Automatic Volumetric Measurement of

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Textile Research Journal Article

Automatic Volumetric Measurement of Nanofiber Webs using


Metaball Approximation Based on Scanning Electron
Microscope Images
In Hwan Sul
Abstract Scanning electron microscope (SEM) i-Fashion Technology Center, Konkuk University, Seoul,
images are an essential tool for identifying the 143-701, Republic of Korea
microstructures of fibrous and polymeric materi-
als. Useful information, including fiber orienta- Young Jun Cho
tion and radius distributions, can be retrieved Hyosong R&D Business Labs., Gyeonggi-do, 431-080,
from the SEM images. However, the image is two Republic of Korea
dimensional and its manual measurement is inevi-
table for acquiring volumetric information. This
Chang Kyu Park1
paper reconstructed the three-dimensional shape of
Department of Textiles and Engineering, College of Eng.,
the fibers by approximating the fibers with spheres
Konkuk University, 143-701, Republic of Korea
known as metaballs. Image analysis techniques were
used to generate metaballs and the volumetric
information was calculated from the sizes and posi-
tions of the metaballs. The proposed method was
tested with nano-scale polymer SEM images and
automatic statistical measurements were compared
with the manual measurements.

Key words scanning electron microscope image,


metaball approximation, orientation vector, radius
distribution, pore size distribution

The scanning electron microscope (SEM) is a device that retrieving the size information, such as the volume fraction
uses high energy electrons to visualize the material surface from two-dimensional cross-section information, e.g. for
topography. It has been used in wide areas, ranging from finding the volume fraction of particles in reinforced mate-
biology to the material sciences. In particular for textiles rials. However, the stereology-based method is classified as
and polymeric materials, we can perform both qualitative an indirect method because it is based on statistical infer-
and quantitative observations on a specimen by taking ence. Moreover, the manual measurement of fiber radii is
SEM images. For qualitative measurements, the SEM can a laborious job. This paper proposes an automatic method
produce the surface status, such as smoothness, and of acquiring statistical values from a single SEM image by
researchers can judge if the synthesis was successful or not. reconstructuring the original micro-fibers in three dimen-
For quantitative measurements, we can measure volumet- sions. We assume that the specimens have three-dimen-
ric sizes, such as fiber orientations, fiber radius distribu- sional web-like fibrous structures of multiple overlapped
tion, and pore size distribution. layers of nano-scaled fibers. Hereafter, these fibers are
Meanwhile, most specimens used for SEMs have a referred to sometimes as fibers, micro-fibers or nanofibers,
three-dimensional microstructure, such as polymers.
Since a SEM image can give only a two-dimensional pic-
ture, researchers should measure fiber sizes manually and
predict their true values using statistical methods, such as 1
Corresponding author: tel: +8224504196; e-mail: cezar@
stereology [1]. A stereology-based method is useful for konkuk.ac.kr

Textile Research Journal Vol 80(11): 995–1003 DOI: 10.1177/0040517509348333 © The Author(s), 2010. Reprints and permissions:
Figures 1, 2, 4–6, 8, 9 appear in color online: http://trj.sagepub.com http://www.sagepub.co.uk/journalsPermissions.nav
TRJ 996 Textile Research Journal 80(11)

and their SEM images are referred to as fiber SEM SEM image into groups of fibers and used metaball-based
images. approximation to build three-dimensional mesh from the
Reconstructing the three-dimensional shape of the two-dimensional fiber images. Once the metaballs’ posi-
material microstructure has been an interesting theme for tion and radii are acquired, the statistical values can be cal-
many material scientists. The well-known technique is a culated easily. In addition, there were some limitations in
serial sectioning-based method using a focused ion-beam our approach. Only the top-most layer of fibers could be
scanning electron microscope (FIB-SEM). Several hun- easily identified, although general fiber SEM images are
dred tomographic SEM image slices with a thickness of a composed of multiple numbers of layers. A high-resolution
few nanometers can be acquired through a focused ion SEM image was preferred because it contains less noise. To
beam (FIB) sectioning process [2]. Next, a three-dimen- simplify the algorithm, we assumed the cross-sections of
sional shape can be reconstructed by stacking the images. fibers to be circular. Our algorithms are described in detail
Sidhu and Chawla reconstructed the three-dimensional in the following.
image of Ag3Sn by serial section [3] and measured the
aspect ratio. Li et al. [4,5] also quantitatively characterized
the microstructural damage of Si particle-reinforced metal Algorithm
matrix composites by using serial sectioning. They calcu-
lated a particle size, a shape, an orientation, a local volume
fraction, and neighbor distances from the three-dimen-
Two-dimensional Image Analysis Procedure
sional reconstructed data. The reconstructed three-dimen- The basic idea of our method with respect to a simple
sional data were also useful as mesh data for further object is illustrated in Figure 1. Suppose that the check-
analysis using finite element analysis [6]. mark-like letter of Figure 1(a) is a kind of fiber and we
The specimens for the slicing methods had only a single intend to measure its dimensions. We want to find simple
object, such as particles, rods, or chromosome [7], so each geometric features, such as a series of circles, to represent
sliced image can be merged and reconstructed relatively the checkmark. To do so, the center line of the checkmark
easily. However, the same technique cannot be simply is acquired first (Figure 1(b)) and the circles are deployed
adapted to the SEM images of the polymeric materials. along the center line (Figure 1(c)). Lastly, the circles are
The major problem is that the polymers have a complex converted to three-dimensional spheres so that the finally
structure with multiple layers of interconnected fibers, so reconstructed mesh can be produced (Figure 1(d)). How-
that the sliced images still contain multiple objects that ever, since the fiber SEM image has many overlapped
cannot be easily isolated. lines, further treatments are needed in fact.
There were also other approaches that used several In the beginning, we need several assumptions to sim-
pieces of the SEM images without slicing them. Samak et plify the problems. The first assumption of our method is
al. [8] used two pieces of the SEM images and visualized that the pixel brightness is proportional to the actual z-
the surface ruggedness of the ceramic and the synthetic coordinate of the specimen. This can be readily proven
foams with a stereo-matching algorithm. They increased from the mechanism of the SEM. The first-hand objective
the visual quality by means of texture mapping, but they of finding z-coordinates from pixel values is to distinguish
did not deal with the quantitative measurement. Palus- each fiber and get the true radii, not to make the final
zyński and Slówko [9] also used four detectors for rela- three-dimensional mesh. Fiber SEM images are composed
tively smooth objects and it seemed to be difficult to apply of hundreds of inter-connected fiber segments. Without
to a complex polymeric structure. Those methods used a separating each fiber, the radii have the wrong values, as
single object specimen and were not concerned with the Figure 2(e) shows. The closer (higher z-coordinate) object
isolation or the qualitative measurements. to the SEM detector has a brighter pixel (high binary value
The SEM is an innovative device for showing the micro- close to 255) and vice versa. Therefore, fibers closest to the
structure of materials. However, mechanical engineers detector will have white colors in the SEM image. Our idea
have no proper tools to analyze the captured SEM image, is to collect pixels of similar colors and identify them as a
except for using a ruler. A fiber SEM image is composed of group of fibers preliminarily. As our objective is the quan-
hundreds of fiber segments and each fiber has a different titative measurement, such as the radius distribution, this
radius. Even a single fiber has different radii depending on grouping does not have to be as perfect as a human eye
its shape. Thus, it is not practical to use only the manual recognizes the fibers. The grouping has only to disassemble
measurement for acquiring a statistically true average the stacked fiber images into individual single layers to
value. This paper presents a novel way to acquire a large ease the whole procedure. To this end, a series of image
amount of measurement data samples from a SEM image by analysis methods were applied to the original SEM images.
separating the overlapped fiber images and reconstructing The second assumption is that we use only the top-most
the three-dimensional mesh. This paper adopted several layers of the fibers. This is an evident shortcoming of our
image analysis techniques to partition the two-dimensional method, but there seems to be no easy way to completely
Automatic Volumetric Measurement of Nanofiber Webs using Metaball Approximation Based I. H. Sul et al. 997 TRJ

Figure 2 Image analysis procedure for sample #PUM1 ((a)


original image; (b) histogram equalized; (c) smoothed; (d)
thresholded; (e) skeletonized; (f) Canny edge; (g) flood filled;
(h) skeletonized from (g); (i) final metaballs).

conventional methods. As a first step, the image was equal-


ized (Figure 2(b)), smoothed (Figure 2(c)), and thresh-
olded (Figure 2(d)) to convert the gray image into a binary
value of 0 or 255. Histogram equalization was used to
enhance the image and to make the inter-fiber interface
clearer. The raw image had similar gray values, but the
equalized image exhibited the gradient difference among
the fibers. The thresholded binary image (Figure 2(d)) was
Figure 1 Diagram of the reconstruction algorithm: (a) origi-
used to get the center lines (skeletons) of the fibers using a
nal two-dimensional picture of a three-dimensional object;
thinning algorithm by Lee et al. [10], as in Figure 2(e).
(b) center line (skeletonization); (c) metaball approximation;
However, the fibers generally had scales on their surface
(d) three-dimensional mesh generation.
and these scales acted as unnecessary noise groups. So, the
images needed to be smoothed (Figure 2(c)) before being
thresholded. Gaussian smoothing with a radius of 3 pixels
identify the fiber layers with image analysis techniques. and a threshold value of 128 was used. However, as seen in
Nonetheless, we believe that our method is more effective Figure 2(d), a simple binary image after thresholding does
than the manual measurement. not distinguish the overlapped fibers, so the skeletons of
In this two-dimensional image analysis procedure, the Figure 2(e) can be said to be a wrong result, especially in
first objective is to distinguish fibers into several groups. the regions where fibers are overlapped. The cause is that
Figure 2(a) shows the original SEM image. As is the usual we tried to treat the bulk image all at once.
case in the fiber webs, the several layers of the fibers are Thus, we rejected Figure 2(e) and skeletonized one
overlapped and it seems difficult to distinguish them with fiber group at each time instead. Herein, we define a col-
TRJ 998 Textile Research Journal 80(11)

lection of pixels with similar colors in the same fiber seg- simple mathematical terms, the total volume V of all the
ment as a fiber group. To dissect the original binary bulk fibers can be shown as:
image to multiple fiber groups, the raw image was edge
detected with the Canny edge algorithm [11] from the histo- f–1
gram-equalized gray image (Figure 2(b)). Figure 2(f) shows V = ∑V i (1)
the edge detection result, which preserves the inter-fiber i=0

boundary well. Figure 2(f) depicts a collection of simply-con-


nected closed-loop polygons. Then, each fiber was painted where f is the number of fibers.
with different colors with a flood-fill algorithm [12]. The out- The volume of each fiber, Vi, is:
of-fiber regions (pore or background regions) were not
painted because they were not necessary. Determination of Vi = Bi,0 傼 Bi,1 傼 … Bi,b–1 (2)
the in/out of the fibers was based on the binary color of Fig-
ure 2(d). Figure 2(g) shows the distinguished fiber groups. where b is the number of balls in fiber i.
Then, the skeleton lines of each group were detected (Fig- The metaball Ri,j is expressed as an equation of a circle
ure 2(h)). In Figure 2(e) the skeleton lines were not in the as follows:
exact centers of the fiber groups. In contrast, those of Fig-
ure 2(g) resided in the correct middle lines. As our grouping 2 2 2 2
is based on the pixel colors, there are some errors in the ( x – X i, j ) + ( y – Y i, j ) + ( z – Z i, j ) – R i, j = 0 (3)
grouping, i.e. a single fiber is broken into several groups, not
into a single group. However, this does not make any dif- where (Xi,j, Yi,j, Zi,j) and Ri,j are the center and the radius of
ference because the metaballs are finally converted into metaball i,j, respectively. Each pixel of skeleton lines
the same mesh in the mesh generation procedure using the becomes the metaball center and the maximum incircle
marching cube method. radii become the metaball radii when they touch the fiber
boundary curves. The binary image after the thresholding
was used to find the incircles. Note that the metaballs can
Three-dimensional Metaball Approximation have different radii depending on their positions, while
they retain the same circular, cross-sectional shape. More-
Once the overlapped fibers were distinguished, the rest of
over, the path of the metaballs can be curvilinear if the
the job was to approximate the fibers with simple geometry,
original fiber group has a curvilinear shape.
such as a sphere. As we have only a single SEM image as our
Now that the positions and the sizes of the metaballs
input data, we cannot make three-dimensional meshes
have been found, the quantitative analysis can be done. The
directly as in the FIB-SEM serial sectioning method. Here
radius distribution can be plotted directly from Ri,j. The ori-
we define another assumption – that the fibers have the
same cross-sectional shape. Generally, the polymeric or entation vector O i, j of a metaball Bi,j was defined as the dis-
fibrous material has circular or elliptical cross-sections. In placement vector between two neighboring metaballs, i.e.,
fact, any kind of cross-section shape can be used, but we
chose the simplest one to ease the calculation. Once the
cross-sections are the same, the individual fibers can be O i, j = { ( X i , j + 1 – X i, j ) , ( Y i, j + 1 – Y i, j ) , ( Z i, j + 1 – Z i, j ) }
(4)
modeled with the same basic geometric feature, such as (0 ≤ j < b)
cylinders or spheres of various sizes and directions. We
approximated the fibers as a bunch of overlapped balls As there can be noises in the images and the metaballs with
with different radii, as seen in Figure 2(i). These balls are the very small radii can affect the statistics, all the statisti-
also known as metaballs [13]. The metaball is an implicit sur- cal values were weight-averaged with the size of the meta-
face used to model a three-dimensional shape effectively. To ball. For example, an average orientation angle θi of a fiber
determine the positions and radii of the metaballs, the skele- Fi was calculated as:
tons of the fibers were needed. The metaballs were assigned
to each pixel of the skeleton lines of Figure 2(h). The radius
b–1
of each metaball was obtained from the biggest incircle
contacting with the fiber boundary. As too many metaballs

i=0
R i, j • θ i, j
θi = ------------------------------ (5)
can slow down the mesh generation process, balls within a b–1

certain range of threshold distance were omitted. Figure ∑


i=0
R i, j
2(i) is the two-dimensional metaball insertion result. They
were given x and y coordinates but not z-coordinates as yet.
The z-coordinate value was calculated from the sum of the where θi,j is the angle between O i, j and the +x-axis.
pixel values within each metaball. Thus, the metaballs in a To show the three-dimensional reconstruction result of
brighter region take higher z-coordinates and vice versa. In the SEM image, the metaballs should be converted to
Automatic Volumetric Measurement of Nanofiber Webs using Metaball Approximation Based I. H. Sul et al. 999 TRJ

Table 1 Specification of the specimens used.


Item PUM1 PUM2
Material PU PU
Thickness (μm) 8–10 12–15
Mean flow pore diameter (μm) 0.6227 0.5805
Bubble point pore diameter (μm) 0.8192 0.7444

Results and Discussion


Test Specimens
To verify the presented method for the actual fiber SEM
images, two kinds of poly-urethane based breathable
nanofiber membrane samples were used. Table 1 shows
the specification of the specimens. The specimens are
made of the same polymer resins, but different electrospin-
ning conditions resulted in different orientations and
radius distributions. Firstly, we need to determine the
proper size of the input SEM image size. Too large a pic-
ture size can delay the calculation time and the possibility
that the noises are accepted as valid metaballs can increase.
By resizing the picture into smaller sizes, the noises were
smoothed and the final three-dimensional reconstruction
mesh contained only the valid fibers. Images were resized to
512 pixels × 512 pixels in this case. The SEM resolution was
the more critical factor for metaball approximation. Figure 4
shows the metaball approximation results in three different
SEM resolutions. The lower the resolution is, the more jag-
ged the skeletonized center lines become, and the final
metaballs yield the more discontinuous lines. In conclu-
Figure 3 Three-dimensional reconstruction result for
sion, the sufficiently high resolution of the SEM images
sample #PUM1: (a) top view; (b) front view.
was advantageous in that the fiber diameter should be
greater than a tenth of the image width. We used images of
resolution 2 × 106 times where 1 μm equals 84 pixels.
three-dimensional mesh data. Marching cube [14] is the Figure 2 shows the image processing steps for the speci-
famous technique to convert the metaball space to triangu- men #PUM1 and Figure 3 shows the final three-dimen-
lar mesh data. To use the marching cube method, the space sional mesh reconstructed as explained earlier. The top-
should be voxelized first because the marching cube most layer of the original image was successfully recon-
method needs voxel data as its input. So the whole three- structed. As we have the three-dimensional mesh and the
dimensional space containing the volume V was voxelized. metaball information, the fiber statistics, such as radius
If a voxel lies inside any metaball, it is assigned a value of 1 distribution and orientation, can be simply acquired. Some
and otherwise 0. The voxels were also used for the calcula- noises exist in Figure 3, but they do not adversely affect the
tion of the fiber volume fraction. The total volume fraction statistics because the statistics are factored by the radius
can be found simply by dividing the number of filled voxels size. An AMD 3.0GHz dual-core personal computer was
by the total voxel numbers. Figures 3(a) and 3(b) show the used and the processing time was about 30 seconds for the
top view and the front view of the final three-dimensional 512 × 512 image.
mesh after the reconstruction. Using the manual measurement, the only information
we could infer was that the radius size would be about
75~135 nm and the average radius would be about 100 nm.
Figures 5(a) and (b) depict the fiber statistics. To get the
same graph with the manual measurement, hundreds of
TRJ 1000 Textile Research Journal 80(11)

Figure 4 Results for various SEM


resolutions (from left: original
image; fiber groups; metaballs;
three-dimensional mesh): (a)
×20,000 (lower resolution); (b)
×200,000; (c) ×2,000,000 (higher
resolution).

iterations would be necessary. As more fibers participated revealed that the radius distribution did not vary, but the ori-
in the calculation compared to the manual measurement, entation graphs showed a different result from Figure 5(b).
the resultant statistical data must be closer to the true val- Such a slight difference cannot be easily identified through
ues. a few repetitions of the manual measurement. The advan-
It is apparent that the volume fraction can also be calcu- tage of the automatic measurement is clearly shown here, in
lated, because the total volume and the fiber volume are that the more general statistical values can be known com-
known from the voxel data. However, we were not able to pared to the manual measurement, although our method
find a material whose exact microscopic volume fraction was confined to the top surface of the specimens. The pore
was known. Instead, the mean flow pore size was chosen to size error with the experimental value was about 12% in
verify the results. The pore size distribution of the this case, under the assumption that the capillary flow
nanofiber membrane was determined using a capillary flow porometer gave a true value of porosity. We postulate that
porometer (Porous Materials Inc., Model CFP-1200AE). the error was caused by using only the surface fiber infor-
As one can see in Table 1, the experimental value of the mation. The error was a little bigger than that in the
mean flow pore diameter was 0.6227 μm and this is similar #PUM1, but we succeeded in acquiring a detailed distribu-
to the result in Figure 5(c) with about 6% error. The pore tion of the pore size in a two-dimensional graph form,
size was also easily calculated by inverting the image of while the porometer gives only the mean and standard
Figure 2(d) and making metaballs for pores. The theoreti- deviation values.
cal pore diameter was chosen as the maximal radius of the The shortcoming of using the high resolution SEM is
pore regions. The result shows that the graph from the pro- that the statistics represent only a small range of the area
posed method may be said to coincide with the expected from the specimen. To get the more general statistics, there
value. are two alternatives. One approach is to carry out the
The same procedure was applied to specimen #PUM2 above-mentioned procedure repeatedly over different
and the results are shown in Figures 6–8. The manual regions. This does not require a special technique but
measurement values were similar to those of #PUM1. The needs only additional SEM filming. The other approach is
final three-dimensional mesh was also found and it was to use low-resolution images. It was not possible to gener-
Automatic Volumetric Measurement of Nanofiber Webs using Metaball Approximation Based I. H. Sul et al. 1001 TRJ

Figure 6 Example of three-dimensional reconstruction


from two-dimensional SEM image for sample #PUM2.

ate valid metaballs at the low resolution as shown in Fig-


ures 4(a) and (b) because the fiber segments were too thin,
but the center lines could be found if a sufficiently larger
size of the input image was used (Figure 9). Center lines
cannot give the radius distribution, but they can offer
knowledge of fiber orientation instead. Combining the
radius distribution result acquired from the high-resolution
image with the fiber orientation acquired from the low-res-
olution image can give more general statistical values for
the given specimens. The main concern of this paper was
to replace the manual measurement with a simple auto-
matic one using a single SEM image, so the verification of
the alternative methods was skipped.

Conclusions
This paper presents a new methodology based on the
Figure 5 Radius distribution and fiber orientation graphs image analysis for measuring fiber statistics from a single
for sample #PUM1: (a) radius distribution; (b) fiber orien- SEM image. The main contribution is that researchers can
tation; (c) pore diameter distribution (mean = 587.69 nm, have more detailed and general distribution information of
st.dev. = 15.9 nm). the fiber radius, the orientation, and the pore size com-
pared with the manual measurement. Unlike the FIB-
TRJ 1002 Textile Research Journal 80(11)

Figure 7 Three-dimensional reconstruction result for


sample #PUM2.

SEM-based method, our method needs only one SEM pic-


ture and the complex fiber webs are decomposed into mul-
tiple fiber groups. The original two-dimensional image is
approximated with sphere-like metaballs after the image
analysis steps. Three-dimensional voxels and mesh are con-
structed from the metaballs for visual representation and
also fiber statistical data are acquired from the metaball
information. The fiber orientation, the fiber radius distri-
butions, and the pore diameter distributions are directly
known from the metaballs and the voxels. Our method is
simple, fast, automatic, and more advantageous than the
manual measurement using the stereology because it can Figure 8 Radius distribution and fiber orientation graphs
also give fiber orientation results. It is also advantageous to for sample #PUM2: (a) radius distribution; (b) fiber orien-
the FIB-SEM method in that the individual fiber is identi- tation; (c) pore diameter distribution (mean = 521.1 nm,
fied, which is an essential step for calculating the statistical st.dev. = 11.1nm).
information.
Automatic Volumetric Measurement of Nanofiber Webs using Metaball Approximation Based I. H. Sul et al. 1003 TRJ

3. Sidhu, R. S., and Chawla, N., Three-dimensional Microstruc-


ture Characterization of Ag3Sn Intermetallics in Sn-rich Sol-
der by Serial Sectioning, Mater. Charact., 52, 225–230 (2004).
4. Li, M., Ghosh, S., Richmond, O., Weiland, H., and Rouns, T.
N., Three Dimensional Characterization and Modeling of
Particle Reinforced Metal Matrix Composites: Part I: Quanti-
tative Description of Microstructural Morphology, Mater. Sci.
Eng. A-Struct. Mater. Prop. Microstruct. Process., 265(1–2),
153–175 (1999).
5. Li, M., Ghosh, S., Richmond, O., Weiland, H., and Rouns, T.
N., Three Dimensional Characterization and Modeling of
Particle Reinforced Metal Matrix Composites Part II: Dam-
age Characterization, Mater. Sci. Eng. A-Struct. Mater. Prop.
Microstruct. Process., 266, 221–240 (1999).
6. Bhandari, Y., Sarkar, S., Groeber, M., Uchic, M. D. Dimiduk,
Figure 9 Skeletonized center lines of sample #PUM2 D. M., and Ghosh, S., 3D Polycrystalline Microstructure
(×3000 SEM, 1024 pixels × 1024 pixels). Reconstruction from FIB Generated Serial Sections for FE
Analysis, Comput. Mater. Sci., 41, 222–235 (2007).
7. Schroeder-Reiter, E., Pérez-Willard, F., Zeile, U., and Wan-
ner, G., Focused Ion Beam (FIB) Combined with High Reso-
However, there were also limitations. Low-resolution lution Scanning Electron Microscopy: A Promising Tool for
SEM images were not applicable because the overlapped 3D Analysis of Chromosome Architecture, J. Struct. Biol., in
fibers could not be easily segmented. It was only possible to press (2008).
get their skeletonized center lines by using a larger image 8. Samak, D., Fischer, A., and Rittel, D., 3D Reconstruction and
file, which required more processing time and might con- Visualization of Microstructure Surfaces from 2D Images,
tain a greater number of noises. We leave the case of the CIRP Annals – Manufacturing Technology, 56(1), 149–152
low-resolution image as a further work. (2007).
9. Paluszyński, J., and Slówko, W., Surface Reconstruction with
the Photometric Method in SEM, Vacuum, 78(2–4), 533–537
Acknowledgements (2005).
10. Lee, T., Kashyap, R. L., and Chu, C., Building Skeleton Mod-
This research is financially supported by the Ministry of els via 3-D Medial Surface/Axis Thinning Algorithms, Graph.
Knowledge Economy (MKE) and Korea Institute for Models Image Process., 56(6), 462–478 (1994).
Advancement in Technology (KIAT) through the Work- 11. Canny, J., A Computational Approach to Edge Detection,
force Development Program in Strategic Technology. IEEE Trans. Pattern Anal. Mach. Intell., 8, 679–714 (1986).
12. Heckbert, P., A seed fill algorithm, in “Graphics Gems”, Aca-
demic Press, Boston, 1990.
13. Blinn, J. F., A Generalization of Algebraic Surface Drawing,
References ACM Trans. Graph., 1(3), 235–256 (1982).
14. Lorensen, W. E., and Cline, H. E., Marching Cubes: A High Res-
1. Russ, J., and Dehoff, R. T., “Practical Stereology”, 2nd edn, olution 3D Surface Construction Algorithm, ACM SIGGRAPH
Plenum Press, New York, 1999. Comput. Graphs., 21(4), 163–169 (1987).
2. Principe, E.L., How to Use FIB-SEM Data for 3-D Recon-
struction, R&D Magazine, 6, 29–30 (2005).

You might also like