Application of image analysis and CAD techniques for detection and modeling of wood features from NDT imaging data

1. Application of image analysis and CAD techniques
for detection and modeling of wood features from
NDT imaging data
B. Sim˜oes1, M. Riggio2
, R. De Amicis1
1Graphitech Foundation
2Dept. of Mechanical and Structural Engineering, University of Trento
September 15, 2011

2. Overview
Agenda
Introduction & Problem definition
Acquisition Techniques
X-ray Computed Tomography (CT)
Close Range Photography (CRP)
3D Surface Reconstruction
Conclusions

3. Introduction
Question:
For what can we use wood structures ?

4. Fire and Weapons
Image courtesy of Peter Hesse

5. Shipbuilding
Travelling & Transportation

6. Wood Bridges
This Victorian Wood Bridge cost only 1300 EUR

7. Archaeological artifacts
Used to express art, social and religious concepts

8. Architecture
All photographs by Cameron Neilson, courtesy of Robert Oshatz

9. Problem definition
Why is nondestructive testing and evaluation of wood important ?
NDT and NDE can provide important information for:
Artistic and historical analysis (e.g. to determine materials,
structures and technologies)
Security (e.g. safe materials for houses, ships, etc.)
Prevention (e.g. fragile state of some public trees), etc.

10. Problem definition
Why is nondestructive testing and evaluation of wood important ?
NDT and NDE can provide important information for:
Artistic and historical analysis (e.g. to determine materials,
structures and technologies)
Security (e.g. safe materials for houses, ships, etc.)
Prevention (e.g. fragile state of some public trees), etc.

11. Problem definition
Why is nondestructive testing and evaluation of wood important ?
NDT and NDE can provide important information for:
Artistic and historical analysis (e.g. to determine materials,
structures and technologies)
Security (e.g. safe materials for houses, ships, etc.)
Prevention (e.g. fragile state of some public trees), etc.

12. Why do we need a geometrical representation ?
Answer:
We want to use Finite Element Analysis (FEA)

13. Problem definition
Why feature modelling isn’t a trivial problem ?
It depends on both the technology involved (e.g. device,
resolution, etc)
raises mobility and price questions
implies different data acquisition pipelines

14. Problem definition
Why feature modelling isn’t a trivial problem ?
Processing techniques used
Affected by user perception
Linked to different geometrical models, probably having
Poor mathematical descriptions (e.g. with limited LOD)
Bad triangles (e.g. large angles decrease the accuracy of the
FEM)
Unsuitable formats for transfer between various FEA

15. Problem definition
Why feature modelling isn’t a trivial problem ?
Processing techniques used
Affected by user perception
Linked to different geometrical models, probably having
Poor mathematical descriptions (e.g. with limited LOD)
Bad triangles (e.g. large angles decrease the accuracy of the
FEM)
Unsuitable formats for transfer between various FEA

16. Problem definition - Conclusion
What are we looking for ?
Automatic procedure for geometrical 3D reconstruction of
features from
CT images (longitudinal-tangential (LT) anatomical direction)
CRP images
Characterization of different types of geometry
Nodes and intra-ring layers, i.e. the interface between
earlywood and latewood of the same year
Confrontation of several mathematical representations
B-spline curves, Composed Bezier curves, Spiro spline, Almost
Vanishing Polynomials
Morphological FEM model

17. Problem definition - Conclusion
What are we looking for ?
Automatic procedure for geometrical 3D reconstruction of
features from
CT images (longitudinal-tangential (LT) anatomical direction)
CRP images
Characterization of different types of geometry
Nodes and intra-ring layers, i.e. the interface between
earlywood and latewood of the same year
Confrontation of several mathematical representations
B-spline curves, Composed Bezier curves, Spiro spline, Almost
Vanishing Polynomials
Morphological FEM model

18. Problem definition - Conclusion
What are we looking for ?
Automatic procedure for geometrical 3D reconstruction of
features from
CT images (longitudinal-tangential (LT) anatomical direction)
CRP images
Characterization of different types of geometry
Nodes and intra-ring layers, i.e. the interface between
earlywood and latewood of the same year
Confrontation of several mathematical representations
B-spline curves, Composed Bezier curves, Spiro spline, Almost
Vanishing Polynomials
Morphological FEM model

19. Problem definition - Conclusion
What are we looking for ?
Automatic procedure for geometrical 3D reconstruction of
features from
CT images (longitudinal-tangential (LT) anatomical direction)
CRP images
Characterization of different types of geometry
Nodes and intra-ring layers, i.e. the interface between
earlywood and latewood of the same year
Confrontation of several mathematical representations
B-spline curves, Composed Bezier curves, Spiro spline, Almost
Vanishing Polynomials
Morphological FEM model

20. Using Isopycnic maps to extract features from CT images

21. Using Isopycnic maps to extract features from CT images
Definition of contour plots
A contour plot is a way of representing three-dimensional data on
a two-dimensional surface.
Definition of Isopycnic maps
A isopycnic map is a set of lines having the same density value
Why we do use isopycnic maps ?
A: We can explore better the CT image space. Each image pixel
was a coordinate value from 0 to 4000 (no need for rescale)

22. Segmentation of CT images
Brief description of the algorithm
Step 0. Generate an isopycnic map for our image
Step 1. Remove isopycnic curves intercepting the image
border or timber boundary.
Step 2. Calculate the flow directions using tangent vectors or
medial axis technique (for closed curves).
Step 3. Remove noisy features using thresholds (for length
and area).
Step 4. Remove the set of isopycnic curves h such that
h ⊂ fconvexhole ∩ gconvexhole, and
f and g are open isopycnic curves (extremities not ending at
the same border or parallel borders)
level(f) differs from level(g)
h is unique

23. Segmentation of CT images
Brief description of the algorithm
Step 5. Repeat Step 4 with a polygon open isopycnic curves f
and g merged with partial timber boundary. Similarly, we
remove intermediary levels
Step 6. Merge similar features according to their direction,
position (avoid intersections) and distance

24. Scenario I

25. Scenario II

26. Scenario II - Final Result

27. CRP Image Segmentation

28. Close Range Photogrammetry (CRP)
Overview
Both optical scanning and photographs provide only
information about external features
We usually deal with bigger images in w.r.t. CT images
Physical damage in timber structures is visible, which affects
the image segmentation process

32. Algorithm Overview
Phase I - Edge Detection
A colour image is converted to a high contrast gray level image,
and then processed using the Canny operator (Canny 1986) to
obtain feature edges.
Phase II - Improved Edge Linking
The objective of this phase is to connect broken edges. Aggregates
edge elements into continuous contours without user interaction.

33. Phase I - Edge Detection
Step 1: Convert the image to gray-scale.
Step 2: Apply a threshold with a cut-off value, in terms of number
of α above the mean.

34. Phase I - Edge Detection
Step 1: Convert the image to gray-scale.
Step 2: Apply a threshold with a cut-off value, in terms of number
of α above the mean.

35. Phase I - Edge Detection
Step 3: Use of Gaussian filter to improve to remove high frequency
noise (smoothing).

36. Phase I - Edge Detection
Step 4: A canny edge detector is applied to the smoothed image.

37. Phase II - Improved Edge Linking
Step 1: Thinning. Each edge segment will be converted into one
pixel wide.
Step 2: End-points marking. The End-points (EPs) of a contour
are defined to be the set of edge pixels which are very likely to be
knots of some contour containing the one being considered
Step 3: Remove noisy edge segments. If the length of one edge
segment is shorter than a threshold value, then the segment is
removed. The threshold value changes dynamically at every
iteration.

38. Phase II - Improved Edge Linking
Step 4: Gap filling. Fills gaps between two distinct EPs. The
connectivity between segments is identified according to the
following properties:
Threshold distance (iterative)
Ratio between threshold distance and segment size
Angle between tangent vectors
Segment topology

39. Phase II - Improved Edge Linking
Step 5: Branch pruning. Removes noisy branches introduced
during thinning step.
We define root points as pixels having at least 3
neighbourhoods.
We prune all branches ending at some root point and having
length shorter than a threshold value

40. 3D Surface Reconstruction

41. 3D Surface Reconstruction
Procedure
Step 1. We extract feature contours using above-mentioned
algorithms
Step 2. We reduce contours to a set of dominant points (e.g.
we use simplification curve algorithms)
Step 3. We interpolate dominant points using the most
appropriated curve algorithm
Step 4. We create a 3D surface by lofting a profile through
the set of curves

42. Interpolation algorithms
Curve interpolation using a) Spiro, b) Bezier, c) Bspline and d)
Akima algorithms

43. Interpolation algorithms
Curve interpolation using Almost vanishing polynomials

44. 3D Skeleton
Skeleton required to generate Gordon Surfaces

45. Algorithm for X-ray CT images
Advantages
Detects possible knots, cracks and other features which
exhibit drastic change of density value
Explores CT images properties
The construction of a voxel-based curve skeleton is completely
automatic; yet, we can refine it using different threshold values
Disadvantages
Others algorithms can have better local precision.

46. Overall Conclusions
Both methods allow spatial modeling of the morphological
wood components, using as input data from low-cost image
acquisition devices
Spiro spline provides the most suitable algorithms to describe
closed shapes (e.g resin pockets)
Composite Bezier yields excellent results where Spiro
algorithm is presented as the worse approximation (e.g. wood
growth layers).
B-Spline provides an excellent representation for indented
layers and a good trade-off in general

47. Overall Conclusions
The Almost vanishing Polynomials (AVP) algorithm can be
used to find an implicit representation of minimal degree that
best approximates the set of points by a tolerance α.
By using AVP, features have a tendency to become more
symmetric, that is, to approximate known shapes such as
circumferences, hyperboles, etc.

48. References
Riggio M., Santini M., De Amicis R., Torrente M. (2010).
Use of X-ray tomography and CAD techniques for ”morphology-based”
wood elements models
in Research in Interactive Design - Vol. 3 Springer Verlag.
De Amicis R., Riggio M., Girardi G., Piazza M. (2011)
Morphology-based macro-scale finite-element timber models
Computer-Aided Design Journal.
Snyder W. (1978).
ALGORITHM 531, Contour Plotting [J6]
ACM Transactions on Mathematical Software.

49. The End
Thank you for your attention!