www.itcon.

org - Journal of Information Technology in Construction - ISSN 1874-4753

AUTOMATIC CRACK CLASSIFICATION ON ASPHALT PAVEMENT

SURFACES USING CONVOLUTIONAL NEURAL NETWORKS AND
TRANSFER LEARNING
SUBMITTED: September 2024
REVISED: December 2024
PUBLISHED: December 2024
GUEST EDITORS: Vito Getuli, Farzad Rahimian, Nashwan Dawood, Pietro Capone, Alessandro Bruttini
DOI: 10.36680/j.itcon.2024.055
Sandra Matarneh, Associate Professor,
Department of Civil Engineering, Al-Ahliyya Amman University, Amman, Jordan;
s.matarneh@ammanu.edu.jo
Faris Elghaish, Lecturer,
School of Natural and Built Environment, Queen's University Belfast; UK
F.Elghaish@qub.ac.uk
David John Edwards, Professor,
Department of the Built Environment, Birmingham City University, UK;
Faculty of Engineering and the Built Environment, University of Johannesburg, South Africa;
david.edwards@bcu.ac.uk
Farzad Pour Rahimian, Professor,
School of Computing, Engineering and Digital Technologies, Teesside University; UK
F.Rahimian@tees.ac.uk
Essam Abdellatef, Assistant Professor,
Department of Electrical Engineering, Faculty of Engineering, Sinai University, Arish;
essam.abdellatef@su.edu.eg
Obuks Ejohwomu, Associate Professor,
Civil Engineering and Management, The University of Manchester; UK
obuks.ejohwomu@manchester.ac.uk
SUMMARY: Asphalt pavement cracks constitute a prevalent and severe distress of surfacing materials and before
selecting the appropriate repair strategy, the type of deterioration must be classified to identify root causes.
Efficient detection and classification minimize concomitant costs and simultaneously increase pavement service
life. This study adopts convolutional neural networks (CNN) for asphalt pavement crack detection using secondary
data available via the CRACK500 dataset and other datasets provided by GitHub. This dataset had four types of
cracks viz.: horizontal, vertical, diagonal and alligator. Five pre-trained CNN models trained by ImageNet were
also trained and evaluated for transfer learning. Emergent results demonstrate that the EfficientNet B3 is the most
reliable model and achieved results of 94% F1_Score and 94% accuracy. This model was trained on the same
dataset by performing transfer learning on pre-trained weights of ImageNet and fine-tuning the CNN. Results
revealed that the modified model shows better classification performance with 96% F1_Score and 96% accuracy.
This high classification accuracy was achieved by a combination of effective transfer-learning of ImageNet weight
and fine-tuning of the top layers of EfficientNet B3 architecture to satisfy classification requirements. Finally,
confusion matrices demonstrated that some classes of cracks performed better than others in terms of
generalization. Further additional advancement with fine-tuned pre-trained models is therefore required. This
study showed that the high classification results resulted from using a successful transfer learning of ImageNet
weights, and fine-tuning.
KEYWORDS: Convolutional Neural Networks; CNN; Deep Learning; Transfer Learning, Multiclass
Classification; Asphalt Pavement.
REFERENCE: S. Matarneh, F. Elghaish, D. J. Edwards, F. P. Rahimian, E. Abdellatef & O. Ejohwomu (2024).
Automatic crack classification on asphalt pavement surfaces using convolutional neural networks and transfer
learning. Journal of Information Technology in Construction (ITcon), Special issue: ‘Managing the digital
transformation of construction industry (CONVR 2023)’, Vol. 29, pg. 1239-1256, DOI: 10.36680/j.itcon.2024.055
COPYRIGHT: © 2024 The author(s). This is an open access article distributed under the terms of the Creative
Commons Attribution 4.0 International (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted
use, distribution, and reproduction in any medium, provided the original work is properly cited.

ITcon Vol. 29 (2024), Matarneh et al., pg. 1239

1. INTRODUCTION
Asphalt pavement cracking is a serious issue that significantly impairs surface performance. A pavement crack is
defined as an unintended discontinuity or a damage compromising the integrity of the pavement surface. These
cracks are frequently attributed to external environmental factors, such as moisture variations, subgrade expansion,
chemical shrinkage, frost, precipitation, flooding, ultraviolet (UV) radiation, and the repeated passage of heavy
loads (Baduge et al., 2023). Surface cracks facilitate water infiltration into the pavement structure, further
worsening the damage and ultimately compromising the pavement's base. Cracks can be classified into various
types, including longitudinal, transverse, diagonal, block, alligator, and irregular patterns (Canestrari and Ingrassia,
2020).
While cracks represent a key indicator of pavement condition assessment (Guo et al., 2022), traditional crack
classification mostly relies on time-consuming, costly and risky manual inspection (Yang et al., 2020). Moreover,
manual inspection is inefficient and unreliable due to human subjectivity and tacit knowledge of inspectors. Indeed,
extant literature reports variations in routine inspection practices and significant differences in crack classification
outcomes (Dais et al., 2021). To augment inspection performance, advancements in AI offer innovative deep
learning (DL) solutions to enhance crack detection accuracy and efficiency while minimizing human error (Tran
et al., 2021). As an established DL technology, convolutional neural networks (CNN) have achieved advanced
human-competitive performance in computer vision activities like object identification, image classification, and
semantic segmentation (LeCun et al., 2015, Krizhevsky et al., 2017). Unlike traditional image analysis methods
that depends on manually defined rules, CNNs automatically extract multi-level feature representations (Liu et al.,
2019). Several studies have highlighted the growing efficacy of CNNs as a robust tool for crack detection
(Gopalakrishnan et al., 2017, Mandal et al., 2018, Huyan et al., 2020).
However, training a CNN requires enormous volumes of properly collected data to ensure that adequate training
data is available to engender accurate outcomes and avoid underfitting (i.e., CNN performs poorly on training data
and new data from the problem domain) (Kim et al., 2018). Transfer learning techniques, which leverage
knowledge from previously trained CNN models, are utilized to mitigate the risk of underfitting (Yang et al., 2020).
These CNNs are employed in transfer learning by utilizing weights pre-trained on the ImageNet dataset. ImageNet
provides a publicly accessible database of images designed to support the training of large-scale object recognition
models (Feng et al., 2019). Various studies have applied transfer learning to establish classifiers for crack detection
viz.: MobileNet (Hou et al., 2021, Hernanda et al., 2022); VGG (Gopalakrishnan et al., 2017, Dung et al., 2019,
Rubio et al., 2019, Guzmán-Torres et al., 2022, Brien et al., 2023); GoogleNet (Jang et al., 2019, Yang et al., 2021,
Elghaish et al., 2022); U-Net (Liu et al., 2019, Yang et al., 2021, Matarneh et al., 2024); Res-Net (Augustauskas
and Lipnickas, 2020, Wang and Guo, 2021, Yoon et al., 2022); Inception (Feng et al., 2019, Ali et al., 2022, Wu et
al., 2021a); AlexNet; YOLO (Liu et al., 2022c, Teng et al., 2022); and EfficientNet (Teng et al., 2022, Liu et al.,
2022b).
Given the aforementioned context of rapidly evolving technology development, this present study introduces a
modified pre-trained CNN model based on EfficientNetB3 to detect different types of asphalt pavement surface
cracks. Therefore, this study investigates the classification performance of different ImageNet pre-trained models
on the Crack500 dataset and employs transfer learning and fine-tuning to enhance the classification accuracy of
the pre-trained models.

1.1 Convolutional neural networks for crack detection

Images can be classified using CNN in three categories namely: 1) semantic segmentation - which offers
information about the specific length, width or location of any crack due to its capability of assigning a class label
to each pixel (Li et al., 2018, Liu et al., 2019, Choi and Cha, 2020, Yang et al., 2021); 2) patch classification –
where each patch is assigned a class label after dividing images into patches; and 3) boundary box regression –
where the box bounds the identified crack and shows its location and boundaries (Zhang et al., 2019). However,
the latter two categories have been extensively used for identifying cracks and have proven positive outcomes
(Cha et al., 2017, Feng et al., 2019). Unlike these two categories which are implemented at block level, semantic
segmentation is implemented at pixel level and has gathered application momentum in recent studies (Tang et al.,
2022).
To identify and categorize cracks in asphalt pavements, Tran et al. (2021) utilized an improved faster R-CNN

ITcon Vol. 29 (2024), Matarneh et al., pg. 1240

(RetinaNet) to train and label asphalt images. The results of this study revealed that the RetinaNet’s classification
accuracy was 84.9% considering both the type and severity level of cracks. Similarly, Li et al. (2022) established
a strategy for pavement surface condition index (PCI) using a genetic algorithm (GA) and CNN algorithm. The
authors classified pavement crack type in five thousand pavement distress images with an accuracy reaching 98%,
and image processing time of 0.047 seconds. To understand the crack features in the intricate fine-grain asphalt
pavement background, Wu et al. (2021a) employed a full CNN to combine features acquired from various scales
of convolutional kernels, the DenseNet and the deconvolution network to accomplish pixel-level recognition.
Emergent results indicated that the adopted method reported significant segmentation results for twelve types of
cracks. Teng et al. (2022) developed DeepLab_v3+, a pixel-level segmentation CNN, to segment cracks by
calculating their length and width to a reported accuracy and F-score values of 80, 97.5 and 78% respectively. Fan
et al. (2022) reported upon the results of a deep residual convolutional neural network (Parallel ResNet) to develop
a pavement crack detection and measurement system with high-performance i.e., precision (94.27%), recall
(92.52%) and F1 (93.08%). Based on a spatial channel hierarchical network, a more accurate and efficient crack
detection process was proposed by Pan et al. (2020) using the Visual Geometry Group 19 (VGG19). To examine
inner cracks in the turbine blade thermal barrier coating, Shi et al. (2022) used VGG19 in addition to a Multi-Scale
Enhanced-Faster R-CNN (MSE-Faster R-CNN) to create prospective crack locations. Results showed that the
suggested approach can precisely locate cracks on different scales (89.8%) and detect them (80.6%).
Recent studies have employed a variety of CNNs, including Tan et al. (2021) who utilized the YOLOv3 to automate
the recognition of sewage pipe deficiencies. The study concentrated on enhancing the model structure, bounding
box prediction, data expansion and loss function. The proposed model achieved a mean average precision (mAP)
value of 92%, which is greater than the level of accuracy achieved in previous related studies. Lu et al. (2022)
utilized YOLOv5 and the multiple sliding windows method to detect defects in ceramic tiles surfaces.
Similarly, Yao et al. (2022) developed twelve distinct attention models for pavement crack detection using
YOLOv5. Their study demonstrated that the model could process images at a rate of 13.15 ms per image while
achieving a precision of 94.4%. Li et al. (2020) employed U-Net with alternately updated clique (CliqueNet),
called U-CliqueNet to separate cracks from the background of tunnel images. The developed network was trained
on a large dataset containing 50,000 images and tested on 10,000 images. The model attained positive results with
92.25% mean pixel accuracy (MPA), 86.96% mean intersection over union (MIoU), 86.32% precision and 83.40%
F1-score. Liu et al. (2019) employed U-Net for concrete crack detection in raw images, achieving precision values
of 0.9 across various complex scenarios. Huyan et al. (2020) constructed a U-shaped model structure by utilizing
convolution, pooling, transpose convolution, and concatenation operations. Their results indicated that the model
achieved 99.01% accuracy, 98.56% precision, 97.98% recall, and 98.42% F-measure with a learning rate of 10^2.
Recently, several studies adopted the generative network along with the DL models. For instance, Mazzini et al.
(2020) proposed a CNN to augment data of highly textured images within the framework of semantic
segmentation. In their study a Generative Adversarial Network (GAN) was used to develop a semantic layout,
and a texture synthesizer (based on a CNN), to produce a new image. This approach was evaluated using the
dataset of German Pavement Distress and the results of evaluation revealed a substantial improvement in prediction
performance. Pei et al. (2021) developed a method using advanced deep convolutional generative adversarial
networks (DCGANs) to solve the problem of crack identification in asphalt pavement small size images. The
outcomes presented that the average precision of the proposed model is 90.32%. Dong et al. (2022) used StyleGAN
and a feature fusion model to solve the issue of realizing precise and effective pixel-level segmentation of
pavement damage. Results showed that the suggested model segmentation could achieve above 0.918 mAP (mean
Average Precision) value for segmentation, and it has evident benefits for the map and cross cracks segmentation.

1.2 Pre-trained CNN for crack detection

Existing CNN models have a tendency for being quite complex and require a huge dataset to prevent overfitting
and therefore, to solve these problems transfer learning is increasingly utilized (Dawson et al., 2023). Transfer
learning is a machine learning technique that uses pre-trained DL models to resolve a new problem that is related
to the original problem the model was initially trained to solve (Pan and Yang, 2010). Recently, transfer learning
in DL has been widely applied across various approaches. For example, to assess the effectiveness of DCNNs in
the pavement cracks classification, Ranjbar et al. (2021) used different pre-trained networks, namely: DenseNet-
201, GoogleNet, AlexNet, SqueezNet, ResNet-18, ResNet-50, ResNet-101 and Inception-v3. In this study, a more

ITcon Vol. 29 (2024), Matarneh et al., pg. 1241

efficient technique for crack segmentation was created using a wavelet transform module with more regularized
metrics. According to the study's findings, retrained classifier models produced accurate results with a confusion
matrix-based performance range of 94% to 99%. Moreover, the constructed wavelet module was capable of clearly
segmenting crack pixels [ibid]. Wu et al. (2021b) used DenseNet, multi-scale CNN and SVM classifier to segment
12 different crack types and produce superior pavement crack segmentation comparing to the most sophisticated
alternative methods Hernanda et al. (2022) employed a DL CNN with a pre-trained SSD MobileNetV2 network
after altering the hyperparameter. Results showed a higher mean average precision (mAP) of 0.0869 and a lower
total loss training of 0.6028. Ha et al. (2022) classified five types of cracks (alligator crack, longitudinal crack,
transverse crack, pothole and patching) by developing a system utilizing SqueezeNet, U-Net and Mobilenet-SSD
models together. Emergent results reported that the system accuracy was 91.2%. Fan et al. (2022) adopted the
Parallel ResNet to develop a deep residual CNN to establish a high-performance system to detect and measure
pavement crack. The highest scores for precision (94.27%), recall (92.52%) and F1 (93.08%) were achieved [ibid].
Another pretrained deep CNN model for crack identification is presented using hybrid images and GoogLeNet
(Jang et al., 2019).
The proposed model was tested and validated utilizing various sizes of concrete sampling and showed that macro-
and micro-cracks were effectively identified utilizing hybrid images. Qu et al. (2022) also modelled crack
segmentation using a CNN network and transformer. To increase the performance of the feature representation,
the authors employed UNet++ and polarized self-attention. In addition, they replaced the last layer of feature
extraction by the transformer. The study results exhibited that the developed model showed its efficacy with F-
score values of 0.856, 0.700 and 0.637 on three different datasets. Xu and Liu (2022) proposed a detection method
under small samples by combining a generative adversarial network (GAN) and a CNN architecture. After training
and analyzing the dataset using the transfer learning technique, results showed that the detection accuracy
improved considerably from 80.75% to 91.61% thus, proving the effectiveness of the extended data.
Recently, several studies compared various CNNs and reported different results (Cha et al., 2017, Feng et al., 2019,
Liu et al., 2022b, Matarneh et al., 2024). Yang et al. (2021) evaluated the performance of three CNNs, namely:
AlexNet, ResNet18 and VGGNet13 and concluded that ResNet18 (accuracy 98.8%) outperforms the other two
models. Another study assessed four CNNs (i.e., Inception-V3, VGG-16, VGG-19 and Resnet-50) and compared
them with the proposed customized CNN model (Ali et al., 2022). The authors developed their CNN model based
on spatial or sequential features and Adam optimizer for crack localization and detection in concrete structures
using eight datasets. The research concluded that all models performed well on a small set of various training data;
though, as the quantity and diversity of the training data increased, generalization performance decreased, and
overfitting resulted. Additionally, the tailored CNN and VGG-16 models showed superior cracking localization
and identification for concrete structures. (Wang and Guo 2021) developed transfer learning-based methods for
fatigue crack initiation sites identification using three CNNs (i.e., feature pyramid network (FPN), ResNet-101
and VGG-16), to extract features along with the backbone model - a faster R-CNN.
The ResNet model outperformed the other two in terms of accuracy and training expense. Pozzer et al. (2021) used
thermographic and regular photos taken from a variety of distances and angles to test the effectiveness of multiple
deep neural network models in concrete cracks identification. The MobileNetV2 performed admirably in detecting
multiclass damages in thermal pictures while in contrast, the VGG 16 model showed improved precision by
reducing the rate of erroneous detections [ibid]. Elsewhere, Elghaish et al. (2022) evaluated four pre-trained CNN
models (i.e., AlexNet, VGG16, VGG19 and GoogleNet) in crack detection for highways and observed that the
accuracies of all pre-trained models are higher (97.72%) than averages and the calculated accuracies for AlexNet
and GoogleNet models by more than 5%.
The application of AlexNet, SE-Net, and ResNet with a variety of configurations was explored by Liu et al. (2022a)
who applied a two-step data pre-processing to decrease the bias of the CNN model. The study used data
augmentation in the beginning to increase the dataset. The original image was then transformed into a binary black-
and-white image using a crack extraction technique; the ResNet with 50 layers exhibited the highest test accuracy.
Loverdos and Sarhosis (2022) evaluated different DL networks (U-Net, DeepLabV3+, U-Net (SM), LinkNet (SM)
and FPN (SM)) to advance automation in brick segmentation and crack detection of masonry walls. Emergent
findings revealed that DL offers superior outcomes than conventional image-processing techniques for brick
segmentation.

ITcon Vol. 29 (2024), Matarneh et al., pg. 1242

Table 1: Comparison between existing studies used transfer learning and optimisation methods for crack detection
and classification. Part One.
Reference Aim Method Results Limitations
Dawson et To evaluate the Used datasets with varying Inception-v3 achieved 92% Models showed overfitting
al., (2023) performance of different sizes (7k, 42k, 104k accuracy on the largest on smaller datasets; findings
CNN architectures for images), trained nine CNN dataset; dataset size strongly highlight challenges for
carbonate rock architectures, and applied affects performance. small geological datasets.
classification using transfer learning.
transfer learning.
Ranjbar et To develop a reliable Retrained pre-trained Achieved reliable crack Some models showed slower
al. (2021) system for pavement DCNNs (e.g., AlexNet, classification with accuracies performance; requires large
crack detection and ResNet) using transfer ranging from 94% to 99%; labeled datasets for training
classification using pre- learning, with wavelet wavelet transform improved despite using transfer
trained deep CNNs and transform for segmentation. learning.
wavelet transform. segmentation.
Hernanda et Optimize road pothole Adjust hyperparameters in Improved detection with Dataset and pre-trained
al. (2022) and crack detection pre-trained CNN and optimal mAP and reduced model dependency may limit
using CNN with SSD compare mAP and loss loss values. Validated adaptability to new
MobileNet V2 and values with prior systems. effectiveness of the approach conditions or datasets.
hyperparameter tuning. in road inspections.
Ha et al. Automate detection, Use SqueezeNet, U-Net, Achieved 91.2% accuracy Limited crack types and
(2022) classification, and and Mobilenet-SSD for crack type and severity specific dataset constraints
severity assessment of models for segmentation detection. Demonstrated reduce generalizability.
road cracks using deep and severity evaluation. enhanced performance for
learning models. pavement management
systems.
Fan et al. Develop a high- Introduced Parallel ResNet Achieved high precision and Dependency on specific
(2022) performance pavement to minimize noise and recall scores (e.g., F1 score datasets and computational
crack detection method accurately identify cracks ~95%). Enhanced ability to complexity for large-scale
addressing noise and in public datasets segment crack features real-world deploymen
topology issues. (CrackTree200, CFD). amidst noise.
Qu et al. Address challenges in Developed CrackT-Net Improved F1 scores across Computational cost of
(2022) long dependencies and using enriched features datasets (e.g., DeepCrack: transformer models and
global context loss in (RF UNet++) with 0.859). Enhanced global reliance on specific datasets.
crack segmentation transformers for better context understanding and
using CNN and feature representation and segmentation capabilities.
transformer techniques. segmentation.
Xu and Liu Develop crack detection Applied GANs for data Increased accuracy from Limited effectiveness for
(2022) methods using small augmentation and CNN 80.75% to 91.61% after highly diverse datasets.
sample datasets with models with transfer GAN-based augmentation. Dependency on GAN-
GANs for data learning to improve Significant improvement in generated data quality.
augmentation and CNN accuracy. crack detection for limited
for detection datasets.

Table 1 provides a comprehensive summary of various studies utilizing transfer learning (TL) techniques for crack
detection and classification. Although TL methods have achieved significant success in this domain, a critical
review of the literature reveals several limitations. Notably, only a limited number of studies have undertaken a
systematic evaluation of multiple pre-trained CNN models, highlighting the need for a more thorough investigation
into their comparative performance in crack classification. Furthermore, prior research has addressed the
optimization of pre-trained CNN architectures, with most efforts confined to applying existing optimization
algorithms to enhance these models. A detailed comparative analysis of pre-trained CNN models, however, can
yield valuable insights into the potential of fine-tuning their architectural layers to improve performance. This
approach not only refines the models but also contributes to the development of robust, automated TL-based
systems for crack classification.

To address these gaps, the present study goes beyond the standard application of transfer learning by emphasizing
comprehensive model evaluation, layer optimization, and generalizability. Unlike prior works that primarily test
pre-existing architectures, this study explores the structural adaptability of pre-trained CNNs, demonstrating how
fine-tuning specific layers can significantly improve model performance. Additionally, the study introduces a
robust methodology for selecting optimal architectures tailored to crack detection, which fills a critical gap in the
literature.

ITcon Vol. 29 (2024), Matarneh et al., pg. 1243

Table 1: Comparison between existing studies used transfer learning and optimisation methods for crack detection
and classification. Part Two.
Reference Aim Method Results Limitations
Matarneh et Optimize pre-trained Compared ten pre-trained DenseNet201 and GWO Inconsistent performance of
al., (2024) CNNs for detecting models, optimized optimizer achieved highest other CNN models (e.g.,
and classifying DenseNet201 with feature accuracy and noise VGG16) and dependency on
pavement cracks with selection and noise- robustness. Highlighted pre-trained models for
a focus on resistance methods. effectiveness of transfer scalability.
DenseNet201 and learning in reducing training
other architectures. time and errors.
Lu et al. Develop an Optimized YOLOv5 with Improved mAP to 96.73%, Limited to ceramic tiles with
(2022) intelligent YOLOv5- Shufflenetv2 backbone and with reduced parameters and specific defect
based system for multiple sliding windows computational cost compared characteristics; effectiveness
defect detection in for feature extraction and to baseline YOLOv5. may vary for other surface
ceramic tiles. classification. types.
Liu et al. Classify asphalt Compared thirteen CNN EfficientNet-B3 achieved the Misclassification occurred
(2022a) pavement crack models using datasets of best accuracy for all image primarily at medium and
severity levels using visible, infrared, and fused types. Infrared imaging high severity levels. Dataset
CNNs and images; applied transfer improved detection in low- biases and varying imaging
thermography. learning. light conditions. conditions may affect
reliability.
Wang and Guo Automate Employed Faster R-CNN ResNet-101 achieved the Limited by small dataset size
(2021) identification of with VGG-16, ResNet-101, highest accuracy (95.9%), and image characteristics like
fatigue crack and FPN as feature balancing precision and low contrast and resolution.
initiation sites in extractors for transfer training costs.
engineering learning.
structures.
Yang et al. Use deep learning to Compared AlexNet, ResNet18 provided superior Dependency on manually
(2021) detect and recognize VGGNet13, ResNet18, and image recognition, while collected and augmented
structural cracks. YOLOv3 for crack YOLOv3 excelled in real- datasets; performance varies
detection and recognition. time crack area detection under adverse environmental
with high precision. conditions like poor lighting.
Pozzer et al. Assess deep learning Trained CNN models MobileNetV2 achieved Limited generalizability to
(2021) methods for detecting (MobileNetV2, VGG16) 79.7% accuracy in thermal diverse conditions; model
defects in damaged on thermal and regular imaging for multiclass performance varies
concrete using images for multiclass damage detection; VGG16 significantly with viewpoint
thermal and regular defect detection (cracks, reduced false positives and distance in real-world
imaging. spalling). effectively. settings.
Louati et al., Optimize CNN Employed co-evolutionary Achieved lightweight, High computational cost for
(2022) architecture and migration-based algorithm optimized CNN architectures the optimization process;
compression jointly (CEMBA) for architectural with reduced parameters limited evaluation on small-
using bi-level design and filter pruning. while maintaining high scale datasets and specific
optimization. accuracy on CIFAR and architectures.
ImageNet datasets.

In summary, this study addresses these gaps by:

• Conducting a systematic comparison of five pre-trained CNN models, evaluating their transfer learning
potential for crack detection to identify the most effective architecture.
• Introducing a novel approach to optimize CNN layers, focusing on refining internal structures to enhance
both accuracy and computational efficiency.
• Evaluating performance using diverse datasets, including varying crack types and environmental settings,
to ensure the model's robustness and generalizability.
• Highlighting fine-tuned transfer learning as a strategy to mitigate overfitting and enhance the adaptability
of pre-trained CNNs for crack detection tasks.

2. METHODOLOGY
This empirical research adopted a positivist philosophy and deductive reasoning (Edwards et al., 2020) to analyze
secondary image data obtained from open-source databases to accurately model the phenomenon under
investigation. Such an approach has been extensively used previously to evaluate risk factors impacting upon
public-private partnerships (Kukah et al., 2022); assess the risk associate with sustainable housing (Adabre et al.,
2022); and model construction machinery stability (Edwards et al., 2019). Therefore, this well-established

ITcon Vol. 29 (2024), Matarneh et al., pg. 1244

approach is deemed suitable for the present study.

2.1 Transfer learning

The primary objective of transfer learning is to improve the learning of the target prediction function by leveraging
knowledge from the source domain and associated learning tasks (Pan and Yang, 2010). Transfer learning typically
employs four key techniques viz.: transferring instances knowledge; transferring feature representations
knowledge; transferring parameter knowledge; and transferring relation knowledge (Pan and Yang, 2010). There
are four basic types of transfer learning used in CNN networks viz.: network transfer learning; adversarial transfer
learning; instance-based transfer learning; and mapping transfer learning (Tan et al., 2018). The use of transfer
learning has proven effective in pavement crack detection (Gopalakrishnan et al., 2017, Joshi et al., 2022, Liu et
al., 2022b, Ranjbar et al., 2021, Xu and Liu, 2022).
When using CNN to categorize the asphalt pavement cracks severity, there are two primary processes (Figure 1).
First, very large picture datasets for the source domain (frequently ImageNet which has 1.2 million images with
1000 categories) are used to train the CNN model (Deng et al., 2009). The pre-trained CNN network (otherwise
known as the CNN trained network), is available via web resources (Paszke et al., 2019). Secondly, the architecture
for the target domain is the pre-trained CNN architecture. Unlike existing studies which focused on utilizing the
existing pre-trained CNN architectures, this study retrained the pre-trained convolutional layers parameters, then
modified the fully connected layers (FC layers) to satisfy the requirements of the stated output labels (four labels
in this case viz.: alligator; diagonal; horizontal; and longitudinal). The modified pre-trained CNN model is then
improved (trained and assessed) using the study dataset.

Figure 1: A transfer learning strategy using a trained CNN for identifying asphalt pavement cracks.

2.2 Performance evaluation metrics

The most typical accuracy measures for the crack classification task involve precision, accuracy, recall, F1 score

ITcon Vol. 29 (2024), Matarneh et al., pg. 1245

and confusion matrix (Yacouby & Axman, 2020). Classification accuracy is defined as the total number of correct
predictions divided by the total number of predictions made for a given dataset (Ilse et al., 2020) and is calculated
as follows:

(1)

Precision as shown in equation (2) is another metric which calculates the total of correct positive predictions made.
Thus, precision is a metric used to calculate the accuracy for the marginal class (Ilse et al., 2020).

(2)

Recall as shown in equation (3) is a metric that measures the proportion of correct positive predictions out of all
possible positive instances. In contrast to precision, which reflects the accuracy of positive predictions, recall
provides insight into the number of missed positive predictions. Therefore, recall gives an indication of the model's
coverage of the positive class (Ilse et al., 2020).

(3)

The F1-score combines both precision and recall into a single metric, effectively capturing both aspects of model
performance. After calculating precision and recall for either a binary or multiclass classification task, these two
metrics are used to compute the F1-score (Ilse et al., 2020) as shown in equation (4).

(4)

The classification problem is frequently related to the multi-class classification, which classifies instances into
three or more classes (Ilse et al., 2020). Precision, recall and F1 score can be classified into three categories for
multi-class classification viz.: ‘macro’, ‘micro’ and ‘weighted’. In macro averaging, the multiclass predictions are
reduced into various sets of binary predictions, the corresponding metric is determined for each of the binary cases,
and the results are averaged for k classes (Luo & Uzuner, 2014) as shown in equations (5), (6) and (7).

(5)

(6)

(7)

Micro averaging treats the entire set of data as an aggregate result and calculates 1 metric rather than k metrics
that get averaged together. Like macro, weighted determines the weighted mean by taking label imbalance into
account (Luo & Uzuner, 2014) as shown in equations (8), (9), and (10).

(8)

(9)

(10)

2.3 Pre-trained models hyperparameters

The CNN model was developed based on web resources (Paszke et al., 2019). This online repository contains an
op-for-op PyTorch reimplementation of various pre-trained models. Five pre-trained models are used in this study;
details about the models are provided in Table 2.

ITcon Vol. 29 (2024), Matarneh et al., pg. 1246

Table 2: The pre trained CNN model details.
No. Model name Total params
1 DenseNet121 7M
2 EfficientNetB0 4M
3 EfficientNetB3 10.8M
4 MobileNet 3.2M
5 MobileNetV2 2.2M

In these pretrained models the raw images are pre-processed by resizing and normalising all images to become
224 × 224. The learning rate (0.001), training epoch (10) and batch size of 64.

2.4 Dataset
The study utilised public datasets that were obtained from the GitHub website. The main dataset used is
CRACK500, consisting of 500 images, each with a size of around 2,000 × 1,500 pixels. The images were taken
on the main campus of Temple University using mobile phones. Yang et al. (2019) augmented the images number
by splitting each image into 16 separate regions without overlap. Only the regions that included more than 1,000
pixels of crack were retained. This method increased the size of the training dataset by incorporating 1,896
additional photos. To evaluate the pre-trained CNN models, additional datasets from GitHub are combined with
the primary dataset to create 2,380 pavement crack images. The dataset's images are divided into four groups (to
represent the four types of cracks: longitudinal, horizontal, diagonal and alligator) as shown in Table 3. A random
selection of 20% of the collected images were pooled to create the test set (see Table 3).
Table 3: Dataset groups summary.
Dataset Longitudinal Horizontal Diagonal Alligator Total
Train 547 546 543 209 1845
Test 160 160 160 55 535
Total 707 706 703 264 2380

3. RESULTS
This section presents and discusses the results of the study, including results of the loss, and accuracy metrics.

3.1 Loss

Figure 2: Loss curves of the training and validation for (a) DenseNet121, (b) EfficientNetB0, (c) EfficientNetB3,
(d) MobileNet, and (e) MobileNetV2.

The loss value varied throughout the first few epochs but as the epoch increased, it decreased and proceeded to

ITcon Vol. 29 (2024), Matarneh et al., pg. 1247

converge (refer to Figure 2). Among all models, DeneseNet121 was the most volatile, whereas the variations of
the other four models were smoother as shown in Figure 2 (a,b,c,d and e). However, the EfficientNet-B3 model
has the smoothest variation among all other models and tended to converge faster than the other four models.

3.2 Accuracy metrics

The purpose of this research was to identify and classify cracks utilising a dataset of pavement images and five
pre-trained CNN models: DenseNet-121, EfficientNet-B0, EfficientNet-B3, MobileNet, and MobileNetV2. These
models were trained and tested, and subsequently validated. Each pre-trained CNN model performance was
measured using several key metrics, including, F1-score, recall, accuracy, and precision, which are commonly
utilised to assess CNN model performance.
Accuracy, for example, serves as an indicator of the model’s ability to correctly identify and categorize cracks
relative to the total number of detection iterations performed. Figure 3 presents the accuracy metric values of
MobileNet started out very high with an accuracy of 0.6475 during the whole training procedure. Contrarily,
DenseNet-121 started with a moderate accuracy of 0.4580, whereas MobileNetV2, EfficientNetB0 and
EfficientNet-B3 started with comparatively low accuracy of 0.3263, 0.3581 and 0.3870, respectively. Conversely,
the EfficientNet-B3 accuracy metrics increased quickly as the epoch augmented and suddenly exceeded that of
DenseNet-121 around Epoch 9.
The recall metric reflects the ability of the model to correctly detect positive instances. A strong relationship exists
between the number of positive instances detected and recall; with the greater recall values indicating a greater
number of correctly identified positive instances.

Figure 3: Accuracy metrics of the training and validation for (a) DenseNet121, (b) EfficientNetB0, (c)
EfficientNetB3, (d) MobileNet, and (e) MobileNetV2.

The results indicate that EfficientNet-B3 realized the highest recall rate at 94%, followed by DenseNet-121 at
92%, EfficientNet-B0 at 89%, MobileNet at 86% and MobileNet V2 at 84%. F1-score is a performance measure
that integrates both precision and recall into a single value, presenting a balanced assessment of a model's accuracy
in categorization tasks. As shown in Table 4, EfficientNet-B3 achieved the highest F1-score at 94%, followed by
DenseNet-121 at 92%, EfficientNet-B0 at 89%, MobileNet at 86%, and MobileNetV2 at 84%.
The accuracy performance parameters for the pre-trained CNNs are showed and compared in Table 4. Notably,

ITcon Vol. 29 (2024), Matarneh et al., pg. 1248

MobileNetV2 and MobileNet exhibited the lowest performance among all accuracy measures when compared to
the other models. In contrast, EfficientNet-B3 outperformed all other models, attaining the highest scores across
all accuracy metrics, with values of 94% for accuracy, precision, recall, and F1-score.
Table 4: Accuracy metrics summary.
Accuracy Precision Recall F1 score

Macro Weighted Macro Weighted Macro Weighted

DenseNet-121 0.92 0.91 0.92 0.91 0.92 0.91 0.92

EfficientNet-B0 0.84 0.91 0.90 0.88 0.89 0.89 0.89

EfficientNet-B3 0.94 0.92 0.94 0.94 0.94 0.93 0.94

MobileNet 0.87 0.84 0.87 0.87 0.86 0.85 0.86

MobileNetV2 0.84 0.86 0.86 0.81 0.84 0.83 0.84

Learning rate = 0.01, batch size = 64, and Max. Epochs = 10

4. DISCUSSION
The classification classes of the evaluated dataset were classified with varying degrees of accuracy using the same
pre-trained CNN models. The confusion matrix illustrates how well-trained models perform while categorizing
various classes. Figure 4 (a,b,c,d,e) shows the normalised confusion matrix, out of these confusion matrices, the
pre-trained models performed very effectively on the diagonal, longitudinal and alligator crack images. However,
misclassifications appeared on the horizontal crack for EfficientNetB0.
Moreover, the confusion matrices show that the alligator crack has less accuracy in the DenseNet121 and
MobileNetV2 models. This could be because the alligator data set is smaller than other datasets.

Figure 4: Normalised confusion matrix for (a) DenseNet121, (b) EfficientNetB0, (c) EfficientNetB3, (d) MobileNet,
and (e) MobileNetV2.

ITcon Vol. 29 (2024), Matarneh et al., pg. 1249

The pre-trained CNNs accuracy is illustrated in Figure 5. The EfficientNetB3 realized the superior accuracy among
the five pre-trained CNNs followed by DenseNet121 and MobileNet.

Figure 5: Accuracy of five pre-train CNN models.

4.1 Proposed approach: modifications in network architecture

Figure 6: The original and modified architecture diagrams of EfficientNet B3.

In this study, some modifications in EfficientNetB3 architecture are proposed to enhance the model performance.
Specifically, the top layers of the model were replaced with dense, batch normalization, and dropout layers, as
suggested by Ali et al. (2022). These modifications were applied to the B3 base architecture, utilizing the Swish
activation function for the dense (fully connected) layers, as recommended by Ramachandran et al. (2018) and

ITcon Vol. 29 (2024), Matarneh et al., pg. 1250

employed by Tan and Le (2019) instead of the conventional ReLU activation function. Figure 6 illustrates the
original EfficientNet-B3 baseline architecture alongside the proposed improvements. The original model was
preserved, while the top layers of EfficientNet B3 architecture were restructured. In the original model, the top
layers—comprising global average pooling 2D, dropout, and a dense layer—were prone to overfitting. In the
modified architecture, these layers were replaced with dense, batch normalization, and dropout layers to improve
performance and prevent overfitting.
4.2 Implementation and performance evaluation metric
In this experiment, a learning rate annealer was employed to reduce the learning rate after a specified number of
epochs if the error rate remained unchanged. The validation accuracy was closely monitored to detect any potential
plateau over a span of three epochs. If a plateau was observed, the learning rate was subsequently decreased by
0.01. After, the EfficientNet B3 convolutional layers received fine-tuning using a Stochastic Gradient Descent
(SGD) optimiser. The modified EfficientNets B3 was tested an evaluated using same dataset (Crack500). The
improved EfficientNet B3 reports have better results than the original EfficientNet B3 architecture. The improved
EfficientNet B3 achieved an F1 score of 96% and accuracy, precision, recall and recall of 96%.

4.3 Contributions to theory and practice

The theoretical contributions of the study lie in its innovative use of convolutional neural networks (CNNs) and
transfer learning to address the critical task of asphalt pavement crack classification. This research addresses gaps
in current methodologies by systematically evaluating five pre-trained CNN models: EfficientNet-B0,
EfficientNet-B3, DenseNet-121, MobileNet, and MobileNetV2, all fine-tuned for crack detection tasks using the
Crack500 dataset. The study's results highlight EfficientNet-B3 as the most effective model, achieving a 96% F1
score and 96% accuracy after applying advanced transfer learning techniques.
Key Theoretical Contributions:
1. Transfer Learning Optimization: By leveraging ImageNet pre-trained weights and fine-tuning the CNN
layers, the study demonstrates the effectiveness of transfer learning for domain-specific tasks. This
approach minimizes the need for extensive domain-specific data, a common challenge in pavement
inspection.
2. Evaluation of Multiple Models: The comparison of five CNN architectures provides a deeper
understanding of their strengths and weaknesses in multiclass classification. The study highlights
EfficientNet-B3’s architecture as optimal due to its ability to balance performance and computational
efficiency.
3. Scalability and Generalization: The study establishes a foundation for future research by demonstrating
that transfer learning techniques can generalize well to different types of cracks (e.g., longitudinal,
diagonal). It sets a roadmap for expanding datasets and testing additional pre-trained models to enhance
robustness further.
4. Integration into Practical Applications: Beyond theoretical advancements, the study offers insights into
integrating AI-driven crack detection into real-world workflows, such as automated road inspections,
reducing manual effort and improving accuracy.

By combining theoretical rigor with practical implications, this research provides a scalable framework that can
be built upon for broader pavement distress classifications and applied in various infrastructure management
scenarios.

5. CONCLUSION
Convolutional neural networks (CNNs) have quickly emerged as a prominent and effective method for crack
detection. While numerous studies have explored crack detection using CNNs, limited research has focused on
leveraging pre-trained models for this purpose. This study applied transfer learning to classify types of asphalt
pavement cracks. The main contributions of this research are: 1) the evaluation of five pre-trained CNN models,
originally trained on the ImageNet dataset, on the collected asphalt pavement crack dataset; 2) the assessment of

ITcon Vol. 29 (2024), Matarneh et al., pg. 1251

the impact of transfer learning on multiclass crack classification using metrics such as Precision, Recall, Accuracy,
F1 Score, and Confusion Matrices; and 3) the selection of the most reliable model for further fine-tuning by
incorporating additional layers to improve its performance.
The results show that the most reliable model is the EfficientNet B3, achieved scores of 94% F1_ Score and 94%
accuracy. Thus, this model was selected and trained on the same dataset by applying transfer learning with pre-
trained weights from ImageNet and fine-tuning the CNN. The modified model showed better classification
performance with 96% F1_Score and 96% accuracy. The pre-trained models performed very effectively on the
diagonal, longitudinal, and alligator crack images. However, misclassifications appeared on the horizontal crack
for EfficientNetB0. Moreover, the confusion matrices show that the alligator crack has less accuracy on two models
namely, DenseNet121 and MobileNetV2. The reason for this could be the that the alligator data set is smaller than
other datasets. EfficientNet B3 had the highest accuracy for both TL and transfer learning with fine tuning.
In future studies, additional crack classes, for example, fatigue cracks, can be examined to further assess the
effectiveness of TL and pre-trained models for the classification of asphalt pavement cracks. A larger dataset can
be collected, encompassing more crack classes and distresses, to test and assess various pre-trained models.
Furthermore, the optimised EfficientNetB0 will undergo additional training to identify a wider range of highway
pavement distress, including joint reflection cracking, slippage cracking, corrugation, shoving, block cracking etc.
The model's performance metrics will be compared across these distress types.
Additionally, a user-friendly interface will be developed to enable novice users without a programming background
to effectively utilise the model. The integration of automated crack detection and classification systems can
potentially replace human intervention due to their superior accuracy, cost-effectiveness, and automatic reporting
functionality. Future studies may explore the use of AI to develop a tool for suggesting corrective actions for
detected and classified cracks.

REFERENCES
Adabre, M. A., Chan, A. P. C., Edwards, D. J., & Osei-Kyei, R. (2022). To build or not to build, that is the
uncertainty: Fuzzy synthetic evaluation of risks for sustainable housing in developing economies. Cities,
125, 103644. https://doi.org/https://doi.org/10.1016/j.cities.2022.103644
Ali, K., Shaikh, Z. A., Khan, A. A., & Laghari, A. A. (2022). Multiclass skin cancer classification using
EfficientNets – a first step towards preventing skin cancer. Neuroscience Informatics, 2(4), 100034.
https://doi.org/https://doi.org/10.1016/j.neuri.2021.100034
Augustauskas, R., & Lipnickas, A. (2020). Improved Pixel-Level Pavement-Defect Segmentation Using a Deep
Autoencoder. Sensors, 20(9), 2557. https://www.mdpi.com/1424-8220/20/9/2557
Baduge, S. K., Thilakarathna, S., Perera, J. S., Ruwanpathirana, G. P., Doyle, L., Duckett, M., Lee, J., Saenda, J.,
& Mendis, P. (2023). Assessment of crack severity of asphalt pavements using deep learning algorithms
and geospatial system. Construction and Building Materials, 401, 132684.
https://doi.org/https://doi.org/10.1016/j.conbuildmat.2023.132684
Brien, D. O., Andrew Osborne, J., Perez-Duenas, E., Cunningham, R., & Li, Z. (2023). Automated crack
classification for the CERN underground tunnel infrastructure using deep learning. Tunnelling and
Underground Space Technology, 131, 104668. https://doi.org/https://doi.org/10.1016/j.tust.2022.104668
Canestrari, F., & Ingrassia, L. P. (2020). A review of top-down cracking in asphalt pavements: Causes, models,
experimental tools and future challenges. Journal of Traffic and Transportation Engineering (English
Edition), 7(5), 541-572. https://doi.org/https://doi.org/10.1016/j.jtte.2020.08.002
Cha, Y.-J., Choi, W., & Büyüköztürk, O. (2017). Deep Learning-Based Crack Damage Detection Using
Convolutional Neural Networks. Computer-Aided Civil and Infrastructure Engineering, 32(5), 361-378.
https://doi.org/https://doi.org/10.1111/mice.12263
Choi, W., & Cha, Y. J. (2020). SDDNet: Real-Time Crack Segmentation. IEEE Transactions on Industrial
Electronics, 67(9), 8016-8025. https://doi.org/10.1109/TIE.2019.2945265

ITcon Vol. 29 (2024), Matarneh et al., pg. 1252

Dais, D., Bal, İ. E., Smyrou, E., & Sarhosis, V. (2021). Automatic crack classification and segmentation on
masonry surfaces using convolutional neural networks and transfer learning. Automation in Construction,
125, 103606. https://doi.org/https://doi.org/10.1016/j.autcon.2021.103606
Dawson, H. L., Dubrule, O., & John, C. M. (2023). Impact of dataset size and convolutional neural network
architecture on transfer learning for carbonate rock classification. Computers & Geosciences, 171, 105284.
https://doi.org/https://doi.org/10.1016/j.cageo.2022.105284
Deng, J., Dong, W., Socher, R., Li, L. J., Kai, L., & Li, F.-F. (2009, 20-25 June 2009). ImageNet: A large-scale
hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition,
Dong, J., Wang, N., Fang, H., Hu, Q., Zhang, C., Ma, B., Ma, D., & Hu, H. (2022). Innovative method for pavement
multiple damages segmentation and measurement by the Road-Seg-CapsNet of feature fusion. Construction
and Building Materials, 324, 126719. https://doi.org/https://doi.org/10.1016/j.conbuildmat.2022.126719
Dung, C. V., Sekiya, H., Hirano, S., Okatani, T., & Miki, C. (2019). A vision-based method for crack detection in
gusset plate welded joints of steel bridges using deep convolutional neural networks. Automation in
Construction, 102, 217-229. https://doi.org/https://doi.org/10.1016/j.autcon.2019.02.013
Edwards, D., Parn, E. A., Sing, M. C. P., & Thwala, W. D. (2019). Risk of excavators overturning. Engineering,
Construction and Architectural Management, 26(3), 479-498. https://doi.org/10.1108/ECAM-03-2018-
0125
Edwards, D. J., Rillie, I., Chileshe, N., Lai, J., Hosseini, M. R., & Thwala, W. D. (2020). A field survey of hand–
arm vibration exposure in the UK utilities sector. Engineering, Construction and Architectural Management,
27(9), 2179-2198. https://doi.org/10.1108/ECAM-09-2019-0518
Elghaish, F., Talebi, S., Abdellatef, E., Matarneh, S. T., Hosseini, M. R., Wu, S., Mayouf, M., Hajirasouli, A., &
Nguyen, T.-Q. (2022). Developing a new deep learning CNN model to detect and classify highway cracks.
Journal of Engineering, Design and Technology, 20(4), 993-1014. https://doi.org/10.1108/JEDT-04-2021-
0192
Fan, Z., Lin, H., Li, C., Su, J., Bruno, S., & Loprencipe, G. (2022). Use of Parallel ResNet for High-Performance
Pavement Crack Detection and Measurement. Sustainability, 14(3), 1825. https://www.mdpi.com/2071-
1050/14/3/1825
Feng, C., Zhang, H., Wang, S., Li, Y., Wang, H., & Yan, F. (2019). Structural Damage Detection using Deep
Convolutional Neural Network and Transfer Learning. KSCE Journal of Civil Engineering, 23(10), 4493-
4502. https://doi.org/10.1007/s12205-019-0437-z
Gopalakrishnan, K., Khaitan, S. K., Choudhary, A., & Agrawal, A. (2017). Deep Convolutional Neural Networks
with transfer learning for computer vision-based data-driven pavement distress detection. Construction and
Building Materials, 157, 322-330. https://doi.org/https://doi.org/10.1016/j.conbuildmat.2017.09.110
Guo, W., Zhang, J., Cao, D., & Yao, H. (2022). Cost-effective assessment of in-service asphalt pavement condition
based on Random Forests and regression analysis. Construction and Building Materials, 330, 127219.
https://doi.org/https://doi.org/10.1016/j.conbuildmat.2022.127219
Guzmán-Torres, J. A., Naser, M. Z., & Domínguez-Mota, F. J. (2022). Effective medium crack classification on
laboratory concrete specimens via competitive machine learning. Structures, 37, 858-870.
https://doi.org/https://doi.org/10.1016/j.istruc.2022.01.061
Ha, J., Kim, D., & Kim, M. (2022). Assessing severity of road cracks using deep learning-based segmentation and
detection. The Journal of Supercomputing, 78(16), 17721-17735. https://doi.org/10.1007/s11227-022-
04560-x
Hernanda, Z. S., Mahmudah, H., & Sudibyo, R. W. (2022, 6-9 June 2022). CNN-Based Hyperparameter
Optimization Approach for Road Pothole and Crack Detection Systems. 2022 IEEE World AI IoT Congress
(AIIoT),

ITcon Vol. 29 (2024), Matarneh et al., pg. 1253

Hou, Y., Li, Q., Han, Q., Peng, B., Wang, L., Gu, X., & Wang, D. (2021). MobileCrack: Object Classification in
Asphalt Pavements Using an Adaptive Lightweight Deep Learning. Journal of Transportation Engineering,
Part B: Pavements, 147(1), 04020092. https://doi.org/doi:10.1061/JPEODX.0000245
Huyan, J., Li, W., Tighe, S., Xu, Z., & Zhai, J. (2020). CrackU-net: A novel deep convolutional neural network for
pixelwise pavement crack detection. Structural Control and Health Monitoring, 27(8), e2551.
https://doi.org/https://doi.org/10.1002/stc.2551
Ilse, M., Tomczak, J. M., & Welling, M. (2020). Chapter 22 - Deep multiple instance learning for digital
histopathology. In S. K. Zhou, D. Rueckert, & G. Fichtinger (Eds.), Handbook of Medical Image
Computing and Computer Assisted Intervention (pp. 521-546). Academic Press.
https://doi.org/https://doi.org/10.1016/B978-0-12-816176-0.00027-2
Jang, K., Kim, N., & An, Y.-K. (2019). Deep learning–based autonomous concrete crack evaluation through hybrid
image scanning. Structural Health Monitoring, 18(5-6), 1722-1737.
https://doi.org/10.1177/1475921718821719
Joshi, D., Singh, T. P., & Sharma, G. (2022). Automatic surface crack detection using segmentation-based deep-
learning approach. Engineering Fracture Mechanics, 268, 108467.
https://doi.org/https://doi.org/10.1016/j.engfracmech.2022.108467
Kim, H., Kim, H., Hong, Y. W., & Byun, H. (2018). Detecting Construction Equipment Using a Region-Based
Fully Convolutional Network and Transfer Learning. Journal of Computing in Civil Engineering, 32(2),
04017082. https://doi.org/doi:10.1061/(ASCE)CP.1943-5487.0000731
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet classification with deep convolutional neural
networks. Commun. ACM, 60(6), 84–90. https://doi.org/10.1145/3065386
Kukah, A. S., Owusu-Manu, D.-G., Badu, E., & Edwards, D. (2022). Evaluation of risk factors in Ghanaian public-
private- partnership (PPP) power projects using fuzzy synthetic evaluation methodology (FSEM).
Benchmarking An International Journal, 30. https://doi.org/10.1108/BIJ-09-2021-0533
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep Learning. Nature, 521, 436-444.
https://doi.org/10.1038/nature14539
Li, G., Ma, B., He, S., Ren, X., & Liu, Q. (2020). Automatic Tunnel Crack Detection Based on U-Net and a
Convolutional Neural Network with Alternately Updated Clique. Sensors, 20(3), 717.
https://www.mdpi.com/1424-8220/20/3/717
Li, J., Liu, T., Wang, X., & Yu, J. (2022). Automated asphalt pavement damage rate detection based on optimized
GA-CNN. Automation in Construction, 136, 104180.
https://doi.org/https://doi.org/10.1016/j.autcon.2022.104180
Li, Y., Li, H., & Wang, H. (2018). Pixel-Wise Crack Detection Using Deep Local Pattern Predictor for Robot
Application. Sensors, 18(9), 3042. https://www.mdpi.com/1424-8220/18/9/3042
Liu, F., Liu, J., & Wang, L. (2022a). Asphalt pavement fatigue crack severity classification by infrared
thermography and deep learning. Automation in Construction, 143, 104575.
https://doi.org/https://doi.org/10.1016/j.autcon.2022.104575
Liu, F., Liu, J., & Wang, L. (2022b). Deep learning and infrared thermography for asphalt pavement crack severity
classification. Automation in Construction, 140, 104383.
https://doi.org/https://doi.org/10.1016/j.autcon.2022.104383
Liu, J., Zhao, Z., Lv, C., Ding, Y., Chang, H., & Xie, Q. (2022). An image enhancement algorithm to improve road
tunnel crack transfer detection. Construction and Building Materials, 348, 128583.
https://doi.org/https://doi.org/10.1016/j.conbuildmat.2022.128583
Liu, Z., Cao, Y., Wang, Y., & Wang, W. (2019). Computer vision-based concrete crack detection using U-net fully
convolutional networks. Automation in Construction, 104, 129-139.
https://doi.org/https://doi.org/10.1016/j.autcon.2019.04.005

ITcon Vol. 29 (2024), Matarneh et al., pg. 1254

Louati, H., Bechikh, S., Louati, A., Aldaej, A., & Said, L. B. (2022). Joint design and compression of convolutional
neural networks as a Bi-level optimization problem. Neural Computing and Applications, 34(17), 15007-
15029. https://doi.org/10.1007/s00521-022-07331-0
Loverdos, D., & Sarhosis, V. (2022). Automatic image-based brick segmentation and crack detection of masonry
walls using machine learning. Automation in Construction, 140, 104389.
https://doi.org/https://doi.org/10.1016/j.autcon.2022.104389
Lu, Q., Lin, J., Luo, L., Zhang, Y., & Zhu, W. (2022). A supervised approach for automated surface defect detection
in ceramic tile quality control. Advanced Engineering Informatics, 53, 101692.
https://doi.org/https://doi.org/10.1016/j.aei.2022.101692
Luo, Y., & Uzuner, O. (2014). Semi-Supervised Learning to Identify UMLS Semantic Relations. AMIA Joint
Summits on Translational Science proceedings AMIA Summit on Translational Science, 2014, 67-75.
Mandal, V., Uong, L., & Adu-Gyamfi, Y. (2018, 10-13 Dec. 2018). Automated Road Crack Detection Using Deep
Convolutional Neural Networks. 2018 IEEE International Conference on Big Data (Big Data),
Matarneh, S., Elghaish, F., Pour Rahimian, F., Abdellatef, E., & Abrishami, S. (2024). Evaluation and optimisation
of pre-trained CNN models for asphalt pavement crack detection and classification. Automation in
Construction, 160, 105297. https://doi.org/https://doi.org/10.1016/j.autcon.2024.105297
Mazzini, D., Napoletano, P., Piccoli, F., & Schettini, R. (2020). A Novel Approach to Data Augmentation for
Pavement Distress Segmentation. Computers in Industry, 121, 103225.
https://doi.org/https://doi.org/10.1016/j.compind.2020.103225
Pan, S. J., & Yang, Q. (2010). A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data
Engineering, 22(10), 1345-1359. https://doi.org/10.1109/TKDE.2009.191
Pan, Y., Zhang, G., & Zhang, L. (2020). A spatial-channel hierarchical deep learning network for pixel-level
automated crack detection. Automation in Construction, 119, 103357.
https://doi.org/https://doi.org/10.1016/j.autcon.2020.103357
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga,
L., Desmaison, A., Köpf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B.,
Fang, L., . . . Chintala, S. (2019). PyTorch: An Imperative Style, High-Performance Deep Learning Library.
ArXiv, abs/1912.01703.
Pei, L., Sun, Z., Xiao, L., Li, W., Sun, J., & Zhang, H. (2021). Virtual generation of pavement crack images based
on improved deep convolutional generative adversarial network. Engineering Applications of Artificial
Intelligence, 104, 104376. https://doi.org/https://doi.org/10.1016/j.engappai.2021.104376
Pozzer, S., Azar, E. R., Rosa, F. D., & Pravia, Z. M. C. (2021). Semantic Segmentation of Defects in Infrared
Thermographic Images of Highly Damaged Concrete Structures. Journal of Performance of Constructed
Facilities, 35(1), 04020131. https://doi.org/doi:10.1061/(ASCE)CF.1943-5509.0001541
Qu, Z., Li, Y., & Zhou, Q. (2022). CrackT-net: a method of convolutional neural network and transformer for crack
segmentation. Journal of Electronic Imaging, 31(2), 023040. https://doi.org/10.1117/1.JEI.31.2.023040
Ramachandran, P., Zoph, B., & Le, Q. V. (2018). Searching for Activation Functions. ArXiv, abs/1710.05941.
Ranjbar, S., Nejad, F. M., & Zakeri, H. (2021). An image-based system for pavement crack evaluation using
transfer learning and wavelet transform. International Journal of Pavement Research and Technology, 14(4),
437-449. https://doi.org/10.1007/s42947-020-0098-9
Rubio, J. J., Kashiwa, T., Laiteerapong, T., Deng, W., Nagai, K., Escalera, S., Nakayama, K., Matsuo, Y., &
Prendinger, H. (2019). Multi-class structural damage segmentation using fully convolutional networks.
Computers in Industry, 112, 103121. https://doi.org/https://doi.org/10.1016/j.compind.2019.08.002
Shi, L., Long, Y., Wang, Y., Chen, X., & Zhao, Q. (2022). Evaluation of Internal Cracks in Turbine Blade Thermal
Barrier Coating Using Enhanced Multi-Scale Faster R-CNN Model. Applied Sciences, 12(13), 6446.
https://www.mdpi.com/2076-3417/12/13/6446

ITcon Vol. 29 (2024), Matarneh et al., pg. 1255

Tan, C., Sun, F., Kong, T., Zhang, W., Yang, C., & Liu, C. (2018, 2018//). A Survey on Deep Transfer Learning.
Artificial Neural Networks and Machine Learning – ICANN 2018, Cham.
Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. ArXiv,
abs/1905.11946.
Tan, Y., Cai, R., Li, J., Chen, P., & Wang, M. (2021). Automatic detection of sewer defects based on improved you
only look once algorithm. Automation in Construction, 131, 103912.
https://doi.org/https://doi.org/10.1016/j.autcon.2021.103912
Tang, W., Huang, S., Zhao, Q., Li, R., & Huangfu, L. (2022). An Iteratively Optimized Patch Label Inference
Network for Automatic Pavement Distress Detection. IEEE Transactions on Intelligent Transportation
Systems, 23(7), 8652-8661. https://doi.org/10.1109/TITS.2021.3084809
Teng, S., Liu, Z., & Li, X. (2022). Improved YOLOv3-Based Bridge Surface Defect Detection by Combining
High- and Low-Resolution Feature Images. Buildings, 12(8), 1225. https://www.mdpi.com/2075-
5309/12/8/1225
Tran, K., Kondrashova, O., Bradley, A., Williams, E., Pearson, J., & Waddell, N. (2021). Deep learning in cancer
diagnosis, prognosis and treatment selection. Genome Medicine, 13. https://doi.org/10.1186/s13073-021-
00968-x
Wang, S. Y., & Guo, T. (2021). Transfer Learning-Based Algorithms for the Detection of Fatigue Crack Initiation
Sites: A Comparative Study [Original Research]. Frontiers in Materials, 8.
https://doi.org/10.3389/fmats.2021.756798
Wu, L., Lin, X., Chen, Z., Lin, P., & Cheng, S. (2021). Surface crack detection based on image stitching and
transfer learning with pretrained convolutional neural network. Structural Control and Health Monitoring,
28(8), e2766. https://doi.org/https://doi.org/10.1002/stc.2766
Wu, Y., Yang, W., Pan, J., & Chen, P. (2021). Asphalt pavement crack detection based on multi-scale full
convolutional network. Journal of Intelligent & Fuzzy Systems, 40, 1495-1508.
https://doi.org/10.3233/JIFS-191105
Xu, B., & Liu, C. (2022). Pavement crack detection algorithm based on generative adversarial network and
convolutional neural network under small samples. Measurement, 196, 111219.
https://doi.org/https://doi.org/10.1016/j.measurement.2022.111219
Yacouby, R., & Axman, D. (2020). Probabilistic Extension of Precision, Recall, and F1 Score for More Thorough
Evaluation of Classification Models. https://doi.org/10.18653/v1/2020.eval4nlp-1.9
Yang, C., Chen, J., Li, Z., & Huang, Y. (2021). Structural Crack Detection and Recognition Based on Deep
Learning. Applied Sciences, 11(6), 2868. https://www.mdpi.com/2076-3417/11/6/2868
Yang, F., Zhang, L., Yu, S., Prokhorov, D., Mei, X., & Ling, H. (2019). Feature Pyramid and Hierarchical Boosting
Network for Pavement Crack Detection. IEEE Transactions on Intelligent Transportation Systems, PP, 1-
11. https://doi.org/10.1109/TITS.2019.2910595
Yang, Q., Shi, W., Chen, J., & Lin, W. (2020). Deep convolution neural network-based transfer learning method
for civil infrastructure crack detection. Automation in Construction, 116, 103199.
https://doi.org/https://doi.org/10.1016/j.autcon.2020.103199
Yao, H., Liu, Y., Li, X., You, Z., Feng, Y., & Lu, W. (2022). A Detection Method for Pavement Cracks Combining
Object Detection and Attention Mechanism. IEEE Transactions on Intelligent Transportation Systems,
23(11), 22179-22189. https://doi.org/10.1109/TITS.2022.3177210
Yoon, H.-Y., Kim, J.-H., & Jeong, J.-W. (2022). Classification of the Sidewalk Condition Using Self-Supervised
Transfer Learning for Wheelchair Safety Driving. Sensors, 22(1), 380. https://www.mdpi.com/1424-
8220/22/1/380
Zhang, J., Lu, C., Wang, J., Wang, L., & Yue, X.-G. (2019). Concrete Cracks Detection Based on FCN with Dilated
Convolution. Applied Sciences, 9(13), 2686. https://www.mdpi.com/2076-3417/9/13/2686

ITcon Vol. 29 (2024), Matarneh et al., pg. 1256