOLO) techniques have
vision automatically captures and interprets pictures of actual been used by scientists to analyze printed circuit boards with
objects. Defect identification is significantly enhanced by incredible accuracy [3]. YOLO is distinct from conventional

978-1-6654-9260-7/23/$31.00 2023
c IEEE 140

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.
 categorization techniques and can categorize many objects in combining techniques such as mosaic data augmentation and
a single image, in contrast to traditional classification the K-means clustering algorithm [13]. The model produced
algorithms. Since convolutional neural networks are the amazing outcomes and complied with real-time demands.
foundation of YOLO, the algorithm has a special approach to Yongjian et al. (2022) employed YOLO-v5 as the foundation
object detection. A single convolutional neural network of a cross-shaft surface anomaly detection method, and as the
predicts several bounding box coordinates and class flaws were smaller in size, the K-Means method was utilized
probabilities for the areas concurrently, making YOLO a very to generate the new anchors [17]. The enhanced YOLOv5
quick method for object detection [18]. The YOLOv5 model had a higher mean average precision, a shorter
inference time, and a smaller model size than the previous
technique adapts incredibly fast, which lowers the cost of
design, which greatly increased the model's detection
experimenting. With YOLOv5, one can quickly learn from
effectiveness.
single photos, sets of photographs, or streaming media.
Several image processing and image subtraction methods
Images may be made lighter or dimmer by using were implemented in the following works for printed circuit
mathematical functions. Subtracting photos may be used for board defect detection as well as other areas.
a variety of tasks, such as contrasting or determining the By applying the primary subtraction technique, one could
variation of related photographs. The template-matching assess the effectiveness of the image processing method
approach is specifically designed to find PCB faults. Another devised by Bonello et al. (2018) and validate the PCB image
approach utilizing OpenCV is the image subtraction binarization's runtime reduction percentage [6]. The approach
approach. The structured similarity index measure can be proposed by R. VijayaGeetha and M. VijayaBharathi (2021)
utilized for a variety of tasks besides image processing since for testing and defect detection on single-layered printed
the method compares and quantifies two signals to determine circuit boards was straightforward to apply and used basic
a matching score, ignoring the difficulty of calculating the mathematical formulas in the image processing tool [7].
SSIM compared to the mean squared error [4]. Munaweera Tanthirige (2022) used the template
comparison approach to build the fault detection system,
The work implements the YOLO-v5 model and an image which compares a reference PCB without any faults to an
subtraction approach along with a structural similarity index assessment picture to find flaws and abnormalities [14]. At
metric to distinguish between PCB photos that are flawed and idealized experiment conditions, the flaw detection method
those that are not. demonstrated incredible precision; however, the model
proved unsatisfactory when used in actual testing
II. RELATED WORKS circumstances.
The section provides a concise yet comprehensive
In contrast to merely evaluating individual image pixels,
summary of the incorporation of computer vision algorithms,
Bergmann et al. (2019) suggested a subjective loss function
including the you-only-look-once algorithm and image
related to the structural similarity index measure that analyses
subtraction techniques.
relations between local picture areas while considering
V. A. Adibhatla et al.'s (2021) YOLO-v5 model for brightness, intensity, and spatial arrangement [15]. On a
detecting PCB flaws was quickly established as the most difficult dataset obtained from real nanofibrous materials, the
effective, precise, and quick model available [1]. The model delivers appreciable performance gains. The approach
detection and localization of many flaws in PCB pictures were also addresses the difficulty of detecting visually modified
accomplished by Abhiroop Bhattacharya and Sylvain G. faulty sections when intensity levels are essentially constant.
Cloutier (2022) with a comprehensive deep-learning
architecture and a single-step object detection model [2]. The III. IMPLEMENTATION
GSC YOLO-v5 model volume was found to be compressed as A publicly available synthetic printed circuit board
a consequence of the research observations of the model, dataset, the HRIPCB [8] dataset, comprising 1386
which incorporates the dual attention mechanism and photographs of six kinds of flaws, namely, missing holes,
lightweight network presented by L. Wu et al. (2022) mouse bites, open circuits, shorts, spurs, and spurious copper,
[5]. created digitally on non-defective PCB images using Adobe
The implementation of YOLO-v5 models for other Photoshop, was utilized in the work.
relevant surface defect detection is as follows:
The dataset was used to train and validate the YOLO-v5
To identify particle board surface flaws, Zhao et al. (2021)
model. The annotations, in the required format, were also
suggested the PB-YOLO-v5 model of the YOLO-v5
available with the dataset.
algorithm [11]. Several modules were developed to decrease
the total dimension of the model and increase the model's
adaptability. The accuracy of textile flaw identification and
defect localization was swiftly and precisely increased with an
enhanced YOLOv5 object detector developed by Wang et al.
(2021) [12]. An approach named Adaptively Spatial Feature
Fusion was used to ameliorate PANet's negative impact on
multi-scale feature fusion and boost the detection performance
of tiny objects.
An upgraded YOLO-v5 algorithm for detecting solar cell
defects was put out by Meng Zhang and Liju Yin (2022), Fig. 1 Sample HRIPCB images [8]

International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE) 141

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.
 To test and make predictions, actual PCB manufacturing The training was performed on 552 images in 16 batches
sector data is also utilized, which had to be modified later to over 100 epochs. The mean average precision increased from
make accurate predictions. 0 to 0.821 on reaching the final epoch. 100 epochs were
completed in 12 minutes and 14 seconds, utilizing 2.08
gigabytes of graphics processing unit memory. The best
weights were chosen, and a set of 138 synthetic images were
used for validation purposes.

TABLE I TRAINING EPOCHS WITH LOSSES AND MEAN AVERAGE PRECISION

Epoch Box loss Object loss Class loss mAP50
1 0.1036 0.01301 0.04108 0
10 0.07726 0.01445 0.0269 0.00994
20 0.05409 0.01404 0.02216 0.281
30 0.05458 0.01282 0.02074 0.427
40 0.04157 0.01266 0.01724 0.472
50 0.04301 0.01172 0.01399 0.596
60 0.03849 0.0116 0.01191 0.683
Fig. 2 Sample image from the PCB manufacturing sector 70 0.03717 0.01081 0.009906 0.719
80 0.03311 0.01045 0.00843 0.775
A. YOLO-v5 90 0.03382 0.01084 0.008292 0.797
100 0.03089 0.01003 0.007015 0.821
An equally sized grid structure is used by YOLO
techniques to split up all the input pictures. Object detection
is the task of every cell. Therefore, the bounding boxes for B. Image Subtraction
the spotted object are determined by those coordinates. The The image subtraction approach used in the study
basic characteristics available for each box are contrasts a template and a test printed circuit board picture by
the coordinates on the axes x and y, the aspect ratio of the turning both actual pictures into grayscale, blurring them with
entity, and a class probability indicating the likelihood that Gaussian blur, and then turning them into binary images
the object is contained in the box. using adaptive thresholding.

The core components of YOLO v5, like all the previous The Gaussian blur is a sort of graphics filter that
single-stage object detector systems, are the head, neck, and determines the alteration to implement for each grid cell in
backbone. the picture using a Gaussian filter. Equation (1), where x is
the distance from the origin in the horizontal axis, y is the
The basic purpose of the backbone is to retrieve distance from the origin in the vertical axis, and σ is the
significant characteristics from an input picture. To extract standard deviation of the Gaussian distribution, yields a
valuable, important characteristics from an input picture in Gaussian function in two dimensions.
YOLO v5, the Cross-Stage Partial Networks are employed as
the backbone. The primary purpose of the neck is to produce ೣమ శ೤మ
ଵ ି
‫ܩ‬ሺ‫ݔ‬ሻ ൌ ݁ మ഑మ (1)
feature pyramids. Feature pyramids enable systems to scale ଶగఙ మ
entities successfully, in particular. The ability to recognize
the same entity in various sizes and scales is beneficial. The The testing picture is eliminated from the template until
final detection step is mostly carried out using the head. The the images have been converted to binary to acquire the
head uses anchor boxes on the attributes and produces final output image and reveal the flaws. The number of faults in
output vectors. Fig. 3 provides the YOLO-v5 architecture the output photo is then determined using contour detection
comprising the model's backbone, neck, and head. [10].

The structural similarity index measure (SSIM) is used to

determine how similar the reference and the testing
pictures are to one another. An SSIM value close to zero
indicates that the images are vastly different from each other,
whereas a value close to unity means the images are
remarkably similar. The range of SSIM lies between zero and
one. Equation (2) could be used in the calculation of SSIM,
where ‫ ݔ‬and ‫ ݕ‬are two ܰ ൈ ܰ sized windows of a picture, ߤ௫
and ߤ௬ are the pixel sample means, ߪ௫ଶ and ߪ௬ଶ are the
variances, ߪ௫௬ is the covariance and ܿଵ and ܿଶ stabilize the
division.

൫ଶఓೣ ఓ೤ ା௖భ ൯ሺଶఙೣ೤ ା௖మ ሻ

ܵܵ‫ܯܫ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ  మ ାఓ మ ା௖ ൯ሺఙ మ ାఙ మ ା௖ ሻ (2)
൫ఓೣ ೤ భ ೣ ೤ మ

Fig. 3 The YOLO-v5 framework [9]

142 International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE)

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.
 The same method is also used in an attempt to find flaws. enough since the model marked many areas as defective
The defective printed circuit board picture is processed using classes even on the non-defective board, which is an
several morphological methods based on its characteristics, implication that the model is overfitted on the synthetic
including enlarging, reducing, and gradient. The identified dataset and does not work well with real data.
contours are looped over and constrained by rectangular
boxes in the final stage.
IV. RESULTS
The validation of the YOLO-v5 model using the HRIPCB
dataset produced appreciable outputs, and the results are
given in Fig. 4(a), 4(b), and 4(c).

Fig. 5 Non-defective areas on non-defective PCB detected as the missing

hole and spurious copper defects

The next step to overcome the challenge was to create

artificial defects on the real, non-defective PCB images. The
approach worked quite accurately since the model was also
trained on digitally fabricated defects and not real ones.
(a) F1-score VS Confidence

(b) Precision VS Confidence

Fig. 6 Detection of various artificially created defects

Results from the image subtraction approach were

moderately precise. The same defective PCB image in Fig. 6
was used as a test image for the experiment. The input and
output images of the same are shown in Figs. 7(a) and 7(b).

The structural similarity index measure was able to detect

and draw bounding boxes over many of the defects, but the
method cannot be considered very feasible. Fig. 8 shows the
output obtained while using the image in Fig. 5 as the
template image.

The outcomes of the image subtraction method and the

YOLO-v5 model on various kinds of PCB defects are
(c) Recall VS Confidence
contrasted in Table II. An observation that can be made from
Fig. 4 Various evaluation metrics VS Confidence the outputs is that many defects are not detected and that
many non-defective areas have been detected as defective.
The testing was initially done on several real, non- The inference proves that both approaches, the you-only-
defective PCB images to check whether the model worked look-once algorithm, and the image subtraction technique,
well or not. Fig. 5 shows that the results were not accurate produce only moderately accurate results with real data.

International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE) 143

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.
 TABLE III SSIM FROM DIFFERENT EXPERIMENTS
Experiment Testing image type SSIM Accuracy

1 Defective PCB image of 0.9980 93.2%

the same template image
2 Different defective PCB 0.1412 32.4%
image

V. CONCLUSION
The work uses two separate methods for finding
anomalies on single-sided printed circuit boards: YOLO-v5
and an image subtraction approach. These techniques might
be used to overcome the challenges associated with the
manual examination of printed circuit board flaws. The
YOLO-v5 model's flaw detection suggests that the model is
overfitted since it performs exceptionally well on synthetic
data but poorly on actual data. Due to the need for high-
resolution pictures to get accurate findings, the image
(a) Non-defective PCB reference image subtraction approach utilized is likewise not particularly
dependable. However, by incorporating the understanding of
object identification algorithms, the technique may be further
expanded. The fundamental issue is the scarcity of publicly
accessible PCB industry data that is required. However, in
terms of computing speed and training accuracy, YOLO-v5
may be regarded as an effective model for detecting printed
circuit board defects.
For improved performance and speed, future work could
implement the most recent version of YOLO, such as the
YOLO-v7 algorithm. Strategies to expand the amount of
publicly available industry data might also be taken into
consideration so that better accuracy could be achieved for
defect detectors in printed circuit boards. Future efforts could
(b) Only two defects were detected out of many different defects.
take into account more varieties of flaws and early detection
as well.
Fig. 7 Results obtained with image subtraction
REFERENCES
[1] V. A. Adibhatla et al., “Applying deep learning to defect detection in
printed circuit boards via a newest model of you-only-look-once,”
Mathematical Biosciences and Engineering, vol. 18, no. 4. American
Institute of Mathematical Sciences (AIMS), pp. 4411–4428, 2021.
[2] Bhattacharya, A., Cloutier, S.G. End-to-end deep learning framework
for printed circuit board manufacturing defect classification. Sci Rep
12, 12559 (2022).
[3] Adibhatla, V.A.; Chih, H.-C.; Hsu, C.-C.; Cheng, J.; Abbod, M.F.;
Shieh, J.-S. Defect Detection in Printed Circuit Boards Using You-
Only-Look-Once Convolutional Neural Networks. Electronics 2020, 9,
1547.
[4] Mudeng, V.; Kim, M.; Choe, S.-w. Prospects of Structural Similarity
Index for Medical Image Analysis. Appl. Sci. 2022, 12, 3754.
Fig. 8 Bounding boxes drawn around defects [5] L. Wu, L. Zhang and Q. Zhou, "Printed Circuit Board Quality
Detection Method Integrating Lightweight Network and Dual
Attention Mechanism," in IEEE Access, vol. 10, pp. 87617-87629,
TABLE II COMPARISON OF ACCURACIES OF THE STRATEGIES IN FINDING
2022.
TWO KINDS OF DEFECTS
[6] Bonello, Daniel & Iano, Yuzo & Neto, Umberto. (2018). A New Based
Technique Missing hole Mouse bite Average accuracy Image Subtraction Algorithm for Bare PCB Defect Detection.
International Journal of Multimedia and Image Processing. 8. 438-442.
YOLO-v5 98.7% 84.2% 87.9% [7] Geetha, R & Bharathi, M.. (2021). Printed Circuit Board Defect
Detection Using Mathematical Expression in Image Processing.
Image 86.1% 72.6% 79.7% [8] Huang, W., Wei, P., Zhang, M. and Liu, H. (2020), HRIPCB: a
subtraction challenging dataset for PCB defects detection and classification. The
Journal of Engineering, 2020: 303-309.
[9] Xu, Renjie & Lin, Haifeng & Lu, Kangjie & Cao, Lin & Liu, Yunfei.
Additionally, two experiments were conducted in which (2021). A Forest Fire Detection System Based on Ensemble Learning.
different sets of reference and testing images were compared Forests. 12. 217.
using the structural similarity index measure so that the faulty [10] Rajput, Muskaan. (2021). Comparative Study of Image Processing and
detection could be analyzed, considering SSIM as a factor. Transfer Learning Techniques for an Automated PCB Fault Detection

144 International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE)

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.
 System. International Journal for Research in Applied Science and [16] Bhatt, P. M., Malhan, R. K., Rajendran, P., Shah, B. C., Thakar, S.,
Engineering Technology. 9. 5090-5097. 10.22214/ijraset.2021.35775. Yoon, Y. J., and Gupta, S. K. (February 9, 2021). "Image-Based
[11] Zhao, Z., Yang, X., Zhou, Y. et al. Real-time detection of particleboard Surface Defect Detection Using Deep Learning: A Review." ASME. J.
surface defects based on improved YOLOV5 target detection. Sci Rep Comput. Inf. Sci. Eng. August 2021; 21(4): 040801.
11, 21777 (2021). [17] Z. Yongjian, D. Weibo, K. Tang, and Y. Xinbo, “An Improved Yolov5-
[12] Wang, Ying & Hao, Zhengyang & Zuo, Fang & Pan, Shanshan. (2021). Based Algorithm for Cross Shaft Surface Defect Detection,” SSRN,
A Fabric Defect Detection System Based Improved YOLOv5 Detector. 2022.
Journal of Physics: Conference Series. 2010. 012191. [18] Wang, Z., Zhu, H., Jia, X., Bao, Y., & Wang, C. (2022). Surface defect
[13] M. Zhang and L. Yin, "Solar Cell Surface Defect Detection Based on detection with modified real-time detector YOLOv3. Journal of
Improved YOLO v5," in IEEE Access, vol. 10, pp. 80804-80815, 2022. Sensors, 2022, 1-10. [8668149].
[14] Munaweera Tanthirige, N. K. D. S. (2022). Printed circuit board defect [19] Mehta, D.; Lu, H.; Paradis, O.P.; MS, M.A.; Rahman, M.T.; Iskander,
detection using image processing [Master’s thesis, Minnesota State Y.; Chawla, P.; Woodard, D.L.; Tehranipoor, M.; Asadizanjani, N. The
University, Mankato]. Cornerstone: A Collection of Scholarly and Big Hack Explained: Detection and Prevention of PCB Supply Chain
Creative Works for Minnesota State University, Mankato. Implants. J. Emerg. Technol. Comput. Syst. 2020, 16, 1–25.
[15] Bergmann, Paul et al. “Improving Unsupervised Defect Segmentation [20] Sachdeva, K., Aggarwal, S., Verma, A., Chawla, S. (2023). Research
by Applying Structural Similarity to Autoencoders.” VISIGRAPP Trends in Image Processing and Defect Detections. In: Khanna, A.,
(2019). Gupta, D., Kansal, V., Fortino, G., Hassanien, A.E. (eds) Proceedings
of Third Doctoral Symposium on Computational Intelligence . Lecture
Notes in Networks and Systems, vol 479. Springer, Singapore.

International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE) 145

Authorized licensed use limited to: Ecole National D'Ingenieurs De Tunis (ENIT). Downloaded on March 10,2025 at 14:18:45 UTC from IEEE Xplore. Restrictions apply.