Individual Yarn Fibre Extraction from Micro CT: Multilevel Machine Learning

Approach

Petr Henyša , Lukáš Čapeka

a Technical University of Liberec - Textile Faculty, Department of Technologies and Structures, Studentská 1402/2, 461 17 Liberec, Czech Republic

Abstract
The internal structure and mechanics of the fibre materials, such as yarn or woven textile, are highly complex. Ex-
ploring the fibre structure is an essential step in material engineering either from the experimental or computational
point of view. In this study, a new method to extract geometrical and morphological parameters of fibre structures
is proposed. The method benefits from standard image analysis and machine learning technique to efficiently extract
fibre segments from microcomputer tomography data. The proposed algorithm is tested on the yarn and woven textile
materials with different resolution and quality. The developed method can extract the individual fibres with varying
accuracy from 73–100% with processing time 2–5s on the tested samples.
Keywords: fibre extraction, multilevel tracing, micro CT, fibre morphology, Gaussian Mixture, computer model

1. Introduction

Extracting the geometrical and topological features of complex fibrous structures, e.g. yarn, from micro-structural
tomographic data is challenging. Hence, understanding the internal structure of yarn is a crucial step for determin-
ing how macro-mechanical parameters are influenced by inter-fibre mechanics and other physical effects. Using an
adequately defined micro-structural model of yarn, a computational model can be established to find out the yarn
micro-mechanics concerning full-resolution models [1, 2] or more efficient multi-scale models [3, 4, 5]. Nevertheless,
prior to any computational modelling, an accurate geometry representation of the yarn structure must be provided. A
few studies have focused on the tracing the individual fibres. Geisselmann et al. used a combination of the distance
transform and skeletonisation methods to separate fibres individually. However, owing to noise and irregularity in
the three-dimensional (3D) data, the fibres might not be continuous and the skeletonisation algorithm may generate
the so-called ‘H’ connectivity (Figure 1) [6, 7]. The challenging step is to properly connect the disconnected part to
obtain the original fibres. Hence, a suitable criterion must be met to connect the appropriate segments. Geiselmann et
al. proposed the use of stochastic models wherein the angle and length of the interconnecting segment were utilised
to control the actual connection probability [6, 7]. Hu et al. used a multilevel tracing algorithm that simplified the
decision criteria of pair selection [8]. Another method, which enjoys the benefits of popular geometrical structures
(e.g. cylindrical structures), employed the Hessian eigenvalues to extract fibres [9]. Jerome et al. proposed another
skeleton-based method suitable for low-density materials [10]. Viguié et al. used a local orientation map combined
with dilatation operations to extract fibres and their contacts [11]. Sencu et al proposed an extraction algorithm, which
used a combination of standard image methods to filter raw CT data and applied an enhanced Bayesian filter and
Bayeisan inference to track fibres in unidirectional composite [12]. Another methods are based on searching fibre
centres according to a certain template with suitable correlation metric [13, 14, 15]. Although fibre-like structures
are specific structures from the perspective of segmentation algorithms, geometrical extraction can be significantly
simplified by reusing several specific properties provided by a certain fibre model. The simplest extraction can be
observed in long and almost homogeneously oriented fibres. This is the case of a wide range of composite fibres

∗ Correspondingauthor
Email address: [email protected] (Lukáš Čapek)

Preprint submitted to Engrxiv December 29, 2020

 Figure 1: H-type connection occurring during skeletonisation of a noisy structure.

and algorithms used the regularity and straight fibre assumption were proposed (a more enhanced approach has been
presented in previous research [16, 17]). The straight fibres allow us to extract centres from individual computer to-
mography (CT) stack cross sections by employing regular Hough-like methods. However, even in these simple cases,
the computational cost can be minimised and more effective methods should be explored. Most algorithms use at
least the basic morphological methods to extract useful information from 3D images. A very effective method can be
observed when combining the distance map and skeletonisation algorithms, which are used to extract the fibre centres
[7]. Enhanced methods, e.g. Hessian eigenvalues-based method, can be used for further extraction of structure-based
information [9]. In recent years, deep neural network variants have been designed to extract fibres from CT [18, 19].
Nevertheless, deep learning methods require effective training, which is difficult to achieve for high dimensional CT
data in a reasonable time. Herein, we propose a new method based on an efficient combination of well established
3D image analysis and a machine learning technique to efficiently extract individual fibres from micro-CT data. Our
method is able to:

• handle spatially complex fibres with many contacts (not possible or untested with methods in [16, 17, 12])
• extract fibres in a reasonable time (compared to [7, 6]),
• use more robust pairing criteria (than used in [8]), and
• extract fibres without time consuming training (compared to deep learning models in [18, 19]).

Proposed method is a fusion of methods introduced in [6, 8], but with enhanced speed and robustness thanks to
effective spatial partitioning of space and machine learning method to check the fibre pairing validity.

2. Materials and Methods

The overall working-flow of the image analysis on the CT data is shown in Figure 2. The analysis is composed
of optimal de-noising of the raw data CT up to final step consisting of re-sampling and proper tracking of the fibre
segments. The steps are discussed in detail in the following subsections.

Figure 2: The work-flow of extraction of fibres from microCT data.

2
2.1. Tested Textile Materials
The tested materials represent planar textile named TEX-2DW and yarn structure named TEX-Y. The planar
textile is made of polyamide multifilament yarns with fineness 167 [dtex], yarn twist 420 [1/m], and warp/weft sets
16/22 [1/cm]. The second structure of yarn is made of polyester with a fibre diameter of ∼12µm, yarn diameter of
∼131µm and fineness of 25 [tex]. Both structures are shown in Figure 3. The series of images are represented by an

Figure 3: Electron microscopy slice of woven structure TEX-2DW with region of interest(yellow window) used for 3D scanning (a). A slice cut
of yarn TEX-Y with visible individual fibre cross-sections made by light microscopy (b).

array with orthogonal indices, where each member represents a grey value in range 0–255. An example of raw grey
structure slices is shown in Figure 4.

Figure 4: Gray slice (z axis) example of TEX-2DW (a) and TEX-Y (b).

2.2. Scanner Setting

The micro CT scanner Rigaku nano3DX was used with tube voltage in range from 20 to 50 kV and current up to
30 mA. The dimension of the TEX-Y is [1427 × 1442 × 1250] with voxel resolution [0.528 × 0.528 × 0.528] µm. The
TEX-2DW has dimension [204 × 1201 × 551] with voxel resolution [1 × 1 × 1] µm.

2.3. Image Preprocessing

The resultant CT stacks were converted to 3D array with metadata regarding to resolution and origins. The noise
was removed by median filter with spherical structure element. Further, the binarisation based on the Otsu’s method
was performed. Once the binarised images were obtained, the Euclidean distance map was computed. It benefits from
the circular cross-section of individual fibres. At fibre centre, a distance is minimal with respect to the background
and peak occurs. Optimally, the peak has the half value of fibre diameter, but due to a noise, the peak filter works in

3
range 4.7–6.3 micro meters for TEX-Y and 4.5–7 for the woven structure TEX-2DW. The fibre diameter is known
either from manufacturer (see subsection 2.1) or can be estimated from distance map histogram. Moreover, the
crossing fibres can be separated by the distance peaks and thus no additional step is required (Figure 5). The new
binarised data is obtained from filter the distance peaks defined by range above. Due to small gaps in the peaks, the
morphological dilatation with small structuring element was used to close such gaps. The final step is to apply the
skeletonisation on binary array and get one-pixel centres of fibre parts. In order to remove the ‘H’ like connections
and another crossings, the pixel were tested for the neighbourhoods by 26-connectivity and those with more than two
neighbourhoods were removed. Finally, the disconnected parts were labelled and sorted ascendantly according to
fibre segment length.

Figure 5: A regular skeletonisation issue (middle) with two fibres (left). The distance map disconnecting two fibre segments

2.4. Segment Vectorisation

For further work with disconnected discrete parts of the fibres, the transformation into continuous representation
is useful. Let call the disconnected parts of fibres as ‘segments’. Each segment represented a structure containing
the unsorted number of pixels, but with known ends, is approximated by the b-spline curve (Figure 6). Once two
candidates for the merging are found, the b-spline curve is also used to connect them by approximating the missing
coordinates between the candidate segments. The spline approximation quality was carefully investigated in depen-
dence on the smoothing factor s (for details, see the Scipy library manual [20]). The smoothing factor influences a
compromise between the accuracy of interpolation and smoothness. The smoothing parameter in the range 0.01-0.05
was usually given good b-spline approximations. The spline is of cubic order and also the first derivative is used to get
a segment direction. The more details of how the spline approximation is constructed can be found in Scipy library
documentation [20].

Figure 6: Fibre segment with cubic spline approximation (a). Merging of two segments by spline curve with smooth transition (b).

2.5. Fibre Reconstruction

Each segment contains two end-points (start/end). The initial connection of two segments is defined by a straight
line, which is called intermediate segment. Segments can be connected only by their end-points. A potential pair of
two segments that belong to a fibre is searched based on geometrical relations inspired by work of Geiselmann [7, 6].
Consider two segments i and j in a pair. We can define Euclidean length li j of intermediate segment and angles αi j , α ji
(see Figure 7). Introducing the weighting factor wr and maximal values lmax , αmax , we can write down the criterion
function as weighted sum of normalized partial members:
αi j + α ji
li j
ci j = (1 − wr ) + wr
<1 (1)
lmax αmax
Every pair of segments satisfying the above criteria is taken into account as a potential candidate. The final pair
is selected according the smallest value of ci j and should also pass the back checking defined later. Increasing the
maximal distance between fibre ends leads to decreasing of the maximal angles, so there is a clear geometrical relation
[7]. But we found that separating both criterion (distance, angles) can provide more flexibility in some cases.
4
Figure 7: Geometrical quantities defined between end-points of segments. The red line represents an intermediate segment. Dashed line represents
true fibre trajectory.

2.6. Back Checking

Once the two segments are merged, the newly constructed segment should be evaluated whether it corresponds
with original CT data. The spline curve of the intermediate segment is discretised onto the original CT data space
and dilated with small structuring element sphere. The mean of grey values in CT data corresponding to newly
occurred intermediate segment (see Figure 6) was evaluated with help of clustering algorithm [21]. Firstly, we let a
Gaussian clustering algorithm detects two density clusters in CT data. The clusters separated the CT data on grey
density belongs to fibre structure and background noise. Secondly, we computed a probability that a newly occurred
intermediate segments belongs to a one of the cluster (see the details in clustering algorithm implementation in [21]).
We found a that a probability level 85 % was optimal for tested examples in this study.

2.7. Algorithms and Implementation Details

The image algorithms work usually with large data collections and in 3D it is even more demanding issue. In
the proposed algorithms in Figure 8 main steps required for the successfully extraction of fibres are introduced. The
input required by algorithm composed of distance map field and raw CT data. Both arrays are encoded in hdf5 format
[22] with uint8/float64 data type. The vectorisation is one of the most computationally intensive task, as it requires to
compute filtering (if required), binarisation and skeletonisation. The initial control parameters are set by the user and
are defined by following python dictionary:

• ”lmax ”: [...,...,...] a list of maximal distance parameters

• ”αmax ”: [...,...,...] a list of alpha angles
• ”wr ”: [...,...,...] a list of weighting constants
From a programming language point of view, a segment is represented by the object allowing recursive merging
of segments and respecting the parent-child relation. As the segment connectivity is defined by spatial coordinates
of the segment end-points, neighbourhood lookup (Kd tree [20]) is constructed before main loop begins in order to
significantly speed up searching of segment pairs. At current point, the pairing of segments runs on the two queues
named as finished and unfinished. Once the pair of two matching segments is found and matched back-checking
criteria, they are merged and partial segments are drop out from queue unfinished and the new segment is pushed
to queue named finished. If necessary the matching criteria can be adapted during each level. The main code is
developed in python language with critical parts (graph construction, searching, merging and b-spline interpolation)
written in C++. The code is written in objective manner.

5
 Figure 8: Algorithm scheme of fibre segmentation. The unfinished, finished, unpaired are queries.

3. Results & Discussion

The model parameters lmax , αmax , wr are experimentally estimated based on the parametric approach and careful
investigating the initial and level-based potentially matched pair of segments. The mean values of grey data for both
models were estimated by Gaussian mixed clustering (see the Figure 9). The clustering algorithm is used to decide
whether the newly created segment corresponds to grey values. The pairs which do not pass the quality criteria are
excluded. It must be pointed out that proposed algorithm could be more improved in control parameter evaluation.
The current model relies on the simple deterministic criteria, which need usually a user initial set-up and due to
multilevel approach can be adapted during the level iteration. Nevertheless, the unique decision on the best suited
segment pair is not always detectable (especially in case of noisy CT data) and mismatches occur. Although the
mismatches are detected by evaluation of the newly created segment which leads to delete the newly created segment,
still some mismatched segments propagate up to end of level iteration or some potentially connectible segments with
long proximity are not correctly detected. Using more advanced criteria can help to improve the decision sensitivity
and should be supported by probabilistic approach either proposed in [7] or some another approach such machine
learning classifier or Fuzzy logic [23]. The optimal model parameters are defined in the Table 1. To decide whether
to start new level, the algorithm checks the number of new pairs for the previous set of control parameters and if
it is non-zero, it continues to another level with the same parameters or if the number of new pairs is zero, the
algorithm increases the control parameters and starts new pairing level. The control parameters can be also controlled
interactively, but it is not efficient. Nevertheless in some cases the manual correction may be required or the weighting
parameter wr can be corrected to get more fine control. Optimal values of parameters αi j , α ji , li j should be as smallest

6
Figure 9: Gaussian mixture cluster classification [21, 24] for woven textile and yarn CT data. Two clusters of grey values were identified for each
materials. The red colour corresponds to a mask of clustered data. The following density clusters corresponding to fibre materials and background
noise were identified: TEX-Y[Cluster 1 (165±12.5); Cluster 2 (210±13.8)]; TEX-2DW:[Cluster 1 (90±10.5); Cluster 2 (85±5.0)]

Table 1: Optimal control parameters and number of levels for both models. The weighting parameter wr was set to 0.5. There is a couple of
numbers: the first means the lmax [mm] and second means αmax . [°]
number of levels
TEX-2DW 0.01; 10 0.01; 10 0.03; 30 0.03; 30 0.05; 40 0.08; 40 0.1; 40
TEX-Y 0.01; 50 0.01; 50 0.01; 50 0.015; 50 0.015; 50 – –

as possible, which indicates that two fibre segments could potentially belong to a fibre. Nevertheless, in some cases
the fibre segment ends are close and excentered a bit. The angles can be close to 90◦ , while distance can be still small.
In such case, the user’s careful intervention can be required. The control parameters are yet adjusted by a user, which
allows us to interactively control segment pairing each level, but on the other hand it requires a certain manual tuning
of those parameters by a user to get valid results. The automatic optimal selection of control parameters and their
levels is intensively developed by authors and will be a part of future study.
The segmentation results of the yarn model TEX-Y are shown in Figure 10. With given level-controlled parame-
ters, the resultant fibres are correctly identified with 100 % success. There are 136 fibres at total (counted from slice
cuts). The mean radius of fibres is 0.0051 mm. The initial number of fibre segments is shown in the most left in Figure
10. The results of fibre extraction for high noise and corrupted CT data for model TEX-2DW are shown in Figure
11. The only 73% of fibres were correctly identified (at total there are 255 fibres counted by careful looking at the CT
data). Lower accuracy is caused by corrupted segments, which cannot be uniquely recognised by the deterministic
approach used. The fibre mean radius of TEX-2DW is 0.0057 mm. The multilevel strategy allows us to adaptively
change pairing criteria (one of the main difference from [8]), which may lead to enhanced segmentation results. The
extracting accuracy can be further improved by clever using of morphological operations and more advanced filtering
techniques. Nevertheless, the most critical part is to have sufficient quality of CT data, which is seen in our results.
The computational sources needed for the fibre segmentation are relative low (136 fibres identified in 2s and 186
fibres identified in 5s) in comparison with stochastic algorithm (83 minutes for 903 fibres) proposed in [7]. Nev-
ertheless, it can depend on fibre-structure complexity and segment-ends proximity. The critical part is the quick
nearest-neighbour lookup, which needs to be implemented efficiently by building optimal graph such KD-tree al-
gorithm [25]. Euclidean distance and skeletonisation algorithms are already implemented efficiently in library [20].
They are not usually the performance bottleneck for moderate size CT.

7
 Figure 10: The initial fibre segments before algorithm applying(left). Segmented yarn with 100% success in all available fibre segments(right).

The proposed algorithm has not been tested on large CT data set containing thousands of fibres. This fact can
be considered as a shortcoming of this study because Euclidean distance and skeletonisation algorithms are computer
memory-intensive. Hence they may become a performance bottleneck for a large data set.

8
Figure 11: Extracted woven fibres and their properties from highly noisy and corrupted micro-CT data: initial fibre segments (left); final fibre
segments with fibre radius is shown (right).

9
4. Conclusion

Building a complex and physically relevant model is of great interest in the scientific community. The fibre-like
structures can be seen in many materials. The inspection of their physics is an important task. Although the computer
tomography can provide an important view of the structure of fibres, often more than is required, especially in the
computational models, where the geometry is of great interest. Due to the complexity of fibre structures with many
contact bonds and overlapping, the proper individual fibre identification is a challenging task. Although a few studies
are dealing with fibre extraction, a consistent family of methods is not well developed. The method relies on the
multilevel strategy, where at each level pairs of segments are searched so it matches the given control criteria. The
method is based on the common morphological image operations enriched by machine learning classifier. The method
was tested on the two fibre structure samples and shown high efficiency and good identification accuracy.

5. Conflict of interest statement

The authors declare no conflict of interest.

References
[1] D. Durville, Finite element simulation of textile materials at the fiber scale, arXiv preprint arXiv:0912.1268 (2009).
[2] N. Naouar, E. Vidal-Sallé, J. Schneider, E. Maire, P. Boisse, Meso-scale fe analyses of textile composite reinforcement deformation based on
x-ray computed tomography, Composite Structures 116 (2014) 165 – 176.
[3] Z. Ullah, L. Kaczmarczyk, S. Grammatikos, M. Evernden, C. Pearce, Multi-scale computational homogenisation to predict the long-term
durability of composite structures, Computers and Structures 181 (2017) 21 – 31. UK Association of Computational Mechanics.
[4] X.-Y. Zhou, P. Gosling, C. Pearce, Z. Ullah, L. Kaczmarczyk, Perturbation-based stochastic multi-scale computational homogenization
method for woven textile composites, International Journal of Solids and Structures 80 (2016) 368 – 380.
[5] A. Vanaerschot, B. N. Cox, S. V. Lomov, D. Vandepitte, Multi-scale modelling strategy for textile composites based on stochastic reinforce-
ment geometry, Computer Methods in Applied Mechanics and Engineering 310 (2016) 906 – 934.
[6] G. Gaiselmann, D. Froning, C. Tötzke, C. Quick, I. Manke, W. Lehnert, V. Schmidt, Stochastic 3d modeling of non-woven materials with
wet-proofing agent, International Journal of Hydrogen Energy 38 (2013) 8448–8460.
[7] G. Gaiselmann, I. Manke, W. Lehnert, V. Schmidt, Extraction of curved fibers from 3d data, in: Forum Bildverarbeitung 2012, KIT Scientific
Publishing, p. 119.
[8] X. Huang, D. Wen, Y. Zhao, Q. Wang, W. Zhou, D. Deng, Skeleton-based tracing of curved fibers from 3d x-ray microtomographic imaging,
Results in physics 6 (2016) 170–177.
[9] M. Teßmann, S. Mohr, S. Gayetskyy, U. Haßler, R. Hanke, G. Greiner, Automatic determination of fiber-length distribution in composite
material using 3d ct data, EURASIP journal on advances in signal processing 2010 (2010) 545030.
[10] J. Lux, Automatic segmentation and structural characterization of low density fibreboards, Image Analysis & Stereology 32 (2013) 13–25.
[11] J. Viguié, P. Latil, L. Orgéas, P. Dumont, S. R. du Roscoat, J.-F. Bloch, C. Marulier, O. Guiraud, Finding fibres and their contacts within 3d
images of disordered fibrous media, Composites Science and Technology 89 (2013) 202–210.
[12] R. Sencu, Z. Yang, Y. Wang, P. Withers, C. Rau, A. Parson, C. Soutis, Generation of micro-scale finite element models from synchrotron
x-ray ct images for multidirectional carbon fibre reinforced composites, Composites Part A: Applied Science and Manufacturing 91 (2016)
85 – 95.
[13] M. W. Czabaj, M. L. Riccio, W. W. Whitacre, Numerical reconstruction of graphite/epoxy composite microstructure based on sub-micron
resolution x-ray computed tomography, Composites Science and Technology 105 (2014) 174–182.
[14] W. Whitacre, M. Czabaj, Automated 3d digital reconstruction of fiber reinforced polymer composites, in: AIAA Guidance, Navigation, and
Control Conference, p. 0342.
[15] G. Fang, C. Chen, S. Yuan, S. Meng, J. Liang, Micro-tomography based geometry modeling of three-dimensional braided composites,
Applied Composite Materials 25 (2018) 469–483.
[16] M. J. Emerson, K. M. Jespersen, A. B. Dahl, K. Conradsen, L. P. Mikkelsen, Individual fibre segmentation from 3d x-ray computed tomog-
raphy for characterising the fibre orientation in unidirectional composite materials, Composites Part A: Applied Science and Manufacturing
97 (2017) 83 – 92.
[17] M. J. Emerson, V. A. Dahl, K. Conradsen, L. P. Mikkelsen, A. B. Dahl, Statistical validation of individual fibre segmentation from tomograms
and microscopy, Composites Science and Technology 160 (2018) 208 – 215.
[18] T. Konopczyński, T. Kröger, L. Zheng, J. Hesser, Instance segmentation of fibers from low resolution ct scans via 3d deep embedding
learning, arXiv preprint arXiv:1901.01034 (2019).
[19] V. Wegmayr, A. Sahin, B. Saemundsson, J. Buhmann, Instance segmentation for the quantification of microplastic fiber images, in: The
IEEE Winter Conference on Applications of Computer Vision, pp. 2210–2217.
[20] E. Jones, T. Oliphant, P. Peterson, et al., SciPy: Open source scientific tools for Python, 2001–. [Online; accessed ¡today¿].
[21] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas,
A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: Machine learning in Python, Journal of Machine Learning
Research 12 (2011) 2825–2830.

10
[22] M. Folk, G. Heber, Q. Koziol, E. Pourmal, D. Robinson, An overview of the hdf5 technology suite and its applications, in: Proceedings of
the EDBT/ICDT 2011 Workshop on Array Databases, ACM, pp. 36–47.
[23] J. Yen, R. Langari, Fuzzy logic: intelligence, control, and information, volume 1, Prentice Hall Upper Saddle River, NJ, 1999.
[24] P. A. Yushkevich, J. Piven, H. Cody Hazlett, R. Gimpel Smith, S. Ho, J. C. Gee, G. Gerig, User-guided 3D active contour segmentation of
anatomical structures: Significantly improved efficiency and reliability, Neuroimage 31 (2006) 1116–1128.
[25] S. Maneewongvatana, D. M. Mount, Analysis of approximate nearest neighbor searching with clustered point sets, Data Structures, Near
Neighbor Searches, and Methodology 59 (2002) 105–123.

11