Original article

Textile Research Journal

2018, Vol. 88(18) 2120–2131

Measuring fiber orientations in ! The Author(s) 2017

Article reuse guidelines:
sagepub.com/journals-permissions
nonwoven web images using corner DOI: 10.1177/0040517517716904
journals.sagepub.com/home/trj
detection by Bézier fitting curves

Jue Hou1,2, Bugao Xu3, Hanchao Gao1 and RongWu Wang1,2,4

Abstract
This paper describes a novel method for measuring fiber orientations in nonwoven web images by using Bézier fitting
curves to detect corners of fiber edges and to separate crossing fiber edges. First, the Canny detector was adopted to
extract fiber edges. Second, Bézier curve fitting was used to fit each fiber edge for calculating the curvature of every point
on the edge. Third, corner points were detected by locating points where the curvatures were minimal on various edges
and below the threshold to divide edges into segments for orientation calculations. Last, a formula calculating the fiber
orientation statistics based on the Euclidean distance was established. The experiment results demonstrated that the
proposed method is robust for analyzing different nonwoven web images, and has a high accuracy for corner detection
and fiber orientation calculation.

Keywords
fiber orientation, corner detection, Bézier curve, nonwoven

Measuring fiber orientations in nonwoven webs is an colleagues proposed a multi-focus image fusion techni-
important task in evaluating performance of nonwoven que to generate a sharp image from sequential images
products. Varying degrees of fiber anisotropy in non- focused at different depths for measuring fiber orienta-
woven webs impart unique material properties to the tions in thick nonwovens.9 However, most of these
construct.1 Two of the main ways that are currently existing image-analysis algorithms regarded a nonwo-
used to measure fiber orientations in a nonwoven web ven web as a single structure without dissecting fibers
are mechanical testing and image analysis. Recently, that were bonded or intersected with others before cal-
image analysis has become more popular for fiber culating fiber orientations.4,9 As a result, two bonding
orientation testing mainly because of its efficiency and or crossing fibers might be traced continuously as the
the availability of high-resolution imaging sensors.2 same fiber, causing a false measurement in fiber
Huang and Bresee and Pourdeyhimi and colleagues orientation.
proposed a direct tracking method to characterize non-
woven web structures.3–5 Xu developed a skeletoniza-
tion algorithm to extract the simplified structure of a
1
nonwoven web from which the orientations and lengths Department of Textile Engineering, Donghua University, Shanghai, China
2
Key Laboratory of Textile Science & Technology, Donghua University,
of fiber segments could be estimated.6 Pourdeyhimi and
Shanghai, China
colleagues utilized a Fourier transform function to 3
Department of Merchandising and Digital Retailing, University of North
measure the fiber orientation in assembled structures.7 Texas, Denton, Texas, USA
4
These analysis methods often used simulation images Henan Clothing Textiles Engineering Research Center, Henan University
for result verification and were applicable to thin non- of Engineering, Zhengzhou, Henan, China
wovens, because it was difficult to acquire a fully
Corresponding author:
focused image for fibers on different layers. Image RongWu Wang, Key Laboratory of Textile Science & Technology,
fusion is a process of combining information from mul- Donghua University, Shanghai, 201620, P.R. China.
tiple images into a single composite image.8 Wang and Email: [email protected].
Hou et al. 2121

Crossing fibers form polygonal areas with corners in

a web image. In other words, most corners are the
Edge detection
crossing points of different fibers. Corners of a polygon As shown in Figure 1a, a nonwoven web image con-
provide a clue for finding the crossing locations of tains abundant fiber edge information. Figure 1b shows
fibers. Thus, corner detection is a crucial step for separ- an enlarged view of a small area of Figure 1a with
ating crossed fibers and calculating fiber orientations. detailed fiber edges, which are the boundaries between
Presently, the technique of corner detection is widely the foreground and the background. At an edge, a
applied in various image-comprehension systems.10 sudden change occurs from the foreground to the back-
Mokhtarian and Mackworth presented a representa- ground, indicating a much higher grayscale gradient
tion theory for planar curves through curvature, and than other neighbor pixels in the direction perpendicu-
proposed a corner detection method through the cur- lar to the edge. The orientation of a fiber can be char-
vature scale space (CSS).11,12 Later, more scholars acterized by the orientations of its edges.
developed a series of ameliorative CSS algorithms for The skeletonization algorithm can be affected by any
corner detection.13–17 Chen and colleagues used a irregularity on boundaries. Because the shapes of inter-
method called chord-to-point distant accumulation sections or hidden portions of fibers in a nonwoven web
(CPDA) to express curvature of discrete curves and are so irregular, the skeletons of these areas may not
compared this with other corner detection methods.18 appropriately represent the orientations of fibers. This
Guo and colleagues modified the triangle-area repre- phenomenon may introduce errors to fiber orientations.
sentation (TAR) algorithm for corner detection to The intersections or hidden portions of fibers occupy a
recognize fiber shapes and other attributes.19 small proportion of the whole nonwoven web; the edge
In this paper, we present a new corner detection detection is used as a primary tool to obtain fiber edge
algorithm to separate tangled fiber segments to information.20 Many conventional filters have been
achieve accurate calculation of fiber orientations in developed to perform edge detection. The Prewitt
a nonwoven web. We first introduce the Canny filter,21 Sobel edge detector,21 Canny edge detector,22
edge detection and edge-thinning algorithms used to Zero Crossing Filter, and Laplace of Gaussian (LoG)
extract fiber edge information, then elaborate the con- filter23 are widely used in image processing. The Canny
tour-fitting algorithm based on Bézier curves for detect- edge detector was demonstrated to outperform other
ing corners of intersecting edges, and finally describe edge detection techniques in many applications.24 The
the method of calculating orientations of fiber edge Canny detector consists of the following steps,25 as
segments separated by corners. The results of fiber shown in Figure 2.
orientation measurements using this algorithm are
compared with those using different corner detection 1. Remove noise in the image using the Gaussian
techniques. kernel convolution operator.

Figure 1. (a) Original nonwoven web image and (b) partially enlarged detail.
2122 Textile Research Journal 88(18)

Figure 2. Result of the Canny detector: (a) original image of nonwoven web; (b) result of the Canny detector; (c) part of the Canny
detector result.

Figure 3. Effect of edge thinning on the image in Figure 2b: (a) Magnification of the Canny detector image; (b) magnification of the
edge-thinning image.

2. Calculate the grayscale gradient of every pixel to The Canny detector exhibits excellent performance,
get potential edges that have a sudden change in preserving most of the valuable edge information in the
intensity. nonwoven web image. Unlike the Canny detector,
3. Thin the potential edges using the non-maximum another commonly used edge detector, LoG operator,23
suppression algorithm.26 often loses certain edge information in dark regions.
Figure 4 shows comparison results of Canny edge
Regular fiber edges should be constituted by 1 px detection and LoG edge detection.
wide lines showing fiber orientations, which means
that every pixel on an edge has one or two neighbors.
However, there is a 45 or 135 edge with uniform areas Corner detection using Bézier contours
on either side, and the non-maximum suppression
could produce thick edges, as shown in Figure 3a.27
Bézier fitting contour
Fiber edges generated by the Canny detector are not In a 2D nonwoven image, segments of edges from dif-
1 px wide. Redundant neighboring pixels should be ferent fibers are connected at crossing points, which
deleted using an edge-thinning process.28 Thinning is are also the corners of edge polygons. A corner is gen-
a morphological operation used to remove pixels that erally defined as the local maximum of the absolute
are not essential for edge connections. The edge-thin- value of the curvature of a curve in topology.27 In
ning effect is shown in Figure 3b. order to compute the fiber orientation distribution in
Hou et al. 2123

Figure 4. Comparison of Canny and LoG edge detectors: (a) the Canny detector image; (b) magnification of the Canny detector
image; (c) the LoG detector image; (d) magnification of the LoG detector image.

a nonwoven web, the directions of all fiber segments fitting here is to locate corner points of joining fibers.
need to be accumulated precisely. We must detect all Therefore, we chose Bézier curves to fit short segments
corners on fiber edges based on changes in curvature. of fiber edges due to its simplicity and efficiency.
A curve at a corner has a much larger curvature, or a The Bézier curve is a parametric curve frequently
smaller angle between its two tangent lines, than at used in engineering science to model a discrete contour.
other points. In a bitmap image, it is difficult to find It contains two endpoints and several control points,
real corners because fiber edges are represented by and its shape can be controlled by changing the relative
discrete pixels, instead of continuous curves. Many locations of the control points. The Bézier curve has
curve-fitting methods can be used to approximate two significant advantages, summarized as follows:
discrete pixels. In general, NURBS curves are more
robust for fitting complex curves,29,30 while Bézier 1. It can easily be expressed as a smooth curve by a
curves are more appropriate for fitting simple curves. parametric equation.
Fiber segments in a nonwoven material are relatively 2. The form of the Bézier curve only relates to the end-
short and straight lines, and the purpose of the curve points and the control point.
2124 Textile Research Journal 88(18)

The Bézier curve can be represented as:31 Bézier curve is written as:30

xðtÞ ¼ a0 þ a1 t þ a2 t2
X
n t 2 ½0, 1 ð2Þ
QðtÞ ¼ Vi Bni ðtÞ, 0t1 ð1Þ yðtÞ ¼ b0 þ b1 t þ b2 t2
i¼0
A nonwoven web image has abundant fiber edges,
which require massive calculations for corner detection.
where Vi is the control point and Bni is the Bernstein A quadratic Bézier curve is found to be an effective
polynomial. The parametric equation of a quadratic choice to model a fiber-fitting contour because fiber
edges are relatively short, straight-line segments.
The main step to calculate the curvatures of pixels
on a fiber edge is to fit the edge with a Bézier curve. The
pixel where a curvature needs to be calculated is called
the current pixel (Figure 5). Fitting a quadratic Bézier
curve is always a convex curve tangent at the endpoints,
with two lines passing the control points.
If the current pixel (i.e. a control point) is a corner
point, there will be a significant deviation between the
fitting curve (Bézier curve) and the original fiber edge.
Thus, it is necessary to divide the fiber edge into two
segments at the current pixel, change the current pixel
to an endpoint, and fit the segments on both sides sepa-
rately. Another endpoint and a control point for each
segment are also needed, and can be decided by the
Figure 5. Endpoints and a control point of a quadratic fitting scale s (the distance between the two endpoints),
Bézier curve. as shown by P0–P1 and P0–P4 in Figure 6.

Figure 6. Fitting a Bézier curve on both sides of a current point (P0 is the current point, P1 and P4 are endpoints, P2 and P3 are
control points).
Hou et al. 2125

Figure 7. The fitting curve deviated from the actual edge (arc AC has an obvious deviation from the actual path ABC).

In order to simplify the algorithm, the scales are set to

Multi-scale calculation
be 5, 10, and 15 pixels, respectively. The deviation of
The fitting scale s can influence the fitting accuracy. every pixel located between the fitting curve and the
If two corners are too close to each other and the dis- actual edge is calculated from 15 to 5 until the deviation
tance between the corners is smaller than s, as shown in is smaller than the threshold. Then, the Bézier curve is
Figure7a, the fitted curve may deviate from the actual fitted under the scale that leads to the smallest deviation.
fiber edge. The fitting process can be summarized as follows:
To choose a proper fitting scale, an evaluation
parameter S of imitative fitting effect can be defined 1. Trace each point that is not an endpoint and use
as follows: Bézier curves to fit the two segments on both sides
Z  of the current point.
 ðx1 ,y1 Þ Z ðx1 ,y1 Þ 
  2. Decrease fitting scale s from large to small until the
S ¼  ðf  lÞ  ð g  l Þ ð3Þ fitting length does not need to be adjusted.
 ðx0 ,y0 Þ ðx0 ,y0 Þ 
3. Draw two fitting curves of the current point and
calculate the curvature. Then, fit two segments on
where the scale s indicates the length of the path both sides of the next point.
between the endpoints of a fitting curve, Ss is the area
enclosed by the actual pathR ðx ,y and the fitting curve,
Þ
as shown in Figure 7b. ðx01,y01Þ ð f  l Þ represents the
region bounded by the actual path and the Curvature expression
R ðx horizontal
,y Þ
line between the two endpoints. Similarly, ðx01,y01Þ ð g  l Þ Curvature is an index to characterize the change
represents the region bounded by the fitting curve of radian. With increasing curvature, the fiber edges
and the horizontal line between the two endpoints. will have a large curvature of the fiber edge.
With the enlarging of curvature of the fiber edge, the Curvature k can be calculated through the curvilinear
Bézier fitting curve and the fiber edgeR will have larger equation k(s) as follows:
ðx ,y Þ
deviations. When Ss exceeds 15% of ðx01,y01Þ ð g  l Þ, the
actual path can be regarded as a segment that is too long x_ ðtÞ y€ ðtÞ  y_ ðtÞ x€ ðtÞ
to be fitted accurately, the fitting length needs to be kðtÞ ¼  3=2 ð4Þ
adjusted. x_ ðtÞ2 þy_ ðtÞ2
2126 Textile Research Journal 88(18)

where t is the parametric of a quadric Bézier curve, and where xn , yn are the coordinates of the tangents’ end-
x_ ðtÞ, y_ðtÞ, x€ ðtÞ, y€ ðtÞ are the first and second derivatives points. The corner is defined as a point where y(x,y) is
of a Bézier curve, respectively. the minimum along a fiber edge.
However, the curves on both sides of the current In Figure 9a, A is the corner point and B and C are
point do not have the same expression and are not C2 adjacent points of A on a fiber edge. Figure 9b displays
continuous. Hence, the first derivative and the second the angles of all points on the edge that are computed
derivative cannot be calculated. at the scale s ¼ 5. For the edge segment, if the angle of
The curvature k has a direct relation with the angle, a point is the minimum among all the points and is
y(x,y), of two tangent lines. A tangent line is the limit smaller than a preset threshold (to be discussed in
case of the secant. Figure 8 shows that when the two the following section), this point can be regarded as a
secant endpoints infinitely close, the secant slope is candidate corner. For example, point A in Figure 9
approximately equal to the tangent slope. The slope of has y ¼ 90 , which is the minimum and smaller
the secant can be easily calculated, avoiding the calcula- than the threshold, and thus A is a candidate corner
tion of curve derivatives. y(x,y) is then defined as follows: on this edge.

0 1
! !
n1  n2
Removal of false corners
ðx, yÞ ¼ cos1 @   A ð5Þ
! ! As shown in Figure 10, the fiber edges are divided
 n1    n2 
into several segments by corners (all of F points and
T points). Partial segments from the same fiber edge are
where
not oriented in exactly the same direction. The angle
between two adjacent segments determines whether to
!
n1 ¼ lim ðx1  x0 , y1  y0 Þ ð6Þ
ðx1 ,y1 Þ!ðx0 , y0 Þ
reserve a corner candidate or not. It is necessary to set
an appropriate angle filtering threshold that balances
! algorithm efficiency and precision, owing to how com-
n2 ¼ lim ðx2  x0 , y2  y0 Þ ð7Þ
ðx2 ,y2 Þ!ðx0 , y0 Þ putational complexity is augmented exponentially with
increasing targets.
Figure 11 shows the preserved corner numbers after
the corner candidates in Figure 10 were filtered using
different angle thresholds. Preserving more corner can-
didates means more segments of fiber edges can be
separated – that is, a higher computational cost.
On the other hand, eliminating more corner candidates
will lead to poor accuracy in orientation calculations.
So setting a proper threshold to filter corner candidates
is important for both the efficiency and accuracy of the
algorithm. From Figure 11, it can be seen that at the
Figure 8. Secant approaches the tangent. threshold of 175, 37% more corners are retained than

Figure 9. Change of angle y along a sequence of points: (a) a sequence of points; (b) angles of two secants.
Hou et al. 2127

Figure 10. All candidate corners; F (yellow) marked false candidates, T (red) marked reserved corners.

anisotropy of nonwoven materials. The fiber edges are

divided into segments that are close to straight lines,
and thus the distance between the two breakpoints can
be calculated using the Euclidean distance. The compu-
tational formula of fiber segment length li ðÞ can be
defined as follows:

8 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
< li ðÞ ¼ R ðx1 ,y1 Þ fds  ðx1  x2 Þ2 þ ð y1  y2 Þ2
ðx2 ,y2 Þ
ð8Þ
:
 ¼ tan1 jjx22 x11jj
y y

where ðx1 , y1 Þ, ðx2 , y2 Þ are the coordinates of a fiber seg-

ment’s endpoints, and  is the angle between the fiber
segment and horizontal line.

Figure 11. Relationship between the number of corners and Fiber web orientations can be calculated from the
threshold. direction distributions of fiber segments. The fiber
web orientation in a certain direction is defined as the
ratio of the accumulated length of fibers in this direc-
at the threshold of 150, and 6.9% more corners than at tion as opposed to the total length of the fibers in all
the threshold of 170. Since there was no significant directions. Therefore, the fiber orientation distribution
increase in corners when the angle threshold varied of a web image can be written as:
from 170 to 175, we set the threshold to be 170 in
this project.
Lð Þ
Fð Þ ¼
Lð0Þ þ Lð1Þ þ    þ Lð178Þ þ Lð179Þ
Pn
Orientation calculation i¼0 li ðÞ
¼  2 ½0, 180o Þ
The fiber orientation distribution in a nonwoven Lð0Þ þ Lð1Þ þ    þ Lð178Þ þ Lð179Þ
web provides important information describing the ð9Þ
2128 Textile Research Journal 88(18)

Figure 12. Common phenomenon of corner detector contrast.

where  is the angle between the fiber segment and In the big and small curvature cases, the four methods
the horizontal line, L() is the sum length of successfully find all the corners. In the other cases, the
fiber edges whose angles are equal to , li() is the three previously reported algorithms ignore or falsely
length of the fiber segment whose angle is equal to , locate some corners. CSS corner detection tends to
and n is the number of fiber edges whose angles detect distinct corners, while both CPDA and TAR
are equal to . use change of chords to represent the point’s curvature
and more likely miss indistinctive corners.
To compare the accuracy and practicability of the
Results and discussion four algorithms, each algorithm was applied to actual
measurement of fiber orientation. The fiber edges have
Comparison tests on corner detection been cut into several segments by corners; these segments
The measurement accuracy of fiber orientation distri- should be close to straight lines. The higher precision of
bution is closely related to corner detection accuracy. In corner detection can ensure that the connecting line
this section, we compare our proposed corner detection between an endpoint and a corner point is closer to an
algorithm with three previously reported contour-based actual fiber edge. The ratio of actual fiber edges covered
algorithms: CPDA, CSS, and TAR. by connecting lines is shown in Table 1, which demon-
Figure 12 displays the corner detection results of strates the significant accuracy made by the proposed
five cases with the four algorithms, in which fiber algorithm for nonwoven fiber images.
edge segments are indicated by black pixels, and cor-
ners are marked by red points. The five cases include a
big curvature curve, a normal curve, a small curvature
Comparison tests on fiber orientation distribution
curve, a curve with multiple angles, and an arc curve. The mechanical properties of a nonwoven, such as
These five cases are the fiber edges extracted from a real breaking strength, are disparate in various orientations,
nonwoven image. The results indicate that the pro- and their distributions should be associated with
posed algorithm can correctly detect all corners in the fiber orientation distribution. The fiber orientation
these cases, outperforming the other three algorithms. distribution in a nonwoven web can be expressed by
Hou et al. 2129

Table 1. The integral covering ratio calculated by four algorithms

Sketch map

Corner detector Covering ratio

Proposed algorithm 92%

CSS detector 85%
CPDA detector 74%
TAR algorithm 73%

Figure 13. Result of different methods measuring fiber orientation.

strengths along different directions. A thermal-bonded orientation distribution calculated by the Bézier fitting
nonwoven was used to perform the tensile test.32 As curve algorithm is highly consistent with the strength dis-
shown in Figure 13, the tested nonwoven has maximum tribution measured by the tensile test. On the contrary, the
strength in the machine direction (near 0 or 180 ), and fiber orientation distributions obtained from the other
minimum strength in the vertical direction (near 90 ). Fifty three algorithms (CSS, CPDA, TAR) show significant
images of the same thermal-bonded nonwoven were cap- differences from the strength distribution, although their
tured for the analysis, and the fiber orientation distribu- overall trends are similar. The comparison proves the pro-
tions of the nonwoven obtained from different algorithms posed algorithm is very well suited for measuring fiber
are presented in Figure 13 as well. It is clear that the fiber orientation distributions in nonwovens.
2130 Textile Research Journal 88(18)

7. Pourdeyhimi B, Dent R and Davis H. Measuring fiber

Conclusion orientation in nonwovens, part III Fourier transform.
In this paper, a new method for measuring fiber orien- Text Res J 1997; 67(2): 143–151.
tation distributions in nonwoven webs was presented. 8. Yao M, Zhao Z, Yao X, et al. Fusing complementary
images for pavement cracking measurements. Meas Sci
The new method is based on the Bézier fitting contour
Technol 2015; 26(2): 025005. DOI: 10.1088/0957-0233/26/
algorithm used to detect corners of fiber edges in web 2/025005.
images and to calculate the orientation angles of sepa- 9. Wang R, Xu B and Li C. Accurate fiber orientation
rate fiber segments. Bézier curves permit using the angle measurements in nonwovens using a multi-focus image
between two adjacent tangents to replace the curvature fusion technique. Text Res J 2013; 84(2): 115–124.
calculation for corner detection with greater precision. 10. Zhang WC and Shui PL. Contour-based corner detection
The experiment results demonstrated that the Bézier via angle difference of principal directions of anisotropic
fitting algorithm can generate fiber orientation distribu- Gaussian directional derivatives. Pattern Recogn 2015;
tions from nonwoven web images that are highly con- 48(9): 2785–2797.
sistent with their strength distributions as measured by 11. Mokhtarian F and Suomela R. Robust image corner
the tensile test. Future work will include performing detection through curvature scale space. IEEE Trans on
Pattern Anal Mach Intell 1998; 20(12): 1376–1381.
more thorough testing on different nonwoven materials
12. Mokhtarian F and Mackworth AK. A theory of multi-
with the presented algorithm, and exploring other fit-
scale, curvature-based shape representation for planar
ting methods, such as NURBS, for more accurate curves. IEEE Trans on Pattern Anal Mach Intell 1992;
curve fitting. 14(8): 789–805.
13. Zhong B and Liao W. Direct curvature scale space:
theory and corner detection. IEEE Trans on Pattern
Declaration of conflicting interests
Anal Mach Intell 2007; 29(3): 508–512.
The authors declared no potential conflicts of interest with 14. Crowder M. An improved curvature scale-space corner
respect to the research, authorship, and/or publication of this detector and a robust corner matching approach for
article. transformed image identification. IEEE Trans on Image
Process 2008; 17(12): 2425–2441.
15. Zhang X, Qu Y, Yang D, et al. Laplacian scale-space
Funding
behavior of planar curve corners. EE Trans on Pattern
The authors disclosed receipt of the following financial sup- Anal Mach Intell 2015; 37(11): 2207–2217.
port for the research, authorship, and/or publication of this 16. Li D, Zhong B, and Ma KK, (eds.). Multi-scale corner
article: This research was supported by the Natural Science detection based on arithmetic mean curvature. IEEE
Foundation of China (Grant No.61172119), the Fundamental China Summit and International Conference on Signal
Research Funds for the Central Universities, China (Grant and Information Processing, 2015.
No. CUSF-DH-D-2017020), Open Fund Program of Henan 17. Zhong BJ and Li C. Robust image corner detection using
Clothing Textiles Engineering Research Center. curvature product in direct curvature scale space.
App Mech Mat 2010; 20–23: 725–730.
18. Chen S, Meng H, Zhang C, et al. A KD curvature based
References corner detector. Neurocomput 2015; 173: 434–441.
1. Ayres C, Bowlin GL, Henderson SC, et al. Modulation of 19. Guo X, Ouyang W and Xu B. Shape recognition of pro-
anisotropy in electrospun tissue-engineering scaffolds: ana- filed fibers using triangle-area representation. Meas Sci
lysis of fiber alignment by the fast Fourier transform. Technol 2013; 24(9): 95401–95410.
Biomaterials 2006; 27(32): 5524–5534. 20. Nevagi UG, Shahapurkar A and Nargundkar S.
2. Yao M, Yu W and Xu B. Evaluating fabric fuzziness using Edge detection techniques: a survey. Int J Innovat Res
laser range sensing. Opt Eng 2008; 47. DOI: 10.1117/ Dev 2016; 5(2): e86174–e.
1.2834715. 21. Cherri AK and Karim MA. Optical symbolic substitu-
3. Huang XC and Bresee RR. Characterizing nonwoven web tion: edge detection using Prewitt, Sobel, and Roberts
structure using image analysis techniques III. Web uni- operators. Applied Optics 1989; 28(21): 4644–4648.
formity analysis. INDA J Nonwovens Res 1993; 5(2): 22. Wang B and Fan SS. An improved CANNY edge detec-
14–21. tion algorithm. Microcomput Inf 2010; 1: 497–500.
4. Pourdeyhimi B, Ramanathan R and Dent R. Measuring 23. Mallick A, Roy S, and Chaudhuri SS. Optimization
fiber orientation in nonwovens: part II Direct tracking. of Laplace of Gaussian (LoG) filter for enhanced edge
Text Res J 1996; 66(12): 747–753. detection: a new approach. In: International conference on
5. Pourdeyhimi B, Kim HS, Pourdeyhimi B, et al. Measuring control, instrumentation, energy and communication,
fiber orientation in nonwovens: the Hough transform. Calcutta, India, 31 January–2 February 2014, pp. 658–661.
Text Res J 2002; 72: 803–809. 24. Gang L and Haralick RM (eds). Two practical issues in
6. Bugao X and Ting YL. Measuring structural characteris- Canny’s edge detector implementation. In: Proceedings
tics of fiber segments in nonwoven fabrics. Text Res J 15th international conference on pattern recognition
1995; 46(1): 41–48. ICPR, Barcelona, Spain, 2000, pp. 676–678.
Hou et al. 2131

25. Canny J. A computational approach to edge detection: 29. Lim SP and Haron H. Cube Kohonen self-organizing
Readings in computer vision. San Francisco, CA: map (CKSOM) model with new equations in organizing
Morgan Kaufmann, 1987, pp.184–203. unstructured data. IEEE Trans Neural Netw Learn Syst
26. Song Y, Glasbey CA, Polder G, et al. Non-destructive 2013; 24(9): 1414–1424.
automatic leaf area measurements by combining stereo 30. Lim SP and Haron H. Surface reconstruction techniques:
and time-of-flight images. IET Comput Vision 2014; a review. Artif Intell Rev 2014; 42(1): 59–78.
8(5): 391–403. 31. Shao L and Zhou H. Curve fitting with Bézier cubics.
27. Mokhtarian F and Suomela R. Curvature scale space Graph Model Im 1996; 58(3): 223–232.
for robust image corner detection. In: International con- 32. Pourdehyhimi B, Dent R, Jerbi A, et al. Measuring fiber
ference on pattern recognition, Brisbane, Australia, 20 orientation in nonwovens part V: real webs. Text Res J
August 1998; pp. 1819–1823. 1999; 69(3): 185–192.
28. Zhang Z, Ma S, Liu H, et al. An edge detection approach
based on directional wavelet transform. Comput Math
Appl 2009; 57(8): 1265–1271.