International Journal of Pavement Engineering

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/gpav20

Determination of the air void content of asphalt

concrete mixtures using artificial intelligence
techniques to segment micro-CT images

Alexis Jair Enríquez-León, Thiago Delgado de Souza, Francisco Thiago

Sacramento Aragão, Delson Braz, André Maués Brabo Pereira & Liebert
Parreiras Nogueira

To cite this article: Alexis Jair Enríquez-León, Thiago Delgado de Souza, Francisco Thiago
Sacramento Aragão, Delson Braz, André Maués Brabo Pereira & Liebert Parreiras Nogueira
(2021): Determination of the air void content of asphalt concrete mixtures using artificial intelligence
techniques to segment micro-CT images, International Journal of Pavement Engineering, DOI:
10.1080/10298436.2021.1931197

To link to this article: https://doi.org/10.1080/10298436.2021.1931197

Published online: 15 Jul 2021.

Submit your article to this journal

Article views: 78

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gpav20
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING
https://doi.org/10.1080/10298436.2021.1931197

Determination of the air void content of asphalt concrete mixtures using artificial
intelligence techniques to segment micro-CT images
Alexis Jair Enríquez-León a, Thiago Delgado de Souza a, Francisco Thiago Sacramento Aragão a
,
Delson Braz b, André Maués Brabo Pereira c and Liebert Parreiras Nogueira d
a
Civil Engineering Program – COPPE, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil; bNuclear Engineering Program – COPPE, Federal
University of Rio de Janeiro, Rio de Janeiro, Brazil; cInstitute of Computing, Fluminense Federal University, Rio de Janeiro, Brazil; dDepartment of
Biomaterials, University of Oslo, Oslo, Norway

ABSTRACT ARTICLE HISTORY

X-ray micro-computed tomography (micro-CT) is an advanced technique able to provide a Received 25 November 2020
comprehensive examination of the volumetric characteristics of asphalt mixtures. A key step for the air Accepted 11 May 2021
void (AV) quantification using micro-CT images is the segmentation, which is a stage of the digital
KEYWORDS
image processing. The most common segmentation technique, the manual threshold (TH) selection, X-ray micro-computed
depends significantly on the operator skills, image homogeneity, and material complexity. These tomography; air void; digital
factors that can limit the reproducibility of the TH procedure. Machine learning and deep learning image processing; threshold;
recently appeared as promising alternatives to solve this challenge. In this paper, images of an asphalt machine learning; deep
concrete (AC) specimen were acquired in a modern high-resolution micro-CT scanner to determine its learning
AV content using four different segmentation tools, i.e. TH, watershed, machine learning, and deep
learning. All methods presented similar results for the total AV content. The advantages and
limitations of using each technique were discussed in terms of computational effort, user-friendliness,
and accuracy of the results. Machine learning and deep learning were identified as powerful tools for
AC segmentation, being accurate and easy to adjust, however taking longer data processing times.

1. Introduction
energies, they concentrated at the top and bottom of the speci-
The air void (AV) content of asphalt mixtures comprises the mens. Additionally, a more homogeneous AV distribution in
impermeable voids existing within the microstructure and the middle of the samples was observed. Regarding field speci-
the permeable voids generated in the material compaction mens, the AVs decreased for deeper locations within the
stage. The compaction energy and method, the mineral skel- asphalt concrete (AC) layer.
eton, and even the shape of the aggregates are factors that Masad et al. (1999b) showed that, in addition to the type of
affect the AV distribution, and, consequently, the mechanical compaction and compaction energy, the nominal maximum
behaviour of the asphalt mixtures (Masad et al. 1999a, aggregate size (NMAS) can influence the AV distribution.
1999b, 2002, Tashman et al. 2002). Tashman et al. (2002) verified that the AVs of SGC specimens
The volumetric measurements of the AV within the asphalt were concentrated in the zones in contact with the moulds. For
mixtures using the maximum and bulk specific gravities field specimens, the authors found a more uniform AV distri-
(AASHTO 2016, ASTM 2019) provide only a global average bution in the horizontal direction, contrary to what happened
of the permeable AV content. Thus, the X-ray computed tom- in the vertical direction.
ography (CT) has spread and advanced in recent decades as a Thyagarajan et al. (2010) performed a detailed study upon
promising non-destructive technique to quantify the AV dis- the AV content of SGC specimens employing CT images.
tribution along asphalt material samples (Masad et al. 1999a, Some key observations about sawing and coring the specimens
1999b, 2002, 2006, Tashman et al. 2002, You et al. 2009, Thya- were obtained: (i) sawing reduced the vertical heterogeneity
garajan et al. 2010, Wang and Xiao 2012, Shaheen 2015, Sadeq but did not influence the horizontal heterogeneity; (ii) sawing
et al. 2018). Besides, it is possible to obtain results of permeable 10 mm from the bottom and top of each specimen was insuffi-
and impermeable AVs using the CT (Chen et al. 2016, Chung cient for reducing the heterogeneity of the samples; and (iii)
et al. 2019). coring the specimen had no significant influence on the lateral
One of the first authors that studied the AV distribution in heterogeneity.
asphalt mixtures samples using CT were Masad et al. (1999a). Although several authors (Masad et al. 1999a, 1999b, 2002,
In their work, specimens from field and manufactured in a 2006, Tashman et al. 2002, You et al. 2009, Thyagarajan et al.
Superpave gyratory compactor (SGC) were analysed. It was 2010, Wang and Xiao 2012, Shaheen 2015, Sadeq et al. 2018)
observed that the AVs were evenly distributed when low com- provided meaningful information about the AC AV content
paction energies were applied. However, for high compaction employing CT, two limitations have been identified: spatial

CONTACT Francisco Thiago Sacramento Aragão [email protected]

© 2021 Informa UK Limited, trading as Taylor & Francis Group
2 A. J. ENRÍQUEZ-LEÓN ET AL.

resolution and segmentation of the images. The spatial resol-

ution limitation can be potentially overcome nowadays with
the use of high-resolution X-ray micro-computed tomography
(micro-CT), which is similar to the CT, but can reach spatial
resolutions of about one micrometer (Landis and Keane
2010). Nevertheless, the application of micro-CT for the
quantification of AV distribution in AC has not been pursued
very intensively in the literature.
With micro-CT, the internal microstructure of the AC
samples can be visualised allowing the performance of quali-
tative analyses. However, for quantitative measurements, the
images must be processed and segmented. Regarding image
segmentation, it is common to use the threshold (TH) tech-
nique for the AV characterisation. The TH value is affected Figure 1. AC gradation.
by several factors, such as operator skills, image homogeneity,
and material complexity, thus, different results can be obtained
because it is difficult to guarantee the reproducibility of this
3.1. Image acquisition
process. Zelelew et al. (2013) presented an algorithm to
enhance AC structure segmentation, indicating that the digital A Zeiss Xradia 510 Versa micro-CT scanner was used for
image processing (DIP) by means of TH could be improved, image acquisition. The micro-CT is a non-destructive tech-
however, at that time, the authors used a CT that could only nique that enables to access digital information about material
identify a minimum aggregate size of 1.18 mm. geometry and properties, and also to visualise its internal
Artificial intelligence (AI) subsets, such as machine and the microstructure (Masad et al. 1999a). The micro-CT device
deep learning, have emerged as promising alternatives to consists of an X-ray emitting source, a detector, and a rotating
improve the reproducibility. AI can be applied to digital stage between them, where the specimen is placed.
image processing and offer an optimisation for the segmenta- The micro-CT scan of the whole AC specimen requires a
tion, as it allows training just some slices or sections of the significant computational effort. In addition, physical limit-
sample to generate a pattern recognition for all other speci- ations in the equipment impose some geometric constraints
mens (Bezerra et al. 2020). to the testing specimens. To overcome these challenges, the
This paper aims to demonstrate the applicability of evaluation of samples smaller than the typically produced in
machine learning and deep learning for AC AV microstructure the compaction moulds can be pursued.
segmentation. For this, high-resolution images of an AC According to Masad et al. (1999b), the centre of the AC
sample were acquired in a modern micro-CT scanner. Four sample is the region presenting less AV variability. Thus, the
different DIP-segmentation techniques, i.e. TH, watershed, AC specimen measuring 100 mm in diameter and 58 mm in
machine learning, and deep learning were compared. Further- height was cored using a drill with an inner diameter of
more, the permeable and impermeable AVs were quantified, 13 mm. Additionally, the specimen was sawn to remove
and later rendered in 3D. The advantages and limitations of 4 mm from its top and 4 mm from its bottom. Therefore,
using each segmentation technique are discussed in terms of the AC sample that was tested in the micro-CT presented
computational effort, user-friendliness, and accuracy of the 13 mm in diameter and 50 mm in height.
results. The electrical parameters applied in the X-ray source were a
voltage of 100 kV, a current of 90 μA, and a power of 9 W. The
intensity of the X-rays was measured when the beams passed
2. Materials through the specimen and the media in which it was con-
In this study, an SGC-compacted AC sample was evaluated. tained. By directing the X-rays through the testing object
The AC constituents were a PG 64S-22 binder and a gneiss along several trajectories, in many directions, a set of 1600
aggregate with bulk specific gravity of 2.720. The AC presented images (projections) for the specimen was produced.
NMAS of 19 mm and was produced following the Superpave The spatial resolution adopted was of 7 μm/pixel, which
mix design (AASHTO 2017) procedures, resulting in an was obtained with an optical magnification of 0.4X, source-
asphalt binder content of 4.7%. The AC specimen measured sample distance of 31 mm, and detector-sample distance of
58 mm in height and 100 mm in diameter and its AV of 274.25 mm. The exposure time was of 4 s for each projection.
4.4% was calculated using the maximum and bulk specific The choice of this resolution was based on Osmari et al. (2020),
gravities (AASHTO 2016, ASTM 2019) of the mixture. The which was one of the first studies using the micro-CT to
AC gradation is presented in Figure 1. characterise asphalt materials.
The software XMReconstructor, which is coupled into the
micro-CT system, was used for the reconstruction, generating
3. Experimental program a set of images (stack) that represents the micro-CT itself. This
The experimental program comprised three main steps, which stack exhibited different grayscale colours across its two-
were image acquisition, DIP, and air void extraction. These dimensional images (slices), and presented 2 MB per slice, or
steps are detailed in this section and summarised in Figure 2. 2 GB for the stack.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 3

Figure 2. Steps to obtain the air void content using the micro-CT (adapted from Tashman et al. 2002, Zelelew et al. 2013).

The sample information consisted of 1011 16bit-slices of 1024 Firstly, an equalisation of the image histograms was accom-
pixels × 1004 pixels (Enríquez-León et al. 2020). To minimise the plished to redistribute the gray shade of the pixels. Thus, the per-
distortion effects generated by optical principles of conical X-ray centage of the pixels became similar in any gray level, and the
propagation, 100 images from the top and 100 from the bottom histograms more homogeneous, as illustrated in Figure 3(b).
were disregarded. Furthermore, some slices were disregarded Then, non-local means denoising filter (NLM) (Buade et al.
because an aggregate particle larger than the cored-specimen 2011) was applied. The filter (sigma = 25 and smoothing factor
diameter was completely occupying some slices. Thus, the final = 185) reduced image noise without blurring the edges. Figure
AC specimen analysis was based on 500 slices. 3(c) shows an example of the NLM application where the post-
application histogram presented the peaks better defined. It is
noteworthy to mention that image sharpness was much better
3.2. Digital image processing (DIP) after the preprocessing, as can be seen in the zoomed regions
DIP covered three main steps: preprocessing, segmentation, from Figure 3(c).
and postprocessing. These steps will be detailed below. It is Figure 4 presents an example of an image after the histo-
noteworthy to mention that the workstation configuration to gram equalisation and NLM application, where two phases
perform DIP comprised an Intel® CoreTM i7-8700 3.2 GHz of the AC could be identified, i.e. aggregates and mastic, in
processor with 32 GB of RAM, and an NVIDIA® GeForce gray, and AV, in dark. To recognise the pixels and later quan-
GTX 1070 Ti 8 GB graphics card. tify the AVs, it was necessary to perform image segmentation.

3.2.1. Preprocessing 3.2.2. Segmentation

The open-source software Fiji/ImageJ (Schindelin et al. 2012) was Preceding the image segmentation, files were converted to 8
selected to perform the preprocessing of the micro-CT images. bits. To separate the mastic and aggregate grayscale regions

Figure 3. Example of image filtering and histogram: (a) original image, (b) after equalisation, and (c) after NLM.
4 A. J. ENRÍQUEZ-LEÓN ET AL.

tool searches for the watershed lines (Gonzalez and Woods

2018).
A sophisticated method known as ‘Watershed Transforms’
was used. This procedure selects the landscapes of the image as
a dataset frontier and applies seed points within them. These
points grow until reaching a ‘barrier’, which is determined
by the image gradients (Pereira et al. 2018). In this paper,
the seeds were preselected considering a TH of 22 (TH22),
while the landscapes of the images were based on a TH30
(inverted selection). The image gradients were generated by
a Euclidean distance map. This process is illustrated in Figure
7.

Figure 4. Example of AC phases identification: (a) preprocessed image and (b) 3.2.2.3. Machine learning (ML). Arganda-Carreras et al.
zoomed preprocessed image.
(2017) presented the so-called ‘Trainable Weka Segmentation’
plugin for Fiji/ImageJ (Schindelin et al. 2012). This plugin is
from the AV, it was necessary to binarize the images. Thus, based on the machine learning (ML) random forest concept
four types of segmentation were used: visual threshold (TH), (FastRandomForest), combining algorithms with manual seg-
watershed (WS), machine learning (ML), and deep learning mentation. This tool works as a link between the ML and
(DL). The TH and the ML segmentations were performed in the manual segmentation characteristics, and the user is not
the software Fiji/ImageJ (Schindelin et al. 2012), while the required to know the programming language.
WS and DL were performed in the software Dragonfly v.1 For the ML, during model training, as manual segmenta-
(2020), from Object Research Systems (ORS). These four tions were performed, it was observed that after the 7th slice,
types of segmentation will be detailed below. the mastic was not misinterpreted as voids. However, ten slices
were chosen for the training. These slices were strategically
selected throughout the sample at locations corresponding to
3.2.2.1. Threshold (TH). The TH consists of modifying the their top, centre, and base.
image grayscale to identify regions of interest. By selecting a The training was based on selecting some regions of interest
TH value, the images were binarized in black and white (Gon- (ROIs) and separating them manually by labels, such as pre-
zalez and Woods 2018). Figure 5 shows the TH application, in sented in Figure 8(a). Later, the stack classifier was also trained,
which two AC phases could be identified, i.e. AV, in black, and and this process was replicated until the AC phases were well
aggregates and mastic, in white. recognised, as shown in Figure 8(b).
It is difficult to guarantee the reproducibility of this seg-
mentation technique because the TH is visually selected by 3.2.2.4. Deep learning (DL). The so-called ‘Segmentation
the operator and it also depends on material complexities, Wizard’ Dragonfly (2020) plugin was used. This tool provides
and on the homogeneity of the images. Figure 6 shows an a guided workflow for implementing DL and ML segmentation
example of a slice that had two TH values visually selected, of multi-dimensional images. Some pre-built neural networks
30 and 35, where the difference of the total AV sizes, segmen- have been tested on the DL. Those included U-Net, FCN-Den-
ted in red, was quite significant. seNet, Sensor3D, and others. Similarly, for ML, it was possible
to apply models with different parameters, including Random
Forest, Extra Trees, Gradient Boosting, and others. The pro-
3.2.2.2. Watershed (WS). Another classical segmentation tech- cedure consists of painting the different features of interest
nique, WS is based on the visualisation of a 3D-image contain- within a limited subset of slices and then training models to
ing spatial coordinates versus intensity. The algorithm of this identify objects according to a predefined set of rules, as

Figure 5. Example of TH application: (a) preprocessed image, (b) after TH, and (c) histogram.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 5

Figure 6. Example of TH image segmentation: (a) preprocessed image, (b) TH30, and (c) TH35.

illustrated in Figure 9. The most promising model can then be

fine-tuned and exported for repeated segmentation tasks (Pro-
vencher et al. 2019).
For the training, four slices of the AC were strategically
selected. Once an efficient classifier was obtained, it could be
applied to the sample stack. The detailed description of neural
networks is out of the scope of this paper, though it is impor-
tant to indicate that the depth of the networks is proportional
to the number of parameters that are provided. As a conse-
quence, the parallel computing capacity of graphic processing
units (GPUs) will be a critical factor (Bezerra et al. 2020).
Figure 8. Example of ML image segmentation: (a) classifier training and (b) resul-
tant image.

3.2.3. Postprocessing
Fiji/ImageJ (Schindelin et al. 2012) and CTAnalyzer (2020) slice was computed to obtain the total AV content of the
were the selected software to perform the postprocessing. To sample using an arithmetic mean.
denoise the effects of the segmentation and micro-CT test, To perform an adequate comparison between the AV calcu-
two steps were accomplished: elimination of edges employing lated using the maximum and bulk specific gravity of the asphalt
the kill borders plugin and morphological operations (MOP). mixture, i.e. permeable, it was necessary to separate the AV. To
The MOP used were erosion (r = 1), dilation (r = 1), geodesic overcome this challenge, the tools of the CTAnalyzer (2020)
reconstruction (dilation e connectivity = 4) (Legland et al. were used. The application of ‘ROI shrink-wrap’ created a
2016), and despeckle 3D (Augusto and Paciornik 2018). The boundary delimiting the external borders of the sample images
3D-images resulting from these processes were also free to recognise the AVs that were connected to the specimen sur-
from total AV sizes smaller than 3 × 3 × 3 voxels (27 voxels). face, and thus, allowing to identify the AV type.

3.3. Air void extraction 4. Results and discussion

CTAnalyzer (2020) was used for the quantification while Drag- Based on the connection of 500 slices, 12 images representing
onfly (2020) for 3D rendering. The AV area fraction of each the AVs were rendered in 3D. These images were divided

Figure 7. Example of WS image segmentation (a) seed points placed in the area Figure 9. Example of DL image segmentation: (a) classifier training and (b) sec-
of interest and (b) grown-up seed points touching a barrier. tion of a resultant image.
6 A. J. ENRÍQUEZ-LEÓN ET AL.

Figure 10. Total AV rendered in 3D using segmentation by: (a) TH, (b) WS, (c) ML, and (d) DL.

Figure 11. Permeable AV rendered in 3D using segmentation by: (a) TH, (b) WS, (c) ML, and (d) DL.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 7

Figure 12. Impermeable AV rendered in 3D using segmentation by: (a) TH, (b) WS, (c) ML, and (d) DL.

into three AV groups: total, permeable, and impermeable. One of the main limitations of the TH method is the sensi-
For each group, one of the four segmentation methods, i.e. tivity of the results. In the interval from 30 to 35, an AV rela-
TH, WS, ML, and DL, was applied. Figures 10–12 illustrate tive error of 17% was obtained and this reinforces the necessity
total, permeable (open), and impermeable (closed) AV, of studying other segmentation techniques, such as, for
respectively. instance, machine learning (ML) and deep learning (DL), to
The visual analysis of the AVs rendered in 3D indicates that obtain more accurate and reproducible results.
there were no significant differences between the images for Figure 14 shows the results of total AV content determi-
the permeable AVs. Also, the open AVs were predominantly nations using TH32, WS, ML, and DL, whose means were
interconnected and accounted for the main part of this AC 4.12%, 4.13%, 4.14%, and 4.16%, respectively. The results of
phase, as expected. On the other hand, a relevant difference total AV content along the specimen presented few variations,
was identified for the impermeable AVs, as discussed later. thus a one-way statistical analysis of variance (ANOVA) was
After the 3D rendering, the total AV contents along the conducted to confirm that they were statistically equivalent.
specimen were determined. The first quantification was per-
formed using the segmentation by means of a TH. In the
image grayscale, a (TH) interval from 30 to 35 was initially
selected. Images containing TH values outside this interval
were not representing the total AV phase, i.e. for TH lower
than 30, some AVs were not recognised, while for values
above 35, they were overestimated.
The results of total AV content along the sample are shown
in Figure 13. The total AV content means were 3.86%, 3.98%,
4.12%, 4.33%, 4.50%, and 4.65% considering TH values of 30,
31, 32, 33, 34, and 35, respectively. After the preliminary analy-
sis of the TH values, many more slices were carefully examined
to fine-tune the initial range, which was from 30 to 35. Thus, a
TH of 32 (TH32) was selected as the representative value to
characterise the total AVs because the visual analysis of the
images by experienced professionals indicated that it better
represented this AC phase. Figure 13. Total AV content along the specimen considering different TH values.
8 A. J. ENRÍQUEZ-LEÓN ET AL.

Figure 14. Total AV content along the specimen considering different segmenta-
tion methods.

Figure 16. Comparison between total, open, and closed AV contents.

Considering the F-distribution and a significance level of 0.05,
the F-statistic was 0.12, while the one-sided critical F was 2.61. different segmentation methods were equivalent. For the
This infers the equivalence among the results. open AV content, the means could be considered equivalent
Figure 15 compares the number of AVs by volume inter- because the one-sided critical F was 2.61 and the F-statistic,
vals, in mm³, determined with the four segmentation tech- 0.89. On the other hand, for the closed AV contents, the
niques. Similar results were obtained for the AV volumes ANOVA result was an F-statistic of 22.75, which indicates
larger than 0.001 mm³. The AVs between 0 and 0.001 mm³ that the mean values may not be considered equivalent.
represented 99.0%, 95.4%, 85.7%, and 95.4% of the total AV Therefore, the difference existing for the Total AV of the
number, considering segmentations by TH, WS, DL, and 3D images shown in Figure 10 can be attributed to the closed
ML, respectively. This trend can also be observed in the 3D- AVs.
images from Figures 10 and 11, where the image rendered To identify the most precise segmentation technique for
using a TH contained numerous small-sized AVs (7067). quantifying closed AVs, it is necessary to improve the micro-
Figure 16 presents a comparison between the total AV con- CT spatial resolution and also to use specimens including a
tent obtained using TH32, WS, ML, DL, and laboratory volu- substance with an atomic number differing from the asphalt
metric-based (LVB) procedure, i.e. calculated using maximum binder and water, such as the iodine, in the experimental pro-
and bulk specific gravities (AASHTO 2016, ASTM 2019) of the gram. This can provide a means for the even more accurate
asphalt mixture. In theory, the comparisons with the labora- determination of permeable and impermeable AVs within
tory measurement must be based on the open (permeable) AC mixtures.
AVs. The open AV content means were 4.40%, 3.86%, As indicated by previous research (Masad et al. 1999a,
3.95%, 3.85%, and 3.92%, for LVB, TH32, WS, ML, and DL, 1999b, Tashman et al. 2002, Thyagarajan et al. 2010), the
respectively. AC samples present considerable variability on AV along
Two one-way ANOVAs using the same aforementioned their height and diameter. Thus, it should be noted that
parameters were conducted to evaluate if the mean of the when comparing the AV characterised by micro-CT-DIP
open and closed (impermeable) AV contents considering and by LVB, they may be closer if the volumetric character-
istics of smaller testing samples is evaluated in the laboratory.
This was not possible in this research because no high pre-
cision analytical hydrostatic scale was available. In addition,
it is also recommended to evaluate the effects of the spatial
resolution used in micro-CT (Cnudde et al. 2009, Augusto
and Paciornik 2018, Pereira et al. 2018) on the AV content
results.
To conclude, the time, user-friendliness, and accuracy
characteristics of each segmentation technique were com-
pared. The time processing for the segmentation of the
stack evaluated was 3, 4, 6, and 10 h for TH, WS, ML, and
DL, respectively. An additional hour was required by each
method to render the specimen in 3D and finally determine
the AV contents. It is noteworthy to mention that depending
on the workstation configurations, micro-CT test parameters,
and image sizes, those processing and analysis times may
vary.
Figure 15. Number of AVs by volume intervals, in mm³, considering different seg-
Regarding user-friendliness, the artificial intelligence (AI)
mentation methods. methodologies, i.e. ML and DL, were the best methods as
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 9

they were easily fine-tuned, using just part of the slices. Con- Acknowledgements
cerning accuracy, the AI methodologies were also better The authors wish to express their gratitude to the Engineering School of
because the segmentations provided total AV contents statisti- Fluminense Federal University for kindly allowing the use of its micro-CT
cally equivalent to the other methods, and, different from the scanner. This study was financed by the Coordenação de Aperfeiçoa-
TH and WS, the AI procedures do not depend significantly mento de Pessoal de Nível Superior (CAPES) - Finance Code 001, by
on the user skills. Considering all aspects, the ML methodology the Conselho Nacional de Desenvolvimento Científico e Tecnológico
(CNPq), and by the Fundação Carlos Chagas Filho de Amparo à Pesquisa
was the most optimised for segmentation. do Estado do Rio de Janeiro (FAPERJ).

Disclosure statement
5. Summary and conclusions
No potential conflict of interest was reported by the author(s).
This paper presented the application of four different segmen-
tation techniques, i.e. threshold (TH), watershed (WS),
machine learning (ML), and deep learning (DL), to determine Funding
the air void (AV) content of an asphalt concrete (AC) sample This study was financed in part by the Coordenação de Aperfeiçoamento
using images acquired in a micro-CT scanner system. The key de Pessoal de Nível Superior - Brasil (CAPES) - Código de Financiamento
findings are summarised as follows: 001, by the Conselho Nacional de Desenvolvimento Científico e Tecnoló-
gico (CNPq), and by the Carlos Chagas Filho Foundation for Research
.
Support of the State of Rio de Janeiro (FAPERJ). The authors wish to
All segmentation techniques have demonstrated applica-
express their gratitude to the Engineering School of Fluminense Federal
bility to the two-dimensional cross-section images of AC University for kindly allowing the use of its micro-CT scanner.
obtained in the micro-CT tests. This means that the artifi-
cial intelligence (AI) techniques, i.e. ML and DL, were suc-
cessfully applied as AC image segmentation tools. ORCID
. The total AV contents obtained using ML and DL were Alexis Jair Enríquez-León http://orcid.org/0000-0002-4657-7900
similar to those obtained with the traditional TH and WS Thiago Delgado de Souza http://orcid.org/0000-0003-3904-2382
approaches. Furthermore, the segmentations were easily Francisco Thiago Sacramento Aragão http://orcid.org/0000-0003-4957-
fine-tuned with the AI procedures. 9474
Delson Braz http://orcid.org/0000-0002-5319-1627
. ML and DL took longer time for data processing, but the André Maués Brabo Pereira http://orcid.org/0000-0003-0991-8515
further segmentation analyses of other AC samples will Liebert Parreiras Nogueira http://orcid.org/0000-0002-7767-464X
just need final fine-tunings and will allow the even faster
identification of the AVs.
. The segmentation by ML and DL also offered fewer user References
interventions, indicating that they can be more reproduci- AASHTO, 2016. T 166-16 standard method of test for bulk specific gravity
ble and less affected by the user skills. (Gmb) of compacted hot mix asphalt (HMA) using saturated surface-dry
. No statistical difference was identified among both the total specimens. Washington, DC: American Association of State Highway
and the permeable AV contents determined by the four and Transportation Officials.
AASHTO, 2017. M 323-17 standard specification for superpave volumetric
different segmentation techniques evaluated. mix design. Washington, DC: American Association of State Highway
. The impermeable AV contents were statistically different and Transportation Officials.
and additional studies are necessary to identify the most Arganda-Carreras, I., et al., 2017. Trainable Weka segmentation: a
adequate segmentation technique for the characterisation machine learning tool for microscopy pixel classification.
of this parameter. Bioinformatics (Oxford, England), 33 (15), 2424–2426.
PMID:28369169. Available from: https://doi.org/10.1093/
. The paper provided meaningful information on different bioinformatics/btx180.
procedures that can be adopted to identify the permeable ASTM, 2019. D 2041-19 standard test method for theoretical maximum
and impermeable AV of AC samples, which can be poten- specific gravity and density of bituminous paving mixtures. West
tially further improved using radiocontrast agents and Conshohocken, PA: American Society for Testing and Materials.
other spatial resolutions for the micro-CT images. Augusto, K.S. and Paciornik, S., 2018. Porosity characterization of iron
ore pellets by X-ray microtomography. Materials Research, 21 (2),
. Considering the different aspects evaluated, i.e. time, accu- e20170621. Available from: https://doi.org/10.1590/1980-5373-mr-
racy, and user-friendliness, the ML technique was the most 2017-0621.
optimised for segmentation among the four methodologies Bezerra, E., Augusto, K.S., and Paciornik, S., 2020. Discrimination of
evaluated. pores and cracks in iron ore pellets using deep learning neural net-
. In the future, manual and operator-dependent segmenta- works. Metals and Materials, 73 (3), 107–203. Available from: http://
dx.doi.org/10.1590/0370-44672019730119.
tion processes, such as the TH technique, will be poten- Buades, A., Coll, B., and Morel, J.-M., 2011. Non-local means denoising.
tially replaced by more objective procedures, such as ML Image Processing On Line Journal, 1, 208–212. Available from: https://
or DL, for which the results are far less dependent on doi.org/10.5201/ipol.2011.bcm_nlm.
the judgement of the operator. Therefore, trained segmen- Chen, F.F., et al., 2016. Clustering analysis for porous media: an appli-
tation models for AC may become powerful tools to help cation to a dolomitic limestone. Journal of Petroleum Science
Engineering, 146, 770–776. Available from: https://doi.org/10.1016/j.
the pavement engineering community on the improve- petrol.2016.07.031.
ment of the quality control and quality assurance Chung, T., et al., 2019. Voxels agglomeration for fast estimation of per-
strategies. meability of micro-CT images. Journal of Petroleum Science
10 A. J. ENRÍQUEZ-LEÓN ET AL.

Engineering, 106577. Available from: https://doi.org/10.1016/j.petrol. effects on the rheological behavior of the material. Transportation
2019.106577. Research Record, 2674 (5), 97–107. Available from: https://doi.org/10.
Cnudde, V., et al., 2009. Porosity and microstructure characterization of 1177/0361198120914607.
building stones and concretes. Engineering Geology, 103 (3-4), 76–83. Pereira, A., et al., 2018. Estimating the effective elastic parameters of nod-
Available from: https://doi.org/10.1016/j.enggeo.2008.06.014. ular cast iron from micro-tomographic imaging and multiscale finite
CTAnalyzer, 2020. v. 1.20.3 [Computer software]. Kontich: Bruker. elements: comparison between numerical and experimental results.
Available from: https://www.bruker.com/products/microtomography/ Metals, 8 (9), 695. Available from: https://doi.org/10.3390/met8090695.
micro-ct-software/3dsuite.html. Provencher, B., Piché, N., and Marsh, M., 2019. Simplifying and stream-
Dragonfly, 2020. 2020.1 [Computer software]. Montreal: Object lining large-scale materials image processing with wizard-driven and
Research Systems (ORS). Available from: http://www.theobjects. scalable deep learning. Microscopy and Microanalysis, 25 (S2), 402–
com/dragonfly. 403. Available from: https://doi.org/10.1017/S1431927619002745.
Enríquez-León, A., et al., 2020. ds-uct-007: Asphalt concrete: X-ray micro- Sadeq, M.A., et al., 2018. Characterization of air voids in W-FAM samples
CT of an asphalt concrete sample [Data set]. Zenodo. Available from: using X-ray CT imaging. Advances in Materials and Pavement
http://doi.org/10.5281/zenodo.3958325. Performance Prediction (AM3P), Conference, Doha, Qatar. Available
Gonzalez, R.C. and Woods, R.E., 2018. Digital image processing. 4th ed. from: https://doi.org/10.1201/9780429457791-18.
New York: Pearson. Schindelin, J., et al., 2012. Fiji: an open-source platform for biological-
Landis, E.N. and Keane, D.T., 2010. X-ray microtomography. Materials image analysis. Nature Methods, 9 (7), 676–682. Available from:
Characterization, 61 (12), 1305–1316. https://doi.org/10.1016/j. https://doi.org/10.1038/nmeth.2019.
matchar.2010.09.012. Shaheen, M., 2015. Evaluation of X-ray computed tomography and finite
Legland, D., Arganda-Carreras, I., and Andrey, P., 2016. Morpholibj: inte- element models for fatigue experimental hot mix asphalt characteriz-
grated library and plugins for mathematical morphology with ImageJ. ation. Thesis (PhD). University of Waterloo, ON, Canada.
Bioinformatics (Oxford, England), 32 (22), 3532–3534. Available from: Tashman, L., et al., 2002. X-ray tomography to characterize air void dis-
https://doi.org/10.1093/bioinformatics/btw413. tribution in superpave gyratory compacted specimens. International
Masad, E., et al., 1999a. Internal structure characterization of asphalt con- Journal of Pavement Engineering, 3 (1), 19–28. Available from:
crete using image analysis. Journal of Computing in Civil Engineering, https://doi.org/10.1080/10298430290029902a.
13 (2), 88–95. Available from: https://doi.org/10.1061/(ASCE)0887- Thyagarajan, S., et al., 2010. The heterogeneity and mechanical response
3801(1999)13:2(88). of hot mix asphalt laboratory specimens. International Journal of
Masad, E., et al., 1999b. Quantifying laboratory compaction effects on the Pavement Engineering, 11 (2), 107–121. Available from: https://doi.
internal structure of asphalt concrete. Transportation Research Record: org/10.1080/10298430902730521.
Journal of the Transportation Research Board, 1681 (1), 179–185. Wang, Z. and Xiao, J., 2012. Evaluation of air void distributions of cement
Available from: https://doi.org/10.3141/1681-21. asphalt emulsion mixes using an X-ray computed tomography scan-
Masad, E., et al., 2002. Characterization of air void distribution in asphalt ner. Journal of Testing and Evaluation, 40 (2), 273–280. Available
mixes using X-ray computed tomography. Journal of Materials in Civil from: https://doi.org/10.1520/JTE103881.
Engineering, 14 (2), 122–129. Available from: https://doi.org/10.1061/ You, Z., Adhikari, S., and Emin Kutay, M., 2009. Dynamic modulus simu-
(ASCE)0899-1561(2002)14:2(122). lation of the asphalt concrete using the X-ray computed tomography
Masad, E., Castelblanco, A., and Birgisson, B., 2006. Effects of air void size images. Materials and Structures, 42 (5), 617–630. Available from:
distribution, pore pressure, and bond energy on moisture damage. https://doi.org/10.1617/s11527-008-9408-4.
Journal of Testing and Evaluation, 34 (1), 15–23. Available from: Zelelew, H.M., et al., 2013. An improved image processing technique for
https://doi.org/10.1520/JTE13112. asphalt concrete X-ray CT image. Road Materials and Pavement
Osmari, P.H., et al., 2020. Determination of volumetric characteristics of Design, 14 (2), 341–359. Available from: https://doi.org/10.1080/
FAM mixtures using X-ray micro-computed tomography and their 14680629.2013.794370.