Detection of Historical Building Crack Using CDNet Model

Priya Mittal

, Bhisham Sharma

, Komuravelly Sudheer Kumar

and Haeedir Mohameed

4,5

Chitkara University Institute of Engineering and Technology, Rajpura, Punjab, India

Centre for Research Impact and Outcome, Chitkara University Institute of Engineering and Technology, Rajpura, Punjab,

India

School of Computer Science and Artificial Intelligence, SR University, Warangal, Telangana, India

Department of Computer Techniques Engineering, College of Technical Engineering, The Islamic University, Najaf, Iraq

Department of Computer Techniques Engineering, College of Technical Engineering, The Islamic University of Al

Diwaniyah, Al Diwaniyah, Iraq

Keywords: Cracks, Building Crack, Deep Learning, CDNet, Detection SDG.

Abstract: It is crucial to address the cracks in ancient monuments to preserve and safeguard them in alignment with the

sustainable development goals (SDG). Thus, a new deep learning-based innovative CNN model, CDNet, is

proposed for crack detection, which overcomes the challenges of manual detection. Our model was trained

and evaluated using the Historical Crack Dataset. It achieved outstanding results, with 99% accuracy, 98.99%

precision, and 99.01% recall. These results beat the performance of both the VGG16 and ResNet-50 models.

The CDNet model can be subsequently deployed to the cloud for real-time applications.

1 INTRODUCTION

The protection and preservation of the cultural

heritage of ancient monuments is one of the issues

that must be addressed by the upcoming generations.

Cracks that result in the loss of historical heritage and

grandeur and the risk to the lives of living beings are

a consequence of the numerous threats that these

ancient structures face from environmental changes

and human practices. Cracks are significant defects in

building infrastructure that can have a significant

economic impact and pose a safety risk if left

unattended. Consequently, it is essential to evaluate

the crack's profundity to ascertain the most suitable

restoration method and forestall significant damage

(Laxman, Tabassum, et al. , 2023). Crack detection

has been conducted manually in the past, a process

that is labor-intensive, time-consuming, and less

precise (Xu, Tang, et al. , 2023).

Due to these constraints, there is a need for the

development of more effective and precise techniques

to identify and evaluate cracks in building structures.

In past years, there has been substantial growth in

the utilization of deep learning methods for the

identification of cracks in civil structures, including

buildings, bridges, dams, and roads. These

advancements have prompted the development of a

novel deep-learning model that can be used to detect

cracks or potential damage to buildings. This model

is capable of analyzing images of structures,

identifying existing cracks, and predicting potential

future damage.

We introduce a novel model, CDNet, that is

specifically designed for the verification of cracks in

building structures. The CDNet model extracts spatial

features from a dataset that comprises images of

cracks in buildings. 80% of the dataset is utilized to

train the model, while the rest 20% is reserved for

evaluating its performance. The F1 score, precision,

accuracy, and recall of the model are assessed while

testing it on unseen data.

In this paper, the working of the CDNet model is

discussed. In Section II, various works done by

researchers are examined. Further, the proposed

model CDNet is discussed in section III. The

subsequent section IV of the paper focuses on the

results of the method's implementation. Finally, the

paper concluded with a summary, which was

followed by references to the numerous works.

2 RELATED WORK

In this part, a detailed review is presented of the crack

detection methods. Tran et al. (Tran, Nguyen, et al. ,

2023) created a deep-learning model called U-Net to

438

Mittal, P., Sharma, B., Kumar, K. S. and Mohameed, H.

Detection of Historical Building Crack Using CDNet Model.

DOI: 10.5220/0013621200004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 438-443

ISBN: 978-989-758-763-4

detect cracks on the bridge deck. The model

demonstrates exceptional performance, with an

accuracy rate of 92.38 percent. Popli et al. (Popli,

Kansal, et al. , 2023) Suggested the use of a robotic

model called Xception to detect cracks in roads. The

newly proposed model achieves an accuracy of 90%.

Tabernik et al. (Tabernik, Šuc, et al. , 2023)

proposed a novel paradigm, SegDecNet++, to

streamline quality control during building and

maintenance processes. The proposed model obtained

a dice score of 81%. Pham et al. (Pham, Ha, et al. ,

2023) concluded the research using Ostu Method to

detect ground cracks along with their length and

width and the accuracy recorded was 86.7% to 99.9%.

Yadav et al. (Yadav, Sharma, et al. , 2024) suggested

a new Convolutional Neural Network based model,

HCTNet to identify cracks in roads ensuring

sustainable road safety. The model achieved the F1

score of 97.20%.

Sun et al. (Sun, Yang, et al. , 2021) suggested

conducting research utilizing the DeepLabV3+ model

to accurately identify cracks and bugholes on the

surface of the concrete. The model attained an

impressive outcome, with a mean average precision

of 95.58%.

Lin et al. (Lin, Li, et al. , 2023) suggested a model,

DeepCrackAt based on the encoder-decoder network

for crack segmentation and recorded an accuracy of

97.41%. Karimi et al. (Karimi, Mishra, et al. , 2024)

proposed the implementation of a YOLO (You Only

Look Only) deep learning model to diagnose damage

in tiles. The CDNet model attained an accuracy rate

of 72%.

Katsigiannis et al. (Katsigiannis, Seyedzadeh, et

al. , 2023) suggested a deep learning-based model,

MobileNetV2 to diagnose the cracks in brickwork

masonry and achieved the F1 score of 100%. Yadav

et al. (Yadav, Prasad, et al. , 2024) proposed the

CCTNet to improve the precision of crack detection

in structures. The model recorded a precision of

99.33% for the proposed dataset. Tasci et al. (Tasci,

Acharya, et al. , 2023) developed a new architecture

inception and concatenation residual (InCR) to

identify damaged buildings. The model outperforms

other models by showing an accuracy of 99.82%.

Reis et al. (Reis, Turk, et al. , 2024) created a

combination model of ResNet152 +SVM to

recognize the cracks in roads after the earthquake.

The hybrid model had the highest level of success,

with accuracy values of 98.68%. Zheng et al. (Zheng,

Lei, et al. , 2020) suggested three deep learning

models out of which RFCN (Richer Fully

Convolutional Network) showed the best results for

identifying the fractures in buildings with a recorded

accuracy as 91%. Akgul et al. (Akgül, 2023)

introduced a novel fusion model called Mobile-

DenseNet for accurately recognizing cracks that

appear on the surface of the concrete. The fusion

model achieved a success percentage of 99.87%.

Roy et al. (Roy, Kukreja, et al. , 2023) developed a

hybrid model of a deep CNN to detect the intensity of

defects in painting in heritage buildings. The hybrid

model resulted in an accuracy of 84.23%. Joshi et al.

(Joshi, Singh, et al. , 2022) created a deep-learning

model to identify surface cracks or defects in various

structures. The model achieved an average precision

of 93.445% in its predictions. ABDELLAOUI et al.

(ABDELLAOUI, Errousso, et al. , 2024) concluded

the study by considering the VGG-based learning

model as the most superior for detecting cracks in the

pavement as the model records an accuracy of 86.5%.

3 ARCHITECTURE OF THE

MODEL

There are several issues with the earlier models that

used deep learning methods that need to be fixed.

Based on the previously employed ResNext and

VGG16 models, a new model is built in order to

optimize the performance of the crack detecting

model through deep learning. This study developed a

CDNet (Crack Detection network) sequential model

for diagnosing building cracks. In this model, 5

convolution blocks.

The sizes of the convolutional

layers, conv1, conv2, conv3, conv4, and conv5, are

16, 32, 64, 128, and 256, respectively. The model

consists of four different stages as shown in Figure 1.

Figure 1: Stages During Crack Detection

3.1 Working of the CDNet model

A 2D convolutional layer processes an input image or

feature map by applying several filters, extracting

Detection of Historical Building Crack Using CDNet Model

439

local features that are useful for tasks including image

recognition, classification, and segmentation. The

CDNet contain 5 convolution layer. The convolution

layers have 16, 32,64,128, and 256 filter.

Furthermore, each convolution layer is followed by

ReLU activation and a maxpooling layer.

In the CDNet each filter detects specific features

such as edges, corners, or textures. Each convolution

layer is followed by ReLu activation functions and a

max-pooling layer. To decrease computational load

and the risk of overfitting, the max-pooling layer

shrinks the feature maps in spatial dimensions while

keeping the most crucial information. The network

can learn more complicated functions because of the

Relu layer's non-linearity. The Relu Layer applies the

Relu function to every input element. Equation 1 is

the mathematical formula for Relu Function.

f(x) = max(0,x) (1)

After every convolutional block, a channel

attention block (CAB) is also included as shown in

Figure 2. Structure of CDNet model to detect cracks

Channel Attention is a mechanism that allows a

neural network to dynamically focus on the most

informative channels of feature maps. It helps the

model to weigh the importance of each channel and

suppress less useful ones. It operates by first

compressing spatial dimensions into channel-wise

summaries using global pooling. These summaries

are processed via thick layers to create channel-

specific attention weights, which are then used to

scale the relevance of each channel. Channels with

higher weights are emphasized, while less relevant

ones are suppressed. This mechanism improves

feature learning and computational efficiency by

ensuring that the network focuses on the most

informative channels, leading to better model

performance.

The experimental setting of the CDNet method is

shown in Table I.

Table 1: Observations

Crack Images 3896

Batch size 8

ochs 50

Channels 5

Trainin

80%

Validation and Testing 20%

4 RESULTS

This section discussed the proposed model CDNet’s

dataset description and quantitative results.

4.1 Dataset

The Historical Building Crack2019 collection

comprises approximately 3886 meticulously curated

images of cracked areas of antiquated structures. The

dataset contains approximately 45 photographs of a

medieval mosque (Masjid) of Egypt. There are 758

photographs of fractured surfaces and 3,138

photographs of non-cracked surfaces in the

collection. The digital camera was used to capture the

images, which had a resolution of 5184 × 3456 pixels.

Some of the sample crack images are displayed in

Figure 3. Also, by utilizing the rotational invariance,

the size of the dataset has been expanded, and

additional data was added through the use of the

ImageGenerator method found in the Keras package.

The crack image was subjected to augmentation

techniques such as rotating clockwise, rotating

anticlockwise, horizontal flip, and vertical flip.

(a) (b)

Figure 3. Sample crack images

4.2 Experimental Setting

Before testing with the CDNet model, the photos

were scaled to 300 by 300 pixels. The model was then

trained with batch size of 8. The initial learning rate

INCOFT 2025 - International Conference on Futuristic Technology

440

was set at 0.001 to improve the performance. We let

the model train and alter its values for 50 epochs.

Parameters that are evaluated to check the

performance are F1 score, Accuracy, Recall,

Precision, and Sensitivity from the confusion matrix.

These parameters are computed by indicators like TP,

TN, FP, and FN. With these data, the necessary

metrics can be computed as follows:

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦  𝑇𝑁  𝑇𝑃/𝑇𝑁  𝑇𝑃  𝐹𝑁 

𝐹𝑃 (2)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛  𝑇𝑃/𝑇𝑃  𝐹𝑃 (3)

𝑅𝑒𝑐𝑎𝑙𝑙  𝑇𝑃/𝐹𝑃  𝑇𝑁 (4)

𝐹1 𝑠𝑐𝑜𝑟𝑒  2 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙 ∗ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛/

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛  𝑅𝑒𝑐𝑎𝑙𝑙 (5)

Kappa is utilized to verify the performance of a

classification model, such as a CNN, by comparing

the predicted labels to the actual labels. Kappa = 1

indicates a perfect match and a negative value

indicates no match or worst agreement.

4.3 Performance evaluation on the

Building Crack dataset

The Building Crack collection consists of 758

pictures illustrating surfaces with cracks and 3138

images exhibiting surfaces without cracks. First input

image is resized to 300x300x3 pixels. After that,

images are fed to the ResNet-50, VGG16, and CDNet

models for training. We utilized batch sizes of 64 over

50 epochs for training of models. In addition, training

process was expedited by using the Adam optimizer,

with a starting learning rate of 0.001. After training,

confusion matrices were created for each model and

may be seen in Figure 4.

(a)

(b)

(c)

Figure 4: Confusion matrix (a) VGG16 (b) ResNet-50 and

From the confusion matrices, Kappa, accuracy,

F1 score, recall, and precision can be calculated using

the formulas described in section B and shown in

Table 2.

Table 2: Performance metric on Building Crack dataset

VGG16 ResNet-50 CDNet

Precision 95.24% 96.15% 98.99%

F1 score 94.53% 95.70% 98.99%

Recall 93.85% 95.27% 99.01%

Accurac

93.52% 95.43% 99%

Kappa 0.90 0.93 0.98

Detection of Historical Building Crack Using CDNet Model

441

4.4 Training and Validation Loss

Valuable insights into the model's learning progress

are provided by training and validation losses. The

model is enhancing its ability to learn from the

training data is indicated by a decrease in loss over

time. Overfitting can be detected by comparing these

losses: it is characterized by a decrease in the training

loss and an increase in the validation loss, which

suggests that the model is memorizing the training

data rather than learning to generalize from it. The

training (TR) and validation (VL) contours for the

proposed technique are depicted in Figure 5, which

utilizes the Historical Building Crack2019 dataset.

Figure 5: Training and Validation Loss over time

4.5 ROC plot

One often used statistic to evaluate building crack

performance is the receiver operating characteristic

(ROC). Taking the true positive rate (TPR) on the y-

axis and the false positive rate (FPR) at the x-axis at

several threshold values generates the ROC curve.

The area under the curve (AUC) measure shows the

likelihood that a positive value selected would rank

better than a random negative instance picked. Figure

Figure 6. Roc-based comparisons of the VGG16, ResNet-

50 and CDNet model

6 shows the comparison of ResNet-50, VGG16, and

the CDNet model in which VGG16 has an AUC of

0.95, ResNet-50 has 0.96 AUC and CDNet boasts the

highest AUC of 0.99.

5 CONCLUSIONS

By leveraging deep learning techniques, the CDNet

model offers a powerful tool for the preservation and

maintenance of historical buildings. Its application

can significantly enhance our ability to identify and

repair cracks, ultimately collaborating with the

preservation of cultural heritage. The proposed

CDNet model shows an accuracy of 99% and a

precision value of 98.99% which is higher than the

previously used VGG16 and ResNet-50 models for

which accuracy of 93.52% and 95.43% were recorded

respectively. Also, the area under the curve for the

CDNet model i.e. 0.99 is greater than the VGG16 and

ResNet-50 models. Despite showing high accuracy,

the model has a high computational cost. So, the

CDNet model needs to be further modified to achieve

low computational cost. Future work will emphasized

on further modifying the model and enlarging the

dataset to include a broader range of structures and

crack types, ensuring even greater accuracy and

reliability in real-world applications.

REFERENCES

Laxman, K.C., Tabassum, N., Ai, L., Cole, C. and Ziehl, P.,

2023. Automated crack detection and crack depth

prediction for reinforced concrete structures using deep

learning. Construction and Building Materials, 370,

p.130709.

Xu, S., Tang, H., Wang, X. and Wang, D., 2023, November.

Assessment of geometric parameters of segmented

crack on concrete building facade using deep learning.

In Structures (Vol. 57, p. 105188). Elsevier.

Tran, T.S., Nguyen, S.D., Lee, H.J. and Tran, V.P., 2023.

Advanced crack detection and segmentation on bridge

decks using deep learning. Construction and Building

Materials, 400, p.132839.

Popli R, Kansal I, Verma J, Khullar V, Kumar R, Sharma

A. ROAD: Robotics-Assisted Onsite Data Collection

and Deep Learning Enabled Robotic Vision System for

Identification of Cracks on Diverse

Surfaces. Sustainability. 2023; 15(12):9314.

Tabernik, D., Šuc, M. and Skočaj, D., 2023. Automated

detection and segmentation of cracks in concrete

surfaces using joined segmentation and classification

deep neural network. Construction and Building

Materials, 408, p.133582.

INCOFT 2025 - International Conference on Futuristic Technology

442

Pham, M.V., Ha, Y.S. and Kim, Y.T., 2023. Automatic

detection and measurement of ground crack

propagation using deep learning networks and an image

processing technique. Measurement, 215, p.112832.

Yadav, D.P., Sharma, B., Chauhan, S., Amin, F. and

Abbasi, R., 2024. Enhancing Road Crack Localization

for Sustainable Road Safety Using

HCTNet. Sustainability, 16(11), p.4409.

Sun, Y., Yang, Y., Yao, G., Wei, F. and Wong, M., 2021.

Autonomous crack and bughole detection for concrete

surface image based on deep learning. IEEE access, 9,

pp.85709-85720.

Lin, Q., Li, W., Zheng, X., Fan, H. and Li, Z., 2023.

DeepCrackAT: An effective crack segmentation

framework based on learning multi-scale crack

features. Engineering Applications of Artificial

Intelligence, 126, p.106876.

Karimi, N., Mishra, M. and Lourenço, P.B., 2024. Deep

learning-based automated tile defect detection system

for Portuguese cultural heritage buildings. Journal of

Cultural Heritage, 68, pp.86-98.

Katsigiannis, S., Seyedzadeh, S., Agapiou, A. and Ramzan,

N., 2023. Deep learning for crack detection on masonry

façades using limited data and transfer

learning. Journal of Building Engineering, 76,

p.107105.

Yadav, Dhirendra Prasad, Bhisham Sharma, Shivank

Chauhan, and Imed Ben Dhaou. 2024. "Bridging

Convolutional Neural Networks and Transformers for

Efficient Crack Detection in Concrete Building

Structures" Sensors 24, no. 13: 4257.

Tasci, B., Acharya, M.R., Baygin, M., Dogan, S., Tuncer,

T. and Belhaouari, S.B., 2023. InCR: Inception and

concatenation residual block-based deep learning

network for damaged building detection using remote

sensing images. International Journal of Applied Earth

Observation and Geoinformation, 123, p.103483.

Reis, H.C., Turk, V., Karacur, S. and Kurt, A.M., 2024,

April. Integration of a CNN-based model and ensemble

learning for detecting post-earthquake road cracks with

deep features. In Structures (Vol. 62, p. 106179).

Elsevier.

Zheng, M., Lei, Z. and Zhang, K., 2020. Intelligent

detection of building cracks based on deep learning.

Image and Vision Computing, 103, p.103987.

Akgül, İ., 2023. Mobile-DenseNet: Detection of building

concrete surface cracks using a new fusion technique

based on deep learning. Heliyon, 9(10).

Roy, P.S., Kukreja, V., Jain, V. and Vats, S., Classification

of Defective Intensity Levels of Paint in Heritage

Buildings using the CNN-SVM.

Joshi, D., Singh, T.P. and Sharma, G., 2022. Automatic

surface crack detection using segmentation-based deep-

learning approach. Engineering Fracture Mechanics,

268, p.108467.

El Arbi ABDELLAOUI, A.L.A.O.U.I., Errousso, H., Filali,

Y., Chekira, C. and Agoujil, S., 2024. Harnessing Deep

Learning Techniques for Enhanced Detection and

Classification of Cracks in Pavement Imagery. Procedia

Computer Science, 236, pp.386-393.

Detection of Historical Building Crack Using CDNet Model

443