FAST INLINE INSPECTION OF APPLES BY A

NEURAL NETWORK BASED FILTERED BACKPROJECTION METHOD

Eline Janssens

Eline Janssens, Luis F. Alves Pereira, Jan de Beenhouwer, Mattias

Van Dael, Pieter Verboven, Bart Nicolaï, Jan Sijbers

iMinds – Vision Lab, University of Antwerp (CDE), Antwerp, BelgiumCentro de Informática, Universidade Federal de Pernambuco, Recife, BrasilBIOSYST-MeBios, KU Leuven, Heverlee, Belgium

11/02/2016

Contents

� X-ray inline inspection

� NN-hFBP

� Results

� Conclusions

2

X-ray Inline Inspection

- New tomographic reconstruction and analysis methods

- Inline inspection of food

Problem

� Inspection of fruit:

Requirements: Good reconstructions

Inexpensive

Fast

4

Outside Inside

Image from: Grant J., Mitcham B., Biasi B. and Chinchiolo S., Late harvest, high CO2 storage increase internal browning of Fuji apples,

California Agriculture 50 (3):26-29, 1996

Inline Inspection

Inspection lines in industry

source

Radiographic inspection

No 3D information

Relative Inexpensive

Fast

5

Inline Inspection

source

Tomographic inspection

3D information

Relative Inexpensive

Slower

Artefacts

Reconstruction

6

Inline Inspection

Proposed inspectionmethod

3D information

Relative Inexpensive

Rotation

Reconstructionsource

Challenges:

� Fast Inspection

� Good Quality

7

E.g.: FBP

Very Fast Reconstruction

Slow acquisition

Lange number of equiangular projections

16 vs 128

projections

Classic Reconstruction algorithms

E.g.: SIRT

Slow reconstruction

Fast acquisition

Small number of projections

Incorporate Prior Knowledge

Analytical Reconstruction Iterative Reconstruction

8

16 vs 128

projections

� Fast inspection � Good quality

� Fast inspection

� Good quality

Inline scanning geometry9

source source

Flatpanel DetectorVirtual Moving

Detector

Neural Network Hilbert transform based

Filtered Back Projection

NN-FBP11

Contribution

2013: Neural Network FBP

Pelt D.M., Sijbers J., Batenburg K.J.,2013a, Fast Tomographic Reconstruction from Highly Limited Data Using Artificial Neural Networks, Proceedings of 1st International Conference on Tomography of Materials and Structures (ICTMS 1), Ghent, Belgium, 109-112.

Parallel beam

Circular scanning geometry

Fan beam: hFBP

Inline scanning geometry

NN-FBP12

Output of a Neural Network: n�,�,�,�� � σ�� ��σ� �� ∙ � � b� � b������

���INPUT

OUTPUT

σ = activation function

b = bias

FILTER

WEIGHTz1

z2

z3

zN

w0

q1

q0

q2

q3

w1

w2

w3⋮

Filters and weights are trained by the neural network

Network trained to reconstruct 1 pixel

n�,�,�,�� � σ�� ��σ� �� ∙ � � b� � b������

���

NN-FBP13

Output of a Neural Network:

FBP� x, y h�τ&�P'(�xcos θ . /sin�θ� � τ&�12∈4'(∈5

q1

q0

q2

q3

w0

w1

w2

w3

z1

z2

z3

zN

⋮

NN-FBP14

n�,�,�,�� � σ�� ��σ� �� ∙ � � b� � b������

���

q1

q0

q2

q3

FBPw0

FBPw1

FBPw2

FBPw3

Output of a Neural Network:

Fan-beam15

Option 1:

q1

q0

q2

q3

FBPw0

FBPw1

FBPw2

FBPw3

Convert fan beam to parallel beam

Fan Beam

Parallel Beam

Fan-beam16

Option 2: Work directly on fan-beam data

q1

q0

q2

q3

hFBPw0

hFBPw1

hFBPw2

hFBPw3

Hilbert transform based

FBP

You J., Zeng GL., Hilbert transform based FBP algorithm for fan-beam CT full and partial scans, IEEE Trans Med Imaging, 26(2), 190-9, February 2007

Fast

Good reconstructions

Directly on Fan-beam data

Filter x Input

Results

Specifications

� Jonagored apples with defects

� Reconstruction 256 x 256

� Simulate inline projections from available circular projections

� Moving detector 287 pixels

� Data from BIOSYST-MeBios, KU Leuven, Belgium

Data Specifications

18

� Training: 100 images from 9 apples

� Validation: 10 images from 9 different apples

� Test: 50 images from 5 apples

� 100.000 random pixels for training and 10.000 for

validation

� 4 Hidden Nodes

Network Specifications

Non-Equiangular Projections19

Reconstructions20

16

projections

256 projections

64

projections

Error images21

16 projections

64 projections

FBP

FBP

SIRT

SIRT

NN-hFBP

NN-hFBP

Comparison Projections NN-hFBP

22

NN-hFBP

RMSE RMSE mask MAD FSIM

#proj EA N-EA EA N-EA EA N-EA EA N-EA

8 16.60 16.66 22.64 22.77 117.15 115.18 0.85 0.85

16 10.65 10.72 14.74 14.85 88.80 87.87 0.88 0.89

32 6.11 6.22 8.54 8.69 60.27 61.08 0.93 0.93

64 4.97 4.99 6.85 6.88 47.28 46.95 0.94 0.94

128 2.96 2.97 3.85 3.85 14.05 14.38 0.97 0.97

256 2.41 2.41 2.93 2.92 3.44 3.72 0.98 0.98

512 2.02 2.11 2.23 2.34 1.56 3.70 0.99 0.99

First Indication: Similar quality achieved for both equiangular and non-equiangular projections

Reconstruction time23

Conclusions

Conclusions

� Conveyor belt tomography with a fixed source-detector

system and an additional object rotation is feasible.

� NN-hFBP outperforms both FBP and SIRT in terms of

image quality.

� The reconstruction time of NN-hFBP is an order of magnitude smaller compared to SIRT.

25

Agency for Innovation by Science and Technology in Flanders (IWT SBO

TomFood)

Thank you for your attention

26