Measurement 211 (2023) 112632

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Computer vision-based statistical crack quantification for

concrete structures
Hyunjin Bae , Yun-Kyu An *
Department of Architectural Engineering, Sejong University, 05006 Seoul, South Korea

A R T I C L E I N F O A B S T R A C T

Keywords: This paper newly proposed a computer vision-based crack quantification algorithm using a statistical approach.
Statistical crack quantification Recently, high-resolution digital images have been effectively utilized for automated crack width and length
Automated crack detection evaluation on concrete structures. However, cracks are often difficult to accurately measure by randomly
Crack separation
distributed and complex shapes. To overcome the technical limitation, a novel statistical crack quantification
Crack map
Deep learning
algorithm is proposed and experimentally validated through concrete structures in this paper. First, cracks on
Computer vision digital images are automatically detected using a deep semantic segmentation network. Then, multi-branched
cracks are separated into single cracks through crack map generation. Each separated crack length and width
is calculated by the Euclidean distance transform algorithm. Finally, crack width is presented as a representative
value with a statistical confidence interval. The quantitative crack evaluation results for width and length were
successfully compared with the actual field measurement values by average differences of 18.07% and − 26.28%,
respectively.

1. Introduction and also suitable for processing vast amounts of data. For instance, Cha
et al. proposed CNN-based crack detection method with a sliding win­
The social interest and demand of civil infrastructural maintenance dow technique [1]. Xu et al. developed a modified CNN-based crack
have been explosively increasing as aging infrastructures gradually in­ detection process with negative sample training such as crack-like for
crease. To come up with such social trend, infrastructure management false alarms reduction [2]. Khani et al. proposed a CNN-based crack
authorities have increased budget and made a long-term plan for detection technique employing patches and grids to enhance block-level
effective structural health monitoring. For example, Korea government detection performance [3]. Jang et al. also proposed a CNN-based crack
has tried to invest a number of smart infrastructure monitoring tech­ detection technique by combination of laser infrared thermography and
nologies to replace conventional expert-dependent structural inspection vision data [4]. As further progressive works, region CNNs have been
which is often unreliable and time-consuming. In reality, it is difficult to used to localize as well as identify cracks [5], and semantic segmenta­
precisely evaluate damage where to hardly access, resulting in ineffi­ tion networks have been proposed to automatically classify the various
cient maintenance budgeting. crack area on concrete structures in pixel-level [6,7]. Bae et al. com­
To tackle the aforementioned technical issues, a number of auto­ bined a deep learning-based super resolution network with the semantic
mated damage evaluation techniques have been developed. One of the segmentation network to enhance the crack detectability by image res­
most widely accepted technologies is the computer vision-based crack olution and quality improvement [8]. More recently, numerous studies
evaluation which makes it possible to automate structural surface crack have been conducted to optimize crack damage detectability through
detection and quantification. In particular, by employing digital camera- various strategies such as network structure modifying [9,10], semi-
mounted unmanned aerial vehicles (UAV) or unmanned robot systems, supervised learning [11,12], and hybrid sensor [13].
high-resolution digital images can be effectively gathered from large- The previous works have successfully addressed the automated crack
scale infrastructures without human access to dangerous in-situ struc­ detection issue of massive image datasets of huge infrastructures with
tures. To classify cracks in the collected data, deep convolutional neural powerful detectability through deep learning algorithms. For detected
networks (CNNs) have been actively paid attention from early studies crack damage, the crack quantification process based on digital image

* Corresponding author.
E-mail address: [email protected] (Y.-K. An).

https://doi.org/10.1016/j.measurement.2023.112632
Received 30 October 2022; Received in revised form 26 January 2023; Accepted 18 February 2023
Available online 24 February 2023
0263-2241/© 2023 Elsevier Ltd. All rights reserved.
H. Bae and Y.-K. An Measurement 211 (2023) 112632

processing has been also intensively proposed as an evaluation process. Thus, the field crack quantity calculation results are often unreliable and
The prior studies have been mainly focused on crack area extraction overestimated, causing waste of maintenance resource and budget.
from digital images and the corresponding area quantification. For In this paper, a computer vision-based statistical crack quantification
example, Dare et al. (2002) proposed a single crack quantification al­ algorithm incorporated with the deep semantic segmentation network is
gorithm based on Bi-linear interpolation of manually detected crack area newly proposed to supplement the representative crack quantification
pixels [14]. Yu et al. (2007) proposed the image processing procedure to index which can be used for concrete structure maintenance in field
simply define crack region with a Dijkstra method on the graph through industry. The proposed technique consists of the following three steps:
image pixels’ nodalization, and quantify them at a fixed working dis­ (1) deep learning-based crack detection, (2) crack separation and (3)
tance [15]. Oh et al. (2009) proposed the robot system to evaluate crack crack length and statistical width calculation. The proposed algorithm
within the fixed working distance based on Fujita, Sobel and Canny’s was experimentally validated using the digital images obtained from a
methods [16]. Jahanshahi et al. (2013) proposed the pinhole camera concrete surfaced building located in South Korea.
algorithm-based crack quantification technique without using scale This paper is organized as follows. Section 2 explains the working
marker [17]. Shan et al. (2016) proposed crack width quantification principle and the detailed description of the statistical crack quantifi­
approaches based on 3D coordinate estimation using a stereovision cation algorithm. Subsequently, Section 3 describes field validation test
camera [18]. Payab et al. (2019) developed a separated individual setup and results. The test results are then thoroughly discussed in
cracks quantification algorithm using a tree graph based on connectivity Section 4. The final conclusion with brief summary and further studies
at pixel-level [19]. Park et al. (2020) proposed crack quantification al­ are addressed in Section 5.
gorithm using the projected laser beam on the surface [20]. Jang et al.
(2021) proposed a Euclidean distance transform (EDT)–based crack 2. Statistical Crack Quantification Algorithm
quantification algorithm at a 100-μm level with high-resolution crack
maps that cover the entire target area [21]. Jereme et al. (2022) com­ Fig. 1 shows the overall flowchart of the proposed computer vision-
bined the shortest and orthogonal projection methods for precise width based statistical crack quantification algorithm.
measurements of cracks with irregular boundaries or high curvature Once the test digital images are obtained by human or unmanned
[22]. Kim et al. (2022) accurately estimated crack width, length, and vehicles including UAV or robot, crack areas in the acquired images are
depth from the efficiently reconstructed 3D model employing stereo automatically detected (VD image) using a deep semantic segmentation
vision from two different focal length lenses [23]. network, if the target areas include actual cracks. Then, only crack pixels
Although a number of computer vision-based crack detection and corresponding to actual cracks are extracted (VC image) through image
quantification techniques have been successfully developed, they have processing algorithm such as conversion, morphological and filtering
some constraints regarding crack characteristic to be widely utilized in operation. Subsequently, the crack map (VM image) showing the con­
field industry. The conventional crack quantification techniques which nectivity of cracks is formed by simplifying the extracted crack pixels.
have focused on simple pixel-level calculation regardless of complex Next, the cracks are separated (VSC images) based on the crack map by
shape could not represent actual crack width index used for reasonable designating the end-point and branch-point. Every single crack is
crack quantity and repair cost estimation. In reality, crack branches and quantified at pixel-level through the EDT algorithm and the pinhole
propagation patterns are all extremely different depending on external camera equation. Here, the crack length is calculated by reflecting the
loading and structural member types even on a single structure. Such crack shape shown in the crack map, and the crack width is computed as
concrete cracks, which have typically multiple branches and complex the range value based on the statistical confidence interval with normal
shape, are often quantified as three or four representative points’ width distribution of the maximum width in the crack slices. At last, the
values which are visually calculated and averaged in field industry. quantification resultant image is finalized (VF image) as the precise

Step (1) Deep learning-based

Step (1) crack detection

Test image Crack area detection using

deep learning ( image)

Crack pixel extraction

( image) Step (2) Crack separation

Crack map generation End-points(n) and Step (3) Crack length

( image) branch-points(j) searching and statistical width
. calculation

No Crack quantification using

j>0?
Euclidean distance transform
Yes

First branch-point Normal distribution fitting

selection of maximum crack width
from crack slices

Branch-point breaking Quantification

( images) Confidence interval
resultant image
(μ+2σ) calculation
( image)

Fig. 1. Flowchart of the statistical crack quantification algorithm.

2
H. Bae and Y.-K. An Measurement 211 (2023) 112632

crack evaluation outcome, which can be used as a crack quantification ratio of image data used for training and verification is about 9:1. The
indicator for establishing an efficient maintenance plan. The details of training and validation datasets are composed of 1955 and 55 images,
each step are explained in the following subsections. respectively. To ensure the network performance, the epoch was set to
150 with a mini-batch size of 10, and an Adam optimizer with the initial
learning rate was set to 10e-3.
2.1. Step (1) Deep Learning-Based Crack Detection Fig. 3 shows the overall generation procedure of the VC and VM im­
ages from the VD image. Since the proposed semantic segmentation
The first step of the proposed algorithm is deep semantic segmen­ network classifies the crack area containing enough crack boundaries to
tation network-based crack area detection. The acquired vision image is catch all crack pixels on the VD image, only actual crack pixel needs to be
used as the test image, and the crack area is automatically detected extracted from the VD image for precise crack quantification as the VC
through a CNN-based deep semantic segmentation network. In this image. Next, the VM image consisted of the simplified crack shape is
study, SegNet [24] is modified and built for crack area detection through sequentially generated from the VC image, so that complex and diverse
transfer learning. The SegNet is constructed based on VGG-16 [25] and crack with multiple branches can be separated as each single crack. In
composed of encoder and decoder as shown in Fig. 2. The convolution reality, the width of each crack branch may have quite different
layer of the decoder and the encoder layer is identical, but the up- depending on crack initiation mechanism, thus it is reasonable that each
sampling and upscale layers are used to maximize the pooling index crack branch needs to be separately quantified.
received from the corresponding encoder layer for pixel-level binary To extract the actual crack pixel from the VD image, several image
classification. As the final step, the crack area pixels in the test image are processing steps such as morphological technique-based area opening
classified into the crack and intact through the softmax layer as the [26], RGB color scale to grayscale image converting [27], median filter
result image as shown in Fig. 2. [28], dilation [29] and skeletonization [30] are carried out, as shown in
As for network construction, CPU of AMD Ryzen™ Threadripper™ Fig. 4. First, dotted or pepper noises detected areas are eliminated on the
3970× and GPU of NVIDIA GeForce RTX 2080 Ti were used in this VD image by an area opening morphology technique [26]. Also, the
study. Note that, the high-performance graphics processing unit RTX detected areas containing fewer pixels than 3% of the input image res­
2080 Ti with 11 GB memory and 4352 cores, was employed to expedite olution are removed, because the continuous crack is assumed to be
the network training and testing. For transfer learning, 500 various considered in this study. The IS image shows the actual crack image in
types of raw crack images are used as training datasets, which were the detection area, and it is then converted to the grayscale (IG) image
collected from web scrapping and field concrete structures by DSLR and which carries only crack intensity information [27]. Subsequently, the
cellphone cameras. The resolution of the training dataset was fixed to median filter is applied to smooth the value of the IG image as the IMF
360 × 480, and the dataset was augmented to 2 K images by reversed image to put out the initial crack (IC) image. The IC image is the
horizontally or vertically with a 50% probability during training. The

Test image Result image

Decoder Crack
Encoder Intact
Upscale &
Conv Softmax Upsampling Pooling

Fig. 2. Architecture of deep semantic segmentation network.

Crack Crack Crack

Intact Intact Intact
0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0
0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0
0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0
0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0
0 0 1 1 1 1 1 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0 1 0 0 0
1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 0 0 0 1 1 1 1 0 0 0
1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 0 1 1 0 0 0 0 1 0 0
1 1 0 0 0 1 1 1 1 1 1 0 0 0 1 1 1 1 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

image image image

Fig. 3. Overall generation procedure of crack area detection (VD), crack pixel (VC) and crack map (VM) images.

3
H. Bae and Y.-K. An Measurement 211 (2023) 112632

image image

image image image image

image image
Fig. 4. Crack pixel extraction process from the detected crack area: IS, IG, IMF, IC, IDL, and ISK images denote the resultants of segmentation, grayscale, median filter,
initial crack area, dilation, and skeletonization resultant images, respectively.

subtracting resultant from the IMF image to the IG image by the back­ disconnected crack may occur as shown in Fig. 4 the IC image. To con­
ground information elimination for crack. For the post-processing, the nect them, the image dilation algorithm is primarily applied to the IC
binarization is performed to the subtraction resultant based on the image by expanding or thickening foreground crack fragments (IDL
threshold value determined by Otsu’s method [31] then, the leftover image) as [29]:
pepper noises are erased using the area opening morphology method­
I DL = I C ⊕ SD . (1)
ology. However, if the difference in luminance is not large, partially

0 1 1 1 0 1 1 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0
0 1 1 1 0 1 1 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0
1 1 1 0 1 1 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0
1 1 1 0 1 1 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0
0 1 1 1 0 1 1 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0
0 0 1 1 0 0 1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1
0 1 1 1 0 1 1 1 0 1 0 1 0 1 0 1 0 0 0 1 0 0 0 1
0 1 0 1 0 1 1 1 0 0 1 1 0 0 1 0 0 0 1 0 0 0 1 0
0 1 0 1 0 1 1 1 0 1 1 1 0 1 0 1 0 0 0 1 0 0 0 1

image image image image

image image
Fig. 5. Crack map image generation process: IF, IT and ISP images denote the filling resultant, thinning resultant and spur pixel pruning resultant images,
respectively.

4
H. Bae and Y.-K. An Measurement 211 (2023) 112632

End-points
Branch-point
1

1
1-1 1-2 1-3
1

1 1 1 1 1 1 1 1 1 1

1 1 1 1

1 1 1 1 1

images
image image
Fig. 6. Crack separation process: VSC images denote the crack separation resultant images.

If the neighborhood of a pixel in the IC image perfectly matches with

1 1 1 0 0 1 1 1 0 0 S, the hit-or-miss transformation at this pixel turns 1, and 0 otherwise.
Finally, the VC image same as the actual crack is created by combining
0 0 1 0 0 0 0 1 0 0
the ISK image and the IC image. The VC image is the finalized crack
0 0 1 0 0 0 0 1 0 0 image, which is used to identify the crack structure and quantitative
evaluation.
0 1 0 1 1 0 1 0 1 1 The shape of the crack is typically complicated and has various
characteristics such as splitting or merging in several branches as shown
0 1 0 0 1 0 1 0 0 1 in VC image. In fact, the crack area in the high-resolution image for
quantifying 100 μm-level has a much more complex shape in pixel-level,
1 : End-points 1 : Branch-point thus the following process is necessary to figure out the main structure
for the quantity calculation of cracks. The VM image to structurally
Fig. 7. Definition of end-point and branch-point. identify cracks in various branches is generated from the VC image
without impairment of the various branches connectivity in the crack
area by the image processing algorithms, as shown in Fig. 5. In the case
where SD and ⊕ are the structuring element of the dilation transform and of the VC image or ISK image, they are not suitable for use in under­
the morphological dilation operator, respectively. standing the crack’s main structure due to undesirable holes and short
SD is selected to be a diamond structuring element in the proposed irregularity spur pixels in the boundary or inside of the crack. To
algorithm. The crack fragments in the IC image would be connected generate the VM image, all holes that may exist inside the crack pixels in
based on the skeletonization (ISK) image from the IDL image. The ISK the VC image, are filled with the fill algorithm [27], resulting in the IF
image is the essential structure of the image without any changes by image, as the first step. Then, skeletonize and thinning algorithms are
reducing all pixels in the IDL image to 1-pixel wide curved lines. To sequentially applied to the IF image so that only the minimum number of
generate ISK image, the morphological thinning algorithm is provided by structural pixels remain, defined as the IT image in Fig. 5. Since spur
follows [30]: pixels may still exist in the generated the IT image, those could be
[ ]M [ ( ( ) )]M removed by the pruning process [32], another sort of thinning method,
I SK = I C ⊗{SK } = I C \ I C ⊙S1 ⋯⊙SK ,(K = 1,…,8,M = 1,2,3,…).
resulting in the ISP image. Here, the skeletonization-thinning-spur
(2)
pruning process can be repeated until there is no change in the ISP
where \, ⨀, M, and S are denote the set difference, the hit-or-miss image, and the iterative number can be different for each image. Finally,
transformation, the number of cyclical iterations until convergence if the iteration algorithm is stopped, the final ISP resultant image is
and the structuring elements, respectively. Note that the K set to eight is defined as the VM image.
a basic number of structuring elements [30].

Fig. 8. Connectivity separation by branch-point removal.

5
H. Bae and Y.-K. An Measurement 211 (2023) 112632

0 0 0 0 0 0 0 0
0 0 1 1 0 0 0 0 0 1 1 0 0 0
0 0 1 0 0 0 0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0 0 1 0 0 0 0
0 1 0 0 0 0 0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0 1 0 0 0 0 0 Crack path
1 0 0 0 0 0 0 1 0 0 0 0 0 0 Basic
1 0 0 0 0 0 0 1 0 0 0 0 0 0
0 1 1 0 0 0 0 0 1 1 0 0 0 0 Overlapped
0 0 1 0 0 0 0 0 0 1 0 0 0 0
0 0 0 1 0 0 0 0 0 0 1 0 0 0 Pixel distance
0 0 0 1 1 0 0 0 0 0 1 1 0 0 0.5 pixel
1-1 0 0 0 0 0 1 0 0 0 0 0 0 1 0
1 pixel
0 0 0 0 0 1 0 0 0 0 0 0 1 0
image 0 0 0 0 0 0 0 0 2 pixel
image
image image
Fig. 9. Pixel distances derive process for crack length calculation: IZ and IP images denote the resultant images of zero padding and the crack distance estimation,
respectively.

(a) (b)
Fig. 10. Pixel structural elements for the optimal pixel distance calculation: (a) basic elements, (b) overlapped elements.

2.2. Step (2) Crack Separation end-point is defined as the pixel that is connected to other pixels at the
maximum of two of eight adjacent pixels and not connected to any pixel
The second step is crack separation based on the connectivity to except directly connected pixels, as displayed in Fig. 7. In the case of the
accurate quantity calculation for dependable evaluation of the detected branch-point, at least three or more pixels surround the one in the eight
cracks. The process of each detected crack separation in the image is as adjacent pixels position as shown in Fig. 7.
follows: first, the end-point and branch-point are extracted on the VM Next, the crack is separated by the first branch-point disconnection
image by considering pixel connectivity [28], as shown in Fig. 6. The as shown in Fig. 8. The eight adjacent pixels to the first branch-point

Fig. 11. Final crack quantification resultant image generation process: IE, IST, and IQ images denote the resultant images of the EDT algorithm, skeletonization and
thinning process (Skel. And Thin.), and crack quantification, respectively.

6
H. Bae and Y.-K. An Measurement 211 (2023) 112632

shown in each VSC image in Fig. 6.

2.3. Step (3) Crack Length and Statistical Width Calculation

The final step is the length and statistical width calculation for each
separated individual crack. For quantitative evaluation of the crack,
length (L) and width (W) are basically computed based on the VMand the
quantification value (IQ) images, respectively corresponding to the
target cracks.
For the crack length estimation, the pixel distance summation be­
tween all pixels in the VM image means the proper length constituting
the crack shape thus all the optimal pixel distances need to be added as
shown in Fig. 9.
In order to obtain the optimal pixel distance for the crack shape,
zeros are firstly padded (IZ image) on the edge of the VM image to match
the pixel structural elements. The basic elements matching result pro­
vides that the basic crack path demonstrated by the straight line, and the
shortest path may then be determined by omitting the duplicate path,
represented as a dotted line utilizing overlapped elements. The basic and
overlapped element types are shown in Fig. 10 (a) and (b), respectively.
Fig. 12. Normal distribution curve fitting resultant of maximum value from For physically computing the actual crack length and width, the scale
separated crack. factor (s) is defined by the following pinhole camera Eq. [33]:
s = dw l/Pf . (4)
changed to zero value to define each remaining crack parts as the cracks
to be quantified individually. If there are multiple branch-points, the where dw, l, P, and f are the working distance between the camera lens
point closest to (0,0) would be defined as the first branch-point. The and target structure, the size of the camera sensor, the image pixel
iterative loop would be stopped when there is no remaining branch- resolution, and the focal length of the camera, respectively.
point, and the number (i) of crack detachments for the target quantifi­ The final crack pixel distance estimation result (IP image) is gener­
cation can be calculated by Eq. (3). ated by pixel distances derive process and the crack length is calculated
i = j + n − 1. (3) by adding up all the pixel distances and multiplying by the scale factor
(s):
where j and n denote the numbers of branch-points and end-points, ∑
L= I P (p, q) × s. (5)
respectively.
As a result, each separated crack is defined as a new single crack, as For the crack width estimation, the quantification algorithm employs

(a) (b) (c)

Branch-point 1-1 1-2 1-3 Crack 1-4

: 317.070 mm
End-point Crack 1-1
: 112.504 mm
: 0.462-0.473 mm
: 0.422-0.435 mm

Crack 1-3
: 148.612 mm Crack 1-6
: 0.432-0.451 mm
1-4 1-5 1-6 : 176.077 mm
: 0.485-0.500 mm

Crack 1-2
: 61.647 mm Crack 1-5
: 0.358-0.379 mm : 84.048 mm
: 0.407-0.432 mm

(d) (e) (f)

Fig. 13. Test results of Case 1: (a) test image, (b) VD image, (c) VC image, (d) VM image with branch-points and end-points, (e) VSC images and (f) VQ image.

7
H. Bae and Y.-K. An Measurement 211 (2023) 112632

Fig. 14. Normal distribution curve fitting resultant of the maximum values from (a) Crack 1–1, (b) Crack 1–2, (c) Crack 1–3, (d) Crack 1–4, (e) Crack 1–5, and (f)
Crack 1–6.

additional statistical procedure to supplement the representativeness width is extracted as the IST image by using the skeletonization and
and secure the reliability of the final quantification resultant index. As a thinning (Skel. And Thin.) algorithm on the VSC image, as shown in
target for applying the statistical approach, the IQ image of each single Fig. 11. It is apart from the VM image for identifying the structure and
crack is generated based on image processing and the EDT algorithm of connectivity of the actual crack in that the IST image is the result of
the VSC image, as shown in Fig. 11. extracting the crack boundary, which physically represents the linear
First, the pixel-level crack width value is computed by pixel distance connection between adjacent pixels within the actual crack area. As
calculation between crack pixel (1) and intact pixel (0) in the Vsc image described in Step (2), in the case of the VM image, algorithms such as fill
using the EDT algorithm. The Euclidean distance (d) of the two pixels and dilation are used to give fluctuations in the actual crack pixel area
that is p and q, is expressed by [34]: for structural analysis.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2 ( )2̅ ( ) For generated IE and IST image, the IQ image can be generated by
d(p, q) = xp − xq + yp − yq p ∈ A, q ∈ AC . (6) using the following Eq. (8).
Q E ST
(8)
where A is the set of black pixels represented by 1 in the VSC image and
Iu,v = s × Iu,v Iu,v .
AC is the white pixel, complement of the black pixel set, A.
where IQ E ST Q E ST
u, v, Iu, v and Iu, v are each pixel value of the I , I , and I images,
Based on the Euclidean distance values for each crack pixel, the
respectively.
Euclidean distance transformed crack (IE) image can be obtained by Eq.
Subsequently, a statistical distribution is used for the IQ image of a
(7).
single crack as the additional process. First, the maximum width value
)
I E (p) = min(d(p, q) | V SC (q) ∈ AC (7) (Wmax) is collected from each region by slicing the IQ image of the target
crack to compose a sufficient population. Then, the collected data are
Then, the simplified shape of the target pixel area to calculate crack

Table 1
Comparison of the crack quantification results between measured values and proposed algorithm values for Case 1.
Crack Lsc (mm) Lour (mm) Difference Wfmax Wour (mm) Difference
(%) (mm) (%)
μ σ μ + 2σ
1–1 97.613 112.504 15.26 0.565 0.422 0.036 0.435 − 23.01
1–2 54.796 61.647 12.50 0.603 0.358 0.034 0.379 − 37.10
1–3 127.863 148.612 16.23 0.870 0.432 0.062 0.451 − 48.16
1–4 261.230 317.070 21.38 1.045 0.462 0.059 0.473 − 54.74
1–5 69.744 84.048 20.51 0.648 0.407 0.055 0.432 − 33.28
1–6 146.640 176.077 20.07 1.048 0.485 0.057 0.500 − 52.27
Average 17.66 Average ¡41.43

8
H. Bae and Y.-K. An Measurement 211 (2023) 112632

Fig. 15. Test results of Case 2: (a) test image, (b) VD image, (c) VC image, (d) VM image with branch-points and end-points, (e) VSC images and (f) VQ imageN

Fig. 16. Normal distribution curve fitting resultant of the maximum values from (a) Crack 2–1, (b) Crack 2–2, (c) Crack 2–3, (d) Crack 2–4, (e) Crack 2–5, and (f)
Crack 2–6.

9
H. Bae and Y.-K. An Measurement 211 (2023) 112632

Table 2
Comparison of the crack quantification results between measured values and proposed algorithm values for Case 2.
Crack Lsc (mm) Lour (mm) Difference Wfmax Wour (mm) Difference
(%) (mm) (%)
μ σ μ + 2σ
2–1 251.807 323.097 28.31 0.747 0.606 0.071 0.620 − 17.05
2–2 24.363 27.642 13.46 1.036 1.011 0.048 1.040 0.34
2–3 375.879 460.786 22.59 1.202 0.999 0.089 1.014 − 15.65
2–4 41.498 47.183 13.70 1.072 0.921 0.101 0.967 − 9.78
2–5 66.032 79.234 19.99 1.009 0.849 0.110 0.888 − 12.00
2–6 98.637 111.276 12.81 0.346 0.288 0.041 0.302 − 12.72
Average 18.48 Average ¡11.14

Crack (mm) (mm)

: 54.796 mm : 61.647 mm : 251.807 mm : 323.097 mm

(a) (b)
Fig. 17. Comparison of crack length: (a) Crack 1–2 in Case 1, (b) Crack 2–1 in Case 2.

Fig. 18. Sample of outlier area from Crack 1–4 in Case 2.

fitted into the normal distribution, and the corresponding mean (μ) and normal distribution and if the crack does not fit in the normal distri­
standard deviation (σ) values are calculated, as shown in Fig. 12. bution due to too small size, it is excluded in the crack quantification
Based on the μ and σ values, a 95% confidence interval (μ ± 2σ) is results.
calculated as the crack width index as shown in Fig. 12. The minimum
and maximum widths are defined as the μ and half-width (μ + 2σ) of the 3. Experimental Validation
95% confidence interval, respectively. The other half of the confidence
interval (μ - 2σ) is not considered due to the probability of under­ 3.1. Experimental Setup
estimating the crack width. From a structural health monitoring point of
view, the underestimation of crack damage may result in a safety haz­ The proposed algorithm was experimentally validated using digital
ard. Note that randomly distributed crack widths typically follow the images obtained from a concrete building structure in South Korea. The

10
H. Bae and Y.-K. An Measurement 211 (2023) 112632

camera model is EOS 5D Mark IV by Cannon. Here, the working distance on the proposed algorithm. However, the largest difference value be­
between the digital camera and the target concrete building surface was tween Wfmax and Wour, − 17.05%, is smaller than the smallest difference
set to 350 mm to evaluate the crack at pixel-level. The physical field of value, − 23.01% in Table 1, and the width difference tends to be reduced
view (FOV) and image resolution of the digital camera were 417.9 × in Table 2. In particular, Crack 2–2 shows that the difference between
313.4 mm2 and 4032 × 3024, respectively. The representative two the two width values is 0.34%, and the Wfmax value even exists within
images of crack areas with complex shapes and various thickness were the width range value of Wour. This case is where the actual measured
selected and defined as Case 1 and Case 2. To quantitatively verify the value by digital vernier calipers and the value from the proposed algo­
proposed algorithm, the crack width value (Wfmax) was calculated by rithm have matched each other.
averaging the values from the four maximum width spots directly
measured along the crack area using a digital vernier calipers in this 4. Discussion
study. The crack length (Lsc) is referred to as the shortest distance be­
tween the crack’s two end-points according to the Guidelines for facility The purpose of this study is to perform a highly accurate quantitative
safety and maintenance by the Ministry of Land, Infrastructure and evaluation based on image processing algorithms and statistical tech­
Transport in South Korea [35]. Note that the crack length and width niques in the automated deep learning-based crack evaluation process.
values calculated based on the proposed algorithm are defined as Lour The algorithm validation results for the test images showed that the
and Wour, respectively. reliability of the quantitative evaluation of crack according to the
existing safety diagnosis guidelines was low. In particular, the Lsc, which
3.2. Experimental Results indicates the length of the crack, is insufficient to reflect the actual
length of the crack shape. A simply estimated straight line is not suitable
Fig. 13 shows the representative test results of Case 1. The original for estimating the true crack length due to the complexity of the real
test image with raw resolution is shown in Fig. 13 (a), and the corre­ crack shape.
sponding VD image is presented in Fig. 13 (b). Cracks in Case 1 contain To help the understanding, the two cracks of Case 1 and 2 are shown
noised, unclear, and vague damage like a common crack condition on in Figs. 17 (a) and (b) which recognized the difference between Lour and
the painted concrete surface building. It can be easily confirmed that the Lsc at 12.50% and 28.31%, respectively. It is proven that the shortest
crack area pixels are successfully extracted from Fig. 13 (c). Three end- distance which connects the branch- and end-points may increase crack
points with the red-dot circles and each single branch-points with the length estimation errors, especially when the crack shape is not straight
blue cross-mark circles are properly computed in the VM image as shown line as shown in Fig. 17 (b). Even if the crack shape looks close to the
in Figure13 (d). Each crack is separated into the six crack segments by straight line as Crack 1–2 in Case 1 of Fig. 17 (a), the Lour reflecting the
breaking the branch-point connection using crack separation process as actual shape is much longer than the Lsc. It is shown that the proposed
depicted in Fig. 13 (e). Finally, each crack segment is quantitatively algorithm resolve the inaccurate problem of the simple straight-line-
evaluated as a different individual as shown in Fig. 13 (f). The statistical based cracks length estimation approach which is widely used in field
distribution resultant used to compute the width value for separated applications.
cracks in Case 1 is displayed in Fig. 14. It turned out that the six crack As for the crack width, the Wfmax, computed from the four spot’s
cases followed the normal distribution, and all the minimum and Wmax values which are directly measured in field using a digital vernier
maximum values are defined as the calculated μ and μ + 2σ, calipers, is overestimated in most validation cases except one case. This
respectively. result raises doubts about whether the representativeness of the Wfmax or
To quantitatively examine the crack quantification results based on Wmax value is appropriate for the quantitative evaluation of the target
the proposed algorithm, the differences between representative values crack area. In particular, the large differences would be the affected
such as Lsc, Wfmax, Lour, and Wour are computed regarding Case 1, which result by the unclear crack area in the blurred image or complex shaped
is summarized in Table 1. crack and the tiny or huge size of outliers such as noise or holes. Such
Table 1 reveals that the average difference index of the Lour is 17.66% outlier regions are easily and generally found in the crack area as shown
and this physically demonstrates that the Lour is around 1.2 times longer in Fig. 18, and most deep learning-based automated crack detection
than Lsc. For the crack width, the difference between Wfmax and Wour is at networks classify them as part of crack regions. Therefore, if there is no
least − 23.01% and up to − 54.74%. In contrast to the crack length consideration of these outliers, the representativeness of Wmax would be
validation results, it is proven that the Wfmax were overvalued by around seriously contaminated when calculating the crack width due to the
1.7 times on average with the Wour. overestimation.
Similarly, the test results of Case 2 are displayed in Fig. 15. In this Furthermore, deep learning-based automated crack detection tech­
case, relatively thick and thin cracks are clearly mixed as shown in niques classify the tiny or huge size of outliers such as noise or holes as
Fig. 15 (a). The various thicknesses cracks are all detected well as shown part of the crack regardless of image data quality as mentioned in the
in Fig. 15 (b), and the actual crack pixels are successfully extracted as introduction. The complex shape of cracks can even degrade the deep
displayed in Fig. 15 (c). The two branch-points and four end-points are learning-based crack detectability and ambiguous crack area is difficult
defined and were successfully detached into the five individual cracks, i. to evaluate precisely as shown in Case 1. For these reasons, the damage
e. 2–1, 2–2, 2–3, 2–4, and 2–5 of Fig. 15 (e), as expected. The quanti­ quantification results could have low reliability without considering the
tative results of the whole six single cracks are elaborately estimated crack’s characteristics such as outliers or complex shapes. It is uneco­
through the length and statistical width calculation step in the proposed nomical to repair the entire area around the crack based on the vague
algorithm as seen in Fig. 15 (f). The statistical distribution results of each quantitative evaluation results and safety factors in field as the con­
crack to support the basis for the minimum to maximum width ventional methods. The proposed algorithm results provide new evi­
computation are shown in Fig. 16. Similarly, all cracks in Case 2 follow a dence for the strong necessity to improve and update the previous crack
normal distribution. quantification approaches. It is strongly reasonable to calculate a width
The comparison of the quantification resultants by the proposed al­ that can represent the thickness value of the crack except for those
gorithm and the presentative values for crack length and width of the six outliers based on our proposed statistical algorithm and a length that
individual cracks in Case 2 are provided in Tables 2. reflects the actual size of the crack shape. Based on the precise quanti­
Similar to Table 1, it can be seen that the Lsc is about 0.85 times tative evaluation, the risk of crack damage can be evaluated in detail,
shorter than the Lour. The length difference of six single cracks in Case 2 and a reasonable maintenance method such as the number of in­
varies from about 30% to about 13%. Likewise, most of the Wfmax was spections, maintenance quantity, and building safety grade calculation
overestimated by about 1.1 times on average than the Wour values based can be established according to the risk level.

11
H. Bae and Y.-K. An Measurement 211 (2023) 112632

5. Conclusion damage types, Comp. Aided Civ. Infrastruct. Eng. 33 (9) (2018) 731–747, https://
doi.org/10.1111/mice.12334.
[6] C.V. Dung, Autonomous concrete crack detection using deep fully convolutional
This study newly proposed a computer vision-based crack quantifi­ neural network, Autom. Constr. 99 (2019) 52–58, https://doi.org/10.1016/j.
cation algorithm by decision-making based on statistical methods. The autcon.2018.11.028.
proposed algorithm separates cracks through image processing tech­ [7] Z. Liu, Y. Cao, Y. Wang, W. Wang, Computer vision-based concrete crack detection
using U-net fully convolutional networks, Autom. Constr. 104 (2019) 129–139,
niques and enables precise quantification based on statistics to calculate https://doi.org/10.1016/j.autcon.2019.04.005.
very detailed damage quantities. The proposed computer vision-based [8] H. Bae, K. Jang, Y.K. An, Deep super resolution crack network (SrcNet) for
statistical crack quantification algorithm was successfully verified improving computer vision–based automated crack detectability in in situ bridges,
Struct. Health Monit. 20 (4) (2021) 1428–1442, https://doi.org/10.1177/
through test images from the various concrete surfaces’ structures. The 1475921720917227.
validation test results reveal that the complex crack shape and the [9] M. Mousavi, A. Bakhshi, Optimized U-shape convolutional neural network with a
outliers that reduce the reliability of precise quantification results must novel training strategy for segmentation of concrete cracks, Struct. Health Monit.
14759217221105647 (2022), https://doi.org/10.1177/14759217221105647.
be considered in the process of automated deep learning-based crack [10] H. Chu, W. Wang, L. Deng, Tiny-crack-net: a multiscale feature fusion network with
evaluation. In the perspective of facility maintenance policy, this pre­ attention mechanisms for segmentation of tiny cracks, Comp. Aided Civ.
cisely calculated crack quantity is able to be used as objective evidence Infrastruct. Eng. 37 (14) (2022) 1914–1931, https://doi.org/10.1111/mice.12881.
[11] W. Wang, C. Su, Semi-supervised semantic segmentation network for surface crack
for the budget for maintenance and reinforcement, it can be used effi­ detection, Autom. Constr. 128 (2021), 103786, https://doi.org/10.1016/j.
ciently as a basic index for reliable budget planning. In especially, the autcon.2021.103786.
precise damage time-series data for the target structure can be used in [12] S. Shim, J. Kim, S.W. Lee, G.C. Cho, Road damage detection using super-resolution
and semi-supervised learning with generative adversarial network, Autom. Constr.
the Digital Twin based on the accurate crack quantification results from
135 (2022), 104139, https://doi.org/10.1016/j.autcon.2022.104139.
the proposed algorithm. [13] K. Jang, H. Jung, Y.K. An, Automated bridge crack evaluation through deep super
As the follow-up study, the proposed algorithm will be extended to resolution network-based hybrid image matching, Autom. Constr. 137 (2022),
detect and precisely quantify multi damage such as rust, rebar exposure, 104229, https://doi.org/10.1016/j.autcon.2022.104229.
[14] P. Dare, H. Hanley, C. Fraser, B. Riedel, W. Niemeier, An operational application of
spalling, efflorescence, etc. Furthermore, we plan that verify our algo­ automatic feature extraction: the measurement of cracks in concrete structures,
rithm with data obtained from various real concrete structures such as Photogramm. Rec. 17 (99) (2002) 453–464, https://doi.org/10.1111/0031-
bridges, dams, roads, and tunnels to expand the scope and improve the 868x.00198.
[15] S.N. Yu, J.H. Jang, C.S. Han, Auto inspection system using a mobile robot for
effectiveness. detecting concrete cracks in a tunnel, Autom. Constr. 16 (3) (2007) 255–261,
https://doi.org/10.1016/j.autcon.2006.05.003.
[16] J.K. Oh, G. Jang, S. Oh, J.H. Lee, B.J. Yi, Y.S. Moon, Y. Choi, Bridge inspection
CRediT authorship contribution statement
robot system with machine vision, Autom. Constr. 18 (7) (2009) 929–941, https://
doi.org/10.1016/j.autcon.2009.04.003.
Hyunjin Bae: Conceptualization, Methodology, Software, Valida­ [17] M.R. Jahanshahi, S.F. Masri, A new methodology for non-contact accurate crack
tion, Writing – original draft, Visualization, Investigation, Data curation, width measurement through photogrammetry for automated structural safety
evaluation, Smart Mater. Struct. 22 (3) (2013), 035019, https://doi.org/10.1088/
Writing – review & editing. Yun-Kyu An: Conceptualization, Method­ 0964-1726/22/3/035019.
ology, Validation, Formal analysis, Resources, Supervision, Writing – [18] B. Shan, S. Zheng, J. Ou, A stereovision-based crack width detection approach for
review & editing, Project administration. concrete surface assessment, KSCE J. Civ. Eng. 20 (2) (2016) 803–812, https://doi.
org/10.1007/s12205-015-0461-6.
[19] M. Payab, R. Abbasina, M. Khanzadi, A brief review and a new graph-based image
analysis for concrete crack quantification, Arch. Comput. Meth. Eng. 26 (2) (2019)
Declaration of Competing Interest 347–365, https://doi.org/10.1007/s11831-018-9263-6.
[20] S.E. Park, S.H. Eem, H. Jeon, Concrete crack detection and quantification using
The authors declare that they have no known competing financial deep learning and structured light, Constr. Build. Mater. 252 (2020), 119096,
https://doi.org/10.1016/j.conbuildmat.2020.119096.
interests or personal relationships that could have appeared to influence [21] K. Jang, Y.K. An, B. Kim, S. Cho, Automated crack evaluation of a high-rise bridge
the work reported in this paper. pier using a ring-type climbing robot, Comp. Aided Civ. Infrastruct. Eng. 36 (1)
(2021) 14–29, https://doi.org/10.1111/mice.12550.
[22] J.C. Ong, M.Z.P. Ismadi, X. Wang, A hybrid method for pavement crack width
Data availability
measurement, Measurement 197 (2022), 111260, https://doi.org/10.1016/j.
measurement.2022.111260.
Data will be made available on request. [23] H. Kim, S.H. Sim, B.F. Spencer, Automated concrete crack evaluation using stereo
vision with two different focal lengths, Autom. Constr. 135 (2022), 104136,
https://doi.org/10.1016/j.autcon.2022.104136.
Acknowledgements [24] V. Badrinarayanan, A. Kendall, R. Cipolla, Segnet: a deep convolutional encoder-
decoder architecture for image segmentation, IEEE Trans. Pattern Anal. Mach.
Intell. 39 (12) (2017) 2481–2495, https://doi.org/10.1109/tpami.2016.2644615.
The research described in this paper was financially supported by a [25] K. Simonyan, A. Zisserman, Very Deep Convolutional Networks for Large-Scale
grant (22TBIP-C160936-02) from Technology Advancement Research Image Recognition, arXiv preprint, arXiv:1409.1556, 2014, https://doi.org/
Program (TARP) funded by Ministry of Land, Infrastructure and Trans­ 10.48550/arxiv.1409.1556.
[26] L. Vincent, Morphological area openings and closings for Grey-scale images, in:
port of Korea Government.
NATO ASI Series Shape in Picture vol. 126, Springer, Berlin, Heidelberg, 1994,
pp. 197–209, https://doi.org/10.1007/978-3-662-03039-4_13.
References [27] W. Burger, M.J. Burge, M.J. Burge, M.J. Burge, Principles of Digital Image
Processing vol. 111, Springer, Berlin, 2009, pp. 110–111 (ISBN 978-1-84800-195-
4.).
[1] Y.J. Cha, W. Choi, O. Büyüköztürk, Deep learning-based crack damage detection
[28] P. Soille, Morphological Image Analysis: Principles and Applications, Springer
using convolutional neural networks, Comp. Aided Civ. Infrastruct. Eng. 32 (5)
Science & Business Media, 2013. ISBN 9783642076961.
(2017) 361–378, https://doi.org/10.1111/mice.12263.
[29] L. Lam, S.W. Lee, C.Y. Suen, Thinning methodologies-a comprehensive survey,
[2] Y. Xu, Y. Bao, J. Chen, W. Zuo, H. Li, Surface fatigue crack identification in steel
IEEE Trans. Pattern Anal. Mach. Intell. 14 (9) (1992) 869–885, https://doi.org/
box girder of bridges by a deep fusion convolutional neural network based on
10.1109/34.161346.
consumer-grade camera images, Struct. Health Monit. 18 (3) (2019) 653–674,
[30] J. Serra, Image analysis and mathematical morphology, vol. 1, Academic Press,
https://doi.org/10.1177/1475921718764873.
1982, pp. 34–407. ISBN 978-0-12-637240-3.
[3] M. Mohtasham Khani, S. Vahidnia, L. Ghasemzadeh, Y.E. Ozturk, M. Yuvalaklioglu,
[31] N. Otsu, A threshold selection method from gray-level histograms, IEEE Trans.
S. Akin, N.K. Ure, Deep-learning-based crack detection with applications for the
Syst. Man Cybern. 9 (1) (1979) 62–66, https://doi.org/10.1109/
structural health monitoring of gas turbines, Struct. Health Monit. 19 (5) (2020)
tsmc.1979.4310076.
1440–1452, https://doi.org/10.1177/1475921719883202.
[32] C. Gonzalez, W. Richard, E. Steven, Digital Image Processing Using MATLAB, 2nd
[4] K. Jang, N. Kim, Y.K. An, Deep learning–based autonomous concrete crack
ed., McGrawHill, New York, 2011. ISBN: 978–007–108478-9.
evaluation through hybrid image scanning, Struct. Health Monit. 18 (5–6) (2019)
1722–1737, https://doi.org/10.1177/1475921718821719.
[5] Y.J. Cha, W. Choi, G. Suh, S. Mahmoudkhani, O. Büyüköztürk, Autonomous
structural visual inspection using region-based deep learning for detecting multiple

12
H. Bae and Y.-K. An Measurement 211 (2023) 112632

[33] Y. Liu, S. Cho, B.F. Spencer Jr., J. Fan, Automated assessment of cracks on concrete [35] Guidelines for safety and maintenance of facilities, Ministry of Land, Infrastructure
surfaces using adaptive digital image processing, Smart Struct. Syst. 14 (4) (2014) and Transport in South Korea, No. 2022-539. https://www.law.go.kr/%ED%96%
719–741, https://doi.org/10.12989/sss.2014.14.4.719. 89%EC%A0%95%EA%B7%9C%EC%B9%99/%EC%8B%9C%EC%84%A4%EB%
[34] R. Fabbri, L.D.F. Costa, J.C. Torelli, O.M. Bruno, 2D Euclidean distance transform AC%BC%EC%9D%98%EC%95%88%EC%A0%84%EB%B0%8F%EC%9C%A0%EC
algorithms: a comparative survey, ACM Comput. Surv. 40 (1) (2008) 1–44, https:// %A7%80%EA%B4%80%EB%A6%AC%EC%8B%A4%EC%8B%9C%EB%93%B1%
doi.org/10.1145/1322432.1322434. EC%97%90%EA%B4%80%ED%95%9C%EC%A7%80%EC%B9%A8, 2022.

13