Image Analysis for Particle Size Measurement
Image Analysis for Particle Size Measurement
R_ P_ KING
Department of Metallurgy. University of the Witwatersrand. Johannesburg (South Africa)
(Received August 18.1983; in revised form January 1.1984)
PARTICLE SIZE AND PARTICLE SHAPE dimensional section through a particle pro-
duces a plane section that has very little direct
The basic principle by which particle size relationship to the size of the particle from
distribution is measured using the image anal- which it came_ A random section has very
yser is the measurement of size- and shape- little chance of including the maximum
related parameters on individu-al particles as chord through the particle or any other
seen in an appropriate two-dimensional directly useful measure of size. Nevertheless,
image. The image can be obtained by pro- the distribution of measured parameter over
jection of the shadow silhouette of the par- the population of particle images in section
ticle or by making a plane section through can be converted to the distribution of particle
the particle after the particles have been sizes over the population of particles. It is
mounted suitably. Most image analysers gen- this conversion procedure that is the major
erate the image in the optical microscope so topic of this paper_
that the most common range of particle sizes Before describing the conversion proce-
measured by this technique is 1 to 1000 pm. dure, it is necessary to discuss briefly the sig-
However, macro optical systems can be used nificance of particle shape because the conver-
for larger sizes and it is feasible to consider sion from section or silhouette measurements
smaller sizes using the electron microscope to size is shape dependent_ This is a source of
as an imaging device_ Because measurements some difficulty because shape is considerably
are made on individual particles, the precise more complex to define than size and it is
definition of particle size is somewhat more only in comparatively recent times that much
significant with this technique than with progress has been made on the development
many others where one or other ‘convenient of the quantitative description of particle
size’ is’measured indirectly via a sedimenta- shape [ 2,3]_ Two attributes of particle shape
tion velocity, sieve separation, per.3eability are of particular interest to us: regularity and
of a powder bed and so on_ Particles of irregu- homot.heticity. Particles that have regular
lar shape do not %YP a single uniquely de- shape are the sphere, the ellipsoid, the cube
fined size and the measure r?sed for size will and a variety of other regular polygons_
be influenced to some estent by the use to Homothetic particles have the same shape and
which the measurement will be put. Allen [l] differ only in a single scale factor as the size
(p_ 104) lists 13 different definitions for the changes. Collections of particles having a
size of an irregular particle. Still others can be single regular shape are obviously homothetic-
defined. Thus, it is always necessary to speci- Early workers in the application of image
fy at the outset just which measure of size is analysis to the problem of size determination
required_ &I additional advantage of the restricted their attention to regular or, at best,
image analysis technique is that it can yield homothetic particles_ This was a severe restric-
virtually any of the measures of size that may tion since the vast majority of particles that
be appropriate. Contrast this with the sieve make up technically important granular mate-
method, for example, which can yield only rials do not satisfy either of these properties_
the sieve.size distribution_ This restriction probably accounts for the
The image analysis method is complicated relatively infrequent use of image analysis for
because the an$yser does not directly measure particle size determination until recent times.
size of the particles in spite of the fact that The early work has been well reviewed by
individual particles in the image are examined. Underwood 143 _ The extension to particles of
Generally speaking, some size-related proper- irregular shape is the subject of this paper_
ty of the particle image is measured on several
thousand particles and the distribution of
measured parameter must be converted to the CONVERSION FROM SECTION DATA TO SIZE
distribution of size over the particle popula- DISTRIBUTION
tion as a whole_ It is not possible to relate
the measured parameter for a single particle It is now necessary to examine some of the
image to the size of that particular particle_ details of the transformation relationships
This is especially true when the image is made that exist for the conversion of sectioned data
by sectioning through the particles_ A two- to three-dimensional size data. It must not be
251
forgotten that there are no direct relation- across a comer, the sectioned area will be
ships for individual particles_ All relationships small_ On the other hand, if the plane passes
esist only for populations of many particles- through the centre of the particle, a large
In addition, any single measurement that is sectioned area will result_ Furthermore_
made on an irregular particle does not have a particles all of size* D drawn for the particle
great deal of absolute significance, but the population will all vary in shape and this will
measurements must be regarded as random add further to the spread of values. Thus,
variables that can be defined precisely only each particle size contributes many different
through their associated probability distri- section measurements and the real popula-
bution functions_ Modem automatic image tion of particle sizes produces two-dimen-
analysers are well equipped to handle data in sional parameter values that overlap over the
this way and they are provided with a variety entire spectrum of measured values. The es-
of statistical packages for the appropriate traction of the size distribution from the con-
analysis of the data. fusion of the section measurements is known
The problem can be simply stated as fol- as a deconvolution operation and there are
lows_ The image analyser can measure the dis- reliable mathematical and numerical tech-
tribution of any feature-specific parameter niques for this operation.
over the population of the particles in section The most convenient representation of the
as they are presented to the image-forming relationships between the distribution of
device_ How can the distribution by volume sectioned area and the three-dimensional
of the particle sizes be calculated from the distribution of particle size is
measured distribution? The measured param-
eter will generally be the cross-sectional area
of each particle in the section plane, or the (1)
masimum diameter of the particle section as
seen in the section plane, or the length of a and, for measurements of random chord
random chord through a section. Other par- length,
ticle-specific measurements that can be used
are the Feret diameter, the diameter of the
circle having the same area of the particle, or
the particle perimeter_ Each of these param-
eters carries a different amount of informa- Equations (1) and (2) show the convolu-
tion regarding the particle size distribution, tion relationship explicitly and the full
and the section area and the random chord derivation of ihese equations is given in King
are believed to be the richest in information [5]_ In the equations, p(A) is the measured
content although no quantitative theoretical distribution of sectioned areas and p(L) the
work is available on this question at the measured distribution of random chord
present time- lengths_ p(A/D) and p(L/D) are the unit. dis-
The link between the distribution of the tributions defined for particles that have size
2-dimensional section measurements and the D_ Thus p(A/D) is the dist.ribution of sec-
three-dimensional size distribution depends tioned areas that will be observed in a popula-
on two important features of the sectioning tion of particles all of which have size D with
process: firstly, there is not an equal chance a similar definition for p(L/D). f(D) is the dis-
that all particles will be sectioned by the tribution by volume of the particle sizes in
plane (larger particles are more likely to be the population and corresponds to the size
sectioned than small particles), and secondly, dist.ribution that would be measured using the
the section param eter is only indirectly re- usual sieving procedure_ ;i’ is the mean of the
lated to the particle size. If a single particle measured area distribution:
of size D is sectioned randomly many times,
it will yield a distribution of values for each (3)
of the measured parameters depending on the
position of the section plane as it cuts *In this paper. particle size means the side dimen-
through the particle. If the plane passes close sion of the smallest mesh through which the particles
to the edge of the particle or merely cuts can pass.
282
and xn is the mean of the unit distribution employed this method, but it does require
for area the separation of the particles and so cannot
be used when particles are imbedded in a solid
&= /- Ap(A/D)dA matrix. The second method that has been
0” successfully used is the synthesis of the unit
The integrals in eqns. (1) and (2) show how distributions by computer simulation of the
each particle size in the D distribution makes particle shape followed by the simulated
many contributions to the measured values of random sectioning of the particle_ The sec-
A or L and the integral represents the summa- tioned area and the random chord length is
tion over all the particle sizes. The lower limit then calculated and the distributions are
on the integrals arises because no particle generated_ This approach has been explored
having D d A “* can contribute a measured fairly extensively by Warren and co-workers
value of A and no particles having D d 0.83L [ 6 - S] , who have generated much useful in-
can contribute a measured value of L. The formation on the unit distributions for a
lower limits are specific to the shape of the variety of particle shapes, and their work has
particles under consideration_ Those used yielded valuable insights into t.he nature of
in eqns. (1) and (2) have been found experi- the unit distributions.
mentally to apply to a range of irregular King [5] has shown that a very useful unit
particle shapes. Further experimental work distribution that is applicable to particles
will be required to establish the precise re- produced by milling operations is
lationship between the lower limits and par-
ticle shape characteristics. Equations (1) and
(2) are preferred to most of the analogous
equations in the literature because eqns. (1)
and (2) apply to particles of irregular as well
as regular shape and are not restricted to Equation (5) can be used in eqn. (2) so that
spherical particles. It is this fact that now the true size distribution function f(D) can
makes image analysis an attractive techique be evaluated by deconvolution from a mea-
for the measrtrement of the particle size dis- sured distribution of the random chord
tribution of granular materials. The numeri- lengths through the particles_
cal deconvolution of eqns. (1) and (2) to ob- Equation (5) is a useful approximation for
tain f(D) from a measured estimate of p(A) or particles that have an irregular angular shape
p(L), knowingp(L/D), is easily within the with sharp comers. The form of eqn (5) is
capacity of the mini- and microcomputers compatible with the principles that have been
that are an integral part of modem image established by Warren and Naumovich [6] _
analysers. Their extensive computer simulations have
shown that different structural features of
the particles lead to definite characteristics in
the unit distribution for the random chord
UNIT DISTRIBUTIONS length. Particles with sharp comers yield dis-
tributions with a finite value at zero chord
The key to the successful application of the length. The value increases with the propor-
image analysis method is a satisfactory knowl- tion of angular edges and their acuteness_
edge of the unit distributions p(A/D) and Faceted bodies have itistributions that ex-
p(L/D). Two approaches have proved fruitful hibit frequency peaks at chord length equal
for the establishment of the unit distribu- to the distance between parallel faces with the
tions- The first is entirely experimental and peak height being related to the relative area
relies on the determination of the distribution of such faces. Particles with rounded surfaces
of A and L in a series of samples that are pre- have greatest chord frequency at the upper
pared by screening in narrow screen intervals. end of the chord size range with a steadily
The intervals are made successively narrower increasing freqr’.ency from zero chord length.
with the ratio R of upper to lower screen size The value of 2/1.20 when L = 0 in eqn. (5)
approaching 1 and the distributions are extra- reflects the presence of many angular edges on
polated to R = 1. King 153 has successfully the particles a-rd the finite value at L = 1-W
reflects the presence of some rounding of In many situations, it is not necessary to
the edges. The distribution function spans the know the complete distribution of particle
chord length intergal from 0 to 1.20 which sizes and a few low-order moments of the
reflects the fact that the greatest chord length distribution suffice_ These are easy to obtain
through a particle that has passed through a from the moments of the measured chord-
square mesh of size D is approsima*tely 1_2L)_ size distribution-
Populations of particles that are described by Let the nth order moments of the chord
eqn. (5) have the shape factor size distribution and the particle size dis-
4iiA tribution be represented by L” and D”
- = O-63 (6) respectiv_ely.
u2
L” = J L”p(L) dL (‘7)
averaged over the entire population. A is the
0
area of a random section through a particle
and U the perimeter of that section. Individ-
D” = J D”f(D) cm (S)
ual particle sections may deviate consider-
0
ably from eqn_ (6) but the population average
should be in the range 0.60 - 0.66 if the The relationships between these moments are
deconvolution procedure described here is to obtained simply by the integration of eqn_
be reliable_ This criterion is easily checked (2), making =use of eqn. (6) for the unit dis-
since the shape factor can be measured auto- tribution.
matically for each particle on modem area- e-l L-ii
scanning image analysers while the chord size p- = (9)
l-2”-‘(1 - n)P, + e-l Z
distribution is being measured_
The averaged shape factor given in eqn. (6) where P n = -0
f’ u”eeu du_ The first two mo-
varies significantly with particle shape r;xd is a ments of the particle size distribution are.
sensitive indicator of shape characteristics- given in terms of the first three moments of
Warren and Durand 191 have calculated the chord size distribution from eqn. (9) by
average values of 4nAJU2 for several regular 7
(13)
ESTIMATION OF SIZE DISTRIBUTION FROM
SECTION MEASUREMENTS and the moment relationships are
n+2 L”
The calculation procedure is detailed here Dn--l=-___~ (14)
3 L
for the conversion of chord intercept data to
the true distribution_ However, the details are The calculation of the full size distribution
presented here only for particles that satisfy is somewhat more complex, requiring the full
the restriction for section shape-factor in the solution of eqn_ (2)_ Various numerical proce-
range 0.60 - 0.66 and hence have unit shard- dures have been advanced in the literature and
size distribution given by eqn- (5)_ Spherical they all suffer to some estent from the nu-
particles represent a particular special case and merical instabiiity associated with numerical
the results are included here for comparison. deconvolution operations_ The method to be
284
described here has been shown to be robust r is the gamma function and is tabulated in
and comparatively accurate for a range of ref. lo_
particle size distributions. It is also easy to use The correction term given by eqn. (16) is
and can be conveniently programmed on the tabulated in Table 1 as a function of the
minicomputers normally incorporated in dimensionless group
automated image analysers. The solution to X XrE
eqn. (2) which gives the cumulative particle _=
Ck PI = e (>
xc b*l
==T--=
b+l (b + 1)z
(19)
+&,~),(16)
l?(b + 1) Y 1_479P
In eqns. (18) and (lQ), the relationships be-
[$ybF) tween the moments of the particle size dis-
tribution and the measured chord length dis-
I(b, y) is the integral given by
tribution as given by eqns. (10) and (12) have
been used. The correction term can be read
(17) from Table 1 at each value of D.
0
TABLE1
Valuesofthe correctionterm C(x, /3)
XlB o’f(fif 1.1 l-2 1.3 1.4 1.5 1.6 1.7 1.8 l-9
0.3 0.012 O-032 0.045 0.053 O-058 O-060 0.061 0.061 0.060
0.6 0.149 0.146 0.135 O-124 0.114 0.107 0.100 0.094 O-089
O-9 O-254 0.195 0.164 0.144 0.129 0.118 0.109 0.102 0.096
l-2 O-204 0.167 0.145 0.130 0.119 0.110 0.102 0.096 0.091
1.5 0.104 O-112 0.109 0.103 0.098 0.093 0.089 0.085 O-081
1.8 O-039 O-065 0.073 0.076 0.076 0.075 0.073 O-071 0.069
2.1 O-012 O-034 0.046 O-053 0.056 0.058 0.058 O-058 O-058
2-4 0.003 0.016 0.028 0.035 0.040 0.043 0.046 0.047 0.047
2.7 0.001 0.007 0.016 0.023 0.028 O-032 0.035 0.037 0.038
3.0 0.000 0.003 0.009 0.015 0.020 0.024 0.027 0.029 0.031
285
TABLE 2
Calculation of the particle size distribution from the measured distribution of chord lengths=’
aOnly every 16th line is shown in Table 2 to condense the table for publication.
bCalibration factor = 3.185 pm/pixel_ (A pixel is a single picture element in the digitised image_)
=C(L. p) evaluated from Table 1_
1
Features segmented in this way will often
0 -1 -1 -1 0 have boundaries that are very jagged on the
-1 -1 3 -1 -1 scale of a couple of picture elements and fre-
quently very small spurious disconnected
h= & -1 3 16 3 -1 (22) features will appear in the segmented image,
-1 -1 3 -I -1 particularly close to the edges of features_
0 -1 -1 -1 o_ These spurious features result from a wide
variety of noise-producing processes in the
See ref_ 15. mounting, sectioning, polishing, imaging and
288
10 M. Abramowitz and I. A_ Stegun. Handbook of 13 A. Rosenfeld and A. C. Kak, Digital Picture Pro-
Mathematical Functions, Dover Publications, 1965. cessing. Academic Press, 1976.
11 A. Savitzky and M. J. E. Golay. Analyt. Chem.. 14 W_ K_ Pratt, Digital Image Processing. Wiley.
36 (1964) 1627. 1978.
12 H. Fischmeister, Digital image analysis in quanti- 15 H. Schwarz, ZBAS ZZ Reference Manual Release
tative metallography_ Computers in Materials 2.0. Kontron Bildanalyse GrnbH 8057, Aching
Technology. Pergamon, 1981, pp_ 109 - 129. b. Miinchen. 1982.