에르셀릭우우르
(Ugur Ercelik)
**iD
김거식
(Keo Sik Kim)
*iD
박형준
(Hyoung-Jun Park)
*iD
김경백
(Kyungbaek Kim)
†iD
-
(Electronics and Telecommunications Research Institute (ETRI), South Korea. Email :
keosikis@etri.re.kr, spacegon@etri.re.kr)
-
(Dept. of Artificial Intelligence Convergence, Chonnam National University, Gwangju,
South Korea. Email: ugurercelik@hotmail.com)
Copyright © The Korean Institute of Electrical Engineers
Key Words
Computer Vision, Weld Defect Detection, Industrial AI, Channel Attention, Spatial Attention, YOLOv8
1. Introduction
Welding technology is common in various industrial applications such as manufacturing,
infrastructure, aerospace and shipbuilding. Therefore, the quality and strength of
weld is vitally important part of industry. This technology ensure strong structural
integrity by combining multiple steel or metal using heat and pressure in industrial
ares [1]-[2]. During the use of these methods, several defects arise such as porosity, spatter,
overlap, and undercut, due to irregularities and complexities in the weld surface
and beads and these may cause quite dangerous results depending on the fields of usage
[3]. In transformer power equipment, major welding defects include oil leakage caused
by poor weld quality such as pinholes, porosity, and lack of fusion, as well as cracks
and leakage resulting from frequent electromagnetic vibrations. These defects generally
originate from various factors, including reduced weld quality, material fatigue,
thermal overload, and prolonged operational conditions. Such issues not only degrade
the reliability of but also shorten its service life and pose significant safety concerns.
In these cases, non-destructive testing(NDT), which obstruct damage to the weld material
and product, is significantly important. Common non-destructive detection(NDT) of
welds contains ultrasonic testing, radiographic, penetration and laser [2]
[6]
[8]. However, these methods have some challenges such as ultrasonic testing results visual
lackness, radiographic testing cause a health problems, and penetration testing is
time-consuming. Also, Laser-based NDT require high equipment costs and high skilled
personel and environmental sensitivity [3]. For these purposes, we propose improved YOLOv8 model with Convolutional Block Attention
Module for automatic small defect detection model to point out defects on weld beads.
The CBAM module handles the defects in terms of enhancing and capturing important
features of defects and emphasize. The proposed system aims to provide a fast, accurate
and automated detection model. In summary, the main contributions consist of the following:
1) A novel small defect detection model (YOLOv8-CBAM) is designed based on YOLOv8
for detecting surface defects on weld beads. By integrating CBAM to neck part of YOLOv8,
adaptive attention mechanism provided. Moreover, feature discrimination and spatial
focus is enhanced.
2) Extensive experiments are conducted on a weld bead defect dataset, and the proposed
model is compared with the baseline YOLOv8 and other state-of-the-art detection methods.
The results demonstrate that our model achieves superior accuracy, particularly in
small defect detection.
3) An ablation study is performed to validate the effectiveness of the CBAM integration,
showing the contribution of channel and spatial attention to feature representation.
4) The study provides practical insights into applying attention-enhanced deep learning
models for industrial quality inspection, which can be further extended to other manufacturing
processes.
The remainder of this paper is structured as follows: Section 2 provides review of
related works, section 3 presents the proposed methodology and contributions, section
4 provides experiments, results and evaluations. Finally, section 5 concludes the
study and outlines future research directions.
2. Related Works
2.1 Defect Detection in Welding
Due to the inability of traditional and human-based methods, the improvement of AI-based
methodologies has been targeted. Liu et al. [1] proposed lighter and faster YOLOv5 model to point out the differences of shape and
multiscale problem on multi defect types. RMF module is integrated to extract parameter-based
and parameter-free multiscale information and EFE module to improve the performance
of detection network. Block et al. [2] proposed a novel dataset and experiments in welding. The LoHi-WELD dataset solve
shortage of vast and reliable datasets by utilizing augmentation methods and CNN,
YOLO-based models. A important contribution was made by Wang et al. [3], who improved Yolo with MSAPF module to improve low accuracy caused by interference
information on weld defect detection. The proposed model aimed to eliminate interference
information and enhance necessary features in each scale in the same time by integrate
multiscale alignment fusion(MSAF) with parallel features filtering (PFF) modules.
2.2 YOLOv8 Based Defect Detection
The development of YOLOv8 has made it easier to identify objects which have complexity
in surfaces. Chen et al. [10] proposed a lightweight YOLOv8-OCHD wood surface defect detection method to handle
the drawbacks of manual defect detection and the challenges of traditional visual
inspection algorithms such as high missed detection rates, slow detection speeds.
Mao et al. [11] presents an enhanced YOLOv8 based framework which integrated novel attention mechanisms
and advanced architectural modules to improve detection accuracy and robustness to
point out fabric defect detection in textile industry. Their framework incorporate
the SimAM attention mechanism with the SPPF module and adopts an optimized Dilation-wise
Residual structure in the backbone. Another research in scope of defect detection
with YOLOv8 proposed by Zhang et al. [12]. They designed a novel GDM-YOLO model, tailored for steel surface defect detection
tasks. In the backbone network, they use Space-to-Depth- Ghost Convolution(SPDG) downsampling
module to aim at minimizing information loss during downsampling operations. Secondly,
their research introduces the C2f-Dilated-Reparam-Block(C2f-DRB) module to leverage
reparameterization and larger kernel convolutions for enhancing feature extraction
capabilities without inference costs.
3. Proposed Attention-Enhanced YOLOv8 with CBAM
3.1 CBAM
CBAM is a lightweight attention module that refines feature maps by sequentially applying
channel attention and spatial attention. The channel attention mechanism helps the
model emphasize ‘what’ is important, while spatial attention guides the model to focus
on ‘where’ the relevant information is located. The channel attention module exploits
global average pooling and max pooling operations, followed by shared multilayer perceptrons(MLPs),
to compute an attention map that highlights the most discriminative feature channels.
Subsequently, the spatial attention module utilizes convolution operations over concatenated
average-pooled and max-pooled feature maps along the channel axis to generate a spatial
attention map. This process allows the network to selectively focus on relevant regions
in the image where defects are more likely to occur. The channel attention map is
computed as:
where $\sigma$ denotes the sigmoid activation function. The refined feature map after
channel attention is:
The spatial attention module emphasize where to focus. It applies average pooling
and max pooling operations along the channel axis and concatenates them. A convolutional
operation with a 7 x 7 kernel is then applied, followed by a sigmoid activation:
The output feature map after spatial attention is:
By sequentially combining both channel and spatial attention, the final CBAM enhanced
feature representation can be expressed as:
This two-step attention mechanism enables the network to adaptively focus on meaningful
features while suppressing irrelevant information, thereby improving detection performance.
Figure 1 shows the architectural design of CBAM.
Fig. 1. CBAM Architecture.
3.2 Integration of YOLOv8-CBAM
To enhance the capability of YOLOv8 in detecting small weld surface defects, we integrated
Convolutional Block Attention Module(CBAM) into both the backbone and the neck of
the network. The integration process is designed to improve feature representation
by selectively focusing on informative regions while suppressing irrelevant background
noise. Figure 2 presents YOLOv8-CBAM Architecture.
Fig. 2. Attention Enhanced YOLOv8-CBAM Architecture.
3.2.1 Backbone Modification
The original YOLOv8 backbone employs a series of C2f modules to perform feature extraction
at multiple scales. In our proposed design, each C2f block was replaced with a C2f_CBAM
block, in which a CBAM unit is appended after the residual feature aggregation. This
modification allows the model to refine intermediate features by applying sequential
channel and spatial attention. The channel attention emphasizes the most discriminative
feature maps, while the spatial attention highlights critical local regions corresponding
to defect patterns.
3.2.2 Neck Modification
In addition to the backbone, CBAM modules were inserted into the neck of YOLOv8 at
several stages of feature fusion. Specifically, after upsampling and concatenation
operations at different pyramid levels, CBAM modules were applied before feeding the
features into subsequent C2f blocks. This integration enables the network to adaptively
recalibrate multiscale features during the feature aggregation process. By doing so,
redundant or noisy features are suppressed, while important spatial cues are preserved
across pyramid levels (P3, P4, P5).
3.2.3 Overall Benefits
The dual integration of CBAM in both the backbone and neck ensures that attention
is applied not only during the feature extraction stage but also during the feature
fusion stage. This comprehensive attention enhancement allows the network to capture
subtle defect regions more effectively, while maintaining robustness agains complex
backgrounds typically observed in weld bead surfaces.
3.2.4 Data Collection and Preprocessing
Weld images which contains defects were obtained by recording videos of welding plates.
The recorded videos were then cropped to extract image sequences corresponding to
weld regions by using a custom OpenCV-based method (Figure 3). The final weld defect dataset consists of 1,539 images, divided into 1,074 training
images, 309 validation images, and 156 testing images. The dataset covers four major
defect categories: overlap, spatter, porosity, and undercut. The defects consists
of small-scale, low-contrast defect regions that often occupy less than 1-3% of the
image area. Figure 4 shows representative examples of four types of small defects, highlighting the challenges
related to scale variation and subtle texture differences. Also, to improve the model’s
robustness and generalization, data augmentation techniques were employed. Basic augmentations
included horizontal flipping (p=0.5), hue adjustment (±0.015), saturation adjustment
(±0.7), and value adjustment (±0.4). In addition, advanced augmentation strategies
such as MixUp (p=0.2) and Mosaic (p=1.0) were applied to further enhance the diversity
of training samples. Each model was trained with parameters of 300 epoch, 20 patience,
and 1024 image size.
Fig. 3. Data Preprocessing.
Fig. 4. Four types of defects on custom weld dataset.
Table 1. Detection results of each defect types (P: Precision, R: Recall)
|
Model
|
Overlap
|
Porosity
|
Spatter
|
Undercut
|
|
P
|
R
|
mAP 50
|
mAP 50-95
|
P
|
R
|
mAP 50
|
mAP 50-95
|
P
|
R
|
mAP 50
|
mAP 50-95
|
P
|
R
|
mAP 50
|
mAP 50-95
|
|
Yolov5
|
83.3
|
90.0
|
93.9
|
74.2
|
69.2
|
70.6
|
77.6
|
40.8
|
70.4
|
68.7
|
76.2
|
43.5
|
73.4
|
74.5
|
80.2
|
42.1
|
|
Yolov8
|
80.8
|
89.8
|
94.1
|
73.3
|
77.5
|
69.9
|
78.2
|
40.5
|
70.6
|
66.9
|
74.8
|
42.2
|
67.8
|
70.7
|
75.8
|
39.1
|
|
Yolov10
|
84.8
|
89.3
|
93.8
|
73.5
|
67.3
|
67.8
|
73.6
|
39.1
|
70.9
|
68.9
|
74.3
|
44.0
|
75.0
|
65.5
|
74.6
|
37.7
|
|
Yolov11
|
85.6
|
87.3
|
94.0
|
74.4
|
72.4
|
62.9
|
73.2
|
39.7
|
65.9
|
73.9
|
76.0
|
43.6
|
69.3
|
74.9
|
75.4
|
38.5
|
|
Yolov8-CBAM
|
93.0
|
89.4
|
95.1
|
75.0
|
72.7
|
77.3
|
77.5
|
40.0
|
71.6
|
74.0
|
76.2
|
44.1
|
72.2
|
70.7
|
75.9
|
37.9
|
4. Experiments
4.1 Metrics
To evaluate the performance of the proposed and baseline models, standard object detection
metrics were employed, including Precision, Recall, mAP50, and mAP50-95. Precision
measures the fraction of correctly detected objects among all actual objects. It evaluates
the ability of the model to find all relevant instances. Formally:
where TP is True Positive (correctly detected objects), FP is False Positive (incorrect
detections).
Recall measures the fraction of correctly detected objects among all actual objects.
It evaluates the ability of the model to find all relevant instances. Formally:
where FN is False Negatives (missed objects). High recall indicates that the model
successfully detects most of the objects in the dataset. Average Precision (AP) summarizes
the precision-recall curve for a given Intersection over Union (IoU) threshold. The
mean Average Precision (mAP) is the average of AP over all classes. IoU between a
predicted bounding box Bp and ground truth Bgt is defined as:
AP at given IoU threshold (e.g., 0.5) is:
where P(R) is precision at recall R.
mAP50 is the mean AP across all classes at IoU threshold 0.5:
mAP50-95 is the mean AP averaged over multiple IoU thresholds from 0.5 to 0.95 in
steps of 0.05:
where N is the number of classes.
Fig. 5. Visualization Results.
Table 2. Detection results of overall (`all` class)
|
Model
|
Precision
|
Recall
|
mAP50
|
mAP50-95
|
|
Yolov5
|
74.1
|
76.0
|
82.0
|
50.0
|
|
Yolov8
|
74.2
|
74.3
|
80.8
|
48.8
|
|
Yolov10
|
74.5
|
72.9
|
79.1
|
48.6
|
|
Yolov11
|
73.3
|
74.8
|
79.7
|
49.0
|
|
Yolov8-CBAM
|
75.2
|
77.6
|
81.0
|
49.0
|
Table 3. Ablation study
|
Model
|
Precision
|
Recall
|
mAP50
|
mAP50-95
|
|
Yolov8
|
74.2
|
74.3
|
80.8
|
48.8
|
|
Yolov8- C2f_CBAM (Backbone only)
|
72.0
|
74.6
|
80.3
|
47.6
|
|
Yolov8-CBAM
|
75.2
|
77.6
|
81.0
|
49.0
|
Table 4. Inference speed analysis
|
Model
|
FPS
|
Inference Time(ms)
|
|
Yolov8
|
38.24
|
6-9 ms
|
|
Yolov8-CBAM
|
32.41
|
11-14 ms
|
4.2 Experimental Results
To validate the effectiveness of the proposed YOLOv8-CBAM model, extensive experiments
were conducted. In Table 1, the results reveal that the integration of CBAM significantly enhances the detection
capability of YOLOv8, especially for complex defect types. For the overlap, YOLOv8-CBAM
achieved the highest precision (93.0%) and mAP@50-95 (75.0%), outperforming all baseline
models. Similarly, for spatter detection, YOLOv8-CBAM reached superior mAP@50 (74.0%)
and mAP@50-95 (44.1%), demonstrating improved small-defect detection. In the case
of porosity, YOLOv8 and YOLOv8-CBAM yielded comparable results, while YOLOv8-CBAM
still offered competitive precision and recall. For undercut, YOLOv5 achieved slightly
higher mAP@50 (80.2%), but YOLOv8-CBAM maintained consistent performance across all
metrics. Also, Table 2 presents the overall detection performance of the evaluated models. YOLOv5 achieved
the highest mAP@50 (82.0%), while YOLOv8 closely followed with 80.8%. However, the
proposed YOLOv8-CBAM model demonstrated the most balanced performance across all metrics,
reaching the highest precision (75.2%) and recall (77.6%). Although its mAP@50(81.0%)
was slightly lower than YOLOv5, it still outperformed YOLOv10 and YOLOv11, and its
competitive mAP@50-95 (49.0%).
4.3 Ablation Study
To further investigate the contribution of the CBAM modules in the proposed YOLOv8-CBAM
model, an ablation study was conducted by systematically enabling and disabling the
attention components:
1. YOLOv8 (baseline): The original YOLOv8 architecture without any CBAM integration.
2. YOLOv8-C2f_CBAM (Backbone only): CBAM modules embedded within the backbone by replacing
standard C2f blocks with C2f_CBAM blocks, while the neck remains unchanged.
3. YOLOv8-CBAM (Backbone+Neck): The full proposed model.
In Table 3, The experimental results present that when CBAM was integrated only into the backbone,
the performance slightly decreased, indicating that the backbone alone is not sufficient
to fully exploit the benefits of the attention mechanism. In contrast, the full proposed
model achieved the best results with 75.2% precision, 77.6% recall, 81.0% mAP50, and
49.0% mAP50-95. Although both YOLOv8 and YOLOv8-CBAM achieved comparable precision
and mAP scores, YOLOv8-CBAM demonstrated a higher recall. This result implies that
YOLOv8-CBAM is able to detect a greater proportion of true weld defects, thus reducing
the likelihood of missing potential defects. In safety-critical domains such as weld
quality assurance, higher recall is essential to minimize the risk of undetected defects,
thereby enhancing the overall reliability and safety of the inspection process. Sample
detection results on test images are presented in Figure 5.
4.4 Inference Speed Analysis
To evaluate the impact of the CBAM integration on real-time usability, we measured
the inference speed(frames per second, FPS) of both the YOLOv8 and the proposed YOLOv8-CBAM
model in Table 4. The standard YOLOv8 model achieved 38.24 FPS, and after integrating the CBAM module,
the inference speed slightly decreased to 32.41 FPS due to additional channel and
spatial attention computations. To measure this, we performed a standardized FPS test
by running each model on a fixed 1024x1024 input. FPS was then computed as the number
of processed frames divided by the elapsed time. Despite increase in computational
cost, the proposed model still maintains real-time performance(≥ 30 FPS), confirming
that the enhancement in detection accuracy does not compromise practical deployability.
5. Conclusion and Future Works
In this study, we proposed YOLOv8-CBAM model by integrating the Convolutional Block
Attention Module (CBAM) into both the backbone and neck of the YOLOv8 architecture
to improve feature discrimination and spatial focus, therefore achieving more reliable
detection of weld surface defects. The experimental results demonstrated that YOLOv8-CBAM
achieved superior performance compared to baseline YOLO models, particularly in terms
of precision and recall, while also maintaining competitive mAP values across different
types. Despite these promising results, there are still directions for future research.
First, additional advanced modules such as Transformer-based or lightweight adaptive
attention blocks can be explored to further enhance the representation power of the
model. Second, extending the detection framework to multi-modal approaches(e.g., combining
ultrasonic signals, infrared, or X-ray images with visual data) may provide more comprehensive
insights for weld quality assessment.
Acknowledgements
This work was supported by an Electronics and Telecommunications Research Institute
(ETRI) grant funded by the Korean government [26ZK1100, Honam region regional industry-based
ICT convergence technology advancement support project].
This work was supported by the Institute of Information & Communications Technology
Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization
program grant funded by the Korea government(MSIT)(IITP-2025-RS-2022-00156287)
References
M. Liu, Y. Chen, J. Xie, L. He, Y. Zhang, 2023, LF-YOLO: A Lighter and Faster YOLO
for Weld Defect Detection of X-Ray Image, IEEE Sensors Journal, Vol. 23, No. 7, pp.
7430-7439
S. Biasuz Block, R. Dutra da Silva, A. Eugnio Lazzaretti, R. Minetto, 2024, LoHi-WELD:
A Novel Industrial Dataset for Weld Defect Detection and Classification, a Deep Learning
Study, and Future Perspectives, IEEE Access, Vol. 12, pp. 77442-77453
G.-Q. Wang, 2023, Yolo-MSAPF: Multiscale Alignment Fusion With Parallel Feature Filtering
Model for High Accuracy Weld Defect Detection, IEEE Transactions on Instrumentation
and Measurement, Vol. 72, pp. 1-14
Y. Gao, P. Phong, X. Tang, H. Hu, P. Xu, 2021, Feature Extraction of Laser Welding
Pool Image and Application in Welding Quality Identification, IEEE Access, Vol. 9,
pp. 120193-120202
Y. Zhang, J. Xiao, Z. Zhang, H. Dong, 2022, Intelligent Design of Robotic Welding
Process Parameters Using Learning-Based Methods, IEEE Access, Vol. 10, pp. 13442-13450
K. Kim, K. S. Kim, H.-J. Park, 2023, Multi-Branch Deep Fusion Network-Based Automatic
Detection of Weld Defects Using Non-Destructive Ultrasonic Test, IEEE Access, Vol.
11, pp. 114489-114496
K. Tian, J. Peng, X. Zhang, Q. Zhang, T. Wang, J. Lee, 2024, Multifeature Fusion Imaging
Based on Machine Learning for Weld Defect Detection Using Induction Scanning Thermography,
IEEE Sensors Journal, Vol. 24, No. 5, pp. 6369-6379
A. Mohammed, M. Hussain, 2025, Advances and Challenges in Deep Learning for Automated
Welding Defect Detection: A Technical Survey, IEEE Access, Vol. 13, pp. 94553-94569
S. Yu, J. Hu, J. Hong, H. Zhang, Y. Guan, T. Zhang, 2025, Optimal Imaging Band Selection
for Laser-Vision System Based on Welding Arc Spectrum Analysis, IEEE Sensors Journal,
Vol. 25, No. 2, pp. 2534-2546
Z. Chen, J. Feng, X. Xzhu, B. Wang, 2025, YOLOv8-OCHD: A Lightweight Wood Surface
Defect Detection Method Based on Improved YOLOv8, IEEE Access, Vol. 13, pp. 84435-84450
Y. Mao, G. Wang, Y. Ma, X. Gui, 2025, Enhancing Fabric Defect Detection With Attention
Mechanisms and Optimized YOLOv8 Framework, IEEE Access, Vol. 13, pp. 96767-96781
T. Zhang, H. Pang, C. Jiang, 2024, GDM-YOLO: A Model for Steel Surface Defect Detection
Based on YOLOv8s, IEEE Access, Vol. 12, pp. 148817-148825
Ji-Soo Kim, Kyungyong Chung, 2024, YOLO Based Crack Detection of Structures with Edge
Detection Process, The Transcations of the Korean Institute of Electrical Engineers,
Vol. 73, No. 12, pp. 2391-2397
저자소개
He received his B.S. degree in Information Systems Engineering from Kocaeli University,
Kocaeli, Republic of Turkiye, in 2022. He is currently pursuing the M.S. degree in
Artificial Intelligence Convergence at Chonnam National University and working as
a student (part-time) researcher in Electronics and Telecommunications Research Institute
(ETRI). His research interests include computer vision based models and systems on
industry and autonomous vehicles, 2d-3d object detection, image classification, neural
networks and deep learning.
He received the B.S, M.S., and Ph.D. degrees in Department of Electronics Engineering
from Chonbuk National University in 2004, 2006 and 2011, respectively. He is presently
the principal researcher of Electronics and Telecommunications Research Institute
(ETRI). His research interests include machine learning, agentic AI, optical imaging
system.
He received the B.S, M.S., and Ph.D. degrees in Department of Electronics Engineering
from Chonbuk National University in 2002, 2004 and 2009, respectively. Since 2013,
he has been a Principal Researcher with the Optical ICT Convergence Research Section,
Electronics and Telecommunications Research Institute (ETRI), South Korea, and he
is currently serving as the Director of the Optical ICT Convergence Research Section.
His current research interests include AI-optical engine convergence, intelligent
optical sensor systems, and image analysis.
Kyungbaek Kim received the B.S., M.S., and Ph.D. degrees in electrical engineering
and computer science from Korea Advanced Institute of Science and Technology (KAIST),
South Korea, in 1999, 2001, and 2007, respectively. He is currently a Professor with
the Department of Artificial Intelligence Convergence, Chonnam National University.
Previously, he was a Postdoctoral Researcher with the Department of Computer Sciences,
University of California, Irvine, CA, USA. His research interests include intelligent
distributed systems, software-defined networks/infrastructure, big data platforms,
GRID/cloud systems, social networking systems, AI applied to cyber-physical systems,
blockchain, and other distributed systems issues. He is a member of ACM, IEICE, KIISE,
KIPS, KICS, KIISC, and KISM.