Quantification of Aggregate Packing in Asphalt Mixtures ...stockman/CV/F09Docs/Readings/aggregate_Analysis_Kutay.pdfStudy of Aggregate Packing: Lab + Image analysis Varying aggregate

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

[email protected]

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