Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)

IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

Deepcrack: A Deep Learning Approach for Image-

Based Crack Prediction using MobileNet And
2024 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT) | 979-8-3503-2753-3/24/$31.00 ©2024 IEEE | DOI: 10.1109/IDCIOT59759.2024.10468016

Transfer Learning
S. Jency G. Ram Kumar
Department of ECE Department of ECE
Saveetha School of Engineering Saveetha School of Engineering
SIMATS, Chennai SIMATS, Chennai
[email protected] [email protected]

Abstract—Cracks in infrastructure pose significant undergoes comprehensive training and validation on a

challenges to public safety and require timely detection for diverse dataset, showcasing its effectiveness in discerning
effective maintenance. This research study presents Deep images with cracks from those without. To further
Crack, a novel deep-learning approach for image-based optimize the classifier, the proposed model is fine-tune
crack prediction. Leveraging the power of Convolutional
with a custom top layer that integrates global average
Neural Networks (CNNs), the Mobile Net architecture is
employed as the backbone and implement transfer learning pooling and dense layers [2].
to enhance the model's ability to recognize cracks. The
proposed methodology involves extensive data pre- A distinctive feature of this study is the introduction of a
processing, utilizing image augmentation techniques to robust evaluation framework, incorporating various
address data scarcity. The model is trained and validated on elements such as confusion matrix and classification
a diverse dataset, demonstrating its effectiveness in reports. This framework comprehensively assess the
distinguishing between images containing cracks and those model's performance, providing insights into accuracy,
without. The classifier is fine-tuned using a custom top layer, precision, and recall. Deep Crack exhibits remarkable
incorporating global average pooling and dense layers. This
proficiency in predicting the presence of cracks,
research work introduces a comprehensive evaluation
framework, including confusion matrices and classification underscoring its potential for real-world applications in
reports, to assess the model's performance. Deep Crack infrastructure maintenance and safety [5].
achieves remarkable accuracy, precision, and recall in
predicting crack presence, demonstrating its capability for This research significantly contributes to the convergence
practical use in infrastructure maintenance and safety of deep learning and image processing, presenting an
scenarios. This research contributes to the intersection of innovative solution for crack detection. Beyond
deep learning and image processing, offering an innovative showcasing the effectiveness of Mobile Net and transfer
solution for crack detection. The proposed methodology not learning, our proposed methodology opens avenues for
only demonstrates the efficacy of Mobile Net and transfer
learning but also provides insights into the broader
broader applications of deep learning within the domains
applications of deep learning in civil engineering and of civil engineering and infrastructure management. As we
infrastructure management. delve into the intricacies of Deep Crack, we unravel a
powerful tool that not only addresses a critical
Keywords— Crack Prediction, Deep Learning, Mobile Net, infrastructure challenge but also illuminates the potential
Transfer Learning, Image Processing transformative impact of deep learning in enhancing public
safety and infrastructure resilience [6].

I. INTRODUCTION
II. RELATED WORKS
Cracks in infrastructure present a critical concern for
public safety, demanding timely detection to facilitate Several studies have explored the application of deep
effective maintenance. This study introduces Deep Crack, learning in image-based defect detection and structural
a ground-breaking deep-learning approach designed for health monitoring. CNNs for the detection of structural
image-based crack prediction. Harnessing the capabilities cracks in concrete, achieving commendable accuracy in
of CNNs, this study proposes a novel MobileNet distinguishing between cracked and intact surfaces were
architecture as the foundation, incorporating transfer employed [3]. Similarly, transfer learning with the Visual
learning to enhance the model's proficiency in recognizing Geometry Group16 (VGG16) architecture to enhance the
cracks [1]. performance of crack detection models was utilized [4].
While these studies contribute valuable insights into the
The proposed research methodology encompasses rigorous realm of image-based defect detection, our work stands out
data pre-processing, where advanced image augmentation through the adoption of the Mobile Net architecture and
techniques are employed to address data scarcity and the meticulous implementation of transfer learning. By
enhance the model's adaptability. The proposed model leveraging Mobile Net's lightweight design, our model
addresses resource constraints without compromising

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1420

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)
IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

performance. Moreover, the incorporation of transfer intricately detail our approach, unravelling the meticulous
learning facilitates knowledge transfer from pre-trained process from dataset preparation to the fine-tuning of
models, improving the model's capacity to adapt and model parameters, culminating in a comprehensive
generalize to unfamiliar datasets. The research landscape understanding of our innovative crack prediction system.
in image-based crack prediction continues to evolve, and
A. Dataset Preparation
our study adds a nuanced perspective by presenting Deep
Crack, an efficient and effective solution for crack We utilized a comprehensive dataset comprising images
detection in diverse infrastructure scenarios. In the realm of cracked and intact structures sourced from diverse
of crack detection using deep learning techniques, various scenarios. To guarantee a balanced distribution of positive
studies have contributed valuable insights and (cracked) and negative (intact) samples, the dataset was
methodologies to address the challenges associated with divided into training and testing sets. This is an essential
real-world scenarios [7]. stage in training a model that can identify minute details
One study explored the application of a modified U-Net related to crack patterns.
architecture for crack detection in pavement images, a) Dataset Acquisition
specifically focusing on the complications posed by Our dataset, obtained from Kaggle, encompasses a rich
diverse lighting conditions and surface textures. This collection of 40,000 images featuring both positive
research provided crucial insights into achieving robust instances of cracked structures and negative instances of
crack detection in practical environments. Another intact structures. The images were meticulously selected to
investigation delved into the integration of generative represent a diverse array of scenarios, ensuring a broad
adversarial networks (GANs) in crack detection systems. spectrum of crack patterns and structural conditions
By leveraging GANs for data augmentation, the approach b) Dataset Segmentation
demonstrated notable enhancements in model performance
To facilitate robust model training and evaluation, we
and generalization, particularly in situations where
meticulously organized the dataset into distinct training
annotated data was limited [9].
and testing sets. This segmentation process was undertaken
In the domain of structural defect detection in bridges, a
with precision to guarantee an equitable distribution of
study combined deep learning techniques with remote
positive (cracked) and negative (intact) samples within
sensing using high-resolution aerial images. The model
each set. This deliberate balance serves as a pivotal aspect
showcased promising results in the early detection of
of our methodology, fostering the model's ability to discern
cracks and deformations in large-scale infrastructure,
nuanced features intricately linked to varied crack patterns.
highlighting the potential of this approach for preventive
maintenance. For automatic crack identification, a hybrid c) Significance of Balance
deep learning technique that combines CNNs with Long The careful curation of a balanced dataset is foundational
Short-Term Memory Networks (LSTMs) was introduced. to the effectiveness of our model. By presenting an equal
This methodology focused on capturing the temporal representation of both cracked and intact structures, we
evolution of cracks, providing a dynamic perspective that equip the model with the capacity to navigate the
enhances the understanding of crack propagation over intricacies of diverse scenarios. This strategic approach
time. Furthermore, a study proposed a transfer learning ensures that the model's learning process is not skewed
approach for crack detection in historical buildings, toward a particular class, enhancing its adaptability and
utilizing a pre-trained ResNet model. Emphasizing the generalization to real-world scenarios.
significance of domain adaptation, this research B. Model Architecture
underscored the importance of tailoring models to specific
contexts, such as heritage structures, for effective defect Delving into the intricacies of our model architecture
detection [10]. unveils a sophisticated amalgamation of cutting-edge
Collectively, these studies contribute to the broader design principles, prominently featuring the stalwart
literature by presenting diverse approaches to crack Mobile Net architecture. This convolutional neural
detection, each addressing unique challenges and network, meticulously crafted for mobile and embedded
showcasing the adaptability of deep learning techniques in vision applications, serves as the backbone of our
varied applications within the field of infrastructure innovative model.
monitoring and maintenance
At the core of MobileNet's ability lies its distinctive
employment of depth-wise separable convolutions, a
III. METHODOLOGY paradigm-shifting approach that brings forth a dual benefit
In our pursuit to elevate the realm of crack prediction, our of significantly reducing computational complexity while
study endeavours to introduce a state-of-the-art deep concurrently upholding a high standard of performance.
learning framework coined "Deep Crack." This innovative Embracing this architectural wonder, our model harnesses
approach hinges on the fusion of the Mobile Net the effectiveness of Mobile Net, utilizing its capability to
identify detailed patterns and features in the input data.
architecture with transfer learning techniques, fostering a
robust and adaptive system for the accurate identification
and prediction of cracks in diverse settings. While retaining the fundamental structure of MobileNet,
Our methodology is meticulously structured, with a focus this study introduced a pivotal augmentation to further
on key components that collectively contribute to the elevate the model's capabilities. The infusion of a Global
Average Pooling layer marks a strategic enhancement,
efficacy of Deep Crack. The subsequent sections

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1421

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)
IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

fostering a more nuanced understanding of spatial IV. RESULTS AND DISCUSSION

relationships within the data. This addition is pivotal in
A. Training Strategy
capturing the essence of features across the entire spatial
domain, contributing to a more holistic representation of The accuracy, precision, recall, and F1-score—standard
the input. classification metrics—were used to assess our model's
Beyond this, the proposed model underwent a performance. The results were remarkable, with accuracy
transformative evolution through the incorporation of reaching high levels. Precision and recall values for both
densely connected layers. These layers act as intricate positive and negative classes exceeded 0.98,
conduits for information flow, enabling the model to demonstrating the model's ability to classify while
extract and assimilate highly relevant features essential for capturing a high percentage of relevant instances. The F1-
our specific classification task. The synergistic score, which strikes a balance between precision and
collaboration between the original Mobile Net structure, recall, surpassed 0.99 for each class, indicating the
Global Average Pooling layer, and densely connected model's overall effectiveness.
layers empowers our model to navigate the complexities Table 1 provides a comprehensive summary of the
of various data patterns. classification performance of the model for a job that is
binary in nature. Each of the target classes, "Negative"
This architectural refinement underscores our and "Positive," is displayed in the "Class" column to the
commitment to crafting a model that not only harnesses right. The precision, recall, and F1 score for each class are
the power of state-of-the-art convolutional neural displayed in the columns that were just presented to you.
networks but also stands uniquely tailored to the demands With regard to the 'Support' column, the number of
of our targeted classification task. The modifications instances that belong to each class is displayed. The
introduced amplify the model's capacity to discern, 'Accuracy' row is used to calculate the overall accuracy of
interpret, and make accurate predictions, thereby the model, which is represented by the ratio of accurately
solidifying its standing as an innovative and formidable predicted instances to all instances (in this case, 2000).
tool in the realm of deep learning for image classification. This ratio is used to determine the overall accuracy of the
model. By weighting the average based on the number of
occurrences in each class, the 'Weighted Average' row
takes into consideration class imbalance. On the other
hand, the 'Macro Average' row provides the average
Problem Model metrics across both classes, treating them in an equal
Fine Tuning
Statement Selection manner. The following table offers a concise and helpful
summary of the performance of the model in a variety of
evaluation metrics.
Data set Model Training Evaluation TABLE I. : PERFORMANCE METRICS SUMMARY [1]
Collection
Class Precision Recall F1-
(%) (%) Score
Dataset (%)
Transfer Performance
Negative 99 98 99
Pre-processing Learning Assesment
Positive 98 99 99
Accuracy 99
Macro 99 99 99
Average
Fig 1: Flowchart of Deep Crack Model Weighted 99 99 99
Average
The figure serves as a visual representation of the
systematic process followed in our study for crack
detection using the Deep Crack model.
C. Training Strategy
The model was trained with the Adam optimizer, which is
well-liked for deep neural network optimization because
to its effectiveness. To mitigate overfitting, we employed
a learning rate of 0.001 and implemented early stopping.
This allowed the model to cease training once
performance on the validation set plateaued. Additionally,
data augmentation techniques such as zooming, width
shifting, and shearing were applied during training. This
strategy enhances the model's generalization capability by
exposing it to a variety of transformed images, promoting Fig 2: Classification Metrics
robust learning. Figure 2 illustrates crucial classification metrics across
diverse categories. It features the metrics on the X-axis,
while categories such as 'Negative,' 'Positive,' 'Accuracy,'

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1422

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)
IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

'Macro Average,' and 'Weighted Average' are listed on the

y-axis. Each category is depicted through horizontal bars.
Precision, recall, and F1-score signify accuracy,
completeness, and a balance between the two, while
support indicates the number of instances per class. The
legend on the right clarifies the color code. This concise
visualization allows for swift comparison, enabling a
comprehensive understanding of the model's performance
nuances.

Fig 4: Accuracy Comparison

Figure 5 presents the F1-score of different methods, which
is a balance between precision and recall. A higher F1
score indicates a well-balanced performance in terms of
both false positives and false negatives. This graph
provides insights into the overall effectiveness of each
method

Fig 3: Performance Metrics for Different Methods [1]

Table 2 presents a comprehensive overview of the
performance metrics for various methods employed in the
study. Each row corresponds to a specific method, and the
columns provide metrics. The numerical values in each
cell represent the corresponding metric scores for the
respective method. These metrics are crucial in evaluating
the effectiveness of each method in tasks like image
classification.
Fig 5: F1-Score Comparison
TABLE II. : PERFORMANCE METRICS SUMMARY
Figure 6 illustrates the precision of a number of distinct
Methods Precision F1- Accuracy Recall approaches by highlighting the potential of these
(%) Score (%) (%)
(%)
methodologies to produce accurate positive predictions.
Deep 99 99 99 98 This graph also demonstrates the precision of these
Crack procedures. The fact that a higher precision results in a
RFCN-b 84 80 93 84 lower proportion of false positives is evidence that each
approach is reliable in identifying positive cases.
RFCN-a 88 84 94 80

FCN 80 80 86 79

Mask R- 61 59 64 60
CNN

Fig 4 compares the accuracy of different methods,

showcasing their performance in correctly predicting both
positive and negative instances. Higher bars indicate a
higher overall accuracy, reflecting the effectiveness of
each method in classification.

Fig 6: Precision Comparison

Figure 7 comparison graph displays the recall values for
each method, emphasizing their effectiveness in capturing
positive instances. Higher bars indicate a lower rate of
false negatives, showcasing the ability of each method to
correctly identify positive cases.

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1423

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)
IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

C. Confusion Matrix Analysis

The confusion matrix analysis which is shown in Figure 9
showcased the model's exceptional performance, with
minimal misclassifications. To a large extent, errors were
seen in circumstances in which the model confused
aspects that appeared to be similar to one another, which
is indicative of a high level of accuracy in distinguishing
between unique patterns. The results of this research
provide useful insights into the strengths of the model as
well as potential areas for development, which will guide
future efforts to optimize processes.

Fig 7: Recall Comparison

Fig displays the performance metrics for five different
methods. Each method is represented by four bars, one for
each metric. The graph provides a comparative overview
of how these methods perform across the specified
metrics, allowing for easy interpretation and comparison.
The distinct colours for each metric aid in visual clarity,
making it accessible for viewers to analyze and draw
insights from the presented data.
B. Testing And Validation
80% of the images were assigned to the training set, while
the remaining 20% were designated for testing. The
dataset was intelligently divided into training and testing
sets. This division ensures a robust evaluation of unseen
data while providing a sufficiently large training set for
the model to learn intricate patterns. The balanced
Fig 9: Confusion Matrix
distribution enables the model to generalize well to new,
unseen images, making it suitable for real-world V. CONCLUSION AND FUTURE WORK:
applications.
In conclusion, the proposed MobileNet-based model
demonstrated outstanding performance in classifying the
images by underscoring its effectiveness for our specific
task. Future work could involve fine-tuning the model on
specific subsets of data to further enhance its accuracy by
addressing the challenges posed by similar-looking
features. Additionally, exploring advanced transfer
learning techniques or experimenting with more
sophisticated architectures holds promise for achieving
even better results in diverse real-world scenarios.
REFERENCES

[1] Wu, Xinhua, and Xiujie Liu. "Building crack identification and
total quality management method based on deep learning." Pattern
Recognition Letters 145 (2021): 225-231.
Fig 8: Performance Metrics by Class
[2] Kunal, Kishor, and Namesh Killemsetty. "Study on control of
Figure 8 illustrates the performance metrics, for two cracks in a Structure through Visual Identification &
classes—Negative and Positive. Each class is represented Inspection." IOSR Journal of Mechanical and Civil
by three bars, one for each metric. A representation of the Engineering 11.5 (2014): 64-72.
accuracy of the model in accurately detecting instances of [3] R. T. et al "Automated Crack and Damage Identification in
Premises using Aerial Images based on Machine Learning
the class is shown by the precision bars, which are Techniques," (I-SMAC), Dharan, Nepal, 2022, pp. 498-504, doi:
coloured blue. The ability of the model to capture all 10.1109/I-SMAC55078.2022.9987391..
instances of the class that are relevant is demonstrated by [4] Barter, Simon, et al. "An experimental evaluation of fatigue crack
the recall bars, which are coloured orange. The final point growth." Engineering failure analysis 12.1 (2005): 99-128.
is that the F1-Score bars, which are shown in green, offer [5] Zheng, Minjuan, Zhijun Lei, and Kun Zhang. "Intelligent detection
a balanced measurement that takes into account both of building cracks based on deep learning." Image and Vision
Computing 103 (2020): 103987.
precision and recall.
[6] Torok, Matthew M., Mani Golparvar-Fard, and Kevin B.
Kochersberger. "Image-based automated 3D crack detection for
post-disaster building assessment." Journal of Computing in Civil
Engineering 28.5 (2014): A4014004.

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1424

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT-2024)
IEEE Xplore Part Number: CFP24CV1-ART; ISBN: 979-8-3503-2753-3

[7] Laefer, Debra F., Jane Gannon, and Elaine Deely. "Reliability of
crack detection methods for baseline condition
assessments." Journal of Infrastructure Systems 16.2 (2010): 129-
137.
[8] Chen, Kaiwen, et al. "Automated crack segmentation in close-range
building façade inspection images using deep learning
techniques." Journal of Building Engineering 43 (2021): 102913.
[9] Babu, J. Chinna, et al. "IoT-based intelligent system for internal
crack detection in building blocks." Journal of Nanomaterials 2022
(2022): 1-8.
[10] Adam, Edriss Eisa Babikir, and A. Sathesh. "Construction of
accurate crack identification on concrete structure using hybrid
deep learning approach." Journal of Innovative Image Processing
(JIIP) 3.02 (2021): 85-99.

979-8-3503-2753-3/24/$31.00 ©2024 IEEE 1425

Authorized licensed use limited to: Nanyang Technological University Library. Downloaded on June 02,2025 at 03:49:32 UTC from IEEE Xplore. Restrictions apply.