© JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

Enhanced Concrete Surface Crack Detection using Deep

Learning
HARSH GUPTA1, NAMITA GOYAL2, VANDANA CHOUDHARY3
1, 2, 3
Department of IT, Maharaja Agrasen Institute of Technology, GGSIP University, Delhi, India

Abstract- This research paper presents a novel ensure the safety of the infrastructure. Traditional
approach to detect cracks in concrete surfaces using methods of detecting cracks in concrete surfaces, such
the ResNet50 deep learning model with advanced as visual inspection, are is labor-intensive and subject
image transforms. The proposed methodology to human error.. With the advancement of deep
demonstrates high accuracy in detecting cracks learning techniques and the availability of large-scale
compared to traditional methods. In this study, we datasets, it is now possible to develop automated crack
have employed transfer learning to train the detection systems with high accuracy and efficiency.
ResNet50 model on a large dataset of concrete In this paper, we present a novel approach for
surface images with cracks. Additionally, we have detecting cracks in concrete surfaces using the
applied advanced image transforms, including ResNet50 deep learning model with advanced image
random rotation, random brightness adjustment, and transforms[1]. We employ transfer learning to train the
random scaling, to enhance the accuracy of the model on a large dataset of concrete surface images
model further. The results of this research provide a with cracks and apply advanced image transforms,
promising solution for automating the process of including random rotation, random brightness
crack detection in concrete surfaces, which is a adjustment, and random scaling, to enhance the
critical step in ensuring the structural integrity of accuracy of the model further. The dataset we use in
infrastructure. The dataset is segregated into two this study is a balanced collection of 40,000 images,
distinct classes: negative and positive, which divided into two classes: negative and positive,
correspond to images of intact concrete surfaces and corresponding to intact concrete surfaces and concrete
concrete surfaces with cracks, respectively. Each surfaces with cracks, respectively. The images are in
class consists of 20,000 images, resulting in a total of RGB format, with a resolution of 227 x 227 pixels,
40,000 images. The images are in RGB format, with providing researchers with a balanced dataset to use
a resolution of 227 x 227 pixels. These specifications for image classification tasks aimed at detecting cracks
provide researchers with a balanced dataset to use in concrete surfaces. The findings of this research
for image classification tasks aimed at detecting provide a promising solution for automating the
cracks in concrete surfaces. process of crack detection in concrete surfaces, which
is a critical step in ensuring the structural integrity of
Indexed Terms- Concrete surface crack detection, infrastructure.[2]
crack detection, deep learning, Image Processing,
Convolutional Neural Network, CNN, Resnet 50 II. RESNET 50

I. INTRODUCTION State-of-the-art performance in a variety of computer

vision tasks, including as picture classification, object
Concrete is one of the most widely used construction recognition, and semantic segmentation, has been
materials worldwide, and its durability is crucial to the attained using the ResNet50 model, a convolutional
safety and longevity of infrastructure. Cracks in neural network (CNN) architecture. With 50 layers,
concrete surfaces can lead to significant structural the ResNet50 model is a deep learning model that has
damage, potentially compromising the safety of a deeper network than its forerunners. The ResNet50
buildings, bridges, and other structures. Early model's architecture is based on the residual block,
detection of cracks is essential for timely repairs, which allows the model to effectively learn the
which can help prevent more severe damage and residual mapping between the input and output of a

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 566

 © JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

layer. This architecture addresses the vanishing augmentation techniques. We compute the mean and
gradient problem that occurs in deep neural networks standard deviation for the dataset and apply the chosen
by enabling the network to learn from residual transformations.[4] We also set up the device to run on
connections. The ResNet50 model has been widely GPU if available. After that, we set up a pre-trained
used for image classification tasks, achieving high ResNet-50 model for transfer learning. We freeze the
accuracy rates in various benchmark datasets such as model parameters and modify the final layer for our
ImageNet, CIFAR-10, and CIFAR-100. The model's classification task. We define the optimizer and loss
high accuracy is attributed to its depth, which enables function for training and set up a learning rate
it to learn complex features and patterns from input scheduler. We load the training and validation datasets
images. In this study, we utilize the transfer learning using data loaders. We then define functions to train
technique to train the ResNet50 model for the task of the model and visualize its performance.[5] The
detecting cracks in concrete surfaces. Transfer training function iterates over the specified number of
learning involves using a pre-trained model on a large epochs and performs forward and backward passes,
dataset to extract features from new data for a different updating the model's weights. It also keeps track of the
task. We used the ImageNet dataset to pre-train the best performing model based on validation
ResNet50 model and fine-tuned it on our concrete accuracy.[6] Finally, we visualize the model's
surface images dataset. The ResNet50 model's ability predictions on a subset of the validation dataset. After
to learn complex features and patterns from images, training the model, we define functions for making
coupled with its pre-trained weights, makes it a predictions on new images. We can predict the class
suitable model for the task of crack detection in of an input image using the trained model.[7]
concrete surfaces. In addition, we applied advanced Additionally, we provide a function to predict on
image transforms to enhance the model's performance image crops, where the input image is divided into
further. Our experiments show that the ResNet50 smaller crops and predictions are made on each crop.
model with advanced transforms achieved high In summary, this code snippet demonstrates the
accuracy in detecting cracks in concrete surfaces, process of loading a dataset, setting up a pre-trained
indicating its effectiveness in this task. model for transfer learning, training the model, and
making predictions on new images. The provided
functions and techniques can be used as a starting
point for similar image classification tasks.

A.Dataset
It is a collection of images specifically curated for the
task of crack detection or classification. The dataset
likely contains a variety of images showing different
types of cracks, such as cracks on walls, pavements, or
other structures.

Fig1.ResNet50 Model The dataset used in this code snippet focuses on crack
detection and classification, containing a collection of
III. PROPOSED METHOD images with and without cracks. It serves as the
foundation for training the model and enabling it to
Iwe start by importing the necessary libraries and accurately classify new images for crack-related
setting up the environment. We then load a dataset of tasks[8]
images containing cracks and visualize random
images with cracks and without cracks.[3] Next, we
create training and validation datasets by moving
images randomly from the training folder to the
validation folder. We define data loaders and
transformations for the dataset, including data

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 567

 © JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

captured under varying lighting conditions, angles,

and backgrounds to simulate real-world scenarios.
Subsequently, the preprocessed dataset was used to
evaluate the performance of the image processing
pipeline.[10]

Fig2 Classification in our datatset

B.Processing
The dataset used for crack detection or classification
underwent several preprocessing steps to enhance its
quality and suitability for the task. These steps
included image scaling to a constant resolution,
applying image enhancement techniques such as
contrast adjustment and noise reduction, and
augmenting the data through transformations like
rotations and flips.[9] Feature extraction techniques
may have been employed to highlight distinctive crack
characteristics, and pixel value normalization ensured
standardized inputs for the model. These
preprocessing techniques aimed to standardize the
dataset, improve image quality, increase dataset
diversity, and potentially extract relevant features. By
preparing the dataset in this manner, the crack Fig4. Model Summary
detection or classification model was trained with
optimized inputs, leading to improved accuracy and A series of experiments were conducted to measure
robustness. These preprocessing steps play a crucial the pipeline's accuracy and processing speed. The
role in enhancing the dataset and optimizing it for the accuracy was evaluated by comparing the predicted
task at hand. outputs with ground truth labels using standard
evaluation metrics such as precision, recall, and F1
score. The processing speed was measured by
recording the time taken by the pipeline to process
each input image.
Fig3.Transforms applied in our dataset

C. System Overview
The user will capture images, which will be
subsequently uploaded through web applications.
These applications will then utilize an API to invoke a
pre-trained model, which will process the images and
provide highly accurate results as output.
Fig 5.Training and Validation Loss
IV. EXPERIMENTATION

In the experimentation phase, a diverse dataset

comprising various object categories was collected
and preprocessed to ensure optimal compatibility with
the pipeline. The dataset encompassed images

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 568

 © JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

The integration of our project into the existing

educational system can provide a valuable alternative
to traditional teaching methods, such as the
conventional blackboard approach. Moreover, the
accessibility of our web application solely requires an
internet connection and a web browser, making it
easily accessible to users.
Fig 6. Training and validation accuracy
In conclusion, the utilization of the ResNet50 model
Preliminary results indicate promising performance of has led to exceptional accuracy in crack detection. Our
the image processing pipeline, with high accuracy findings reaffirm the cognitive advantages of visual
achieved in object recognition tasks. The average aids and underscore the simplicity and versatility of
processing time per image was within acceptable our web application, which holds promise for
limits, ensuring real-time performance. Further enhancing education beyond conventional
analysis and experiments are underway to optimize the instructional techniques.
pipeline's parameters and explore its scalability for
large-scale deployments.[12][13][14][15][16] CONCLUSION

V. RESULT AND ANALYSIS In conclusion, our experiment focused on the

development of a web application utilizing the
The above-mentioned experimental findings show the ResNet50 model to detect concrete surface cracks.
ResNet50 model's excellent effectiveness in precisely Despite the limitations imposed by a relatively small
identifying concrete surface cracks., achieving an dataset and limited computational resources, our
impressive accuracy rate of 99%. Despite limitations results were highly promising. The ResNet50 model
in terms of dataset size and computational resources, showcased exceptional accuracy, achieving an
the ResNet50 model proved to be highly effective in impressive rate of 99% in accurately identifying
identifying cracks within the images. cracks.

The experiment demonstrated the effectiveness of

visual aids, such as images and videos, in enhancing
memory retention, making our web application a
valuable tool for educational purposes. Its user-
friendly interface ensures accessibility to users of all
ages and cognitive abilities, providing an alternative to
traditional teaching methods.

Overall, our findings highlight the potential of the

ResNet50 model and our web application in improving
crack detection and enhancing the educational
experience.
Fig 7. New way of detecting cracks
FUTURE WORK
Previous research has consistently shown that visual
aids, including images, videos, and animations, have a
Several areas of future work can be identified based on
significant impact on memory retention. To cater to
the results and findings of our experiment. Firstly,
users of all ages and cognitive abilities, our web
expanding the dataset used for training the ResNet50
application was meticulously designed with a user-
model would be beneficial to improve its performance.
friendly interface.
By including a larger and more diverse collection of
concrete surface crack images, the model can learn a

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 569

 © JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

wider range of features and patterns, potentially [7] https://en.wikipedia.org/wiki/Deep_learning

enhancing its accuracy even further. [8] https://www.kaggle.com/datasets/arunrk7/surfac
e-crack-detection
Additionally, incorporating more powerful processors
[9] https://www.tensorflow.org/
or leveraging parallel computing techniques can
accelerate the computation process and enable the [10] https://www.kaggle.com/code/gcdatkin/concrete
utilization of more complex models. This would allow -crack-image-detection/notebook
for the exploration of deeper convolutional neural [11] https://www.mdpi.com/2071-1050/14/13/8117
network architectures or the incorporation of other [12] https://fastapi.tiangolo.com/
advanced deep learning techniques, potentially [13] https://fastapi.tiangolo.com/tutorial/
leading to improved performance and accuracy.
[14] https://www.python.org/
Furthermore, conducting comparative studies with [15] https://www.w3schools.com/js/
other pre-trained models or exploring ensemble [16] https://reactjs.org/
methods could provide insights into the performance [17] https://www.analyticsvidhya.com/blog/2021/05/
variations and identify potential alternatives for crack convolutional-neural-networks-cnn/
detection.
[18] https://towardsdatascience.com/convolutional-
neural-networks-explained-9cc5188c4939
Lastly, integrating the web application with existing
educational systems and conducting user studies to [19] https://towardsdatascience.com/convolutional-
evaluate its effectiveness in enhancing the learning neural-networks-explained-
experience would be valuable. Understanding the 9cc5188c4939#:~:text=A%20CNN%20typicall
usability, impact, and user satisfaction of the y%20has%20three,and%20a%20fully%20conne
application can guide further improvements and cted%20layer.
facilitate its adoption in educational institutions. [20] https://www.ripublication.com/ijaer18/ijaerv13n
8_63.pdf
Overall, these future directions can contribute to the [21] He, K., Zhang, X., Ren, S., & Sun, J. (2016).
ongoing development of our crack detection system, Deep Residual Learning for Image Recognition.
making it more robust, accurate, and accessible for In Proceedings of the IEEE Conference on
various real-world applications and educational Computer Vision and Pattern Recognition
contexts. (CVPR) (pp. 770-778).
[22] Simonyan, K., & Zisserman, A. (2015). Very
REFERENCES Deep Convolutional Networks for Large-Scale
Image Recognition. In International Conference
[1] https://www.sciencedirect.com/science/article/pi on Learning Representations (ICLR).
i/S187705091930866X
[23] Deng, J., Dong, W., Socher, R., Li, L. J., Li, K.,
[2] https://www.frontiersin.org/articles/10.3389/fear & Fei-Fei, L. (2009). Imagenet: A Large-Scale
t.2021.817785/full Hierarchical Image Database. In IEEE
[3] https://www.techtarget.com/searchenterpriseai/d Conference on Computer Vision and Pattern
efinition/convolutional-neural-network Recognition (CVPR) (pp. 248-255).
[4] https://towardsdatascience.com/a- [24] Krizhevsky, A., & Hinton, G. (2009). Learning
comprehensive-guide-to-convolutional-neural- Multiple Layers of Features from Tiny Images.
networks-the-eli5-way-3bd2b1164a53 Technical report, University of Toronto.
[5] https://en.wikipedia.org/wiki/Convolutional_ne [25] Zeiler, M. D., & Fergus, R. (2014). Visualizing
ural_network and Understanding Convolutional Networks. In
[6] https://www.geeksforgeeks.org/convolutional- European Conference on Computer Vision
neural-network-cnn-in-machine-learning/ (ECCV) (pp. 818-833).

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 570

 © JUN 2023 | IRE Journals | Volume 6 Issue 12 | ISSN: 2456-8880

[26] Iandola, F. N., Han, S., Moskewicz, M. W.,

Ashraf, K., Dally, W. J., & Keutzer, K. (2016).
SqueezeNet: AlexNet-level accuracy with 50x
fewer parameters and <0.5MB model size. arXiv
preprint arXiv:1602.07360.
[27] Russakovsky, O., Deng, J., Su, H., Krause, J.,
Satheesh, S., Ma, S., ... & Fei-Fei, L. (2015).
ImageNet Large Scale Visual Recognition
Challenge. International Journal of Computer
Vision (IJCV), 115(3), 211-252.

IRE 1704685 ICONIC RESEARCH AND ENGINEERING JOURNALS 571