International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

ISSN (Print): 2394-5443 ISSN (Online): 2394-7454

Research Article
[Link]

Enhancing road crack detection using hybrid PSO-WOA feature selection and
random forest classifier
Pallav Kumar1* and Shivangi Mishra2
Assistant Professor, Department of Civil Engineering, MIT Muzaffarpur, Muzaffarpur, 842003, Bihar, India1
Assistant Town Planning Supervisor, Urban Development and Housing Department, Patna, 800015, Bihar, India2

Received: 05-October-2024; Revised: 10-September-2025; Accepted: 16-September-2025

©2025 Pallav Kumar and Shivangi Mishra. This is an open access article distributed under the Creative Commons Attribution
(CC BY) License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Abstract
Road crack detection plays a vital role in maintaining infrastructure safety and preventing accidents. A robust approach
was presented for road crack detection using a hybrid feature extraction method and a feature selection technique,
coupled with a powerful classification algorithm. The hybrid approach combines texture-based features, including
multiscale local binary pattern (MSLBP), discrete cosine transform (DCT), speeded-up robust features (SURF), and
convolutional neural network (CNN) features, to capture both local and global characteristics of road crack images. To
further enhance classification accuracy and reduce dimensionality, a hybrid particle swarm optimization and whale
optimization algorithm (PSO-WOA) is employed for feature selection. Hybrid PSO-WOA ensures that the selected
features have high relevance to the classification task while minimizing redundancy among them. By leveraging PSO-
WOA, the most discriminative features for accurate road crack detection are identified. Subsequently, the selected
features are utilized to train a random forest (RF) classifier, a robust and ensemble-based algorithm known for handling
complex datasets and maintaining strong generalization capabilities. The RF classifier accurately classifies road crack
images into relevant categories, enabling effective detection and localization of cracks on road surfaces. Extensive
experiments on a dataset of road crack images were conducted to evaluate the performance of the proposed approach.
Various performance metrics, including accuracy, sensitivity, specificity, precision, and F1-Score, were measured to
assess the quality of the approach. The results demonstrated that the hybrid feature extraction, PSO-WOA-based feature
selection, and RF classification method significantly enhanced road crack detection accuracy and reliability compared to
existing techniques.

Keywords
Road crack detection, Hybrid feature extraction, PSO-WOA optimization, Random forest classifier, Image classification,
Infrastructure safety.

[Link] These methods typically apply high-resolution

The role of road infrastructure is to ensure safe and imagery systems for detecting surface defects and can
efficient transportation systems. The problem is that, perform this task much faster and more accurately
with time, road surfaces eventually break down due than manual methods. In machine learning, a
to traffic and environmental effects and tend to get combination of feature extraction and classification
worn out, which affects traffic safety and the stability techniques is utilized to improve detection
of the infrastructure. For instance, traditional manual performance [3].
inspection methods are very expensive, time-
consuming, and prone to errors [1]. Recently, road Although advancements in automated road crack
obstruction detection solutions based on image detection have been achieved, several challenges
processing and machine learning have been proposed remain, including issues related to image quality,
as efficient approaches for detecting road cracks [2]. varying lighting conditions, and different crack types.
Real-world crack images are often highly complex
and subtle, which makes feature extraction
challenging and frequently results in reduced
accuracy [2]. Another challenge is feature
*Author for correspondence redundancy, where features are not entirely irrelevant

1360
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

but still negatively affect classification performance, techniques has been proposed to improve detection
thereby degrading the overall model quality [3]. In performance and reduce dependence on high-quality
this study, a hybrid feature extraction method and an images [9].
efficient feature selection technique were employed
to enhance accuracy and minimize redundancy in the The primary objective of this paper is to enhance the
feature extraction process. accuracy of road crack detection through the
application of a hybrid feature extraction process and
A crucial issue in the domain of road crack detection a hybrid feature selection approach based on PSO
is the inherent complexity and variability of real- and the whale optimization algorithm (WOA). The
world conditions. Challenges include changing study also aims to address the challenges posed by
lighting conditions, the quality of captured images, variations in image quality and feature redundancy by
and the diverse types of cracks that may appear on employing an efficient classification strategy for road
road surfaces. Although combinations of machine cracks.
learning and deep learning strategies have been
applied, they are not immune to these barriers. The A combination of feature extraction methods—
shortage of labeled datasets and the high multiscale local binary pattern (MSLBP), discrete
computational cost remain significant obstacles [4]. cosine transform (DCT), speeded-up robust features
Furthermore, noise in images often hinders the (SURF), and CNN—along with their hybridization is
detection process, resulting in false positives and proposed for automatic road crack detection. A
compromising the accuracy of automated detection hybrid PSO-WOA mechanism is employed to
systems. achieve optimal feature selection with minimal
redundancy and high discriminability. The selected
Recent research has highlighted several challenges in features are subsequently classified using a random
the development of future feature extraction and forest (RF) classifier to accurately categorize road
classification methods, emphasizing the need for images. Extensive experiments demonstrate that the
scalable and adaptable approaches that can handle a proposed method achieves higher accuracy and
wide variety of road cracks and environmental reliability compared to existing approaches.
conditions. Several studies have focused on
improving the precision of road crack detection The remainder of this paper is organized as follows:
through the use of combined image features, Section 2 presents a review of related work. Section 3
including texture-based features and frequency- describes the proposed methodology for feature
domain features [5]. Furthermore, hybrid approaches extraction, selection, and classification. Section 4
incorporating optimization techniques such as provides the experimental results, followed by a
particle swarm optimization (PSO) and genetic discussion in Section 5. Finally, Section 6 concludes
algorithms (GA) have been proposed to minimize the paper with a summary of the findings and
feature redundancy and improve classification potential directions for future research.
accuracy [6].
[Link] review
Also, recent developments with deep learning, and The task of segmenting cracks is a challenging
convolutional neural networks (CNNs), have problem that has been extensively studied by
demonstrated notable increases in accuracy of crack researchers using a variety of image processing and
detection. Nevertheless, the classification of CNN machine learning techniques. The primary objective
models with small data can also be a problem, and is to establish a clear distinction between the
not all of them can effectively generalize to other pavement background and the damaged areas. In
types of cracks and to other environmental factors most cases, binary images are utilized, where cracks
[7]. As road crack detection models continue to are highlighted as the focus regions and assigned a
advance, there is an increasing need for larger and value of 1 (or 255 in the case of 8-bit depth images).
more diverse datasets, as well as more
computationally intelligent approaches for real-world Wang et al. [10] introduced a new mechanism,
applications. Another important area of research known as comp-potential crack exponents (PCrCs),
involves the development of real-time detection which allows them to carry out a binary segmentation
systems capable of operating effectively under – especially those pertaining to transverse and
varying environmental conditions [8]. Additionally, longitudinal cracks. This approach involves filtering
the use of multi-modal data and sensor fusion the intensity levels between the pavement and cracks,
1361
Pallav Kumar and Shivangi Mishra

which then eradicate disconnected parts effectively. different types of cracks or within situations marked
Other elements define the cracks with the help of the by a large variance in crack morphology. Cheng et al.
shape metric (SM) designed by the authors, giving [16] use U-NET architecture, which might have
thus a lower rate of false positives than the current trouble in delineating accurately cracks at places with
machine learning-based assessments show. However, overlapping features or presence of noise. The deep
this approach could have difficulty identifying very residual network for pixel-level classification is
complex forms of cracks and would need adopted by Bang et al. [17], but this kind of system
considerable computational resources to process. may have difficulty in training because of complexity
Alternatively, Amhaz et al. [11] suggested a unique or small amount of labeled data. Additionally,
concept called minimum path selection (MPS), which strategies such as data fusion, for example by Xu et
is based on the premise that the minimum path of a al. [18], provide millimeter accuracy in
crack will not be the lowest cost function within all reconstruction of defects, but suffer in the merging of
the possible paths in image. This entails two phases: data across different sensors or challenging scenarios
detection of crucial pixels delineating the starting and involving occlusions or contrast across levels of
peak of crack paths, and based on these points, all the surface reflectance properties.
paths between them are detected, and afterward, post-
processing is performed to remove short isolated In the field of pavement defect classification utilizing
paths as signs of image artifacts. Although helpful in only one machine learning algorithm, there are
most cases, this approach might fail at detecting different methods, where the particularities of
cracks in areas of high noise and result in false analyzed defects define the peculiarity of the
negative [Link], the authors determine objectives. As an example, Hoang and Nguyen [19]
the magnitude of the cracks by the values of perform a classification algorithm analysis
neighboring pixels located along the detected paths. (supervised) in order to identify the best method of
Another new technique is proposed by Lins and discriminating longitudinal, transverse, mesh cracks,
Givigi [12], who attempted to design a real-time and healthy pavements. They use image processing
crack detection system based on the application of methodologies to create a binary image and further
particle filtering, which should be implemented in the validate artificial neural networks (ANN), SVMs, and
Red, Green, Blue (RGB) color space. By iterating, RF algorithms as a mode of classification. Amid their
the algorithm is adjusted to trace crack pixels, attempts, the authors discovered that the SVMs
allowing reconstruction of cracks and their length provided the most desirable outputs. However, this
estimation. However, the use of this method might approach might have limitations in (in) contexts of
prove to be problematic in differentiating cracks from complex crack patterns or differing variability of
other features in a highly complex urban area and crack patterns. Ibrahim et al. [20] make a similar
thus may result in false detections. comparison in terms of the efficiency of the k-nearest
neighbors algorithm and its fuzzy counterpart for the
Besides image processing techniques, several authors longitudinal and transverse cracks classification.
have applied the use of supervised machine learning Even though only focusing on these types of cracks,
methods in their models. For example, two their approach might encounter difficulties in
probability maps for segmenting cracks are used by generalization to other types of cracks or in situations
Ai et al. [13], using intensity levels and neighborhood of great difference in appearance of strips. In
information, backed with support vector machines addition, reducing the image attributes to delta-x and
(SVM). However, in regions characterized by delta-y parameters might compromise critical
complex textures, this technique may have a difficult information with potential to reflect on the
time differentiating cracks under differences in classification accuracy.
lighting conditions. Likewise, extreme learning
machines were also used by Wang et al. [14], used The system introduced by Ahmadi et al. [21] used
for image segmentation, which can be initialized image processing methods from thresholds to multi-
without any expert knowledge, and improving stage, intensity matrices, and the LAB color model in
processing time. However, the results could be attempting to denoise and beget a binary image.
influenced by datasets that are not balanced non- Nevertheless, the strategy might be conditioned by
linear cracks patterns. Cha et al. [15] propose a CNN constraints in correctly classifying intricate patterns
for analyzing cracks putting out regions of interest of cracks or in certain cases where vast variations are
from cement pavement images. Though novel, such a witnessed in the pattern of cracks. Vasculature is
method may have challenges generalizing across easily identified during ascent while on descent the
1362
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

reverse process takes place. Diverse algorithms from Deep learning has become prominent, with CNNs
classification, namely SVMs, decision trees and k- automatically detecting crack features [26], but they
nearest neighbors, were used, while the k-nearest require large, labeled datasets. generative adversarial
neighbors performed at the highest level. network (GANs) combined with transfer learning
Additionally, the application of deep learning were proposed to generate synthetic images and
methods in serving pavement defect classification is improve detection in limited-data scenarios [27],
limited by the need for massive data. Kim et al. [22] though synthetic data risks inaccurate
decided to develop a mechanism to distinguish good representations. Multiscale feature fusion [28] and
pavement from bad pavement with the two phases of edge computing with IoT devices [29] offer real-time
segmentation and classification; they compared the detection and processing, while integrating visual and
SURF method and convolved neural networks. Albeit thermal imaging [30] enhances detection in tough
this technique could face difficulties in identifying conditions. Aerial images with semantic
cracks within the midst of complicated urban features segmentation also show promise for large-scale crack
or noise. Although RF provided the best monitoring [31].
performance, this strategy may struggle with
incorporation of data from various sensors or in cases The discussed methods prove extensive diversity of
with occlusion, or disparate surface reflectance methods used in the road crack detection, ranging
properties. Besides, it was used by Li et al. [23] for from simple image processing methods to complex
crack detection in the form of 3D PaveVision 3D machine learning models. Whenever there is the use
images using convolved neural networks, but this of potential crack exponents and minimum path
method may have difficulties training the network selection to detect a crack, they are extremely useful
effectively because of computational complexity and approaches to crack detections but find themselves
in situations lacking labeled data. helpless in complex pattern when such a crack is to
be identified, in areas of high noise- a factor that
In the segment/classification of segmented or binary renders processing to level of accuracy and efficiency
images into different types of cracks, very few impossible. On the other hand, machine learning
authors in literature come to propose the use of models such as SVMs, extreme learning machines,
several classification models, while those who do so and CNNs have proven to deliver better results on
offer interesting methodologies. Li et al. [24] used a crack segmentation and classification, but they are
system for classifying defects into mesh-type and faced with problems, such as a need for a large
linear cracks while the later ones were further divided dataset, computationally, and problem of failing to
into longitudinal and transverse cracks. However, generalize across different types of cracks. GANs and
they fail to consider the fact that there is a lot of multi-scale feature fusion; belong to methods that
crack-free surfaces on roads. Moreover, the promise limited and real-time data, but are limited by
interpretability of their model of neural network for the need for accurate data and complex training.
discrimination between transverse and longitudinal Despite advances in artificial intelligence, issues still
cracks is constrained. Such an approach could be remain with crack detection in other environments,
complicated with accurate classification of such discrepancies between various lighting, and non-
complex crack patterns, or where there is a lot of correlation between assessed crack results across
variability of crack morphology. Cubero-fernandez et roads. Other improvements can as well be made by
al. [25] use also two models of for classification establishing the combination of various classification
which make use of image processing algorithms for models and optimizing the data fusion methods with
illumination and projective integrals extraction. The the aim of increasing the accuracy of the detection
application of two individual classifiers is process and decreasing the number of calculations.
characterized by relatively weak fault tolerance
(accuracy values of 72% and 67%, respectively for [Link] methodology
transverse and longitudinal cracks). This method Road crack defect detection focuses on identifying
could possibly have difficulty in correctly identifying cracks in road surfaces for safety and maintenance.
cracks in a region with complicated surfaces or This section outlines a feature-based methodology
changes in lights. In addition, the use of decision using the SURF algorithm, known for its scale,
rules for mesh type crack classification may lead to rotation, and translation invariance. The process
inaccuracies in cases of overlapping features and on begins with acquiring road surface images and
fusing cases. converting them to grayscale. These images are
resized into blocks (3×3, 4×4, 8×8, and 16×16)
1363
Pallav Kumar and Shivangi Mishra

before applying the SURF algorithm to extract hybrid PSO-WOA algorithm refines the feature set
keypoints from textured regions. This technique for better classification accuracy. These selected
improves feature capture in cracked areas while features train a RF classifier to distinguish between
reducing issues in smooth regions. Extracted normal and cracked road surfaces. The proposed
keypoints for each block size are organized in a system, detailed in Figure 1, involves preprocessing,
dictionary and then undergo dimensionality reduction SURF-based feature extraction, dimensionality
using DCT and MSLBP. This results in a resource reduction, feature selection, and classification,
vector that retains essential texture information. The yielding a binary prediction of road cracks.

Input Road Crack Image from

Dataset

Preprocessing

RGB2YCBCR Conversion

Feature Extraction
Discrete Cosine Transform (DCT)
Speeded-Up Robust Convolutional Neural
+ Multi-Scale Local Binary Pattern Features (SURF) Network (CNN)
(MSLBP)

Feature Selection using

Hybrid PSO-WOA Algorithm

Classification using Random Forest Classifier

10-Fold Cross Validation

Crack forest
Crack 500 Crack tree Deep crack Gaps
dataset

Majority Voting

Crack Crack Not

Detected Detected

Figure 1 Proposed method for road crack defect detection

The reason for choosing SURF, DCT-MSLBP, CNN oriented FAST and rotated BRIEF (ORB) because it
and hybrid PSO-WOA algorithms for this study is offers superior computational efficiency and the
their respective ability to manage difficulties ability to detect keypoints across scales, rotations,
involved in road crack detection. SURF was chosen and translations—crucial parameters given the
over scale-invariant feature transform (SIFT) and variability of road crack images. Although SIFT
1364
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

offers highly effective feature matching, it is 3.2RGB2YCBCR conversion

extremely resource-intensive, whereas ORB, while RGB2 Luminance, Chrominance Blue, Chrominance
faster has comparatively lower accuracy than the one Red (YCbCr) conversion [32] transforms images
demonstrated by SURF for the given application. The from the RGB color space to the YCbCr color space,
fusion of DCT and MSLBP augments the texture separating luminance (Y) from chrominance (Cb and
features and frequency domain properties of images Cr). This separation allows for better compression, as
of crack in the roads, helping in the identification of the human eye is more sensitive to brightness than
the cracks from the background under differing color.
lighting conditions. The conversion steps are:
1. Normalize RGB values to [0, 1].
CNN-based feature extraction boosts performance by 2. Apply formulas from the following Equation 1,
enabling automatic learning of deep hierarchical Equation 2, and Equation 3:
concepts required for dealing with complicated high- 𝑌 = 0.299𝑅 + 0.587𝐺 + 0.114𝐵 (1)
order nonlinearities found in crack images; this is a 𝐶𝑏z − 0.1687𝑅 − 0.3313𝐺 + 0.5𝐵 + 128
significant benefit when dealing with large/diverse (2)
datasets. The hybrid PSO-WOA was chosen due to 𝐶𝑟 = 0.5𝑅 − 0.4187𝐺 − 0.0813𝐵 + 128
the efficiency of utilizing the same as an optimization (3)
technique for the feature selection problem. PSO-
WOA integrates the search capability at a local level Where, 𝑅, 𝐺, and 𝐵 are the red, green, and blue color
along with WOA’s ability to explore the global channels, and 𝑌, 𝐶𝑏, and 𝐶𝑟 are the resulting YCbCr
search space and thereby eliminates redundancy and channels.
ensures that the most discriminative features are 3. Quantize YCbCr values using a quantization
selected for the classifier which is highly essential to matrix.
improve the accuracy of crack detection. This hybrid
approach is more effective at balancing exploration 3.3Feature extraction
and exploitation, especially as regards high- Feature extraction is a vital process in image
dimensional feature spaces wherein simple processing that converts raw image data into useful
algorithms might be challenged as far as convergence features for classification or analysis. This study
or local minima trap is concerned. presents a hybrid method combining SURF, DCT-
MSLBP, and CNN techniques to identify digital
3.1Image preprocessing image falsification. A summary of the methods used
Color filters arranged in small mosaics are integrated is provided below.
into the capture sensors of each pixel to generate a 3.3.1 SURF with K-means
digital image. These filters aid in the color definition Figure 2 shows the description of the proposed
process. SURF with k-means algorithm:

Feature extraction for individual

Query image
color bands

Keypoints pooling algorithm

Extract those keypoints that

match the image

Clustered features of K-means clustering of

training keypoints matching

Figure 2 Block diagram for proposed feature extraction approach

1365
Pallav Kumar and Shivangi Mishra

SURF algorithm: The k-means algorithm efficiently The k-means algorithm assigns feature vectors to the
clusters SURF features extracted from images into a nearest centroid and updates centroids until
predefined number of representative groups. The assignments stabilize or a maximum number of
mathematical formulation of the k-means algorithm iterations is reached. After convergence, feature
is given as follows: vectors form a histogram using the bag-of-visual-
• Initialization: Randomly choose K cluster centers words approach for training the RF classifier.
from the SURF feature space. Variations like mini-batch and hierarchical k-means
• Assignment: Assign each SURF feature to the can impact performance, so experimenting with
nearest cluster center. hyperparameters and initializations is recommended
• Update: Calculate the cluster centers again, based for optimal results.
on the mean of all the SURF features assigned to 3.3.2 DCT features extraction of each MSLBP block
that particular cluster. To extract features from the DCT, the first three steps
• Iterate steps 2 and 3 until convergence (when the are implemented. The DCT is applied to each local
cluster centers no longer change or after reaching binary pattern (LBP) block, calculating the standard
the maximum number of iterations). deviation of the coefficients. For each block, a vector
After clustering SURF features into K representative is created with two features from the chrominance
clusters, a histogram of cluster assignments for each matrices. These vectors are concatenated as the
image is created as a feature vector for training the image is processed block by block. The total features
RF classifier. This histogram counts the occurrences equal 2 multiplied by the number of blocks. Figure 3
of SURF features per cluster and can be normalized illustrates the algorithm.
to ensure independence from the number of features. • The suspicious RGB color image with dimensions
These feature vectors are then used as inputs for the 𝑀 × 𝑁 × 3 is originally defined, as shown in
RF classifier, trained using standard techniques: Equation 6.
𝐼 = 0.299 ∗ 𝑟 + 0.587 ∗ 𝑔 + 0.114 ∗ 𝑏 (6)
Given a set of 𝑁 feature vectors {𝑥1 , 𝑥2 , . . . , 𝑥𝑁 }
where each feature vector 𝑥𝑖 is a 𝑑-dimensional Where the 𝐼 component is obtained by a weighted
vector, 𝑘-means clustering aims to partition the sum of the red, green, and blue color channels using
feature vectors into 𝐾 clusters {{𝐶1 , 𝐶2 , . . . , 𝐶𝐾 } by coefficients 0.299, 0.587, and 0.114 respectively.
minimizing the sum of squared distances between • To reduce the dimensionality of the resulting 𝐼
each feature vector 𝑥𝑖 and its assigned cluster component, it is partitioned into overlapping
centroid 𝑐𝑖 , as represented by Equation 4: square blocks of fixed size 𝑏 × 𝑏, where 𝑏 can take
values 3, 4, 8, 𝑜𝑟 16. The blocks are denoted as
min 𝐽(𝐶1 , 𝐶2 , … , 𝐶𝐾 , 𝑐1 , 𝑐2 , … , 𝑐𝐾 ) = ∑‖𝑥𝑖 − 𝑐𝑖 ‖2 𝐵𝑖𝑗 , where 𝑖 and 𝑗 represent the starting point of
𝑖
the row and column of the block, and the number
(4)
of generated blocks is (𝑀 − 𝑏 + 1) × (𝑁 − 𝑏 +
𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑐𝑖 𝑖𝑛 𝐶𝑗 , 𝑓𝑜𝑟 𝑖, 𝑗 = 1,2, … , 𝐾
1), as shown in Equation 7.
𝐵(𝑥, 𝑦) = 𝐼(𝑥 + 𝑖, 𝑦 + 𝑖) (7)
Where, ‖𝑥𝑖 − 𝑐𝑖 ‖2 is the Euclidean distance
between the feature vector 𝑥𝑖 and the centroid 𝑐𝑖 of
𝑥, 𝑦 𝜀{0, … , 𝑏 − 1}, 𝑖 𝜀 {0, … , 𝑀
its assigned cluster, and 𝐶𝑗 is the set of feature vectors
− 𝑏}, 𝑎𝑛𝑑 𝑗 𝜀 {0, … , 𝑁 − 𝑏}
assigned to cluster 𝑗. The centroids 𝑐𝑖 are computed
as the mean of the feature vectors in each cluster, as • DCT is applied to each 𝑏 × 𝑏 block, following the
given by Equation 5: first 3 steps of the previous section, and the
1
𝑐𝑖 = ∑𝑥𝑖 𝐶𝑗 (𝑥𝑖 ) (5) standard deviation of the resulting DCT values is
|𝐶𝑗 |
computed, as described in Equation 8.
For clustering SURF features , k-means algorithm is 𝐶 = 𝐷𝐶𝑇(𝐵(𝑥, 𝑦)) (8)
setup with the hyperparameters below:
• Number of Clusters: 5 clusters 𝑥, 𝑦 𝜀{0, … , 𝑏 − 1}, 𝑖 𝜀 {0, … , 𝑀
• Initialization (init): K-means++ − 𝑏}, 𝑎𝑛𝑑 𝑗 𝜀 {0, … , 𝑁 − 𝑏}
• Max Iterations: 300 iterations
• Convergence Tolerance: 0.0001 • Next, a feature vector is extracted from each block
• Random State: 42 by taking the first 𝑁𝑐 coefficients of the DCT
values, where 𝑁𝑐 is a set of integers from 1 to 𝑛,

1366
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

with 𝑛 being the length of the DCT vector, as convolution at layer 𝑙 is represented as shown in
shown in Equation 9. Equation 10:
𝑍 = 𝐶(1: 𝑁𝑐) (9) 𝑙
𝐹𝑖,𝑗,𝑘 = ∑𝐻−1 𝑊−1 𝑙−1 𝑙 𝑙
𝑚=0 ∑𝑛=0 𝐼𝑖+𝑚,𝑗+𝑛 ⋅ 𝑊𝑚,𝑛,𝑘 + 𝑏𝑘 (10)

• These feature vectors are structured into a matrix Where 𝑊 𝑙 are the filter weights and 𝑏 𝑙 is the bias.
𝑀 with dimensions (𝑀 − 𝑏 + 1) × (𝑁 − 𝑏 + 1). and 𝐻 and 𝑊 represent the height and width of the
Matrix 𝐴 is then obtained by ordering the rows of filter. This operation captures local patterns such as
𝑀 lexicographically. edges, textures, and more complex features at deeper
• The matrix A is subjected to a matching method layers.
that involves comparing pairs of neighboring Layer Configuration: The architecture has several
vectors using their Euclidean distance. When the convolutional layers. For each layer, 32 filters of size
distance between two adjacent vectors is below a 3×3 with ReLU activation function are used. The
pre-determined threshold value, 𝑁𝑑 , the block following layers exponentially increase the number
being analyzed is regarded as a false candidate of filters with 64 in the second and 128 in the third
pair. The equivalent locations (𝑥𝑖 , 𝑦𝑖 ) and (𝑥𝑗 , 𝑦𝑗 ) convolutional layer. The kernel size is uniform, 3×3.
and their displacement vectors [|𝑥𝑖 − 𝑥𝑗 |, |𝑦𝑖 −
𝑦𝑗 |] are recorded. The most common offset value Activation and Pooling: After convolution, an
activation function like ReLU is applied, as shown in
is calculated from the displacement vectors, and
Equation 11:
each pixel in the image is compared to this value.
𝑓(𝑥) = max(0, 𝑥) (11)
If the pixel is found to be within the threshold
distance of the common offset value, it is
This brings about non-linearity, and therefore, the
exchanged with the equivalent black pixel.
network will be in a position to pick more complex
patterns. Max-pooling is followed by reducing the
spatial size of feature maps using 2×2 pooling, as
shown in Equation 12:
𝑙 𝑙
𝑃𝑖,𝑗,𝑘 = max 𝐹𝑖+𝑚,𝑗+𝑛,𝑘 (12)
𝑚,𝑛

That is why this step helps to retain the most

meaningful features and decrease the level of
computational complexity, avoiding overfitting.
Fully Connected Layer: The outputs of the
convolutional and pooling layers are represented as a
one-dimensional vector after flattening the output
feature maps. The last fully connected layer yields a
vector 𝑣, that captures the deep features extracted
from the image, as shown in Equation 13:
𝑣 = flatten(𝑃𝐿 ) (13)

This feature vector is used for additional

classification activities.
Figure 3 Conceptual description DCT-MSLBP The CNN adopted in this study has three
convolutional layers and subsequent fully connected
3.3.3 CNN based deep feature extraction layer. The number of filters increases as the depth of
CNNs are highly effective for extracting deep the network also does, in the first layer 32 filters, the
features from images. Given an input image: second and 64, and the third - 128 filters. The kernel
𝐼 ∈ ℝ𝑀×𝑁×3 , where 𝑀 × 𝑁 are the image dimensions size is retained at 3×3, and the last layer features
and 3 represents the RGB channels, CNNs vector is finally passed into the classifier for crack
automatically learn hierarchical features by applying detection.
convolutional operations.
Convolutional Layers: The core operation in CNNs The hyperparameters for CNN training are as
is convolution, which generates feature maps. The follows:

1367
Pallav Kumar and Shivangi Mishra

• Number of Filters: 32 (first layer), 64 (second 4. WOA Phase:

layer), 128 (third layer) a. Calculate the best position found so far, as
• Kernel Size: 3×3 shown in Equation 15:
• Activation Function: ReLU 𝑥𝑏𝑒𝑠𝑡 = arg max (𝑓(𝑥𝑖 (𝑡))) (15)
• Max Pooling Size: 2×2 b. For each whale 𝑖:
• Learning Rate: 0.001 (for Adam optimizer) i. Calculate the distance 𝐷 between 𝑥𝑖 (𝑡) and
• Batch Size: 32 𝑥𝑏𝑒𝑠𝑡 .
• Epochs: 50 ii. Generate a random number 𝑟 between 0
and 1.
3.3.4 Combined feature set iii. Update the whale's position, , as shown in
The extracted features from SURF, DCT-MSLBP, Equation 16:
and CNN are concatenated into a single unified 𝑥𝑖 (𝑡 + 1) = 𝑥𝑏𝑒𝑠𝑡 − 𝐴 ∙ 𝐷 ∙ 𝑒𝑥𝑝(𝐵 ∙ 𝑟) ∙ cos(2𝜋𝑟)
feature vector, 𝐹. Each feature type captures unique (16)
characteristics of road crack images: local and global Where A, B are constants controlling the movement.
texture properties (SURF), frequency-domain and 5. PSO Phase:
texture descriptors (DCT-MSLBP), and deep learning a. Calculate the best local position found by
features (CNN). whale 𝑖, as shown in Equation 17:
Let the individual feature sets be:
𝑝𝑖 = arg max (𝑓(𝑥𝑖 (𝑡))) (17)
• 𝐹𝑆𝑈𝑅𝐹 : SURF features
• 𝐹𝐷𝐶𝑇−𝑀𝑆𝐿𝐵 : Combined DCT and MSLBP features b. Calculate the best global position found by
• 𝐹𝐶𝑁𝑁 : CNN-based deep features any whale, as shown in Equation 18:
𝑝𝑔 = arg max (𝑓 (𝑥𝑗 (𝑡))) 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑗 (18)
The combined feature vector 𝐹 is represented as
c. For each whale 𝑖:
shown in Equation 14:
i. Calculate a random number 𝑟1 and 𝑟2
𝐹 = [𝐹𝑆𝑈𝑅𝐹 , 𝐹𝐷𝐶𝑇−𝑀𝑆𝐿𝐵 , 𝐹𝐶𝑁𝑁 ] (14)
between 0 and 1
ii. Update the velocity of the whale, as shown
This vector 𝐹 is used for feature selection.
in Equation 19:
3.4Features selection using hybrid PSO-WOA 𝑣𝑖 (𝑡 + 1) = 𝑤 ∙ 𝑣𝑖 (𝑡) + 𝑐1 ∙ 𝑟1 ∙ (𝑝𝑖 − 𝑥𝑖 (𝑡)) + 𝑐2 ∙
algorithm 𝑟2 ∙ (𝑝𝑔 − 𝑥𝑖 (𝑡)) (19)
The hybrid PSO-WOA algorithm is used to select the Where 𝑤 is the inertia weight, and 𝑐1 , 𝑐2 are
most relevant features from the combined set of acceleration constants.
features (𝐹) extracted using SURF, DCT-MSLBP, iii. Update the whale's position, as shown in
and CNN-based deep features. Equation 20:
1. Initialization 𝑥𝑖 (𝑡 + 1) = 𝑥𝑖 (𝑡) + 𝑣𝑖 (𝑡 + 1) (20)
a. Initialize a population of humpback whales
and particles with random positions and 6. Feature Update:
velocities. Each whale represents a potential a. Convert the updated whale positions 𝑥𝑖 (𝑡 +
feature subset derived from the combined 1) into binary feature subsets, selecting
SURF, DCT-MSLBP, and CNN features. relevant features from combined feature set
b. Whale positions: 𝑥𝑖 (0) for each whale 𝑖 in the 𝐹.
population 𝑁. 7. Objective function evaluation:
c. Whale velocities: 𝑣𝑖 (0) for each whale 𝑖 in a. Evaluate the fitness of the updated feature
the population 𝑁. subsets 𝑥𝑖 (𝑡 + 1) using the RF classifier. The
2. Feature Encoding: Convert each whale's position objective function 𝑓(𝑥𝑖 ) is defined as shown
𝑥𝑖 (𝑡) into a binary feature subset, where each bit in Equation 21:
corresponds to a feature selected from the full set
𝑓(𝑥𝑖 ) = −𝐿 (𝑅𝑎𝑛𝑑𝑜𝑚 𝐹𝑜𝑟𝑒𝑠𝑡(𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑦(𝑥𝑖 ))) (21)
of features (SURF, DCT-MSLBP, and CNN).
3. Objective Function Evaluation: Evaluate the 8. Termination criteria:
fitness of each whale's feature subset 𝑥𝑖 (𝑡) using a. The algorithm continues to iterate until a
the objective function 𝑓(𝑥𝑖 (𝑡)),which measures predefined stopping criterion is met, such as
the classification performance (accuracy, F1- the maximum number of iterations or a
Score) using the RF classifier. convergence threshold.

1368
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

The following hyperparameters are used for PSO- utilizing distinct subsets of the training dataset.
WOA: This process is known as bootstrap aggregating or
• Population Size: 50 individuals bagging. Each tree is exposed to different
• Number of Iterations: 100 variations of the data, reducing the risk of
• Inertia Weight (𝑤): 0.5 overfitting to a specific subset.
• Acceleration Coefficients (𝑐1 , 𝑐2 ): 1.5 for both Given a training dataset 𝐷 with 𝑁 samples:
• Search Space Bounds: [0, 1] a. For each decision tree 𝑇𝑖 :
b. Randomly select 𝑁 samples with replacement
Figure 4 illustrates the working flow for integrating from 𝐷. This forms a new dataset 𝐷𝑖 for
PSO and WOA in feature selection. It identifies the training the 𝑖 𝑡ℎ tree, as shown in Equation 22.
initialization of populations, subsequent evaluation 𝐷𝑖 = {(𝑥1 , 𝑦1 ), (𝑥2 , 𝑦2 ), … , (𝑥𝑁 , 𝑦𝑁 )} (22)
and updating of positions, and the combination of Where 𝑥𝑁 represents the features of the 𝑁 𝑡ℎ
PSO and WOA to optimize and choose the most sample and 𝑦𝑁 is its corresponding label.
relevant features to the classification. 2. Random Feature Selection: During the
construction of each decision tree, the algorithm
Initialize PSO and WOA Populations randomly selects a subset of features. This
introduces diversity among the trees and prevents a
single dominant feature from overly influencing
Generate Initial Positions of Particles
the decision. The number of features selected at
each split is a hyperparameter.
Evaluate Fitness of Each Particle Given a dataset 𝐷𝑖 for training tree 𝑇𝑖 :
a. Randomly select a subset of features with size
𝑚 (where 𝑚 is a hyperparameter) from the
total 𝑀 features in 𝐷.
Yes b. Construct the decision tree 𝑇𝑖 using 𝐷𝑖 and the
Maximum Iterations Reached? selected features, as shown in Equation 23.
𝐹𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 = {𝐹1 , 𝐹2 , … , 𝐹𝑚 } (23)
Where 𝐹𝑚 represents the 𝑚𝑡ℎ selected feature.
No 3. Constructing Decision Trees: The creation of
Update PSO Velocity and Position every decision tree within the ensemble entails a
recursive process of data partitioning, contingent
on the chosen attributes. At each node, the
Switch to WOA algorithm opts for the attribute that most
effectively segregates the data through a
designated metric (such as Gini impurity or
Update WOA Position information gain).
Let's consider the construction of a single decision
tree 𝑇𝑖 :
Select Features a. At a node 𝑁, choose the feature 𝐹 that best
splits the data based on a splitting criterion
Figure 4 Flowchart illustrating the hybrid integration
(e.g., Gini impurity or information gain).
of PSO and WOA for feature selection
b. Split the data into child nodes 𝑁𝑙𝑒𝑓𝑡 and
3.5Random forest (RF) classifier 𝑁𝑟𝑖𝑔ℎ𝑡 based on the chosen feature 𝐹 and its
The RF classifier is a powerful ensemble learning threshold value.
method used for classification tasks, such as road c. Recursively apply the splitting process to the
crack detection. It combines multiple decision trees, child nodes until a stopping criterion is met
each trained on a unique dataset subset, to improve (e.g., maximum depth or minimum samples
prediction accuracy. RF effectively handles high- per image).
dimensional data and reduces overfitting. Splitting criterion (e.g., gini impurity) is represented
Algorithm Overview and Mathematical by Equation 24:
Formulations: 𝐺𝑖𝑛𝑖(𝐷) = 1 − ∑𝐾 𝑘=1(𝑝𝑘 )
2
(24)
1. Bootstrap Aggregating (Bagging): Within a RF Where 𝐾 is the number of classes, 𝑝𝑘 is the
framework, diverse decision trees undergo training proportion of samples in class 𝑘 in node 𝑁.

1369
Pallav Kumar and Shivangi Mishra

Gini impurity for split is given by Equation 25: In automated road crack detection, the RF classifier
|𝑁𝑙𝑒𝑓𝑡 | uses features from the hybrid PSO-WOA selection
𝐺𝑖𝑛𝑖𝑠𝑝𝑙𝑖𝑡 (𝑁, 𝐹, 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑) = ∙ 𝐺𝑖𝑛𝑖(𝑁𝑙𝑒𝑓𝑡 ) +
|𝑁| process, effectively representing relevant information
|𝑁𝑟𝑖𝑔ℎ𝑡 |
∙ 𝐺𝑖𝑛𝑖(𝑁𝑟𝑖𝑔ℎ𝑡 ) (25) from road crack images. Its ensemble nature and
|𝑁|
feature randomness help it handle complex patterns,
Where 𝑁𝑙𝑒𝑓𝑡 and 𝑁𝑟𝑖𝑔ℎ𝑡 are the child nodes after avoid overfitting, and accurately predict road cracks.
splitting based on feature 𝐹 and threshold. By aggregating multiple decision tree predictions, it
4. Voting and Classification: Once all trees are built, ensures reliable detection.
each tree independently classifies an input sample.
The final classification is determined by a majority 4. Results
vote among the trees. The class with the most
4.1Dataset
votes becomes the predicted class for the input
The Crack Segmentation Dataset [33] contains about
sample.
11,200 images from 12 crack segmentation, with
For an input sample 𝑋𝑖𝑛𝑝𝑢𝑡 :
image prefixes indicating their origins. Images
a. Each decision tree 𝑇𝑖 classifies 𝑋𝑖𝑛𝑝𝑢𝑡 based without cracks can be identified using the "noncrack"
on its trained model. pattern. All images are resized to 448×448 and
b. The class predictions from all trees are organized into "images" and "masks" folders, with
collected. separate "train" and "test" folders for training and
c. The final class prediction is the class with the testing images. The splitting process ensures similar
highest frequency among the predictions is proportions of each dataset in both folders. Table 1
given by the Equation 26. shows the sample images taken from the dataset. To
𝑁𝑡𝑟𝑒𝑒𝑠
𝑦𝑓𝑖𝑛𝑎𝑙 = arg max ∑𝑖=1 𝐼(𝑦𝑝𝑟𝑒𝑑𝑖 = 𝑐) (26) manage the multi-source data, all datasets used in this
c
research were merged during training. For data
Where 𝑦𝑝𝑟𝑒𝑑𝑖 is the predicted class by the 𝑖 𝑡ℎ tree, sources, the dataset contains images from Crack
Forest dataset, Crack 500, Crack Tree, Deep Crack
𝑁𝑡𝑟𝑒𝑒𝑠 is the total number of trees, and 𝐼(⋅) is the
and GAPs. To training, a complete training set was
indicator function.
jointly created by combining these datasets for
additional variety in the road crack images. During
The RF classifier employs the following
the cross-validation, the 10-fold cross-validation was
hyperparameters:
used to measure the effectiveness of the model. All
• Number of Trees (n_estimators): 100 trees
datasets were represented in each fold, so the model
• Maximum Depth (max_depth): None (trees are received exposure to a diversified set of road crack
grown until all leaves are pure) images during both the training and test phases. This
• Minimum Samples for Splitting strategy is instrumental in evaluating the model’s
(min_samples_split):2 generality level about road crack conditions of
• Minimum Samples for Leaf (min_samples_leaf):1 various types. The cross-validation approach of
• Criterion: Gini impurity performance was reported, with combined dataset as
• Bootstrap Sampling (bootstrap): True evaluated rather than an individual source.

Table 1 Sample images from dataset [33]

Crack forest dataset

Crack 500

1370
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

Crack tree

Deep crack

GAPs

4.2Performance evaluation parameter The assessment included three phases:

Following are the evaluation parameters, as shown in 1. Analyzing feature frameworks across datasets.
Equation 27 to Equation 35: 2. Testing model generalization by training on one
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (27) dataset and testing on another.
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
𝑇𝑃 3. Comparing the proposed framework with existing
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (28) techniques using a benchmark dataset.
𝑇𝑃+𝐹𝑃
𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = (29)
𝑇𝑃+𝐹𝑁
𝑇𝑁 Figure 5 shows the hybrid feature distribution,
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = (30) indicating better performance with reduced skewness
𝑇𝑁+𝐹𝑁
𝐹𝑃+𝐹𝑁
𝐸𝑟𝑟𝑜𝑟 𝑅𝑎𝑡𝑒 = (31) compared to other features. Skewness indicates data
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
𝐹𝑃 distribution. The hybrid feature extraction creates a
𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑅𝑎𝑡𝑒 (𝐹𝑃𝑅) = more balanced distribution, making it stable for
𝐹𝑃 + 𝑇𝑁
(32) classifier training. No single feature can distinguish
2𝑇𝑃 between the four classes, but multivariate analysis
𝐹 − 𝑆𝑐𝑜𝑟𝑒 = (33)
2𝑇𝑃+𝐹𝑃+𝐹𝑁 with machine learning effectively identifies road
𝑀𝑎𝑡𝑡ℎ𝑒𝑤𝑠 𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 𝐶𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡 (𝑀𝐶𝐶) = cracks. Table 2 compares feature extraction methods
(𝑇𝑃×𝑇𝑁)−(𝐹𝑃×𝐹𝑁)
(34) with the RF classifier for road crack detection.
√(𝑇𝑃+𝐹𝑁)(𝑇𝑃+𝐹𝑃)(𝑇𝑁+𝐹𝑁)(𝑇𝑁+𝐹𝑃)
Methods evaluated include SURF, MSLBP+DCT,
𝐾𝑎𝑝𝑝𝑎 𝑆𝑡𝑎𝑡𝑖𝑠𝑡𝑖𝑐𝑠 =
𝑂𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦−𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 CNN features, and a hybrid approach. The hybrid
(35) method achieved the highest accuracy (99.07%), with
1−𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦
excellent error rate (0.0093), sensitivity (99.17%),
Where TP – True positive and specificity (99.68%). CNN features follow with
TN – True negative accuracy (98.53%) and low error rate (0.0147).
FP – False positive MSLBP+DCT performs well at 97.22%, while SURF
FN – False negative achieves 97.50% but has a higher false positive rate
(0.0083).
4.3Experimental results and analysis
This section evaluates the methodology's Metrics like Matthews correlation coefficient (MCC)
effectiveness across various datasets, focusing on and Kappa statistics also favor the hybrid approach,
classifying images as normal or cracked. The reflecting strong classification agreement. Overall,
experiment was performed on the Dell Precision the hybrid feature extraction method proves most
5820 workstation with Intel Xeon W-2225 CPU; 32 effective for road crack detection. Figure 6 shows the
GB RAM; NVIDIA Quadro RTX 4000 GPU. The comparative analysis of accuracy with various
algorithms were applied using MATLAB R2020b; features on several dataset using RF classifier. Table
the Image Processing Toolbox, Parallel Computing 3 presents the performance of the crack classification
Toolbox, and the Statistics and Machine Learning framework across different training datasets ("Crack
Toolbox. A 10-fold cross-validation was used for Forest," "Crack 500," "Crack Tree," "Deep Crack,"
reliability. and "GAPs") and performance measures

1371
Pallav Kumar and Shivangi Mishra

("Accuracy," "Sensitivity," "Specificity," "F1-Score," accuracy (0.95) and specificity (0.9833) but a lower
and "MCC"). The "Crack Forest" dataset is not used MCC (0.8212), indicating weaker correlation. The
for testing. For "Crack 500," accuracy ranges from "Deep Crack" dataset demonstrates high specificity
0.9167 to 0.9389, with sensitivity from 0.9087 to (0.978) and F1-Score (0.9236), highlighting balanced
0.9286, specificity from 0.9744 to 0.9806, F1-Score precision and recall. Overall, the table reveals the
from 0.8603 to 0.9266, and MCC from 0.8562 to framework's strengths and weaknesses in detecting
0.8906. The "Crack Tree" dataset shows high road cracks.

Figure 5 Comparative analysis of various methods in a boxplot

Table 2 Comparative analysis for average outcome of all dataset with different feature extraction using RF classifier
Parameters SURF MSLBP+DCT CNN Features Hybrid Features
Accuracy 0.975 0.9722 0.9853 0.9907
Error Rate 0.025 0.0278 0.0147 0.0093
Sensitivity 0.975 0.9722 0.9853 0.9917
Specificity 0.9917 0.9907 0.9951 0.9968
Precision 0.9753 0.9775 0.9861 0.9917
False Positive Rate 0.0083 0.0093 0.0049 0.0032
F1-Score 0.9749 0.9721 0.9853 0.9915
MCC 0.9668 0.9642 0.9808 0.9885
Kappa Statistics 0.9333 0.9259 0.9608 0.9753

100
99
98
97
Accuracy in %

96
95 CNN
94 MSLBP+DCT
93
SURF
92
91 MRMR Hbrid
90
89
Crack Forest Crack 500 Crack Tree Deep Crack GAP
Datasets

Figure 6 Comparative analysis of accuracy with various features on various dataset trained and tested with RF
classifier
1372
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

Table 3 Performance measures of the proposed crack classification framework with cross-training and testing
Performance measures Testing dataset
Training dataset Crack forest Crack 500 Crack tree Deep crack GAPs
dataset
Crack Forest Accuracy -- 0.9417 0.9444 0.925 0.8889
Sensitivity -- 0.9417 0.9444 0.925 0.8889
Specificity -- 0.9806 0.9815 0.975 0.963
F1-Score -- 0.942 0.9456 0.9266 0.8865
MCC -- 0.9259 0.9305 0.9058 0.8607

Crack 500 Accuracy 0.9167 -- 0.931 0.925 0.9083

Sensitivity 0.9244 -- 0.9306 0.925 0.9083
Specificity 0.9733 -- 0.9792 0.975 0.9694
F1-Score 0.9147 -- 0.9209 0.9266 0.9089
MCC 0.8906 -- 0.9057 0.9058 0.8838

Crack tree Accuracy 0.95 0.9286 -- 0.9333 0.8611

Sensitivity 0.95 0.9286 -- 0.9333 0.8611
Specificity 0.9833 0.9762 -- 0.9778 0.9537
F1-Score 0.9513 0.9295 -- 0.9335 0.8595
MCC 0.9371 0.909 -- 0.9147 0.8212

Deep Crack Accuracy 0.9354 0.907 0.9302 -- 0.907

Sensitivity 0.9354 0.8736 0.9087 -- 0.8736
Specificity 0.978 0.9664 0.9744 -- 0.9664
F1-Score 0.9327 0.8875 0.9236 -- 0.8875
MCC 0.9143 0.8672 0.9051 -- 0.8672
GAPs Accuracy 0.9417 0.9389 0.8621 0.8994
Sensitivity 0.9417 0.9087 0.8611 -- 0.8994
Specificity 0.9806 0.9744 0.9521 -- 0.9641
F1-Score 0.9421 0.9236 0.8603 -- 0.8875
MCC 0.9243 0.9127 0.829 0.8562

Performance Metrics with Variance Indicators: In deviations (calculated during ten-fold cross-
order to increase the reliability of reported validation) are given for each feature extraction
performance metrics we also present the standard method. The Hybrid Features method used the best
deviations for accuracy, F1-Score and MCC also performance in terms of accuracy, F1-Score, and
based on 10-fold cross validation. The results are as MCC, standard deviations showed the stability of the
follows: In Table 4, the means for accuracy, F1- results.
Score, and the MCC, as well as the standard

Table 4 Performance metrics with standard deviations for road crack detection using different feature extraction
methods
Metric SURF MSLBP+DCT CNN Features Hybrid Features
Accuracy 97.50% (± 0.015) 97.22% (± 0.017) 98.53% (± 0.012) 99.07% (± 0.005)
F1-Score 97.49% (± 0.015) 97.21% (± 0.016) 98.53% (± 0.011) 99.15% (± 0.004)
MCC 0.9668 (± 0.015) 0.9642 (± 0.018) 0.9808 (± 0.010) 0.9885 (± 0.005)

1373
Pallav Kumar and Shivangi Mishra

Confidence Intervals (CI) for Performance Metrics: performance metrics through a range in which the
For completeness 95% CI for the reported metrics is actual metrics values would fall. Once again the
also calculated, as shown below. These intervals Hybrid Features method is superior to other methods,
represent the lack of certainty in the performance with a narrow confidence interval, hence high
metrics because of cross-validations. These 95% CI confidence in its classifying capability.
in Table 5 give an idea about the accuracy of the

Table 5 Performance metrics with 95% CI for road crack detection using different feature extraction methods
Metric SURF MSLBP+DCT CNN features Hybrid features
Accuracy 97.50% (95% CI: 97.22% (95% CI: [97.20%, 98.53% (95% CI: 99.07% (95% CI:
[97.48%, 97.52%]) 97.24%]) [98.52%, 98.54%]) [99.05%, 99.09%])
F1-Score 97.49% (95% CI: 97.21% (95% CI: [97.19%, 98.53% (95% CI: 99.15% (95% CI:
[97.47%, 97.51%]) 97.23%]) [98.51%, 98.54%]) [99.13%, 99.17%])
MCC 0.9668 (95% CI: [0.965, 0.9642 (95% CI: [0.963, 0.9808 (95% CI: [0.979, 0.9885 (95% CI: [0.987,
0.968]) 0.965]) 0.982]) 0.990])

In Table 6, hybrid features exhibited the best results approach implying that deep learning features may
in ensuring both the false positives and false not be able to detect some of road cracks types
negatives were at a minimum, with a minimum false effectively as hybrid approach. MSLBP + DCT and
positive rate of 0.0093 and a false negative rate of SURF methods showed higher false positive and
0.0085. This also implies that the hybrid approach is false negative rates than in the case of a hybrid
considerably more reliable in distinguishing cracks method, proving their disparities with regard to fine
compared to non-cracks. CNN Features was also crack details within crack detection, and discerning
good performing but had higher rates of false from background textures.
positives and negatives compared to the hybrid

Table 6 Misclassification analysis of feature extraction methods

Feature extraction FP FN FPR FNR
method
SURF 23 13 0.025 0.0278
MSLBP + DCT 25 14 0.0278 0.0285
CNN Features 18 8 0.0147 0.0175
Hybrid Features 15 7 0.0093 0.0085

Where False Positives (FP): Number of non-cracked features, and PSO-WOA-based feature selection
images incorrectly classified as cracked. illustrate high accuracy as well as generalization. The
performance of hybrid feature extraction method had
False Negatives (FN): Number of cracked images the best accuracy, 99.07% and F1-Score of 99.15%,
incorrectly classified as non-cracked. showing major enhancement compared to past
False Positive Rate (FPR): Calculated as False methods. The high accuracy of the system owes to
Positives / (False Positives + True Negatives). the fact that the feature extraction process is complete
False Negative Rate (FNR): Calculated as False and the images of road cracks are captured at both
Negatives / (False Negatives + True Positives). local and global [Link] PSO-WOA feature
selection technique further reduces these features
5. Discussion only using the most discriminative features in
This research work proposes a strong technique for classification. This helps to reduce dimensionality
road crack detection by integrating advanced feature while at the same time improving classification
extraction techniques, feature selection techniques accuracy as we are able to see from the results from
and a classification algorithm that has been shown to our RF classifier, which exhibited good
perform better in comparison to other methods using performancewith a minimal error rate (0.0093).
evaluation metrics.
Contrary to this, conventional approaches (e.g.,
5.1Overall result analysis SURF and MSLBP+DCT) showed worse results in
The outcomes achieved by using the proposed hybrid precision and sensitivity. While CNN features
approach that combines SURF, DCT-MSLBP, CNN performed well, the hybrid approach outperformed
1374
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

them by overcoming the noise problem, varying improvement in precision-recall balance. The "CNN
image quality and complicated crack pattern issues. [35]" method has an accuracy of 88.3% with a high
The comparative analysis reveals that the hybrid F1-Score of 87.5%, suggesting a need to enhance
approach performs superior to individual feature accuracy. "DenseNet161 [36]" and "ResNet152 [36]"
extraction techniques and thus is more robust and show accuracies of 90.87% and 95.70%, respectively,
reliable classification system. with F1-Scores indicating areas for improvement.
"VGG19 [36]" achieves 97.66% accuracy and an F1-
5.2Comparative study Score of 89.14%. Notably, the proposed method
Table 7 compares various methods for detecting road excels with 99.07% accuracy and an F1-Score of
cracks on the "GAPs" dataset, focusing on 99.15%, demonstrating its superior performance in
"Accuracy" and "F1-Score." The "RCDNet [34]" road crack detection. Overall, the table highlights the
achieves 97.32% accuracy and an F1-Score of proposed method's effectiveness compared to existing
66.42%, indicating good accuracy but room for techniques.

Table 7 Comparative analysis of GAPs dataset with previous research works

Method Accuracy F1-Score
RCDNet [34] 97.32% 66.42%
CNN [35] 88.3% 87.5%
DenseNet161 [36] 90.87% 84.77%
ResNet152 [36] 95.70% 82.59%
VGG19 [36] 97.66% 89.14%
Proposed 99.07% 99.15%

As compared to previous studies, the proposed used for feature selection, increases high
method is considerably superior to existing computational complexity. This may constrain
approaches for the detection of road crack. An such real-time applications, particularly, in large
analysis was carried out by comparing the scale road inspection projects that need rapid
performance on accuracy, sensitivity, specificity, and processing of large volumes of images.
and F1-Score. For example, in the “GAPs” dataset, • The findings reported here rely heavily upon the
the accuracy of the proposed method was equal quality and span of the datasets used for training
99.07% and F1-Score was equal 99.15%, which were and testing. Even though 10-fold cross-validation
better than methods like RCD net (97.32% accuracy), technique gives results that are considered rather
and CNN (88.3% accuracy). The performance of stable, the model may face difficulties when
other techniques such as DenseNet161 and deployed on other road crack datasets.
ResNet152 which produced accuracies of 90.87% • Although the proposed system works nicely on
and 95.70% respectively, was also less than our different datasets, there are cases when the
suggested technique. The VGG19 method which detection accuracy can be affected by the diversity
returned 97.66% accuracy lagged by F1-Score by of crack type, including very fine or overlapping
89.14% behind the hybrid method. cracks. The method could also fail with cracks in
low contrast images or images disturbed by poor
In addition to that, the proposed method also proved lighting conditions.
its supremacy when trained on a variety of datasets • While the approach manifests well in controlled
while providing a high and comparable performance environment, the actual application of this
for various types of road cracks. These results creation in the real-world road inspection
support that the ability of the hybrid method to possibility is necessary to be further validated
accommodate different types of cracks and images when presented in different environmental
conditions is an important contribution towards its conditions, such as dynamic weather and traffic
greater accuracy. and differences in road surface.
• The imbalance between cracked and non-cracked
5.3Study limitations images in the data set could affect performance
Although the proposed approach offers major level of the model. Even when a combination of
progress in road crack detection, there are some features and a strong classifier is employed, the
aspects that should be recognized: model may have trouble in the situations in which
• The hybrid approach, with the highlight of the non-cracked images are much more prominent
CNN used for feature extraction and PSO-WOA than those that are cracked.
1375
Pallav Kumar and Shivangi Mishra

Finally, the proposed method provides a positive could focus on incorporating advanced deep learning
answer to the automatization of the road crack architectures for feature extraction and classification.
detection using the promising approach and superior Additionally, exploring real-time implementations
performance over existing approaches. Nevertheless, and the deployment of the proposed approach in
the problem indicated would have to be solved, practical road inspection scenarios could enhance its
especially in the real-time applications and complex real-world applicability. The integration of sensor
crack types, for further applications. Future work can data and multi-modal approaches could also
target increasing the computational efficiency of the contribute to more robust and accurate road crack
system and investigate multi-modal data inputs to detection. Continuous refinement and optimization of
increase the model’s robustness in a wide range of the proposed approach could lead to even higher
operational environments. A complete list of detection accuracy and efficiency, ultimately
abbreviations is listed in Appendix I. contributing to safer road infrastructure and accident
prevention.
6. Conclusion and future work
A robust approach for road crack detection was Acknowledgment
presented, integrating advanced feature extraction None.
techniques, an efficient feature selection method, and
a powerful classification algorithm. The hybrid Conflicts of interest
The authors have no conflicts of interest to declare.
feature extraction approach—combining MSLBP,
DCT, SURF, and CNN features—effectively captures Data availability
both local and global characteristics of road crack The dataset [33] used in this work, which incorporates
images. To enhance classification accuracy and images from the Crack Forest, Crack 500, Crack Tree,
reduce dimensionality, a hybrid PSO-WOA Deep Crack, and Gaps datasets, is publicly accessible. It
algorithm is employed for feature selection, aiming to can be obtained from the following source:
identify the most informative and relevant features [Link]
while minimizing redundancy. Finally, the RF segmentation-dataset
classifier, known for its robustness and strong
generalization capabilities, is utilized to achieve Author’s contribution statement
accurate road crack classification. Pallav Kumar: Conceptualization, Methodology, Writing
– original draft. Shivangi Mishra: Data collection, Data
analysis, Writing – review and editing.
The simulation results section provides a
comprehensive analysis of the proposed method's References
performance. Comparative evaluations of different [1] Shi Z, Jin N, Chen D, Ai D. A comparison study of
feature extraction techniques combined with the RF semantic segmentation networks for crack detection in
classifier showcase the superiority of the hybrid construction materials. Construction and Building
approach. The proposed hybrid feature extraction Materials. 2024; 414:134950.
achieves the highest accuracy (99.07%) and F1-Score [2] Inam H, Islam NU, Akram MU, Ullah F. Smart and
(99.15%), demonstrating its exceptional capability to automated infrastructure management: a deep learning
accurately detect road crack defects. The analysis approach for crack detection in bridge images.
Sustainability. 2023; 15(3):1-35.
also reveals that individual feature extraction [3] Ashraf A, Sophian A, Shafie AA, Gunawan TS, Ismail
methods, such as CNN features, exhibit specific NN. Machine learning-based pavement crack
strengths, further emphasizing the advantages of the detection, classification, and characterization: a
hybrid approach. Moreover, the performance review. Bulletin of Electrical Engineering and
evaluation using cross-training-and-testing datasets Informatics. 2023; 12(6):3601-19.
demonstrates the robustness of the proposed [4] Zuo H, Li Z, Gong J, Tian Z. Intelligent road crack
classification framework. Metrics such as accuracy, detection and analysis based on improved YOLOv8.
sensitivity, specificity, F1-Score, and MCC provide In 8th international conference on advanced
insights into the framework's performance under algorithms and control engineering (ICAACE) 2025
(pp. 1192-5). IEEE.
various scenarios. The comparative analysis with [5] Guo W, Zhong L, Zhang D, Li Q. Pavement crack
previous research methodologies validates the detection using fractal dimension and semi-supervised
proposed approach's efficacy, positioning it as a learning. Fractal and Fractional. 2024; 8(8):1-16.
leading contender in road crack defect detection. In [6] Xu C, Zhang Q, Mei L, Chang X, Ye Z, Wang J, et al.
terms of future scope, this research provides the basis Cross-attention-guided feature alignment network for
for several promising directions. Further exploration
1376
 International Journal of Advanced Technology and Engineering Exploration, Vol 12(130)

road crack detection. ISPRS International Journal of [21] Ahmadi A, Khalesi S, Bagheri M. Automatic road
Geo-Information. 2023; 12(9):1-25. crack detection and classification using image
[7] Ma N, Song Z, Hu Q, Liu CW, Han Y, Zhang Y, et al. processing techniques, machine learning and
Vehicular road crack detection with deep learning: a integrated models in urban areas: a novel image
new online benchmark for comprehensive evaluation binarization technique. Journal of Industrial and
of existing algorithms. IEEE Transactions on Systems Engineering. 2018:85-97.
Intelligent Vehicles.18082. 2025:1-16. [22] Kim H, Ahn E, Shin M, Sim SH. Crack and noncrack
[8] Zhu G, Shen SL, Yao J, Wang M, Zhuang J, Fan Z. classification from concrete surface images using
Automatic lightweight networks for real-time road machine learning. Structural Health Monitoring. 2019;
crack detection with DPSO. Advanced Engineering 18(3):725-38.
Informatics. 2025; 68:103610. [23] Li B, Wang KC, Zhang A, Yang E, Wang G.
[9] Wang Z, Qiu D, Wu R, Shi Y, Niu W. Research on Automatic classification of pavement crack using deep
road crack detection based on RGB-LPC-GPR data convolutional neural network. International Journal of
fusion. The International Archives of the Pavement Engineering. 2020; 21(4):457-63.
Photogrammetry, Remote Sensing and Spatial [24] Li L, Sun L, Ning G, Tan S. Automatic pavement
Information Sciences. 2025; 48:1543-50. crack recognition based on BP neural network.
[10] Wang T, Gopalakrishnan K, Smadi O, Somani AK. PROMET-Traffic &Transportation. 2014; 26(1):11-
Automated shape-based pavement crack detection 22.
approach. Transport. 2018; 33(3):598-608. [25] Cubero-fernandez A, Rodriguez-lozano FJ, Villatoro
[11] Amhaz R, Chambon S, Idier J, Baltazart V. Automatic R, Olivares J, Palomares JM. Efficient pavement crack
crack detection on two-dimensional pavement images: detection and classification. EURASIP Journal on
an algorithm based on minimal path selection. IEEE Image and Video Processing. 2017; 2017(1):1-11.
Transactions on Intelligent Transportation Systems. [26] Pandey V, Mishra SS. A review of image-based deep
2016; 17(10):2718-29. learning methods for crack detection. Multimedia
[12] Lins RG, Givigi SN. Automatic crack detection and Tools and Applications. 2025; 84:35469-511.
measurement based on image analysis. IEEE [27] Lu KL. Comparison of deep learning methods and a
Transactions on Instrumentation and Measurement. transfer-learning semisupervised generative-
2016; 65(3):583-90. adversarial-network combined framework for
[13] Ai D, Jiang G, Kei LS, Li C. Automatic pixel-level pavement crack image identification. Pattern
pavement crack detection using information of multi- Recognition and Image Analysis. 2025; 35(1):19-30.
scale neighborhoods. IEEE access. 2018; 6:24452-63. [28] Liu Z, Gu X, Chen J, Wang D, Chen Y, Wang L.
[14] Wang B, Li Y, Zhao W, Zhang Z, Zhang Y, Wang Z. Automatic recognition of pavement cracks from
Effective crack damage detection using multilayer combined GPR B-scan and C-scan images using
sparse feature representation and incremental extreme multiscale feature fusion deep neural networks.
learning machine. Applied Sciences. 2019; 9(3):1-21. Automation in Construction. 2023; 146:104698.
[15] Cha YJ, Choi W, Büyüköztürk O. Deep learning‐ [29] Kim Y, Yi S, Ahn H, Hong CH. Accurate crack
based crack damage detection using convolutional detection based on distributed deep learning for IoT
neural networks. Computer‐Aided Civil and environment. Sensors. 2023; 23(2):1-12.
Infrastructure Engineering. 2017; 32(5):361-78. [30] Huang S, Chen H, Yan L, Zou X, Li B, Bi Y. A
[16] Cheng J, Xiong W, Chen W, Gu Y, Li Y. Pixel-level review of the progress in machine vision-based crack
crack detection using U-Net. In TENCON 2018 (pp. detection and identification technology for asphalt
462-6). IEEE. pavements. Digital Transportation and Safety. 2025;
[17] Bang S, Park S, Kim H, Yoon YS, Kim H. A deep 4(1):65-79.
residual network with transfer learning for pixel-level [31] Zaheer Q, Qiu S, Hassan SSM, Ai C, Wang J.
road crack detection. In proceedings of the Intelligent multitasking framework for boundary-
international symposium on automation and robotics preserving semantic segmentation, width estimation,
in construction 2018 (pp. 1-4). IAARC Publications. and propagation modeling of concrete cracks. Journal
[18] Xu G, Chen F, Wu G, Li X. Active solution of of Infrastructure Systems. 2025; 31(3):04025009.
homography for pavement crack recovery with four [32] Li M. A better color space conversion based on
laser lines. Scientific Reports. 2018; 8(1):1-10. learned variances for image compression. In
[19] Hoang ND, Nguyen QL. A novel method for asphalt proceedings of the conference on computer vision and
pavement crack classification based on image pattern recognition workshops 2019 (pp. 1-5). IEEE.
processing and machine learning. Engineering with [33] [Link]
Computers. 2019; 35(2):487-98. k-segmentation-dataset. Accessed 10 August 2025.
[20] Ibrahim A, Osman MK, Yusof NA, Ahmad KA, [34] Anand S, Gupta S, Darbari V, Kohli S. Crack-pot:
Harun NH, Raof RA. Characterization of cracking in autonomous road crack and pothole detection. In
pavement distress using image processing techniques digital image computing: techniques and applications
and k-nearest neighbour. Indonesian Journal of (DICTA) 2018 (pp. 1-6). IEEE.
Electrical Engineering and Computer Science. 2019; [35] Zhang C, Nateghinia E, Miranda-moreno LF, Sun L.
14(2):810-8. Pavement distress detection using convolutional

1377
Pallav Kumar and Shivangi Mishra

neural network (CNN): a case study in Montreal, Appendix I

Canada. International Journal of Transportation S. No. Abbreviation Description
Science and Technology. 2022; 11(2):298-309. 1 ANN Artificial Neural Network
[36] Minhas MS, Zelek JS. Defect detection using deep 2 CI Confidence Intervals
learning from minimal annotations. In proceedings of 3 CNN Convolutional Neural Network
the 15th international joint conference on computer 4 DCT Discrete Cosine Transform
vision, imaging and computer graphics theory and 5 GA Genetic Algorithm
6 GAN Generative Adversarial Network
applications 2020 (pp. 506-13). Science and
7 LBP Local Binary Pattern
Technology Publications.
8 MCC Matthews Correlation Coefficient
9 MSLBP Multiscale Local Binary Pattern
10 ORB Oriented FAST and Rotated BRIEF
Dr. Pallav Kumar is currently working 11 PSO Particle Swarm Optimization
as an Assistant Professor in the 12 RF Random Forest
Department of Civil Engineering at 13 RGB Red, Green, Blue
MIT Muzaffarpur. He has more than 14 SIFT Scale-Invariant Feature Transform
seven years of teaching and research 15 SURF Speeded-Up Robust Features
experience. He obtained his [Link]. and 16 SVM Support Vector Machine
[Link]. degrees in Civil Engineering 17 UAV Unmanned Aerial Vehicle
from NIT Agartala and his Ph.D. in 18 WOA Whale Optimization Algorithm
Transportation Engineering from SVNIT Surat. He has 19 YCbCr Luminance, Chrominance Blue,
published Chrominance Red
AuthorPeighteen
to research papers in SCI, Scopus, and
other peer-reviewed journals, along with sixteen
international conference papers, five national conference
papers, and one book chapter. He has also successfully
completed two research projects. Dr. Kumar is a member of
several professional bodies in India, including the Institute
of Urban Transport, the Transportation Research Group,
the Indian Society for Technical Education, and the
Institution of Engineers (India).
Email: pallav318@[Link]

Dr. Shivangi Mishra is working as an

Assistant Town Planning Supervisor in
the Urban Development and Housing
Department, Patna, Bihar, India. She
has more than seven years of teaching
and research experience. She holds a
Ph.D. degree in Civil Engineering from
NIT Patna and an [Link]. degree in
Geoinformatics and its Applications from MANIT Bhopal.
Author’
Email: shivangicivil0038@[Link]

1378