X-FAST Interne kwaliteit van poreuze levensmiddelen: X-stralentomografie toegepast op product en proces Pieter Verboven
X-ray technology • Imaging methods o o
Radiography (2D) Tomography (3D)
X-ray imaging • Features o
Absorption depending on atomic number • Good contrast of internal density differences •
Air-solid phase • Fat-water-protein
o o o
Good penetration in all foods Macro to (sub-)micron resolution in 3D Little or no sample preparation
Challenges • • • •
Quality, size and resolution of CT images Amount of data Cumbersome analysis Unknown relationship between microstructure and food properties / quality attributes • Few online techniques
Applications • • • • • •
Fruit and vegetables Bakery products Dairy food Confectionary & chocolate Meat products Packaged food
European Synchrotron Radiation Facility (ESRF)
Jonagold
Braeburn
Kanzi
Conference
Jonagold
Braeburn
Kanzi
Conference
Ho et al. (2010) Journal of Experimental Botany, 61 (10), 2745-2755
X-ray micro- and nano-CT 35,3 µm
22,6 µm
10,7 µm
5,2 µm
3,3 µm
1,3 µm
0,7 µm µm 0.8
250 µm
Foamed gel samples CT images 250 x 250 µm2
0.5 0,5µm µm
Climate controlled micro-CT • Bruker micro-CT 1172 • Cooling stage o
to -15°C +/- 0.5°C
Ice cream
Microstructure-process relationships • Sugar foam: effect of mixing time
o
2 min
4 min Relative volume percentage (%)
16 14 12 10 8 6
8 min
4 2 0 10
100
Statistical analysis
Food structure engineering • MAP verpakken van brood (Tetra-project i.s.m. Campus Geel) korst
kruim
5 mm
5 mm
• Berekende [O2] profielen kruim
korst
2.6mm 4.5mm (mol m-3 )
inline X-ray imaging (2D)
Increasing degree of granulation in oranges
Inline CT (3D)
Inline CT (3D) 1. Recording X-ray projections on conveyer system X-ray source
X-ray detectors
2. Fast reconstruct of 3D images
TomFood, KU Leuven, UGent & UA
3. Detection algorithms
Inline detection algorithms of internal disorders Detection rate
• Detect intensity shifts • Watersoaked tissues • Deydrated areas and cavities
False Correct negative (%) (%)
Apples (n=74)
96
4
False positive (%)
0
• Inline 3D shape and size measurement
Foreign materials
3D inline CT development • Translation+rotation tomography X-ray
• 3D imaging + X-ray inspection 3D
Conveyor belt mock up Source: transmission tube (opening angle 180°)
Translation axis Detector
Translation+rotation stage
Translation axis Detector
Van Hoorebeke et al. U Gent Novel techniques for inspection and engineering of food (micro)structure based on X-ray computed tomography
Inline-CT Limited amount of data in a limited amount of time
KU Leuven U Antwerpen U Gent
60°
90°
120°
Novel techniques for inspection and engineering of food (micro)structure based on X-ray computed tomography
Inline-CT Reconstruction 16 projections
128 projections
Novel techniques for inspection and engineering of food (micro)structure based on X-ray computed tomography
Advanced reconstruction Number of radiographs FBP
Limited angle
SART
FBP
100
60°
50
90°
25
120°
SART
SART init.
In collaboration with Antwerp University Novel techniques for inspection University and engineering of & foodGhent (micro)structure based on X-ray computed tomography
In-line CT Analyse/Compare: Quality versus Throughput Throughput (samples/s) 100 cm 80 cm More projections
60 cm
Larger detector
40 cm
Feature size (mm/lp) Novel techniques for inspection and engineering of food (micro)structure based on X-ray computed tomography
3D imaging + X-ray inspection Concept: • Compare the radiograph with a
•
•
The reference sample contains no defects Large deviations in the residual image indicate a defect
No need to develop individual algorithms for different defects!
simulate
•
reference sample
measure
simulated radiograph of a reference sample
sample
simulated radiograph
radiograph
compare
residual image
Simulated radiograph
Adapted from Palenstijn, W.J., Batenburg, K.J., Sijbers, J., 2013. The ASTRA tomography toolbox, in: 13th International Conference on Computational and Mathematical Methods in Science and Engineering, CMMSE 2013. pp. 1139–1145.
Line integrals through 3D volume between detector pixels and source Scan geometry should match the radiographs scanner geometry How to determine the reference sample?
simulated
measured
The reference sample •
Estimate shape and pose with 3Dvision
•
Assume a model is known
•
Shape / surface
•
•
rigid (e.g. CAD)
•
deformable (e.g. SSM)
Internal density distribution •
uniform
•
non-uniform
3D
3D scanner fit shape model radiograph scanner compare with radiograph simulated from model
Truth
98.7
scan 1
scan 2
Prediction
Potential with perfect reference
scan 3
SCAN
1
2
3
4
5
1
19.54
0
0
0
0
2
0
21.17
0
0
0
3
0
0
20.2
0
0
4
0
0
0
19.54
0.98
5
0
0
0
0.33
18.24
scan 4
scan 5
Rigid shape model Internally uniform •
•
2 shapes •
Toroid
•
Ellipsoid
5 defect shapes •
Variable intensity
1. Apply random rotation and generate point cloud to simulate 3D-vision
2. Fit model to point cloud (Iterative closest point / Procrustes)
3. Generate radiographs 4. Classify • •
Added noise Naïve bayesian classifier on sum of pixel values in residual images.
Classification (N = 5000)
Morphable bodies: Statistical shape model PC1
• •
-3 sd
Generate a set of shapes with corresponding shapes
+0 sd
+3 sd
-3 sd
Apply PCA on point coordinates Results in a mean shape and displacement vectors (principal components) for each point
PC2
•
+0 sd
+3 sd
•
i.e. a new shape is a linear combination of the mean shape and the principal components
Different scale in images: my mistake…
Estimating shape and pose •
Iterate •
Rigid registration (iterative closest point)
•
Least squares optimization of every point in model to nearest neighbour in point cloud
•
Until error is acceptable
•
Sensitive to local minima
•
Alternative:
Schneider, D.C., Eisert, P., 2009. Fitting a morphable model to pose and shape of a point cloud, in: Vision, Modeling, and Visualization Workshop 2009. Proceedings. pp. 93 – 100.
random
Non-rigid shape model Internally uniform 1. Generate random shape 2. Apply random rotation and generate point cloud to simulate 3D-vision
3. Fit model to pose and shape of point cloud
4. Generate radiographs 5. Classify • •
Added noise Naïve bayesian classifier on sum of pixel values in residual images.
fit
residual
To do • Add defects to random shapes • Generate large datasets to asses classification rates • Implement non-uniform density distribution
X-FAST doelstellingen • Product- en procesontwikkeling: ontwikkelen van een testfaciliteit voor microstructuuranalyse met X-stralentomografie o onderzoeken van de relatie tussen structuur en sensorische of functionele eigenschappen • Online niet-destructieve kwaliteitscontrole: o online meetmethode van de interne samenstelling en dimensies o detectie van defecten en vreemde voorwerpen o
Praktische aanpak • Is de microstructuur, het defect of het vreemd voorwerp • • • •
meetbaar met X-stralen? Is dit accurater/sneller/gemakkelijker/betrouwbaarder dan met een andere methode? Zijn er structurele verschillen op basis van een verschillende samenstelling of bereidingswijze? Zijn deze structuurverschillen bepalend voor de beoogde producteigenschappen? Kan dit op een kostefficiënte manier worden geïmplementeerd in de productontwikkeling of kwaliteitscontrole?
X-FAST • Werkplan 3 WP1. Microstructuur 1.1 Cases 1.2 X-stralentomografie 1.3 Analyse 1.4 Adviezen Mijlpalen WP2. Online inspectie 2.1 Cases 2.2 X-stralen inspectie 2.3 Analyse 2.4 Adviezen Mijlpalen
6
9
12
15
18
M1.1
21
24
27
30
33
36
39
42
M1.2
M2.1
45
48
M1.3
M2.2
M2.3
Toepassingen • Ontbijtkoek o
Inline defecten meten • Gaten vorm/grootte • Breuk/beschadiging in niet-doorzichtige verpakking • Samenstelling • Inpasbaarheid op lijn • Oorzaken defecten
o
Vreemde voorwerpen • Plastic
o
Optimalisatie 2D scanner
Toepassingen • Fruitpuree Vreemde voorwerpen Glas Rubber Detectie 2D/3D Inline toepassen
kiwi glas
• • • •
2D rubber
o
framboos
aardbei
Toepassingen • Fruitpuree Vreemde voorwerpen
glas
• Glas • Rubber • Detectie 2D/3D
3D rubber
o
Toepassingen • Chocolade o
Vreemde voorwerpen • Metaal • Plastic • Haren
• Productsamenstelling % vulling o % toevoegingen • Productvorm o afwijkingen o
Discussie en actiepunten
Thanks to www.TomFood.be
