Quantification of Aggregate Packing in Asphalt Mixtures using 3D Image Processing and Artificial Neural Networks M. Emin Kutay, Ph.D., P.E. Assistant Professor Department of Civil & Environmental Engineering
Quantification of Aggregate Packing in Asphalt Mixtures using 3D Image Processing and Artificial Neural
Networks
M. Emin Kutay, Ph.D., P.E.Assistant Professor
Department of Civil & Environmental Engineering
Outline
IntroductionAggregate packing in HMA
3D X‐ray CT imaging and analysis methods Challenges in processing of AC images
Use of ANN‐based pattern recognition tool to process 3D X‐ray CT images
0.075 1.18 4.75 9.5 19 0
20
40
60
80
100
Sieve Size (mm)
Pe
rce
nt P
ass
ing
(%
)
Aggregate characteristics in Hot Mix Asphalt (HMA) pavements
Fine graded HMA
Coarse graded HMA
Stone Matrix Asphalt
Study of Aggregate Packing: Top‐down vs. Bottom‐up
(Lab) (Lab) SPECIMENSPECIMEN
Individual Individual aggregateaggregate3D shapes3D shapes
XX--ray CTray CTImage Image ProcessingProcessing
OUTPUT:OUTPUT:Contact points, Contact points, Orientation, Orientation, Spatial distributionSpatial distribution
(Digital) (Digital) SPECIMENSPECIMEN
Individual Individual aggregateaggregate3D shapes3D shapes
TopTop--downdown(Experimentation)(Experimentation)
BottomBottom--upup(Simulation)(Simulation)
NumericalNumericalSimulationSimulation
Study of Aggregate Packing: Lab + Image analysis
Varying aggregate type, Varying aggregate type, compaction characteristicscompaction characteristics
Specimen Prep.Specimen Prep.(Lab compacted / (Lab compacted / field core)field core)
Measurement of 3D Measurement of 3D internal geometryinternal geometry
XX--ray CT imagingray CT imaging
Rigorous image Rigorous image processing & analysisprocessing & analysis
Separation of aggregatesSeparation of aggregates
Individual aggregate props:Individual aggregate props:3D orientation, angularity, 3D orientation, angularity, Specific surface area Specific surface area ……etc.etc.
Calculation of contact Calculation of contact pointspoints
TopTop--downdown
Study of Aggregate Packing: Numerical Simulation
Varying aggregate type, Varying aggregate type, CompactionCompactioncharacteristicscharacteristics
Digital specimen Digital specimen Preparation Preparation
(numerical simulation)(numerical simulation)
Measurement of 3D Measurement of 3D Aggregate shapesAggregate shapes
XX--ray CT imaging, ray CT imaging, LADAR, projection moire LADAR, projection moire interferometryinterferometry
BottomBottom--upup
Individual aggregate packing Individual aggregate packing props: 3D orientation, props: 3D orientation,
spatial distribution, contact spatial distribution, contact points points ……etc.etc.
Study of Aggregate Packing: Top‐down vs. Bottom‐up
(Lab) (Lab) SPECIMENSPECIMEN
Individual Individual aggregateaggregate3D shapes3D shapes
XX--ray CTray CTImage Image ProcessingProcessing
OUTPUT:OUTPUT:Contact points, Contact points, Orientation, Orientation, Spatial distributionSpatial distribution
(Digital) (Digital) SPECIMENSPECIMEN
Individual Individual aggregateaggregate3D shapes3D shapes
TopTop--downdown(Experimentation)(Experimentation)
BottomBottom--upup(Simulation)(Simulation)
NumericalNumericalSimulationSimulation
XX‐‐ray Computed Tomography setupray Computed Tomography setup
X-Ray Source
Detector
Specimen
AlgorithmsCT# (in)
CT# (Out)
Vertical shift
3D reconstruction from image slicesX-ray CT
Characteristics of aggregates packed in an asphalt mixture
3D Contact points (or influence zone)
3D OrientationSegregation: Spatial distribution of
different sizes3D Angularity, sphericity,
specific surface area and texture
We want to study: Given : gradation and compaction
level Variation of contact points for
different aggregates: rectangular, flat , elongated and round
Locking point and aggregate degradation
Effect of number of contact pointson HMA performance
Mix design considering contact points and packing
X‐ray CT Image 3D Analysis Tool
Challenge in processing of X-ray CT images: separation of aggregates
IdealIdeal RealReal
Thresholding and labeling (2D description)
Intensity Distribution
181
0 255
Grayscale image Thresholded binary image
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Thresholding and labeling(2D description)
Binary image Labeled image
Separation of aggregates:Separation of aggregates:3D image processing steps3D image processing steps
Calculation of aggregate properties
3D Contact points 3D Orientation 3D Volume, Angularity, Specific surface area
Segregation: Spatial distribution of different sizes
Calculation of contact points
Shortest distance between surface voxels
Validation of the algorithms
Orig
inal
gra
datio
n (%
)
Sieve Size (mm)
0
20
40
60
80
100
0 0.3 1.18 2.36 4.75 9.5 12.5
Original gradationImage-based gradation
Internal structure change with compaction – (e.g., locking point)
NdesignN1 N2 N4 N5 Gyrations (N)N6
Height (mm)
Good Poor
Good Poor
Image A
Rigorous Image Processing:Image noise filters (Gaussian, median …etc) H-max regional maxima filterWatershed transformThresholding
Good binary image
Image B Poor binary image
Use of ANN to recognize aggregates in X‐ray CT images Trained to recognize coarse aggregates (>4mm) ANN Architecture was similar to those used in the
field of face detection
Preparation of Training Dataset for the ANNManually crop and save boxes from the X‐ray CT images
Two types of training datasets were prepared: Aggregates: boxes encompassing only one whole coarse aggregate (the box may also include other smaller size material)
Non‐aggregates: random voxels which does not include whole aggregate
p gANN
X-ray CT images of 11 different specimens were used to obtain: 20 aggregate and 20 non-aggregate input
boxes for each specimen A full dataset of 440 boxes.
Of the 11 specimens: 7 of them were utilized in training the
ANN (i.e., total 7 × 40 = 280 input boxes) and
4 of them were used for testing the ANN (i.e., total 4 × 40 = 160 input boxes).
Preparation of Training Dataset for the ANN (Cont'd) Wide variety of aggregate sizes and shapes
A standard input dimension was needed Resizing (3D interpolation) → 20×20×20 voxels Reshaping → 8000×1 voxels
Structure of the ANN A feed-forward (backpropagation) network
1 hidden layer of 160 neurons Output layer had 1 neuron
H160x8000W
H160x1b
1x8000p
H160x1n
o1x160W
ob
on
H160x1a
Hidden layer Output layerInput
y
tansig tansig
)(tansig HHH bpWa )(tansig oHo by aW
Output
Training ANN
0 5 10 15 20 25 30 3510
-2
10-1
100
101
Epochs
Mea
n S
quar
e E
rror
(M
SE
)
Training
Validation
performance goal
H160x8000W
H160x1b
1x8000p
H160x1n
o1x160W
ob
on
H160x1a
Hidden layer Output layerInput
y
tansig tansig
)(tansig HHH bpWa )(tansig oHo by aW
Output
Performance of the ANN
ANN was tested using the input dataset not utilized in the training
After the ANN was trained: Weights and bias vectors were used to
calculate the output scalars (y) of the testing images
The output was compared with the known target values (yt)
Performance of the ANN
-1 0 1-1
-0.5
0
0.5
1
Kno
wn
y t
ANN simulated y
Specimen: acl2
Correct = 100%
-1 0 1-1
-0.5
0
0.5
1
ANN simulated y
Specimen: bcm1
Correct = 100%
-1 0 1-1
-0.5
0
0.5
1
ANN simulated y
Specimen: 9.5C25
Correct = 95%
Incorrect = 5%
incorrect
-1 0 1-1
-0.5
0
0.5
1
ANN simulated y
Specimen: 19F50
Correct = 95.24%
Incorrect = 4.76%
incorrect
Ongoing Work on ANN Search within an X-ray CT
image of coarse aggregates with different sizes Get spatial distribution of
different sizes (segregation)
A further ANN to get the boundaries of the aggregate within each box
M. Emin Kutay, PhD, PE Michigan State University
Department of Civil & Environmental Engineering
3D numerical simulation of compaction3D numerical simulation of compactionusing Dissipative Particle Dynamics (DPD)using Dissipative Particle Dynamics (DPD)
Animation -1
Animation -2Animation-3
Original Gaussian smoothed Hmax filtered
Watershed transformed image
Watershed transformed image
Watershed transformed image
Not good Not good Good
Original Gaussian smoothed Hmax filtered
Watershed transformed imageInverted Watershed lines
Filtered watershed transformation
Thresholding IsuesThresholding Isues……Different intensities
Dynamic thresholding in 2DDynamic thresholding in 2D
0 100 200 3000
100
200
300
400
500Original image Intensity histogram
peak
# of
pix
els
Grayscale intensity
Dynamic thresholding in 3DDynamic thresholding in 3D
0 100 200 3000
100
200
300
400
500
Intensity histogram
peak
# of
pix
els
Grayscale intensity