This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

materials

Michel Coster, Jean-Louis Chermant

*

LERMAT, URA CNRS 1317, ISMRA, 6 Boulevard Mar´ echal-Juin, 14050 Caen Cedex, France

Abstract

The scope of this paper is to present the main tools of image analysis to investigate materials and, specially, civil engineering ones.

First the acquisition methods are described. The dierent operators for ®ltering, segmentation and binary image processing are

presented and illustrated on dierent images. The in¯uence of the observation ®eld on these operators and the bias correction is also

introduced. Then the problem of the parametrical characterization is presented: stereological parameters and functions related to size

distributions, dispersion and anisotropy. Finally, the model methods based on image analysis are recalled. Some annexes illustrate

this paper to precise main basic notions to understand the morphological tools. Ó 2001 Elsevier Science Ltd. All rights reserved.

Keywords: Automatic image analysis; Image processing; Mathematical morphology; Acquisition; Anisotropy; Dispersion; Extraction; Filtering;

Granulometry; Measurements; Modelling; Morphological parameters; Segmentation

1. Introduction

Image processing and analysis are very powerful tools

when morphological information is required to under-

stand the behaviour of a material, or to help a physician

in its diagnosis, or to sort out objects, etc. Civil engi-

neering materials are also concerned with this type of

characterization. In this paper, we shall brie¯y present

the main principles of image processing and analysis

applied to (civil engineering) materials, their constraints,

their procedures, and the main tools to be used. This

must help in the understanding of the various papers of

this special issue on image analysis applied to civil en-

gineering materials.

2. Main principles

Image analysis applied to materials, and more spe-

cially civil engineering ones, can be divided into three

main steps:

· image acquisition,

· extraction of the pertinent characteristics and image

segmentation,

· extension of this information via parameters or func-

tions, and/or modelling.

During such analysis, problems of statistical and

stereological nature have to be solved.

Statistical problems are related to the size of the images

compared to the structure [1], their position and number.

Due to the observation scale used, the morphology of a

material presents generally a random character, and can

only be analysed in a frame of measurements of reduced

size. This frame has to be large enough to observe the

largest features and the size of the pixel small enough to

detect the smallest details. Acorrect magni®cation will be

the result of a compromise. Thereafter the minimum

number of frames of measurements to be analysed must

be determined and a sampling plan chosen. For this

purpose geostatistic methods can be used [2].

Stereological problems are also to be solved as in situ

analysis (3D). Generally, this is not possible on such

materials as one mainly works on a plane section (2D).

So stereological parameters and relationships can be

used to access 3D characteristics from 2D analysis [3±5].

3. Sampling preparation and choice of the image source

The sampling preparation and the choice of the image

source will depend on the nature of the material: either

bulk or particulate material.

Cement & Concrete Composites 23 (2001) 133±151

www.elsevier.com/locate/cemconcomp

*

Corresponding author. Tel.: +33-231-452-664; fax: +33-231-452-

660.

0958-9465/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.

PII: S 0 9 5 8 - 9 4 6 5 ( 0 0 ) 0 0 0 5 8 - 5

3.1. Bulk materials

Regarding the probe for bulk materials, one must

know if the material is opaque or transparent to the

photons (visible light or X-rays) or electrons (scanning,

transmission electron microscopies). Civil engineering

materials are relatively opaque to photons of the visible

wavelength, but X-rays can penetrate a certain thick-

ness. Using an optical microscope one can use either a

thin slice as in petrography, or work by re¯ection on a

polished section. Then an image of a 2D section is ob-

tained from a 3D structure. A 3D reconstruction from

X-rays observation can be performed using tomography

equipments as for biology, but the X-ray penetration

remains low and this type of investigation is rarely used

[6±9].

Of course, the two main modes of acquisition are

scanning electron microscopy (SEM) and optical sys-

tems using visible light. For this last mode, many types

of equipments exist and their choice will depend on the

magni®cation required. For higher magni®cation, a

black and white or colour camera connected to a mi-

croscope will be utilized.

Macrographic bench or a ¯at scanner are used for the

macroscopic scale. The progress in multimedia systems

leads to the development of high performance scanners

with more than 4000 pixels per inch: they compete with

the optical microscopes in the ®eld of low magni®cation

with a size of image much larger. In both cases colour or

grey tone images can be used.

Regarding the SEM, several types of signals can be

used: secondary, back scattered electrons, X-ray im-

ages, etc. Using a numerical SEM, several images of

the same area are obtained, corresponding to multi-

modal images. This can enable to select, for example,

speci®c phases and to perform a threshold without

diculties [10].

3.2. Particulate materials

If the analysis is performed on a particulate ma-

terial, the morphological investigation will mainly

concern the outline of the particles, as they have to be

set on an adequate support. However the lack of

depth of focus of the optical microscopes (confocal

microscopy is little used in material science) is to be

taken into account. With optical microscopy, one only

accesses re¯ected or transmitted images to analyse the

2D silhouette of particles. The observation by SEM is

only by re¯ection. Due to its depth of focus, more

morphological information can be attained than by

optical microscopy. Nevertheless, it is a pseudo-relief

which is analysed, which limits this type of charac-

terization.

To investigate the internal structure of a particulate

material, it must be embedded in a resin to obtain a

polished section.

4. Acquisition and nature of images

The acquisition of an image corresponds to a trans-

formation of a real image in number: so, a numerical

image is obtained the nature of which will closely de-

pend on its source. Practically two types of images must

be distinguished: grey tone images and colour images

classically de®ned from three channels. SEM multimo-

dal images are closed to colour images, as several

functions can be associated to each pixel. Nevertheless,

the comparison cannot go any further as the colour

image is in fact de®ned in a vectorial space (see Ap-

pendix A): that will possibly be taken into account in the

image processing.

5. Extraction of features and image segmentation

The aim of this step is to improve the quality of im-

ages or to evidence some speci®c characteristics. That

includes the parameter extraction and the restoration of

images, which are noisy, or with defects.

So a ®ltering process must be used to improve the

quality of such images.

The segmentation consists of constructing a symbolic

representation of the images, i.e., to de®ne a cartogra-

phy of the image to describe the homogeneous areas

according to one or some attributes (features) chosen a

priori.

5.1. Filtering

Whatever is the acquisition mode, image processing

must be used on the one hand to eliminate the noise

related to the acquisition system, and on the other hand

to increase the dierence in the contrast between the

phases or the object classes which are to be extracted.

Generally, noise appears for a frequency much larger

than that of the pertinent information. So low-pass ®lter

will eliminate that noise.

Sometimes it is more interesting to extract the con-

tour (outline) of objects than the objects themselves.

Then it will be necessary to enhance these outlines which

are generally recognizable by their large variation in the

colour or the grey tone levels. They are high-pass ®lters.

Outside their role in the elimination of noise or en-

hancement of outlines, ®lters can be classi®ed as a

function of their mathematical properties. One mainly

distinguishes:

134 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

· linear ®lters,

· rank ®lters,

· morphological ®lters.

5.1.1. Linear ®lters

Linear ®lters possess a characteristic of additivity

(hence their names), speci®c to their mathematical

structure. Convolution products and ®ltering by Fourier

transform (FT) belong to this class [11].

The convolution product is de®ned by:

g(i) = f (i) + v(i) =

¸

n

j=0

f (i ÷j)v(j)

¸

n

j=0

v(j)

for

¸

n

j=0

v(j) = 0; (1)

g(i) = f (i) + v(i) =

¸

n

j=0

f (i ÷j)v(j)

for

¸

n

j=0

v(j) = 0: (2)

In these relationships, f (i) is the value of the pixel i of

the image source, and g(i) that corresponding to the

®ltered image. j is a classi®cation index of the coecient

v(j) of the neigbourhood function. When the sum of

these coecients is nul, high-pass (gradient or Lapla-

cian) ®lters are de®ned. Generally, the neighbourhood

function corresponds to 3 ×3 or 5 ×5 matrix. Fig. 1(a)

presents a composite structure and Fig. 1(b) the eect of

a simple mean ®lter of 5 ×5 size on image 1a.

5.1.2. Rank ®lters

With rank ®lters, values of pixels (i ÷j) close to

common pixel i of image f are classi®ed. Thus a dis-

tribution is de®ned and a characteristic position dis-

tribution is attributed to the pixel i of image g.

Generally the median value is taken (®ltering by the

median), i.e., that which corresponds to the grey tone

level such as the number of darkest and brightest pixels

in the investigated neighbourhood will be the same.

These ®lters do not possess particular properties, but

Fig. 1. (a) Metallographic surface of a composite. (b) Eect of a 5 ×5 mean ®lter on image 1a. (c) Eect of an opening ®lter on image 1a. (d) Filtering

by reconstruction of the opened image by the closed one of size 1 on image 1a.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 135

they are relatively ecacious. When the maximum or

the minimum value of the distribution is chosen, rank

®lters are equivalent to dilations or erosions. Fig. 2(a)

shows the eect of a median ®lter on image 1a.

5.1.3. Morphological ®lters

The methods of morphological ®ltering using opera-

tors of mathematical morphology [12,13] are divided in

two classes: morphological ®lters [14] and ®lters based

on morphological residues [15]. The standards of mor-

phological ®lters are the morphological opening and

closing. These operators are themselves constructed

from the two main basic operators of mathematical

morphology: erosion and dilation [5,12], (see Appendix

B). The morphological opening is the erosion of the

function f by a structuring element B of size k, followed

by a dilation with the same transposed structuring ele-

ment (a structuring element in mathematical morphol-

ogy is a morphological elementary tool of a given simple

geometry (circle, triangle, hexagon, bi-point, etc.) which

is used to transform the image). It is de®ned by

g = f

B(k)

= ((f (x)

´

B(k)) B(k)): (3)

The morphological closing corresponds to the recip-

rocal operation

g = f

B(k)

= ((f (x)

´

B(k)) B(k)): (4)

These operators verify the properties of increasing and

idempotence which are characteristic of morphological

®lters (see Appendix B). These ®lters are low-pass ®lters,

but contrary to convolution products, they do not act by

a symmetrical way on the image: openings (O) suppress

the clear pics on an image (Fig. 1(c)), while closings (C)

®ll all the dark parts. From these basic elements, one can

construct ®lters which correspond to opening and clos-

ing combination: in series (alternate sequential ®lters:

Fig. 2. (a) Eect of an order median ®lter, on image 1a. (b) Eect of a Laplacian ®lter, on image 2a. (c) Eect of a black top hat transformation of

size 20 on image 2a. (d) Treatment of image 2c by a threshold between two grey tone levels (0±58) and binary morphology (erosion-reconstruction on

two phases).

136 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

for example OCOC) or in parallel with sup and inf

(morphological centres or morphological contrasts). At

last a reconstruction can also be used (Fig. 1(a)), (see

Appendix C).

Filters constructed from morphological residues

correspond to high-pass ®lters of the linear domain. The

morphological gradient [16] is de®ned by

g(x) = Mgrad(f (x); B(k))

=

(f (x)

´

B(k)) ÷(f (x)

´

B(k))

2k

: (5)

Another very useful morphological residue is the top-

hat transformation [17]: it is made from the dierence

between the initial image and the opened image (``white

hat'') or between the closed and initial image (``black

hat''), (Fig. 2(c)). The corresponding relationships are:

WTop hat

B(k)

(f (x)) = f (x) ÷f

B(k)

(x); (6)

BTop hat

B(k)

(f (x)) = f

B(k)

(x) ÷f (x): (7)

The top-hat transformation is also generalized in re-

placing simple openings by combined morphological

®lters.

6. Image segmentation

Attributes, which enable to de®ne conditions of seg-

mentation, can be related to textures, colours, grey tone

levels or topography. They are associated either with the

area itself or with its outline. In the ®rst case one will

speak of orientated region segmentation and in the

second one of orientated contour segmentation.

6.1. Threshold

The threshold is the most simple and useful case of

segmentation utilized in material science. If the case of

colour images is, for instance, excluded, then threshold

corresponds to the transformation of a grey tone image

in a binary one. This step is extremely important as it

corresponds to the maximum loss of information. Put

simply, one can say that the ®le is eight times smaller. So

one must pay a lot of attention to the choice of the

method to be used. Two main categories can be de®ned:

· threshold from histogram in grey tone levels,

· threshold by contrast analysis.

The ®rst category de®nes regions as a function of the

grey tone level, and the second one mainly concerns the

transition of grey tone levels on the outlines of the re-

gions.

6.1.1. Threshold from grey tone level histogram

Several methods of threshold belong to this class.

6.1.1.1. Manual threshold or threshold by prede®ned lev-

els. In this case the attribute of segmentation is a domain

of grey tone level, de®ned by an upper and lower value.

All image pixels with grey tones belonging to this do-

main will enable to create a binary image representing

the set X which must be extracted. It can be de®ned by

the following expression, where X is the union of the

pixels of this attribute:

X = ¦x : max Pf (x) P min¦: (8)

For a same image, several domains can be de®ned. A

binary plane will correspond to each domain (Fig. 2(d)):

that is the multiple threshold. The upper and lower

thresholds can be de®ned by the operator or calculated

from a threshold algorithm. In the following para-

graphs, the most useful methods will be indicated, lim-

iting ourselves only to the determination of a unique

threshold (a multi-threshold is also possible, see Ap-

pendix D).

6.1.1.2. Threshold by maximization of the interclass

variance. In this approach, the histogram is divided in-

to a certain number of classes, P, being a priori a set.

The relationship for two classes is given in Appendix D.

When the position of the class limit varies, the value of

the variance changes and the position(s) which maxi-

mize(s) that value is (are) determined. This type of

threshold is suciently ecient when two classes are

relatively more or less equivalent [18].

6.1.1.3. Threshold by entropy maximization. In this case

the histogram is divided in several classes. The entropy is

de®ned in Appendix D. This threshold by entropy

maximization gives good results when there are few

objects [19]. An example is given in [1].

6.1.1.4. Threshold by moment-preserving. The method

proposed by Tsai [20] consists of assigning a well-de-

termined grey tone level z(j) to each class, so that the

®rst statistical moments will be identical to that of the

initial image. In the case of a system with two classes,

one uses the three ®rst moments as it is indicated in

Appendix D. This method gives good results with low

contrast images.

6.1.1.5. Threshold by contour information. The threshold

by information on the contour is also possible from an

algorithm computing the image contrast [21]. One con-

siders two points x and y of the image and their grey

tone level associated to f (x) and f (y). As one seeks for

characterizing a contour, x and y will be all the close

pixels of which the values are within the limits of the

threshold t. They de®ne a set K(t) and the mean contrast

for the threshold t is given by

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 137

¯

C(t) =

¸

K(t)

min([f (x) ÷t[; [f (y) ÷t[)

card(K(t))

: (9)

The threshold retained gives the maximum mean con-

trast. This type of threshold is ecient if the contours

are related to only one transition, so to a two-phased

system. An example is also given in [7].

6.2. Watershed: a segmentation by morphological oper-

ator

The watershed is the main segmentation method us-

ing morphological operators [16,22]. To describe how it

works, one can imagine that a grey tone image can be

assimilated to a relief. Any relief possesses lines of crests

and catchment basins, i.e., zones for which the rain

water ¯ows towards a minima. These catchment basins

are separated by a watershed line. The principle of

construction is as follows: the low parts (minima) of the

relief are ®lled (inundated) progressively but the lakes

formed are prevented from joining by constructing a

``dyke'' which is nothing else than the watershed line.

This construction method enables to classify the water-

shed among the geodesic operators (see Appendix C).

All these operators use two source images to form seg-

mentation: the minima which constitute the image of

markers (Fig. 3(c)) and the reference geodesic image

which represents the relief which will be inundated. In

that example, the binary image (Fig. 3(a)) is converted

into a grey tone image by the distance function

(Fig. 3(b)). The quality of the segmentation will mainly

depend on the choice of the markers. When the minima

are directly used, it leads to an over-segmentation. To

improve the result, extended minima will be utilized

using either a sucient depth in their neighbourhood

(h min), (Fig. 3(c)) or minima brought together by di-

lation, or a combination of these two possibilities.

Fig. 3. (a) Binary image of a sintered glass. (b) Image of the distances to the border of the binary image 3a. (c) Binary image of dilated h-minima of

image 3b (image of markers). (d) Result of the segmentation after application of the watershed method.

138 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

7. Binary image treatment

The watershed line and the segmentation by thresh-

old give binary images which can operate as they are,

but which can also require an image processing either by

®ltering or segmentation: it is often necessary to elimi-

nate artefacts, to ®ll holes to separate objects which are

touching, to extract useful information for the analysis.

For this step, mathematical morphology [5,12] is a very

powerful tool. The most often useful processing will

then be presented.

7.1. Main operators

7.1.1. Binary ®ltering

Morphological ®lters used with images in grey tone

levels have their equivalent in binary images (opening,

closing, etc.). One has two main basic operators:

Y = X

B(k)

= ((X

´

B(k)) B(k)); (10)

Y = X

B(k)

= ((X

´

B(k)) B(k)): (11)

The morphological opening is most often replaced by

the opening-reconstruction which has the advantage of

not eliminating the smallest objects without modifying

the remaining objects. For that purpose, the step of the

Euclidean dilation is replaced by an in®nite geodesic

dilation of the eroded image (image of the markers) with

regard to the initial image X:

Y = X

X;B(k)

((X

´

B(k)) B

X

(·)): (12)

7.1.2. Elimination of the particles interesting the frame of

measurements

When size distributions have to be performed particle

per particle or when one has to investigate the shape of

an object in 2D, that object must be totally included in

the frame of measurements in order to avoid a bias.

Mathematical morphology enables to rapidly eliminate

objects intersecting that frame [5,12,23,24]. For that

purpose one considers the image frame as a marker M

and this frame is then reconstructed in the image X to be

analysed. Objects totally included are obtained by the

dierence between the initial and reconstructed images:

Y = X=(M

´

B

X

(·)): (13)

7.1.3. Hole ®lling

In many cases, the result of segmentation leads to

particles with holes, which has no physical meaning.

Hence using geodesic operators of mathematical mor-

phology, one must reconstruct the complementary of the

objects to be analysed from the frame already used thus

in inverting (reversing) the resultant image:

Y = (M

´

B

X

C(·))

C

: (14)

7.2. Other operators

These three previous operators constitute only a small

fraction of the possibilities of mathematical morphology

in segmentation. So it is not possible to give an ex-

haustive list. The skeleton Sk(X) and the skeleton by

in¯uence zone (SKIZ(X)) are also often used.

7.2.1. Skeleton, Sk

If X is a connected set, dX its border and s a point of

X, the skeleton X, noted Sk(X), is the union of the

centre of the maximum balls included in X. Mathe-

matically speaking it means

s ÷ Sk(X) == ¬y

1

; y

2

÷ oX; y

1

= y

2

: d(s; oX)

= d(s; y

1

) = d(s; y

2

): (15)

The skeleton by maximum disk (opened skeleton) is

obtained from an algorithm developed by Lantuejoul

[23]

Sk(X) =

¸

k>0

¸

l>0

(X k

´

B)=(X k

´

B)

lB

: (16)

Nevertheless in the discrete space, this skeleton by

opening is not topologically similar to the initial set. One

prefers to often use the skeleton by thinning (see Ap-

pendix B) which leads to object of 1 pixel thickness and

with the same topology. So it corresponds to a simpli®ed

representation of objects which can be used for a sort-

ing. This skeleton possesses characteristic points: ex-

tremities and multiple points, which can present an

interest, for example for branching information [25].

Nevertheless, one must pay attention to the use of the

skeleton as it is sensitive to noise.

7.2.2. Skeleton by in¯uence zone, SKIZ

The skeleton by in¯uence zone (SKIZ) separates an

image in several regions. Each region or in¯uence zone,

Y

i

, is associated to an object X

i

. It corresponds to the set

of the nearest points of that object. Its mathematical

de®nition is given by

Y

i

=

¸

(y : d(y; x

i

) < d(y; x

j

); x

i

÷ X

i

and x

j

÷ X

j

): (17)

The complementary set of these regions is the SKIZ:

SKIZ =

¸

i

Y

i

¸ ¸

C

: (18)

For example, this operator is used to correctly represent

the grain boundary network. It can also be used to in-

vestigate the neighbourhood of objects [24]: it can be

very important to know, for example, the environment

of certain crystals or cells in biology [26]. In fact it is the

binary version of the watershed line.

An illustration of these dierent operators is given by

the sequence of the images in Fig. 4.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 139

Sorting of dendritic particles from others can be

made by selecting the triple points of the pruned skele-

ton (centre of the skeleton cross of image 4d). Using

these triple points as markers, only the dendrites are

reconstructed.

To investigate cracks in concrete, it is necessary to use

the skeleton function after their detection, as it can be

seen in several papers of this issue [7,27,28].

7.3. Problem of local knowledge

As it has been already said, the structure of materials

is always known in a frame of a size smaller than that of

the plane section observed. The use of dierent opera-

tors of ®ltering or of segmentation introduces a bias on

the borders of the image. To be rigorous one must

eliminate certain zones as around the image. Mathe-

matical morphology allows to solve this problem in

de®ning the image zone (surface area) which is known

without a bias. Hence, if one uses a local operator

(morphological operator or convolution window

(mask), the result will be known without bias in a frame

Z

/

deduced from the initial frame (mask) by an opera-

tion of erosion with the structuring element B associated

to the morphological operator or to the ®ltering func-

tion of the convolution product. This rule applies

Z

/

= Z B the theorem of the frame (mask) of mea-

surements [12].

8. Measurements

After the segmentation and binary image processing,

one can express the information contained in the image

in parametric or function form. Characteristics related

to objects or phases can be classi®ed according to sev-

eral criteria: ratio, size, shape and dispersion of the

phases. The partial knowledge of a 3D structure through

Fig. 4. (a) Threshold image of a cast iron. (b) Cleaned image by opening-reconstruction and hole ®lling. (c) Elimination of the particles intersecting

the frame of measurements, and associated skeletons. (d) Clipped skeletons and particles reconstructed from markers.

140 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

a part of a 2D section introduces constraints on the

choice of the parameters when one wishes to access 3D

space. In this case parameters or functions only with

stereological properties must be used.

8.1. Stereological parameters

The stereological parameters [3±5] verify a certain

number of properties, as proposed by Hadwiger [29] in

the case of sets corresponding to a ®nite union of convex

bodies. These properties can easily be explained, par-

ticularly in the case of stationary systems.

If W (X) is a measurement on the set X, this mea-

surement must not depend on the position of X. One says

that W (X) is invariant under translations and rotations.

As any optical system enables one to obtain images of

various magni®cations, the measurement must not de-

pend on the magni®cation. It is then said that the

measurement is compatible with similarities.

Although any experimentalist would try to work on

the clean image, residual noise remains. So the mea-

surement must mathematically be robust. That can be

guaranteed if the measurement veri®es the property of

continuity. The stability will still be possible if the

measurement is upper semi-continuous or lower semi-

continuous.

As the measurement is made only on a part of the set,

the magnitudes will be estimated by means, and conse-

quently by sums. In the case of sets, the measurement

will verify the C-additivity. This last property is a lim-

iting condition. So it can be shown that a numbering of

objects does not correspond to a C-additive measure-

ment.

In the case of a stationary system and to be inde-

pendent on the size of the frame of measurements, the

measurements will be given as a function of the frame

size of the de®ned space: this will be termed local pa-

rameters opposed to global or absolute parameters.

In the space with n dimensions, there only exist n ÷1

basic parameters, which verify these properties. They are

gathered in Table 1. Some remarks can be made.

The norm (integral of mean curvature in R

2

) is gen-

erally called integral of mean curvature. To each point

of the boundary corresponds two main curvatures. If

these two curvatures are positive, the surface is convex.

If they are negative, the surface is concave. If the main

curvature is positive and the other one negative, one has

a surface like a saddle: then for each point one can de-

®ne a mean curvature which is integrated to obtain the

norm.

The speci®c connectivity numbers in R

3

; R

2

and R

1

are the Euler±Poincar´e characteristics (EPC), [12,30]. In

R

1

it corresponds to the number of segments per unit

length, intersected by a straight line D on the object X.

In R

2

it corresponds to the number of objects less the

number of holes contained in these objects per unit

surface. In R

3

, it is more complicated as this number is

calculated from the topological nature of surfaces

making the objects (number minus genus).

There exist stereometric relationships between the

parameters of a same column [3±5]. For a stationary set,

there is:

V

V

(X) = A

A

(X) = L

L

(X) = P

P

(X); (19)

S

V

(X) = 4N

L

(X); L

A

(X) = pN

L

(X); (20)

M

V

(X) = 2pN

A

(X): (21)

One notes that the Euler±Poincar´e characteristics are

only accessible in their own space. From these basic

parameters, secondary parameters, which verify the

same properties, can be de®ned. The most used pa-

rameters are related to the size: there are the mean free

path, L

1

(X), and the mean surface area of objects

without holes,

¯

A(X

i

). They are calculated from

L

1

(X) =

V

V

(X)

N

L

(X)

and

¯

A(X

i

) =

V

V

(X)

N

A

(X)

: (22)

Table 1

Stereological parameters in the local case

Space Parameters of quality Measure of boundary Measure of curvature

R

3

Volume fraction Speci®c surface area Speci®c norm Speci®c connectivity

number

V

V

(X) S

V

(X) M

V

(X) N

V

(X)

R

2

Surface fraction Speci®c perimeter Speci®c connectivity

number

A

A

(X) L

A

(X) N

A

(X)

R

1

Linear fraction Speci®c connectivity

number

L

L

(X) N

L

(X)

R

0

Point fraction

P

P

(X)

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 141

It is also to be noted that the measure of N

L

(X) is a

function of the orientation of the analysis straight line D.

Several examples are given in this issue and more par-

ticularly in [7,31].

8.2. Stereological parameters and frame of measurements

The problems of local knowledge also exist in the case

of stereological measurements. Two approaches are

possible to obtain results without a bias. In the ®rst case,

one can consider that the measurement is associated to a

hit or miss transformation. In these conditions, the the-

oremof the frame of measurements will alsobe applied. In

the second case, one will use the notion of shell [32]. For a

measurement of the connectivity number, N

2

(X), the shell

Q will be made of two adjacent sides of the rectangular

mask Z. The non-biased value of N

A

(X) will be given by

N

A

(X) =

N

2

((X ¨ Z) Q) ÷1

A(Z)

: (23)

In the case of the connectivity number N

L

(X), the

frame of measurements Z

/

is a segment of line (D ¨ Z)

and the shell Q

/

is one of the extremities of this segment.

The non-biased measurement is given by

N

L

(X) =

N

1

((X ¨ Z

/

) Q) ÷1

L(Z

/

)

: (24)

8.3. Parameters related to individual analysis

When the phase X corresponds to many objects iso-

lated one from the others, measures of size (Feret di-

ameters, etc.), of shape (as the ratio length/thickness or

surface/(perimeter

2

), etc., can be performed [3±5]. But all

these parameters do not possess any stereological char-

acter, which limits their use only to a comparison and

not to an estimation of some 3D characteristics of the

material.

8.4. Morphological functions

The mean stereological parameters are often insu-

cient to access a precise evaluation of the morphological

characteristics. Thus very dierent structures can give

the same value of the mean free path. So a more accu-

rate description can be obtained using functions de-

pending on a parameter.

One can distinguish:

· the granulometric distribution functions to character-

ize the size,

· the covariance functions to characterize the phase

dispersion,

· the distance functions to also characterize the phase

dispersion but by another approach,

· the rose of intercepts or of directions to characterize

the structure anisotropy.

8.4.1. Granulometric functions

Granulometric functions are the most useful. To

undertake granulometric measures, there exists two

main categories of methods: measurement by individual

analysis or by morphological opening.

8.4.1.1. Individual granulometric analysis. Individual

granulometric analysis is very well known: the size of

each object is measured. Then they are classi®ed ac-

cording to their size to establish the distribution. To

perform such an analysis, each object must be totally

visible. So one is constrained to eliminate objects inter-

secting the frame of measurements, which introduces a

bias as the largest objects are (most often) eliminated

from the measure. To correct this bias, one assigns to

the counting a coecient which is nothing else than the

reciprocal probability of inclusion of the object in the

frame. This probability is easily obtained from the ratio

surface area of the frame of measurements Z/rectangle

to the object [24]. An example is given in [6].

8.4.1.2. Opening granulometry. The granulometric anal-

ysis by opening is a more general method than the

previous one, as it can be used independently on the

nature of the set X. The granulometric distribution in

measure is given by

G(k) =

A

A

(X ¨ (Z B(2k))) ÷A

A

(X

B(k)

¨ (Z B(2k)))

A

A

(X ¨ (Z B(2k)))

(25)

B(k)) is a convex structuring element of size k. It can

be a square, an hexagon, a segment, etc. It can also

be performed on grey tone images (Fig. 5(a) and (b)).

In this case, the measurement is made on the volume

under the image, which replaces the surface fraction.

The use of a linear structuring element, `, leads to

linear granulometries which are particularly interesting

as the estimate in R

2

remains valid for R

3

[33,34]. Of

course erosion by a segment is sensitive to the orien-

tation a of the segment `. That will be useful to make

estimates on the structure anisotropy. One can obtain

granulometry in number and in measure directly

from the erosion which gives the function P(`),

(probability that a segment ` is included in the

objects). One has

P(`) = A

A

((X `) ¨ (Z `)): (26)

It induces

G(`) =

P(0) ÷P(`) ÷`P

/

(`)

P(0)

for granulometry in measure; (27)

F (`) =

P

/

(0) ÷P

/

(`)

P

/

(0)

for granulometry in weight: (28)

142 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

By integration of the P(`) function one obtains the star

function in R

3

which corresponds to the mean volume of

the set X seen from any point of X. One has

St

3

(X) =

4

P(0)

·

0

`

2

P(`)d`: (29)

8.4.2. Covariance function and dispersion

The covariance function is obtained by erosion of the

analysed structure by a bi-point h. One distinguishes the

(simple) covariance, for which two points at distant h

are tested on the same phase, X, and the crossed or

rectangular covariance for which each point belongs to

dierent phases, X and Y.

8.4.2.1. Simple covariance. The simple covariance is

given by

C(h) = A

A

((X h) ¨ (Z h))

= A

A

((X ¨ X

~

h

) ¨ (Z ¨ Z

~

h

)): (30)

To obtain a measure of covariance, it is simpler to take

the intersection of X and of its translated by

~

h . The

value at the origin is nothing else than A

A

(X). The slope

is equal to N

L

(X). For h ÷ · the covariance tends

towards a horizontal asymptote equal to A

A

(X)

2

. Be-

tween the origin and the value at the in®nite, the beha-

viour of the covariance is related to the state of

dispersion of the phase X. Perfectly random structures

without attraction or repulsion eect will give a regu-

larly decreasing curve. A repulsion eect between the

dierent parts of X will give a curve passing below the

asymptote, while with an attraction case it passes above.

For a periodic structure the C(h) curve will oscillate. As

for the P(`) function, the covariance function is a di-

rectional analysis related to the orientation of

~

h .

8.4.2.2. Crossed covariance. The crossed or rectangular

covariance corresponds to

C(X=Y ; h) = A

A

((X ¨ Y

~

h

) ¨ (Z ¨ Z

~

h

)): (31)

The rectangular covariance enables to investigate the

dispersion of phase X with regard to Y in polyphased

systems. But in this case the asymptote corresponds to

A

A

(X) + A

A

(Y ), while the value at the origin is null and

the slope is equal to 1=4S

V

(X; Y ). It has been used by

several authors of this issue [1,7,31].

8.4.3. Distance functions and dispersion

A more physical approach of the dispersion uses the

distance function. Several distance functions can be

de®ned. The three main functions are described in this

issue: the distance between the nearest neighbours, the

mean distance between neighbours, and the distance

function for which the value of each pixel is given by its

distance to the phase under investigation [35]. This type

of characterization is very important for materials of

civil engineering and gives rise to European and Amer-

ican standards and to many papers ([35,36]).

8.4.4. Rose and anisotropy

Anisotropy investigation can be performed using

analysis tools sensitive to orientation. If N

L

(X; a) is a

function, as it corresponds to diametral variation in the

direction a, this value can be plotted in polar coordi-

nates as a function of a; thus the rose of intercepts can

be constructed [4,5]. One example derived from this

method is given in [37].

The rose of intercepts is particularly useful when

there is one dominant orientation. When there are sev-

eral directions, it is better to use the rose of directions

which corresponds to the ratio of length of boundaries

Fig. 5. (a) Image of a sintered glass. (b) Opening of size 30 on image 5a and mask Z

/

inside of which the result is known without bias.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 143

orientated in a given direction [4,5]. An example is also

given in [7].

Unfortunately numerical images do not lend them-

selves to an angular investigation as it is dicult to

convert an image into a square or hexagonal grid ac-

cording to any angle without using interpolations which

can induce errors. The ideal is to have a rotating optical

or captor device.

9. Modelling

As it has been indicated a structure of a material can

only be characterized by a large number of parameters

or functions. Moreover, the majority of these parametric

characterizations are not of stereological nature. So the

modelling of a structure, in favourable cases, can solve

that problem.

To model a structure, several strategies are at our

disposal: (i) either one can use the physical process of

formation of the structure or associated to a physical

behaviour, (ii) or use probabilistic models as Boolean

models. The ®rst strategy is illustrated here by the paper

of Boussa et al. [38], while the second one corresponds

to the works of Dequiedt et al. [39] and Jeulin et al. [40].

In the following only the second strategy will be ap-

proached.

To observe a morphology, it requires that the set X

does not ®ll all the space. So there will be at least a

medium of two components: a set X and its comple-

mentary X

C

. Objects constituting that set can be as well

points than straight lines or any sub-sets. The obtained

set will be a random topologically closed set (random

closed set, RACS), [15,41±44]. This restriction is neces-

sary to preserve good properties, but it is not in practice

inconvenient. Thus a RACS will remain a RACS after

an erosion, a dilation.

The choice of a model depends on a certain number

of prior knowledge. In the case of bulk materials, the

number of morphologically discernable phases is the

®rst element. So there will be two main categories: mo-

nophased materials, which will be described from a

space tessellation, and polyphased materials described

from polyphased RACSs. Inside these categories other

elements must be taken into consideration such as

shape, dispersion, size distribution, etc.

9.1. Choquet's capacity

A RACS can be characterized by a probability of

events corresponding to morphological measurements

such as the probability of inclusion of a compact K in a

set or its complementary. That is the role attributed to

the Choquet's capacity, T(K), de®ned by [41,45]

T(K) = Pr(X ¨ K = O): (32)

One also can de®ne T(K) from the probability Q(K) that

the intersection between X and K is empty:

Q(K) = 1 ÷T(K) = Pr(X ¨ K = O) = Pr(K · X

C

):

(33)

As a distribution function de®nes a random variable,

the knowledge of the Choquet's capacity for a compact

K enables to completely de®ne a probabilistic model. It

is evident that all the possible compacts cannot be

tested. One will use the simplest ones: the segment of line

(the more isotropic (hexagon or square) structuring el-

ement), `, the bi-point, h, through the covariance. It can

be seen from Eq. (33) that the Choquet's capacity could

be tested in using only operations of erosion on the

considered compact.

9.2. Point Poisson process

The starting point of all probabilistic models is the

point Poisson process. The probability that n points of a

Poisson process of density h belong to a set Z is given by

Pr(Z) =

(hmes(Z))

n

n!

exp(÷hmes(Z)): (34)

This point process is the basis of all derived models

such as clusters or hard core models [46].

9.3. Classi®cation of probabilistic models

The probabilistic models can be classi®ed into several

categories:

· tessellation models,

· polyphased models.

9.3.1. Tessellation models

9.3.1.1. Voronoi tessellation. The most known tessella-

tion is that of Voronoi. To construct it, one sets points p

i

according to a Poisson process of density h. To each

point p

i

corresponds an in¯uence zone X

i

de®ned by

X

i

= ¦x : d(x; p

i

) < d(x; p

j=i

)¦; (35)

where d(x; p) is the distance from x to p. This in¯uence

zone is a convex polygon in R

2

and a convex polyhedron

in R

3

. Fig. 6 illustrates such a model. There also exists

other models such as the Poisson tessellation or the

Johnson±Mehl [47] model which subdivide the space.

Among those, only the Poisson tessellation possesses a

stereological character and is easy to test from the

Choquet's capacity.

9.3.1.2. Poisson tessellation. The space tessellation in R

2

according to a Poisson process is made from Poisson

straight lines [48]. They are constructed as follows. If D

144 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

is a straight line with an orientation between x and

x ÷dx and passing through the origin of the plane, on

that straight line one realizes a point Poisson process of

density k

2

dx. To each point one sets a Poisson straight

line perpendicular to D. In the case of an isotropic

mosaic, k is constant and the value of x is chosen ac-

cording to a uniform law of probability. The R

2

space is

then divided into an in®nite number of random poly-

gons called Poisson polygons (Fig. 7).

9.3.2. Polyphased models

The most known two-phased RACS model is the

Boolean model [49]. To each point of the Poisson pro-

cess of density h, a primary grain is set. The Boolean

model is the union of all these primary grains. The main

morphology will largely depend on these primary grains

which opens up many possibilities (Fig. 8(a) and (b)).

The Boolean scheme models the same properties in a

sub-space (stereological characteristic). Moreover, if X

/

is a primary grain of the Boolean scheme, one has

Q(K) = exp(÷hE(mes(X

/

¯

K))): (36)

Thus it is easy to test it by the Choquet's capacity.

Other polyphased models exist, but they are beyond

the scope of this general paper [44]. One can quote, for

example, dead leave models [50] and Stienen's model

[51].

10. Concluding remarks

The scope of this paper was to give a short overview

of automatic image analysis, and on mathematical

morphology, in order to facilitate a better understanding

of the papers included in this special issue on the use of

that technique in the ®eld of civil engineering materials.

Other more detailed and complementary information

can be found in the main books related to this subject

[3±5,12,14,52±64].

Fig. 7. Poisson mosaic.

Fig. 8. (a) Boolean model with circular disks of normal distribution. (b) Boolean model with Poisson grains.

Fig. 6. Voronoi tessellation in R

2

.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 145

Acknowledgements

This work has been performed in the frame of the

``P´ ole Traitement et Analyse d'Images'' of Basse-Nor-

mandie (P´ ole TAI): Image Processing and Analysis P´ ole

of Basse-Normandie.

Appendix A. De®nition of images

Images are de®ned by two main characteristics:

· a support belonging to the metric space of a given di-

mension (generally 2, sometimes 3),

· one or several functions.

If x is a point of the support Z, the functions asso-

ciated to this point will be de®ned by Fi(x). When the

images have to be processed on a computer, the support

is represented by a grid of points regularly spaced. In R

2

these points are distributed according to a square or

hexagonal grid. In R

3

one generally uses a cubic grid. So

the numerical supports are represented by n lines and

m columns. That gives the dimension of the image.

The most common values today are 256 ×256 and

512 ×512.

The functions associated to each point are themselves

extended to numbers. Then dierent types of images

must be distinguished:

· binary images, related to sets where f (x) takes either

the value 1 if the point belongs to the set, or 0 if not,

· grey tone images, which can be most often coded on 8

bits (black=0 and white =255); one can also use in-

teger on 16 bits or real numbers,

· multimodal and colour images where each mode or

each primary colour will be represented by a grey

tone image.

At that level, multimodal images must be distin-

guished from colour ones. For multimodal images the

dierent modes are independent. For colour images the

primary colour de®nes a vectorial space. The most

common mode of vectorial representation is the mode

RGB (red, green, blue). Nevertheless other representa-

tions also exist. The mode HSL, where the primary

colours are replaced by hue, saturation and luminance.

The mode YIQ is characteristic of the American televi-

sion standard (Y: luminance component; I and Q:

chrominance components).

The ®ltering of colour images by operators without

good vectorial properties leads to a problem. That is the

case of many ®lters. Hopefully with image analysis, one

does not try to reconstruct the ®ltered colour image, but

to treat separately each plane to extract the pertinent

information.

Taking into account the real nature of materials

in civil engineering, one has to use sometimes such

images.

Appendix B. Elements of mathematical morphology

Examples given to illustrate mathematical morphol-

ogy will only concern binary images. The approach to

functions is made via the sub-graph. All the morpho-

logical operators are based on the notion of structuring

element, noted B

x

, when it is located in a point x of an

image.

Erosion. The eroded of the set X is a set of points x of

the image such as B

x

is totally included in X. It is written

as

X

´

B = ¦x : B

x

· X¦: (B:1)

Fig. 9 presents an example of erosion by an hexagon.

The small details (smaller than the size of B

x

) disappear.

In the case of images in grey tone levels, the de®nition of

the erosion by a ¯at structuring element attributes to a

current pixel the smallest value of pixels in the domain

de®ned by the structuring element. One has

g(x) = f (x)

´

B = inf(f (y) : y ÷ B

x

): (B:2)

Dilation. Dilation is the dual operation of erosion

with regard to the complementation. It is the set of

points x such as B

x

has a not-empty intersection with X.

It is written as

X

´

B = ¦x : B

x

¨ X = O¦: (B:3)

Fig. 10 illustrates this operator. Near objects are con-

nected and holes ®lled. In the case of images in grey tone

levels, the de®nition of dilation by a ¯at structuring el-

ement attributes to a current pixel the highest value of

pixels of the domain de®ned by the structuring element.

One has

g(x) = f (x)

´

B = sup(f (y) : y ÷ B

x

): (B:4)

As it was indicated earlier the opening and closing

operators can immediately be deduced (see Fig. 11).

Fig. 9. Initial image.

146 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

Hit or miss transformation. The hit or miss transfor-

mation is more general. The structural element B is

constituted of two sub-elements T

x

(a) and T

x

(b), re-

spectively, related to X and X

C

. The point x will be at 1

if T

x

(a) is included in X and if T

x

(b) is included in X

C

.

One has

X~B = ¦x : (T

x

(a) · X) and (T

x

(b) · X

C

)¦: (B:5)

These operators are used to extract particular points such

as extremities and multiple points of the skeleton. They

are also used with thinning and thickening operators.

Thinning. The thinning operation consists of sup-

pressing points of an object having a given neighbour-

hood con®guration. It is de®ned by the following

operator:

X (B = X=(X~B): (B:6)

It is used to obtain homotopic skeleton and the cen-

tres of objects.

Thickening. Thickening is the dual operation of

thinning: one adds to the object the points having a

given neighbourhood con®guration. It is de®ned by the

following operator:

X B = X (X~B): (B:7)

It is used to have the skeleton by in¯uence zone or to

obtain a convex hull. The thinning and thickening op-

erations de®ned for binary images have their equivalent

in grey tone images.

Appendix C. Geodesic operators

Erosion and dilation de®ned in Appendix B were

performed in the Euclidean space. The square structur-

ing element used is the extension in the digital space Z

2

of the circular disk B of size k de®ned in R

2

. If B is

Fig. 12. Comparison between Euclidean disk and geodesic disk with a

set Xr as reference.

Fig. 13. Initial sets. Reference set Xr is in grey, Y set in white.

Fig. 11. Dilation of a square of size 6 on image Fig. 9.

Fig. 10. Erosion by a square of size 5 on image Fig. 9.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 147

centred on x, then the corresponding closed disk or ball

is de®ned by

B

x

= ¦y : d(x; y)k¦: (C:1)

d(x; y) is the Euclidean distance between x and y, and B

a Euclidean disk (in R

2

) or ball (in R

3

).

Geodesic erosion and dilation are performed with

geodesic disks or balls for which the notion of Eu-

clidean distance is replaced by the geodesic distance

one. The geodesic distance is de®ned with regard to a

reference set, noted Xr. In the case of the geodesic

distance the shortest way to go from one point x to a

point y is not the segment xy , but an arc xy

_

totally

included in Xr. There exists properties associated with

the notion of distance, such as the triangular inequal-

ity. There also exists in the frame of the geodesic dis-

tance, but one adds the following property: if the path

between point x and point y is not totally included in

Xr, then d(xy) has an in®nite value. Fig. 12 presents

how a disk B can be modi®ed with regard to the ref-

erence set (see Fig. 13).

The de®nitions concerning erosion and dilation

practically remain the same for the corresponding geo-

desic operations. One only has to replace the Euclidean

disk or ball by a geodesic one:

Y

´

B

Xr

= ¦x : (B

Xr

)

x

· Y ¦; (C:2)

Y

´

B

Xr

= ¦x : (B

Xr

)

x

¨ Y = O¦: (C:3)

Figs. 14 and 15 illustrate the eects of erosion and

dilation of a set Y with regard to the set Xr (Fig. 13).

One notes that the border of the set Y is not aected by

the erosion and dilation operations when it is common

to Xr.

The main application of the geodesic operation on

sets is the reconstruction. To de®ne a reconstruction

Fig. 14. (a) Euclidean erosion of the set Y by an hexagon of size 10. (b) Geodesic erosion of the set Y by an hexagon of size 10.

Fig. 15. (a) Euclidean dilation of the set Y by an hexagon of size 10. (b) Geodesic dilation of the set Y by an hexagon of size 10.

148 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

operation one must have a reference Xr in which the set

Y will be reconstructed by a geodesic dilation of in®nite

size. It is written as

Build (Y : Xr) = Y

´

B(·)

Xr

= (Y

´

B(1) ¨ Xr)

n=·

:

(C:4)

So it can be obtained from a dilation of size 1 followed

by an intersection by Xr. The process is iterated until

idempotence, i.e., until the image remains unchanged.

This operator will be found in the case of hole ®lling,

of the elimination of particles intersecting the frame of

measurements, and with the opening operator by re-

construction.

Fig. 16 illustrates this operator.

Appendix D. Threshold and data analysis

Entropy for two classes:

H(t) = H(C(1)) ÷H(C(2))

= ÷

¸

t

i=0

n(i)

N(1)

ln

n(i)

N(1)

÷

¸

m

i=t÷1

n(i)

N(2)

ln

n(i)

N(2)

:

Interclass variance for two classes:

V (t) = (M

1

(1) ÷M

1

)

2

¸

t

i=0

n(i)

N(1)

÷(M

1

(2) ÷M

1

)

2

×

¸

m

i=t÷1

n(i)

N(2)

:

System of equations for the moment conservation:

M

k

=

¸

t

i=0

n(i)

N(1)

z(1)

k

÷

¸

m

i=t÷1

n(i)

N(2)

z(2)

k

:

References

[1] Ringot E, Bascoul A, Mowret M. Image analysis: at tool for the

characterization of hydration of cement in concrete-metrological

aspects of magni®cation on measurement. Cem Concr Comp

2001;23:201±6.

[2] Matheron G. Les variables r´egionalis´ees et leur estimation. Paris:

Masson; 1965.

[3] DeHo RT, Rhines FN. Quantitative microscopy. New York:

McGraw-Hill; 1968.

[4] Underwood EE. Quantitative stereology. Reading, MA: Addi-

son-Wesley; 1970.

[5] Coster M, Chermant JL. Pr´ecis d'analyse d'images. Paris: Les

Editions du CNRS; 2nd ed. Paris: Les Presses du CNRS, 1989.

[6] Redon C, Chermant L, Coster M, Chermant JL. Measurements

of orientation in discretized space by Fourier transform: auto-

matic investigation of ®ber orientation in a reinforced concrete.

Acta Stereol 1998;17:93±8.

Parameter Meaning

m maximum value of the grey tone

levels

M

k

moment of order k of an image

(k = 1: mean)

N number of pixels of an image

n(i) number of pixels with a grey level i

t threshold to determine

C(1) class of pixels with grey tone levels

between 0 and t

C(2) class of pixels with grey tone levels

between t ÷1 and m

N(1), N(2) number of pixels of the class C(1),

C(2)

M

1

(1),

M

1

(2)

mean of pixels of the class C(1) and

C(2)

Fig. 16. (a): Erosion (in white) of size 12 on the Y set (in grey) to obtain markers for reconstruction. (b) Result of opening by reconstruction (in

white).

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 149

[7] Chermant JL, Chermant L, Coster M, Dequiedt AS, Redon C.

Some ®elds of applications of automatic image analysis in civil

engineering. Cem Concr Comp 2001;23:157±69.

[8] Slate FO, Olsefski S. X-rays for study of internal structure

and microcracking of concrete. J Am Concr Inst 1963;(May):

575±88.

[9] Bentz DP, Martys NS, Stutzman P, Levenson MS, Garbozi EJ,

Dunsmuir J, Schwartz LM. X-rays microtomography of an

ASTM C109 mortar exposed to sulfate attack. Mater Res Soc

Symp Proc 1995;370:77±82.

[10] Quenec'h JL, Jeulin D, Coster M, Chermant JL. Approach of

liquid phase sintering process by probabilistic models. In: Jeulin

D, editor. Advances in theory and applications of random sets.

Singapore: World Scienti®c; 1997. p. 231±49.

[11] Pratt WK. Digital image processing. New York: Wiley; 1991.

[12] Serra J. Image analysis and mathematical morphology. London:

Academic Press; 1982.

[13] Chermant JL, Coster M. Role of mathematical morphology in

®ltering, segmentation and analysis. Acta Stereol 1994;13:125±36.

[14] Serra J. In: Serra J, editor. Image analysis and mathematical

morphology, vol 2: Theoretical advances. London: Academic

Press; 1988 [chapter 5].

[15] Serra J. Mathematical morphology course. Paris: Ecole des Mines

de Paris; 1998.

[16] Beucher S. Segmentation d'image et morphologie math´ematique.

Th`ese de Docteur-Ing´enieur. Paris: Ecole des Mines de Paris;

1990.

[17] Meyer F. Contrast feature extraction. In: Chermant JL, editor.

Quantitative analysis of microstructures in materials science,

biology and medecine, Pract Met 1978;S8:374±80.

[18] Otsu N. A threshold selection method from gray-level histogram.

IEEE Trans Syst Man Cybernet 1979;9:62±6.

[19] Kapur JN, Sahoo DK, Wong AKC. A new method for gray-level

picture thresholding using the entropy of the histogram. Comp

Vision Graph Image Process 1985;29:273±89.

[20] Tsai W.-H. Moment-preserving thresholding: a new approach.

Comp Vision Graph Image Process 1985;29:377±93.

[21] Kohler R. A segmentation system based on thresholding. Comp

Vision Graph Image Process 1981;15:319±38.

[22] Meyer F, Beucher S. Morphological segmentation. J Visual

Comp Image Represent 1990;1/1:21±46.

[23] Lantuejoul C. La squelettisation et son application aux mesures

topologiques des mosa`õques polycristallines. Th`ese de Doctorat of

the Ecole des Mines de Paris; 1978.

[24] Lantuejoul C. On the estimation of mean values in individual

analysis of particles. Microscopica Acta 1980;5:266±73.

[25] Coster M, Chermant L, Chermant JL. Separation of dierent

shapes by sequential image transformation. Stereol Iugosl

1981;3(Suppl):175±83.

[26] Herlicoviez D, Chermant L, Coster M, Chermant JL. Contribu-

tion of image analysis to a study of muscle changes in hypothy-

ro

`

id rats. Acta Stereol 1983;2:107±12.

[27] Ringot E, Bascoul A. About the analysis of microcracking in

concrete. Cem Concr Comp 2001;23:261±6.

[28] Ammouche A, Riss J, Breysse D, Marchand J. Image analysis for

automatic study of microcracks in concrete. Cem Concr Comp

2001;23:267±78.

[29] Hadwiger H. Vorslesungen ` uber inhalt, ober¯`ache und isoperi-

metrie. Berlin: Springer; 1957.

[30] Nagel W, Ohser J, Pischang K. An integral-geometric approach

for the Euler±Poincar´e characteristic of spatial image. J Microsc

2000;198:54±62.

[31] Dequiedt AS, Coster M, Chermant L, Chermant JL. Study of

phase dispersion in concrete by mage analysis. Cem Concr Comp

2001;23:215±26.

[32] Pinnamanemi BP, Lantuejoul C, Jernot JP, Chermant JL.

Unbiased estimation of Euler±Poincar´e characteristic. Acta

Stereol 1989;8:101±6.

[33] Haas A, Matheron G, Serra J. Morphologie math´ematique et

granulom´etries en place. Ann Mines 1967;11:735±58.

[34] Haas A, Matheron G, Serra J. Morphologie math´ematique et

granulom´etries en place. Ann Mines 1967;12:767±82.

[35] Dequiedt AS, Coster M, Chermant L, Chermant JL. Distances

between air-voids in concrete by automatic methods. Cem Concr

Comp 2001;23:247±54.

[36] Pleau R, Pigeon M, Laurencot JL. Some ®ndings on the

usefulness of image analysis for determining the characteristics

of the air-voids system on hardened concrete. Cem Concr Comp

2001;23:237±46.

[37] Stroeven P, Stroeven AP, Dalhuisen DH. Image analysis of

``natural'' concrete samples by automatic and manual procedures.

Cem Concr Comp 2001;23:227±36.

[38] Boussa H, Tognazzi-Lawrence C, La Borderie C. A model for

computation of leakage through damaged concrete structures.

Cem Concr Comp 2001;23:279±87.

[39] Dequiedt AS, Coster M, Chermant JL, Jeulin D. Towards a

model of concrete mesostructure. Cem Concr Comp 2001;23:289±

97.

[40] Jeulin D, Monnaie P, Peronnet F. Gypsum morphological

analysis and modeling. Cem Concr Comp 2001;23:299±311.

[41] Matheron G. Random sets and integral geometry. New York:

Wiley; 1975.

[42] Jeulin D. Mod`eles morphologiques de structure al´eatoires et de

changement d'´echelle. Th`ese de Doctorat `es Sciences, University

of Caen; 1991.

[43] Jeulin D. In: Jeulin D, editor. Advances in theory and application

of random sets. Singapore: World Scienti®c; 1997.

[44] Chermant JL, Coster M. Material models and model materials.

In: Wojnar L, editor. STERMAT, Sixth International Conference

on Stereology and Image Analysis in Material Science, 20±23 Sept

2000; Krakow, Poland. Krakow: Fotobit Design; 2000: p. 17±36.

[45] Choquet G. Theory of capacities. Ann Inst Fourier 1953;5:131±

295.

[46] Stoyan D, Kendall WS, Mecke J. Stochastic geometry and its

applications. 2nd ed. NewYork: Wiley; 1995.

[47] Johnson WA, Mehl RF. Reaction kinetics in processes in

nucleation and growth. Trans Am Inst Min Eng 1939;135:416±58.

[48] Miles RE. Poisson ¯ats in euclidean spaces. Part 2: Homogeneous

Poisson ¯ats and the complementary theorem. Adv Probab

1971;3:1±43.

[49] Matheron G. El´ements pour une th´eorie des milieux poreux.

Paris: Masson; 1967.

[50] Jeulin D. Dead leaves models: from spaces tessellation to random

functions. In: Jeulin D, editor. Advances in theory and

application of random sets. Singapore: World Scienti®c; 1997.

p. 137±56.

[51] Stienen H. The sectionning of randomly dispersed particles, a

computer simulation. Mikroscopie (Wien) 1980;S37:74±8.

[52] Saltykov SA. Stereometric metallography. Leipzig: VEB Deut-

scher Verlag f` ur Grund-stondustrie; 1974 in German.

[53] Saltykov SA. Stereometric metallography. Moscow: Metallurgia;

1976 in Russian.

[54] Weibel ER. Stereological methods, vols. 1±2. London: Academic

Press; 1980.

[55] Elias H, Hyde DM. A guide to practical stereology. Basel: Karger

Pub Comp; 1983.

[56] Russ JC. Practical stereology. NewYork: Plenum Press; 1986.

[57] Exner HE, Hougardy HP. Quantitative image analysis of

microstructure. Oberursel: DGM-Informationsgesellschaft Ver-

lag; 1988.

150 M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151

[58] S´axl I. Stereology of objects with internal structure. Prague:

Academia; 1989.

[59] Russ JC. Computer-assisted microscopy. The measurement and

image analysis. NewYork: Plenum Press; 1990.

[60] Chassery JM, Montanvert A. G´eometrie discr`ete en analyse

d'images. Paris: Herm`es; 1991.

[61] Schmitt M, Mattioli J. Morphologie math´ematique. Paris:

Masson; 1993.

[62] Coquerez JP, Philipp S. Analyse d'image: ®ltrage et segmentation.

Paris: Masson; 1995.

[63] S´axl I, Pelikan K, Rataj J, Besterci M. Quanti®cation and

modelling of heterogeneous systems. Cambridge: International

Science; 1995.

[64] Wojnar L. Image analysis. Applications in materials engineering.

Boca-Raton, FL: CRC Press; 1998.

M. Coster, J.-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 151

the morphological investigation will mainly concern the outline of the particles. With optical microscopy. or work by re¯ection on a polished section.134 M. SEM multimodal images are closed to colour images. To investigate the internal structure of a particulate material. which limits this type of characterization. The progress in multimedia systems leads to the development of high performance scanners with more than 4000 pixels per inch: they compete with the optical microscopes in the ®eld of low magni®cation with a size of image much larger. Acquisition and nature of images The acquisition of an image corresponds to a transformation of a real image in number: so. So low-pass ®lter will eliminate that noise. Due to its depth of focus. etc. Outside their role in the elimination of noise or enhancement of outlines. The segmentation consists of constructing a symbolic representation of the images. 3. Using a numerical SEM. the two main modes of acquisition are scanning electron microscopy (SEM) and optical systems using visible light. Using an optical microscope one can use either a thin slice as in petrography.-L. For this last mode. 4. i. J. They are high-pass ®lters. In both cases colour or grey tone images can be used. as they have to be set on an adequate support. Macrographic bench or a ¯at scanner are used for the macroscopic scale. one must know if the material is opaque or transparent to the photons (visible light or X-rays) or electrons (scanning. a black and white or colour camera connected to a microscope will be utilized.e. Nevertheless. So a ®ltering process must be used to improve the quality of such images. X-ray images. ®lters can be classi®ed as a function of their mathematical properties. the comparison cannot go any further as the colour image is in fact de®ned in a vectorial space (see Appendix A): that will possibly be taken into account in the image processing..1. back scattered electrons. 5. For higher magni®cation. Civil engineering materials are relatively opaque to photons of the visible wavelength. noise appears for a frequency much larger than that of the pertinent information.1. Then it will be necessary to enhance these outlines which are generally recognizable by their large variation in the colour or the grey tone levels. which are noisy. Regarding the SEM. but X-rays can penetrate a certain thickness. Of course.2. The observation by SEM is only by re¯ection. but the X-ray penetration remains low and this type of investigation is rarely used [6±9]. to de®ne a cartography of the image to describe the homogeneous areas according to one or some attributes (features) chosen a priori. Chermant / Cement & Concrete Composites 23 (2001) 133±151 3. for example. Nevertheless. 5. A 3D reconstruction from X-rays observation can be performed using tomography equipments as for biology. Filtering Whatever is the acquisition mode. speci®c phases and to perform a threshold without diculties [10]. Then an image of a 2D section is obtained from a 3D structure. or with defects. it must be embedded in a resin to obtain a polished section. transmission electron microscopies). image processing must be used on the one hand to eliminate the noise related to the acquisition system. as several functions can be associated to each pixel. One mainly distinguishes: . one only accesses re¯ected or transmitted images to analyse the 2D silhouette of particles. Sometimes it is more interesting to extract the contour (outline) of objects than the objects themselves. Bulk materials Regarding the probe for bulk materials. This can enable to select. Practically two types of images must be distinguished: grey tone images and colour images classically de®ned from three channels. Coster. several images of the same area are obtained. many types of equipments exist and their choice will depend on the magni®cation required. and on the other hand to increase the dierence in the contrast between the phases or the object classes which are to be extracted. more morphological information can be attained than by optical microscopy. a numerical image is obtained the nature of which will closely depend on its source. several types of signals can be used: secondary. Generally. Extraction of features and image segmentation The aim of this step is to improve the quality of images or to evidence some speci®c characteristics. corresponding to multimodal images. However the lack of depth of focus of the optical microscopes (confocal microscopy is little used in material science) is to be taken into account. it is a pseudo-relief which is analysed. That includes the parameter extraction and the restoration of images. Particulate materials If the analysis is performed on a particulate material.

that which corresponds to the grey tone level such as the number of darkest and brightest pixels in the investigated neighbourhood will be the same. Rank ®lters With rank ®lters. the neighbourhood function corresponds to 3 Â 3 or 5 Â 5 matrix. Chermant / Cement & Concrete Composites 23 (2001) 133±151 135 · linear ®lters. These ®lters do not possess particular properties. .e. values of pixels i À j close to common pixel i of image f are classi®ed. Generally the median value is taken (®ltering by the median). When the sum of these coecients is nul. (c) Eect of an opening ®lter on image 1a. (a) Metallographic surface of a composite. 5. speci®c to their mathematical structure. 1. 1 j0 In these relationships. (d) Filtering by reconstruction of the opened image by the closed one of size 1 on image 1a. and g i that corresponding to the ®ltered image. 1(b) the eect of a simple mean ®lter of 5 Â 5 size on image 1a.2. Coster. The convolution product is de®ned by: Pn j0 f i À jv j Pn g i f i Ã v i j0 v j n X for v j T 0.M. · rank ®lters. · morphological ®lters. J.1.1. high-pass (gradient or Laplacian) ®lters are de®ned.-L. 5. Thus a distribution is de®ned and a characteristic position distribution is attributed to the pixel i of image g. f i is the value of the pixel i of the image source. 1(a) presents a composite structure and Fig. Convolution products and ®ltering by Fourier transform (FT) belong to this class [11].1.. Linear ®lters Linear ®lters possess a characteristic of additivity (hence their names). Generally. Fig. j is a classi®cation index of the coecient v j of the neigbourhood function. i. (b) Eect of a 5 Â 5 mean ®lter on image 1a. but g i f i Ã v i for n X j0 n X j0 f i À jv j 2 v j 0: Fig.

hexagon. while closings (C) ®ll all the dark parts. followed by a dilation with the same transposed structuring element (a structuring element in mathematical morphol- ogy is a morphological elementary tool of a given simple geometry (circle. From these basic elements. The standards of morphological ®lters are the morphological opening and closing. Morphological ®lters The methods of morphological ®ltering using operators of mathematical morphology [12. they do not act by a symmetrical way on the image: openings (O) suppress the clear pics on an image (Fig.13] are divided in two classes: morphological ®lters [14] and ®lters based on morphological residues [15].12]. (b) Eect of a Laplacian ®lter. 1(c)). etc. bi-point. These operators are themselves constructed from the two main basic operators of mathematical morphology: erosion and dilation [5. (see Appendix B). . Chermant / Cement & Concrete Composites 23 (2001) 133±151 they are relatively ecacious. Coster. The morphological opening is the erosion of the function f by a structuring element B of size k. (a) Eect of an order median ®lter.-L. triangle. one can construct ®lters which correspond to opening and closing combination: in series (alternate sequential ®lters: Fig. When the maximum or the minimum value of the distribution is chosen. (c) Eect of a black top hat transformation of size 20 on image 2a.) which is used to transform the image).136 M. on image 1a.1. rank ®lters are equivalent to dilations or erosions. (d) Treatment of image 2c by a threshold between two grey tone levels (0±58) and binary morphology (erosion-reconstruction on two phases). These ®lters are low-pass ®lters. Fig. 2. but contrary to convolution products. 2(a) shows the eect of a median ®lter on image 1a.3. 5. J. on image 2a. It is de®ned by g fB k f x É B k È B k: 3 The morphological closing corresponds to the reciprocal operation g f B k f x È B k É B k: 4 These operators verify the properties of increasing and idempotence which are characteristic of morphological ®lters (see Appendix B).

Threshold by entropy maximization. B k f x È B k À f x É B k : 2k 5 6. being a priori a set. P. can be related to textures. one uses the three ®rst moments as it is indicated in Appendix D. which enable to de®ne conditions of segmentation. where X is the union of the pixels of this attribute: X fx : max P f x P ming: 8 For a same image. In the following paragraphs. Manual threshold or threshold by prede®ned levels. 6. 2(c)). 6. one can say that the ®le is eight times smaller.2.1. . Threshold The threshold is the most simple and useful case of segmentation utilized in material science. The relationship for two classes is given in Appendix D. the value of the variance changes and the position(s) which maximize(s) that value is (are) determined. One considers two points x and y of the image and their grey tone level associated to f x and f y. The upper and lower thresholds can be de®ned by the operator or calculated from a threshold algorithm. All image pixels with grey tones belonging to this domain will enable to create a binary image representing the set X which must be extracted. This threshold by entropy maximization gives good results when there are few objects [19]. At last a reconstruction can also be used (Fig. 2(d)): that is the multiple threshold. so that the ®rst statistical moments will be identical to that of the initial image.3.-L. 6. 6. The morphological gradient [16] is de®ned by g x Mgrad f x. They are associated either with the area itself or with its outline. Threshold by maximization of the interclass variance. several domains can be de®ned. Put simply. They de®ne a set K t and the mean contrast for the threshold t is given by Another very useful morphological residue is the tophat transformation [17]: it is made from the dierence between the initial image and the opened image (``white hat'') or between the closed and initial image (``black hat''). Image segmentation Attributes. An example is given in [1]. Two main categories can be de®ned: · threshold from histogram in grey tone levels.1. In this case the attribute of segmentation is a domain of grey tone level. As one seeks for characterizing a contour. The method proposed by Tsai [20] consists of assigning a well-determined grey tone level z j to each class. 6.1. In the case of a system with two classes.1.1. In this case the histogram is divided in several classes. then threshold corresponds to the transformation of a grey tone image in a binary one. The ®rst category de®nes regions as a function of the grey tone level.1. (see Appendix C). for instance. This method gives good results with low contrast images.1.M.1. excluded.1.1. see Appendix D). In this approach. 6. When the position of the class limit varies.1. The entropy is de®ned in Appendix D. limiting ourselves only to the determination of a unique threshold (a multi-threshold is also possible. The threshold by information on the contour is also possible from an algorithm computing the image contrast [21].1. Coster.4. grey tone levels or topography. Threshold by moment-preserving. BTop hatB k f x f B k x À f x: 6 7 The top-hat transformation is also generalized in replacing simple openings by combined morphological ®lters. (Fig. 1(a)). · threshold by contrast analysis. de®ned by an upper and lower value. Threshold from grey tone level histogram Several methods of threshold belong to this class. A binary plane will correspond to each domain (Fig. 6. and the second one mainly concerns the transition of grey tone levels on the outlines of the regions. So one must pay a lot of attention to the choice of the method to be used. the most useful methods will be indicated.1. If the case of colour images is. In the ®rst case one will speak of orientated region segmentation and in the second one of orientated contour segmentation. x and y will be all the close pixels of which the values are within the limits of the threshold t. This type of threshold is suciently ecient when two classes are relatively more or less equivalent [18].1.5. Filters constructed from morphological residues correspond to high-pass ®lters of the linear domain. J. the histogram is divided into a certain number of classes. Threshold by contour information. colours. This step is extremely important as it corresponds to the maximum loss of information. It can be de®ned by the following expression. The corresponding relationships are: WTop hatB k f x f x À fB k x. Chermant / Cement & Concrete Composites 23 (2001) 133±151 137 for example OCOC) or in parallel with sup and inf (morphological centres or morphological contrasts).

The quality of the segmentation will mainly depend on the choice of the markers. jf y À tj card K t : 9 The threshold retained gives the maximum mean contrast.2. Watershed: a segmentation by morphological operator The watershed is the main segmentation method using morphological operators [16. 3(c)) or minima brought together by dilation. All these operators use two source images to form segmentation: the minima which constitute the image of markers (Fig. 3. (c) Binary image of dilated h-minima of image 3b (image of markers). 3(c)) and the reference geodesic image which represents the relief which will be inundated. Coster. (d) Result of the segmentation after application of the watershed method. .e. To improve the result. extended minima will be utilized using either a sucient depth in their neighbourhood (h min). or a combination of these two possibilities. An example is also given in [7].. To describe how it works. 3(a)) is converted into a grey tone image by the distance function (Fig. (Fig. (b) Image of the distances to the border of the binary image 3a. 3(b)). Chermant / Cement & Concrete Composites 23 (2001) 133±151 P C t K t min jf x À tj. i. the binary image (Fig. Any relief possesses lines of crests and catchment basins. Fig. When the minima are directly used. 6. These catchment basins are separated by a watershed line. This type of threshold is ecient if the contours are related to only one transition.22]. (a) Binary image of a sintered glass. This construction method enables to classify the watershed among the geodesic operators (see Appendix C). The principle of construction is as follows: the low parts (minima) of the relief are ®lled (inundated) progressively but the lakes formed are prevented from joining by constructing a ``dyke'' which is nothing else than the watershed line. zones for which the rain water ¯ows towards a minima. so to a two-phased system. one can imagine that a grey tone image can be assimilated to a relief. In that example.-L.138 M. J. it leads to an over-segmentation.

the step of the Euclidean dilation is replaced by an in®nite geodesic dilation of the eroded image (image of the markers) with regard to the initial image X: Y XX . which has no physical meaning. y1 T y2 : d s. this skeleton by opening is not topologically similar to the initial set. 7. mathematical morphology [5. This skeleton possesses characteristic points: extremities and multiple points. Skeleton by in¯uence zone. Objects totally included are obtained by the dierence between the initial and reconstructed images: Y X = M È BX I: 13 Nevertheless in the discrete space. The most often useful processing will then be presented. Mathematically speaking it means s P Sk X ( A Wy1 . Binary ®ltering Morphological ®lters used with images in grey tone levels have their equivalent in binary images (opening. noted Sk X .). Mathematical morphology enables to rapidly eliminate objects intersecting that frame [5.2. One prefers to often use the skeleton by thinning (see Appendix B) which leads to object of 1 pixel thickness and with the same topology.1. the environment of certain crystals or cells in biology [26]. is associated to an object Xi . Its mathematical de®nition is given by [ Yi y : d y. Skeleton. dX its border and s a point of X. It corresponds to the set of the nearest points of that object. one must pay attention to the use of the skeleton as it is sensitive to noise. that object must be totally included in the frame of measurements in order to avoid a bias. etc. Main operators 7.1. xi < d y. y2 P oX .M.1. For that purpose. to ®ll holes to separate objects which are touching. Chermant / Cement & Concrete Composites 23 (2001) 133±151 139 7. the skeleton X. So it is not possible to give an exhaustive list. xj . Coster. Sk If X is a connected set. but which can also require an image processing either by ®ltering or segmentation: it is often necessary to eliminate artefacts. Each region or in¯uence zone. In fact it is the binary version of the watershed line. An illustration of these dierent operators is given by the sequence of the images in Fig.1. Binary image treatment The watershed line and the segmentation by threshold give binary images which can operate as they are.12. xi P Xi and xj P Xj : 17 The complementary set of these regions is the SKIZ: " #C [ SKIZ Yi : 18 i 7. Nevertheless. which can present an interest. for example.1.2. J. .1. y1 d s.B k X É B k È BX I: 12 7. oX d s. y2 : 15 The skeleton by maximum disk (opened skeleton) is obtained from an algorithm developed by Lantuejoul [23] [ \ X É kB= X É kBlB : Sk X 16 k>0 l>0 The morphological opening is most often replaced by the opening-reconstruction which has the advantage of not eliminating the smallest objects without modifying the remaining objects. to extract useful information for the analysis. For this step. one must reconstruct the complementary of the objects to be analysed from the frame already used thus in inverting (reversing) the resultant image: Y M È BX C IC : 14 For example.24]. One has two main basic operators: Y XB k X É B k È B k. SKIZ The skeleton by in¯uence zone (SKIZ) separates an image in several regions. for example for branching information [25]. B k Y X X È B k É B k: 10 11 7. this operator is used to correctly represent the grain boundary network. Hole ®lling In many cases. Yi . 7.23. So it corresponds to a simpli®ed representation of objects which can be used for a sorting. closing.12] is a very powerful tool. For that purpose one considers the image frame as a marker M and this frame is then reconstructed in the image X to be analysed. Hence using geodesic operators of mathematical morphology.2. Elimination of the particles interesting the frame of measurements When size distributions have to be performed particle per particle or when one has to investigate the shape of an object in 2D. It can also be used to investigate the neighbourhood of objects [24]: it can be very important to know. the result of segmentation leads to particles with holes.2. 4. is the union of the centre of the maximum balls included in X.-L. The skeleton Sk X and the skeleton by in¯uence zone (SKIZ(X)) are also often used.2. 7. Other operators These three previous operators constitute only a small fraction of the possibilities of mathematical morphology in segmentation.3.

as it can be seen in several papers of this issue [7. only the dendrites are reconstructed. 8. Mathematical morphology allows to solve this problem in de®ning the image zone (surface area) which is known without a bias. and associated skeletons. it is necessary to use the skeleton function after their detection. 7. if one uses a local operator (morphological operator or convolution window (mask). Using these triple points as markers. size. Problem of local knowledge As it has been already said.140 M. (d) Clipped skeletons and particles reconstructed from markers. shape and dispersion of the phases.-L. the structure of materials is always known in a frame of a size smaller than that of the plane section observed. (b) Cleaned image by opening-reconstruction and hole ®lling.27.3. To be rigorous one must eliminate certain zones as around the image. (a) Threshold image of a cast iron. (c) Elimination of the particles intersecting the frame of measurements. one can express the information contained in the image in parametric or function form. J. Coster. This rule applies Z H Z É B the theorem of the frame (mask) of measurements [12]. Measurements After the segmentation and binary image processing.28]. the result will be known without bias in a frame Z H deduced from the initial frame (mask) by an operation of erosion with the structuring element B associated to the morphological operator or to the ®ltering function of the convolution product. The use of dierent operators of ®ltering or of segmentation introduces a bias on the borders of the image. Hence. The partial knowledge of a 3D structure through . To investigate cracks in concrete. 4. Characteristics related to objects or phases can be classi®ed according to several criteria: ratio. Chermant / Cement & Concrete Composites 23 (2001) 133±151 Fig. Sorting of dendritic particles from others can be made by selecting the triple points of the pruned skeleton (centre of the skeleton cross of image 4d).

So the measurement must mathematically be robust. That can be guaranteed if the measurement veri®es the property of continuity. residual noise remains. Stereological parameters The stereological parameters [3±5] verify a certain number of properties. this measurement must not depend on the position of X. In R2 it corresponds to the number of objects less the number of holes contained in these objects per unit surface. the measurement will verify the C-additivity. These properties can easily be explained. the measurement must not depend on the magni®cation. In this case parameters or functions only with stereological properties must be used. If W X is a measurement on the set X. It is then said that the measurement is compatible with similarities. which verify these properties. A Xi . the surface is concave. There exist stereometric relationships between the parameters of a same column [3±5]. MV X 2pNA X : 19 20 21 One notes that the Euler±Poincar characteristics are e only accessible in their own space. In the case of sets. The most used parameters are related to the size: there are the mean free path. So it can be shown that a numbering of objects does not correspond to a C-additive measurement.M. and the mean surface area of objects without holes. If they are negative. 8. If these two curvatures are positive. In e R1 it corresponds to the number of segments per unit length.1. This last property is a limiting condition. From these basic parameters. it is more complicated as this number is calculated from the topological nature of surfaces making the objects (number minus genus). R2 and R1 are the Euler±Poincar characteristics (EPC). secondary parameters. Table 1 Stereological parameters in the local case Space R3 Parameters of quality Volume fraction VV X R2 Surface fraction AA X R 1 In the space with n dimensions. the magnitudes will be estimated by means. To each point of the boundary corresponds two main curvatures. The speci®c connectivity numbers in R3 . Chermant / Cement & Concrete Composites 23 (2001) 133±151 141 a part of a 2D section introduces constraints on the choice of the parameters when one wishes to access 3D space. LA X pNL X . For a stationary set. Some remarks can be made. intersected by a straight line D on the object X . In R3 . and consequently by sums. In the case of a stationary system and to be independent on the size of the frame of measurements. They are calculated from L1 X VV X NL X and VV X A Xi : NA X 22 Measure of boundary Speci®c surface area SV X Speci®c perimeter LA X Speci®c connectivity number NL X Measure of curvature Speci®c norm MV X Speci®c connectivity number NA X Speci®c connectivity number NV X Linear fraction LL X R 0 Point fraction PP X .-L. As the measurement is made only on a part of the set. as proposed by Hadwiger [29] in the case of sets corresponding to a ®nite union of convex bodies. L1 X . One says that W X is invariant under translations and rotations. Coster. the measurements will be given as a function of the frame size of the de®ned space: this will be termed local parameters opposed to global or absolute parameters. Although any experimentalist would try to work on the clean image. one has a surface like a saddle: then for each point one can de®ne a mean curvature which is integrated to obtain the norm. The norm (integral of mean curvature in R2 ) is generally called integral of mean curvature. there is: VV X AA X LL X PP X . which verify the same properties. the surface is convex. The stability will still be possible if the measurement is upper semi-continuous or lower semicontinuous. particularly in the case of stationary systems. If the main curvature is positive and the other one negative. As any optical system enables one to obtain images of various magni®cations. [12. SV X 4NL X . can be de®ned.30]. J. there only exist n 1 basic parameters. They are gathered in Table 1.

It can be a square. That will be useful to make estimates on the structure anisotropy. It can also be performed on grey tone images (Fig. For a measurement of the connectivity number.142 M. An example is given in [6]. 26 In the case of the connectivity number NL X . Then they are classi®ed according to their size to establish the distribution. 8. etc. In this case.34]. To undertake granulometric measures. So one is constrained to eliminate objects intersecting the frame of measurements. one assigns to the counting a coecient which is nothing else than the reciprocal probability of inclusion of the object in the frame. measures of size (Feret diameters. The granulometric analysis by opening is a more general method than the previous one. 8. Coster.). The non-biased measurement is given by N1 X Z H Q À 1 : 24 NL X L Z H 8.1. To correct this bias. can be performed [3±5]. Individual granulometric analysis. the measurement is made on the volume under the image.1. Several examples are given in this issue and more particularly in [7. 8.1. an hexagon.-L. Two approaches are possible to obtain results without a bias. But all these parameters do not possess any stereological character. This probability is easily obtained from the ratio surface area of the frame of measurements Z/rectangle to the object [24]. Parameters related to individual analysis When the phase X corresponds to many objects isolated one from the others. a segment.1. In these conditions. as it can be used independently on the nature of the set X. In the second case. The granulometric distribution in measure is given by G k AA X Z É B 2k À AA XB k Z É B 2k AA X Z É B 2k 25 B k is a convex structuring element of size k. 8. Of course erosion by a segment is sensitive to the orientation a of the segment `. there exists two main categories of methods: measurement by individual analysis or by morphological opening. which replaces the surface fraction.3. One can obtain granulometry in number and in measure directly from the erosion which gives the function P `. of shape (as the ratio length/thickness or surface/(perimeter2 ). · the covariance functions to characterize the phase dispersion.2.31]. One can distinguish: · the granulometric distribution functions to characterize the size. etc. the frame of measurements Z H is a segment of line (D Z) and the shell QH is one of the extremities of this segment. which introduces a bias as the largest objects are (most often) eliminated from the measure. Morphological functions The mean stereological parameters are often insucient to access a precise evaluation of the morphological characteristics. `. the shell Q will be made of two adjacent sides of the rectangular mask Z. which limits their use only to a comparison and not to an estimation of some 3D characteristics of the material. One has P ` AA X É ` Z É `: It induces G ` P 0 À P ` `P H ` P 0 for granulometry in measure.. the theorem of the frame of measurements will also be applied. Thus very dierent structures can give the same value of the mean free path. The non-biased value of NA X will be given by NA X N2 X Z Q À 1 : A Z 23 8.2. Individual granulometric analysis is very well known: the size of each object is measured. etc. In the ®rst case. · the distance functions to also characterize the phase dispersion but by another approach. (probability that a segment ` is included in the objects).4.4.4. J. one will use the notion of shell [32]. 27 P H 0 À P H ` F ` P H 0 for granulometry in weight: 28 .4. each object must be totally visible. So a more accurate description can be obtained using functions depending on a parameter. The use of a linear structuring element. Granulometric functions Granulometric functions are the most useful. N2 X . Stereological parameters and frame of measurements The problems of local knowledge also exist in the case of stereological measurements. Opening granulometry. To perform such an analysis. 5(a) and (b)). Chermant / Cement & Concrete Composites 23 (2001) 133±151 It is also to be noted that the measure of NL X is a function of the orientation of the analysis straight line D. leads to linear granulometries which are particularly interesting as the estimate in R2 remains valid for R3 [33. · the rose of intercepts or of directions to characterize the structure anisotropy. one can consider that the measurement is associated to a hit or miss transformation.

Crossed covariance. while the value at the origin is null and the slope is equal to 1=4SV X . One distinguishes the (simple) covariance. Between the origin and the value at the in®nite. But in this case the asymptote corresponds to AA X Ã AA Y . One has Z I 4 St3 X `2 P `d`: 29 P 0 0 8.2. A repulsion eect between the dierent parts of X will give a curve passing below the asymptote. the behaviour of the covariance is related to the state of dispersion of the phase X. the covariance function is a directional analysis related to the orientation of ~ . For h 3 I the covariance tends 2 towards a horizontal asymptote equal to AA X .4. for which two points at distant h are tested on the same phase. The simple covariance is given by C h AA X É h Z É h AA X X~ Z Z~: h h 30 To obtain a measure of covariance. 8.31]. Coster. 5. h AA X Y~ Z Z~: h h 31 The rectangular covariance enables to investigate the dispersion of phase X with regard to Y in polyphased systems. Distance functions and dispersion A more physical approach of the dispersion uses the distance function. When there are several directions.2. (b) Opening of size 30 on image 5a and mask Z H inside of which the result is known without bias.4. If NL X . this value can be plotted in polar coordinates as a function of a. Simple covariance. while with an attraction case it passes above. a is a function.36]).2. Y . the mean distance between neighbours. (a) Image of a sintered glass. The rose of intercepts is particularly useful when there is one dominant orientation. This type of characterization is very important for materials of civil engineering and gives rise to European and American standards and to many papers ([35. J. The three main functions are described in this issue: the distance between the nearest neighbours. as it corresponds to diametral variation in the direction a.M.4. For a periodic structure the C h curve will oscillate. One example derived from this method is given in [37]. The slope is equal to NL X . Chermant / Cement & Concrete Composites 23 (2001) 133±151 143 Fig. Covariance function and dispersion The covariance function is obtained by erosion of the analysed structure by a bi-point h.4.1.4. h 8. 8. As for the P ` function.2. Several distance functions can be de®ned. The crossed or rectangular covariance corresponds to C X =Y . X.7. 8. it is better to use the rose of directions which corresponds to the ratio of length of boundaries . Perfectly random structures without attraction or repulsion eect will give a regularly decreasing curve. and the distance function for which the value of each pixel is given by its distance to the phase under investigation [35]. it is simpler to take the intersection of X and of its translated by ~ . X and Y. thus the rose of intercepts can be constructed [4. By integration of the P ` function one obtains the star function in R3 which corresponds to the mean volume of the set X seen from any point of X. It has been used by several authors of this issue [1. Rose and anisotropy Anisotropy investigation can be performed using analysis tools sensitive to orientation. and the crossed or rectangular covariance for which each point belongs to dierent phases. The h value at the origin is nothing else than AA X .5].3.4.-L.

The ideal is to have a rotating optical or captor device. An example is also given in [7]. This in¯uence zone is a convex polygon in R2 and a convex polyhedron in R3 . One will use the simplest ones: the segment of line (the more isotropic (hexagon or square) structuring element). Moreover. in favourable cases.2. Voronoi tessellation. The space tessellation in R2 according to a Poisson process is made from Poisson straight lines [48]. Inside these categories other elements must be taken into consideration such as shape. the majority of these parametric characterizations are not of stereological nature. pjTi g. 9. p is the distance from x to p.2. which will be described from a space tessellation.1. · polyphased models.3. Poisson tessellation. The most known tessellation is that of Voronoi. So there will be two main categories: monophased materials. [40]. but it is not in practice inconvenient. They are constructed as follows. dispersion. (ii) or use probabilistic models as Boolean models. Modelling As it has been indicated a structure of a material can only be characterized by a large number of parameters or functions. Among those. Tessellation models 9. 35 where d x. To each point pi corresponds an in¯uence zone Xi de®ned by Xi fx : d x.41±44]. T K. So the modelling of a structure. several strategies are at our disposal: (i) either one can use the physical process of formation of the structure or associated to a physical behaviour. and polyphased materials described from polyphased RACSs. That is the role attributed to the Choquet's capacity. through the covariance. J. can solve that problem. [39] and Jeulin et al. It is evident that all the possible compacts cannot be tested.1. To model a structure. Unfortunately numerical images do not lend themselves to an angular investigation as it is dicult to convert an image into a square or hexagonal grid according to any angle without using interpolations which can induce errors. one sets points pi according to a Poisson process of density h. The ®rst strategy is illustrated here by the paper of Boussa et al. the bi-point. Choquet's capacity A RACS can be characterized by a probability of events corresponding to morphological measurements such as the probability of inclusion of a compact K in a set or its complementary. Chermant / Cement & Concrete Composites 23 (2001) 133±151 orientated in a given direction [4. size distribution. it requires that the set X does not ®ll all the space. Fig. So there will be at least a medium of two components: a set X and its complementary X C .-L. (33) that the Choquet's capacity could be tested in using only operations of erosion on the considered compact. [15. a dilation. 9.1. pi < d x. Objects constituting that set can be as well points than straight lines or any sub-sets.3. de®ned by [41. 9. only the Poisson tessellation possesses a stereological character and is easy to test from the Choquet's capacity. Point Poisson process The starting point of all probabilistic models is the point Poisson process. the number of morphologically discernable phases is the ®rst element.144 M. In the following only the second strategy will be approached. In the case of bulk materials.3. Classi®cation of probabilistic models The probabilistic models can be classi®ed into several categories: · tessellation models. The choice of a model depends on a certain number of prior knowledge. There also exists other models such as the Poisson tessellation or the Johnson±Mehl [47] model which subdivide the space. the knowledge of the Choquet's capacity for a compact K enables to completely de®ne a probabilistic model. while the second one corresponds to the works of Dequiedt et al. It can be seen from Eq. h. Coster. The probability that n points of a Poisson process of density h belong to a set Z is given by Pr Z h mes Z exp Àh mes Z: n! n 9.5]. [38]. The obtained set will be a random topologically closed set (random closed set. 9. T K Pr X K T Y: 32 One also can de®ne T K from the probability Q K that the intersection between X and K is empty: Q K 1 À T K Pr X K Y Pr K & X C : 33 As a distribution function de®nes a random variable. To observe a morphology.3. If D .1. This restriction is necessary to preserve good properties. To construct it. 6 illustrates such a model. Thus a RACS will remain a RACS after an erosion. `.45] 34 This point process is the basis of all derived models such as clusters or hard core models [46].1. 9. RACS). etc.

-L. one has Q K exp ÀhE mes X H È K: 36 Fig. 6. k is constant and the value of x is chosen according to a uniform law of probability. if X H is a primary grain of the Boolean scheme. for example. Coster. Thus it is easy to test it by the Choquet's capacity. is a straight line with an orientation between x and x dx and passing through the origin of the plane.52±64]. a primary grain is set. dead leave models [50] and Stienen's model [51]. 9.14.2.12.3. To each point one sets a Poisson straight line perpendicular to D. J. The main morphology will largely depend on these primary grains which opens up many possibilities (Fig. In the case of an isotropic mosaic. Other polyphased models exist.M. Poisson mosaic. . (a) Boolean model with circular disks of normal distribution. 8(a) and (b)). The R2 space is then divided into an in®nite number of random polygons called Poisson polygons (Fig. on that straight line one realizes a point Poisson process of Fig. Moreover. Chermant / Cement & Concrete Composites 23 (2001) 133±151 145 density k2 dx. 7. in order to facilitate a better understanding of the papers included in this special issue on the use of that technique in the ®eld of civil engineering materials. Other more detailed and complementary information can be found in the main books related to this subject [3±5. Voronoi tessellation in R2 . Concluding remarks The scope of this paper was to give a short overview of automatic image analysis. 7). One can quote. Fig. 8. To each point of the Poisson process of density h. and on mathematical morphology. 10. but they are beyond the scope of this general paper [44]. Polyphased models The most known two-phased RACS model is the Boolean model [49]. The Boolean scheme models the same properties in a sub-space (stereological characteristic). (b) Boolean model with Poisson grains. The Boolean model is the union of all these primary grains.

Fig. 11). One has g x f x É B inf f y : y P Bx : B:2 Dilation. which can be most often coded on 8 bits (black 0 and white 255). · grey tone images. The eroded of the set X is a set of points x of the image such as Bx is totally included in X. blue). Dilation is the dual operation of erosion with regard to the complementation. Appendix B. The ®ltering of colour images by operators without good vectorial properties leads to a problem. one has to use sometimes such images. Coster. the functions associated to this point will be de®ned by Fi x. 9 presents an example of erosion by an hexagon. Near objects are connected and holes ®lled. J. In R3 one generally uses a cubic grid. Nevertheless other representations also exist. That is the case of many ®lters. the support is represented by a grid of points regularly spaced. The mode HSL. Fig. Elements of mathematical morphology Examples given to illustrate mathematical morphology will only concern binary images. 10 illustrates this operator. The most common mode of vectorial representation is the mode RGB (red. one does not try to reconstruct the ®ltered colour image. In R2 these points are distributed according to a square or hexagonal grid. multimodal images must be distinguished from colour ones. Initial image. It is written as X È B fx : Bx X T Yg: B:3 Fig. For colour images the primary colour de®nes a vectorial space. I and Q: chrominance components). one can also use integer on 16 bits or real numbers. . When the images have to be processed on a computer. · one or several functions. The functions associated to each point are themselves extended to numbers. At that level. Then dierent types of images must be distinguished: · binary images. The small details (smaller than the size of Bx ) disappear. sometimes 3). · multimodal and colour images where each mode or each primary colour will be represented by a grey tone image. It is the set of points x such as Bx has a not-empty intersection with X.146 M. The mode YIQ is characteristic of the American television standard (Y: luminance component. The approach to functions is made via the sub-graph. Taking into account the real nature of materials in civil engineering. but to treat separately each plane to extract the pertinent information. the de®nition of dilation by a ¯at structuring element attributes to a current pixel the highest value of pixels of the domain de®ned by the structuring element. If x is a point of the support Z. Chermant / Cement & Concrete Composites 23 (2001) 133±151 Acknowledgements This work has been performed in the frame of the ``P^le Traitement et Analyse d'Images'' of Basse-Noro mandie (P^le TAI): Image Processing and Analysis P^le o o of Basse-Normandie. All the morphological operators are based on the notion of structuring element. The most common values today are 256 Â 256 and 512 Â 512. when it is located in a point x of an image. or 0 if not. noted Bx . saturation and luminance. Hopefully with image analysis. Erosion. That gives the dimension of the image. In the case of images in grey tone levels. In the case of images in grey tone levels. De®nition of images Images are de®ned by two main characteristics: · a support belonging to the metric space of a given dimension (generally 2.-L. green. One has g x f x È B sup f y : y P Bx : B:4 As it was indicated earlier the opening and closing operators can immediately be deduced (see Fig. It is written as X É B fx : Bx & X g: B:1 Appendix A. where the primary colours are replaced by hue. So the numerical supports are represented by n lines and m columns. For multimodal images the dierent modes are independent. related to sets where f x takes either the value 1 if the point belongs to the set. the de®nition of the erosion by a ¯at structuring element attributes to a current pixel the smallest value of pixels in the domain de®ned by the structuring element. 9.

One has X ~B fx : Tx a & X and Tx b & X C g: B:5 Thickening. Erosion by a square of size 5 on image Fig. 13. Dilation of a square of size 6 on image Fig. The square structuring element used is the extension in the digital space Z 2 of the circular disk B of size k de®ned in R2 . The thinning operation consists of suppressing points of an object having a given neighbourhood con®guration. Appendix C.M. The hit or miss transformation is more general. They are also used with thinning and thickening operators. respectively. Fig. . 9. Hit or miss transformation. J. Geodesic operators Erosion and dilation de®ned in Appendix B were performed in the Euclidean space. related to X and X C . Fig. 11. The structural element B is constituted of two sub-elements Tx a and Tx b. Coster. Reference set Xr is in grey.-L. Y set in white. If B is It is used to obtain homotopic skeleton and the centres of objects. Chermant / Cement & Concrete Composites 23 (2001) 133±151 147 Fig. It is de®ned by the following operator: X B X = X ~B: B:6 It is used to have the skeleton by in¯uence zone or to obtain a convex hull. Initial sets. Thickening is the dual operation of thinning: one adds to the object the points having a given neighbourhood con®guration. It is de®ned by the following operator: X B X X ~B: B:7 These operators are used to extract particular points such as extremities and multiple points of the skeleton. Thinning. 9. 10. Fig. Comparison between Euclidean disk and geodesic disk with a set Xr as reference. The thinning and thickening operations de®ned for binary images have their equivalent in grey tone images. The point x will be at 1 if Tx a is included in X and if Tx b is included in X C . 12.

-L. Chermant / Cement & Concrete Composites 23 (2001) 133±151 Fig. (b) Geodesic dilation of the set Y by an hexagon of size 10. but an arc xy totally included in Xr.148 M. 14. The main application of the geodesic operation on sets is the reconstruction. ykg: C:1 d x. There also exists in the frame of the geodesic distance. centred on x. noted Xr. J. 14 and 15 illustrate the eects of erosion and dilation of a set Y with regard to the set Xr (Fig. and B a Euclidean disk (in R2 ) or ball (in R3 ). 13). In the case of the geodesic distance the shortest way to go from one point x to a _ point y is not the segment xy . then the corresponding closed disk or ball is de®ned by Bx fy : d x. Coster. To de®ne a reconstruction Fig. (a) Euclidean dilation of the set Y by an hexagon of size 10. Geodesic erosion and dilation are performed with geodesic disks or balls for which the notion of Euclidean distance is replaced by the geodesic distance one. (b) Geodesic erosion of the set Y by an hexagon of size 10. y is the Euclidean distance between x and y. 13). . Fig. then d xy has an in®nite value. such as the triangular inequality. There exists properties associated with the notion of distance. The geodesic distance is de®ned with regard to a reference set. 15. 12 presents how a disk B can be modi®ed with regard to the reference set (see Fig. but one adds the following property: if the path between point x and point y is not totally included in Xr. Y È BXr fx : BXr x Y T Yg: C:2 C:3 Figs. The de®nitions concerning erosion and dilation practically remain the same for the corresponding geodesic operations. (a) Euclidean erosion of the set Y by an hexagon of size 10. One notes that the border of the set Y is not aected by the erosion and dilation operations when it is common to Xr. One only has to replace the Euclidean disk or ball by a geodesic one: Y É BXr fx : BXr x & Y g.

Fig. New York: McGraw-Hill. 1965.-L. Coster. until the image remains unchanged. [5] Coster M. Paris: Les Presses du CNRS. . Quantitative stereology. Bascoul A. Paris: Les e Editions du CNRS. (b) Result of opening by reconstruction (in white). The process is iterated until idempotence. Acta Stereol 1998. Reading. This operator will be found in the case of hole ®lling. Quantitative microscopy. Chermant JL. Les variables rgionalises et leur estimation. of the elimination of particles intersecting the frame of measurements. 1970. M1 2 Meaning maximum value of the grey tone levels moment of order k of an image (k 1: mean) number of pixels of an image number of pixels with a grey level i threshold to determine class of pixels with grey tone levels between 0 and t class of pixels with grey tone levels between t 1 and m number of pixels of the class C 1. 1968. and with the opening operator by reconstruction. i. 16 illustrates this operator. 2nd ed.M. [4] Underwood EE. 1989. N 2 M1 1. Image analysis: at tool for the characterization of hydration of cement in concrete-metrological aspects of magni®cation on measurement. Prcis d'analyse d'images. Threshold and data analysis Parameter m Mk N n i t C 1 C 2 N 1. Rhines FN. Paris: e e Masson. MA: Addison-Wesley.e. [6] Redon C.. [2] Matheron G.17:93±8. It is written as Build Y : Xr Y È B IXr Y È B 1 XrnI : C:4 So it can be obtained from a dilation of size 1 followed by an intersection by Xr. Coster M. Entropy for two classes: H t H C 1 H C 2 X t m X n i n i n i n i À ln ln À : N 1 N 1 N 2 N 2 i0 it1 Interclass variance for two classes: V t M1 1 À M1 2 Pm n i : Â it1 N 2 Pt n i M1 2 À M1 2 N 1 i0 Appendix D. Chermant / Cement & Concrete Composites 23 (2001) 133±151 149 Fig. Chermant JL. J. (a): Erosion (in white) of size 12 on the Y set (in grey) to obtain markers for reconstruction. operation one must have a reference Xr in which the set Y will be reconstructed by a geodesic dilation of in®nite size. C 2 mean of pixels of the class C 1 and C 2 System of equations for the moment conservation: Pm Pt n i k k i0 n i z 1 it1 z 2 : Mk N 1 N 2 References [1] Ringot E. Mowret M. 16. Measurements of orientation in discretized space by Fourier transform: automatic investigation of ®ber orientation in a reinforced concrete. [3] DeHo RT. Cem Concr Comp 2001. Chermant L.23:201±6.

Matheron G. Ann Mines 1967. editor. 1957. Wong AKC. Chermant JL. 17±36. Riss J. London: Academic Press. Coster M. [23] Lantuejoul C. Chermant / Cement & Concrete Composites 23 (2001) 133±151 [32] Pinnamanemi BP. Some ®elds of applications of automatic image analysis in civil engineering. 2000: p. [7] Chermant JL. Stereological methods. Digital image processing. 1976 in Russian. [39] Dequiedt AS.23:227±36. e e Paris: Masson. Acta Stereol 1994. p. Marchand J. Part 2: Homogeneous Poisson ¯ats and the complementary theorem. 1980. [47] Johnson WA. A segmentation system based on thresholding. Some ®ndings on the usefulness of image analysis for determining the characteristics of the air-voids system on hardened concrete. editor. In: Jeulin D.-L. . Monnaie P. e Thse de Docteur-Ingnieur. Poisson ¯ats in euclidean spaces. J Microsc e 2000. A threshold selection method from gray-level histogram. Comp Vision Graph Image Process 1981. 1991. Bascoul A. Dead leaves models: from spaces tessellation to random functions. Coster. vols. Cem Concr Comp 2001.23:299±311. Chermant L.370:77±82. In: Chermant JL. Approach of liquid phase sintering process by probabilistic models. IEEE Trans Syst Man Cybernet 1979. About the analysis of microcracking in concrete. 1988 [chapter 5]. Thse de Doctorat of e the Ecole des Mines de Paris. [18] Otsu N. Coster M. Image analysis of ``natural'' concrete samples by automatic and manual procedures. Ann Mines 1967. [37] Stroeven P. [15] Serra J. Singapore: World Scienti®c. Microscopica Acta 1980. Pischang K. [17] Meyer F. Quantitative image analysis of microstructure. Paris: Ecole des Mines de Paris.23:215±26. Coster M. 1978.13:125±36. Coster M. Moment-preserving thresholding: a new approach. editor.3:1±43. Morphological segmentation. [38] Boussa H. [45] Choquet G.3(Suppl):175±83. Jeulin D. Serra J. [19] Kapur JN. [24] Lantuejoul C. Laurencot JL. Elments pour une thorie des milieux poreux. Stereometric metallography. [51] Stienen H. 20±23 Sept 2000. biology and medecine. Image analysis for automatic study of microcracks in concrete. [50] Jeulin D. Ohser J. e [34] Haas A.23:247±54. Cem Concr Comp 2001. Dunsmuir J. Morphologie mathmatique et e granulomtries en place. Random sets and integral geometry. Jernot JP. 1975. Role of mathematical morphology in ®ltering. Matheron G. Jeulin D.198:54±62.29:377±93. Cem Concr Comp 2001. Poland. [56] Russ JC. Thse de Doctorat s Sciences. Chermant JL. [29] Hadwiger H.S37:74±8. J Visual Comp Image Represent 1990. editor.23:289± 97.23:157±69. In: Jeulin D. [30] Nagel W.1/1:21±46. u [53] Saltykov SA. Distances between air-voids in concrete by automatic methods. Tognazzi-Lawrence C. Study of phase dispersion in concrete by mage analysis. Leipzig: VEB Deutscher Verlag fr Grund-stondustrie.-H. [12] Serra J. Cem Concr Comp 2001. [13] Chermant JL. [25] Coster M. Serra J. Comp Vision Graph Image Process 1985. [57] Exner HE.5:131± 295. Separation of dierent shapes by sequential image transformation. ober¯che und isoperia metrie. Chermant L.23:279±87. Cem Concr Comp 2001. New York: Wiley.23:267±78. Pract Met 1978. Mecke J. Stereometric metallography.150 M. Ann Inst Fourier 1953. Vorslesungen uber inhalt. [11] Pratt WK. e [35] Dequiedt AS. La squelettisation et son application aux mesures topologiques des mosa õques polycristallines. Sixth International Conference on Stereology and Image Analysis in Material Science. Breysse D. Advances in theory and application of random sets. 2nd ed. [27] Ringot E. Trans Am Inst Min Eng 1939. Gypsum morphological analysis and modeling. Image analysis and mathematical morphology. In: Serra J. Chermant JL. [43] Jeulin D. Segmentation d'image et morphologie mathmatique. [9] Bentz DP. Adv Probab 1971. 1967. Contribution of image analysis to a study of muscle changes in hypothyid ro rats. Chermant L. Levenson MS. Acta e Stereol 1989. In: Wojnar L. Chermant JL.29:273±89. London: Academic Press. Coster M. [21] Kohler R. 1983. Modles morphologiques de structure alatoires et de changement d'chelle. An integral-geometric approach for the Euler±Poincar characteristic of spatial image. Acta Stereol 1983. 137±56. Krakow. On the estimation of mean values in individual analysis of particles. [48] Miles RE. 1997. 1997. Redon C.12:767±82. vol 2: Theoretical advances. [14] Serra J.S8:374±80. Beucher S. A guide to practical stereology. 1998. 1974 in German. Contrast feature extraction. Theory of capacities. Sahoo DK. Stochastic geometry and its applications. [31] Dequiedt AS. [44] Chermant JL.23:237±46. p. Hyde DM. A new method for gray-level picture thresholding using the entropy of the histogram. Coster M. Oberursel: DGM-Informationsgesellschaft Verlag. 231±49.8:101±6. London: Academic Press. Mathematical morphology course. Comp Vision Graph Image Process 1985. In: Jeulin D. [49] Matheron G. STERMAT. Quantitative analysis of microstructures in materials science. [8] Slate FO. Stroeven AP. Singapore: World Scienti®c. Martys NS. Practical stereology. 1982. Moscow: Metallurgia. Cem Concr Comp 2001. Morphologie mathmatique et e granulomtries en place. [10] Quenec'h JL. Krakow: Fotobit Design. Advances in theory and application of random sets. Chermant JL.9:62±6. Unbiased estimation of Euler±Poincar characteristic. X-rays for study of internal structure and microcracking of concrete. Hougardy HP. [52] Saltykov SA. [28] Ammouche A. Mater Res Soc Symp Proc 1995. Mehl RF. Paris: Ecole des Mines de Paris. Stutzman P. Chermant JL. [20] Tsai W. [36] Pleau R. Material models and model materials. 1995. Coster M. NewYork: Wiley.11:735±58. Chermant L.5:266±73. [55] Elias H. New York: Wiley. 1±2. e e 1990. Cem Concr Comp 2001. 1997. Cem Concr Comp 2001. Kendall WS. Peronnet F. segmentation and analysis. Reaction kinetics in processes in nucleation and growth. Towards a model of concrete mesostructure. NewYork: Plenum Press. [40] Jeulin D.23:261±6. [33] Haas A. Schwartz LM. [46] Stoyan D. Coster M. A model for computation of leakage through damaged concrete structures. J. 1988. 1986. Cem Concr Comp 2001. Dalhuisen DH. editor.2:107±12.(May): 575±88. X-rays microtomography of an ASTM C109 mortar exposed to sulfate attack. The sectionning of randomly dispersed particles. Chermant JL. a computer simulation. Cem Concr Comp 2001. Chermant L. Singapore: World Scienti®c. [41] Matheron G. Garbozi EJ. editor. La Borderie C. e e [42] Jeulin D.135:416±58.15:319±38. Advances in theory and applications of random sets. Lantuejoul C. Mikroscopie (Wien) 1980. Berlin: Springer. Olsefski S. [16] Beucher S. Pigeon M. Dequiedt AS. Image analysis and mathematical morphology. J Am Concr Inst 1963. Basel: Karger Pub Comp. 1991. [26] Herlicoviez D. [22] Meyer F. University e e e of Caen. [54] Weibel ER. Stereol Iugosl 1981.

Rataj J. Philipp S. 1990. [64] Wojnar L. Cambridge: International Science. Computer-assisted microscopy. 1993. NewYork: Plenum Press. Paris: e Masson. . Mattioli J. Paris: Masson. Pelikan K. FL: CRC Press. [59] Russ JC. Gometrie discrte en analyse e e d'images.M. Morphologie mathmatique. The measurement and image analysis. 1995. Quanti®cation and a modelling of heterogeneous systems. Paris: Herms. Chermant / Cement & Concrete Composites 23 (2001) 133±151 [58] Sxl I. 1991. [60] Chassery JM. Stereology of objects with internal structure. 1989. J. Analyse d'image: ®ltrage et segmentation. 1998. Boca-Raton. [63] Sxl I. Applications in materials engineering. 1995. Coster. Image analysis. Prague: a Academia.-L. e [61] Schmitt M. Montanvert A. Besterci M. 151 [62] Coquerez JP.

- Arithmetic Fractions and Percentages (1).doc
- PURCHASE ONLYCorrosion Inhibitor
- Composite Structure (31-38)
- Fly Ash for Cement Concrete
- Comparison of Thermal Power Plant and Nuclear Power Plant
- What is Anodising - How Anodising Works - The Anodising Process
- Floatation Check
- Dynamic Analysis and Structural Design of Turbine Generator Foundations @
- PURCHASE ONLYDesign of Turbo-Generator Foundations @
- CIVIL Eengineering
- Horizontal Curves
- HeliScope rajesh
- CV-Norm-29.04.12
- TSO
- Smart Structures Presentation
- Chapter 1
- Assignment 1 Finished Report
- Concrete 1

- A Review of Blood Vessel Extraction Techniques and Algorithms
- 01308101
- Segmentation
- major.pptx
- IJCTT-V3I4P113
- FUZZY BASED GRAPH CUT CLASSIFICATION OF HYPER-SPECTRAL IMAGE
- Tema 14.pdf
- Final Synopsis
- e i Gen Movements
- Signal Processing IEEE 2015 Projects
- 10.1.1.23
- Adaptive Background Estimation and object detection applying in Automated visual surveillance
- Good Practice Guide - Callipers and Micrometers
- GAUGE R&R FOR AN OPTICAL.pdf
- 18.IJAEST Vol No 6 Issue No 2 Development of Vision Based Iris Recognition System 277 281
- ruhl_michael_w_200908_ms
- 56
- MG2.en
- CorbettPA2Webinar
- UCM092233
- Simulado Ingles
- 14 March 2016
- Measurement Issue and Criteria
- Ermer (2006)
- Medical Electronics
- KSOU Courses
- ECE_Communication DIP (Abstracts)
- GEOMETRIC TOLERANCE
- LayTec X Link In-line Flyer
- Benders Decomposition

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd