MORPHOLOGICAL SCALE-SPACE OPERATORS

REIN VAN DEN BOOMGAARD

University of Amsterdam, Kruislaan 403, 1098 SJ Amsterdam, The
Netherlands
and
HENK HEIJMANS
CWI, P.O. Box 94079, 1090 GB Amsterdam, The Netherlands

Abstract. This paper proposes a general algebraic construction technique for image scale-
spaces. The basic idea is to rst downscale the image by a factor using an invertible scaling,
t
then apply an image operator at a unit scale, and nally resize the image to its original scale.
It is then required that the resulting one-parameter family of image operators satises some
semigroup property. In this paper we focus on the morphological erosions. In this case,
classical tools from convex analysis play an important role.
Key words: Convex functions, Erosion, Inmal convolution, Parabolic structuring function,
Scale-Space, Scaling, Semigroup, Slope transform, Subpolynomial functions, Young Fenchel
conjugate

1. Introduction
Scale-space is an accepted and often used formalism in image processing and
computer vision. Today, this formalism is so important because it makes the
choice at what scale visual observations are to be made explicit. Before the
popularization of the scale-space concept, the choice for an observation scale
was often hidden somewhere in the denition of the operators.
The notion of linear scale-space has a very long history in image process-
ing[1, 2, 3]. Weickert et.al.[4] only recently `discovered' that the concept of
linear Gaussian scale-space dates back to the sixties.
In mathematical morphology the notion of scale (or size) dependent observa-
tions was pioneered by Matheron[5] in his study of granulometries. Jackway[6]
and van den Boomgaard[7] showed that a morphological analogue of the Gaus-
sian linear scale-space does exist: the parabolic erosions and dilations. Van
den Boomgaard and Smeulders[8] also showed that the morphological parabolic
scale-space is the solution of a partial dierential equation, just like the Gaus-
sian linear scale-space is the solution of the diusion equation[2].
In this paper we look at the denition of scale-spaces from an algebraic point
of view. Let f be the image at scale zero and let T (s) be the operator such that
T (s)f is the observation at scale s. The family of operators fT (s)gs>0 is collec-
tively known as a scale-space. The atlas principle, introduced by Koenderink
states that a scale-space can be build incrementally, can be mathematically
formulated as a semigroup property of the one-parameter family of scale-space
2 REIN VAN DEN BOOMGAARD AND HENK HEIJMANS
operators T (t).
We want our scale-space operator to be a macroscopic one, in the sense
that T (t) only needs the zero scale image without the need to calculate all
intermediate images T (s)f for s < t. We thus take a quite dierent approach
to scale-space as the approach that has so elegantly been put forward by Alvarez
et. al.[9] who take the evolution of the zero scale image modeled with a partial
dierential equation to be the starting point. By concentrating on the PDE
description one runs into the risk of dening a sequence of images indexed
with a continuous `time' parameter where the time parameter cannot be linked
intuitively with the notion of scale.
In this paper we will look at morphological scale-space operators using con-
vex structuring functions. For this class of functions convex function theory
provides a valuable tool in the analysis of morphological (scale-space) oper-
ators. This paper therefore contains a short introduction to convex function
theory in section 3.

2. Scalings, semigroups and scale-space

In the presented approach to the construction of scale-space operators the scale
at which the operators observe the image is explicitly captured in a scaling
operator. The actual image processing operator works at a xed `unit' scale.
In this paper we consider the scale to be represented with a positive real scalar
(denoted as the set T = (0; 1)).
Denition 1 (Scaling) A family S = fS (t) j t > 0g of operators on L is
called a scaling if (i) S (1) = id and (ii) S (t)S (s) = S (ts) for s; t > 0.
A scaling thus forms a commutative group with inverse S (t),1 = S (1=t) and
unity element S (1).
In a scale-space construction it is the ordering of scales that is much more
important than the actual value. The scale is the free parameter in the sys-
tem and we are thus free to reparameterize the scale at will as long as the
reparameterization preserves the ordering of the scale values.
Denition 2 (Anamorphic scalings) Two scalings S and S 0 are said to be
anamorphic if there exists an increasing bijection on T such that S ( (t)) =
S 0 (t) for all t 2 T .
It is obvious that such a bijection must satisfy the conditions (1) = 1 and
(st) = (s) (t) for s; t 2 T : The most important example is given by (t) = tp ;
where p > 0.
In this paper we only consider images that are mappings from the continuous
plane Rd to the real values R (R extended with the values +1 and ,1). This
complete lattice of images is denoted as L = F un(Rd; R ). The scalings we
will look at are a combination of a spatially isotropic scaling and a grey value
scaling:
Denition 3 (Image scalings) The image scaling S (t) : L ! L is dened
as S  (t)f = t1, f ( t
 ):
 MORPHOLOGICAL SCALE-SPACE OPERATORS 3
For  = 1 we obtain a pure spatial scaling whereas for  = 0 we obtain a pure
grey value scaling. For  = 21 there is a perfect balance between spatial scaling
and grey value scaling.
The atlas principle introduced by Koenderink[2] states that it should be
possible to build a scale-space incrementally. If f denotes the `zero scale' image,
the image observed at scale s is given as T (s)f: Observing this image at scale
t results in T (t)T (s)f and according to the atlas principle we should have that
T (s)T (t) = T (t)T (s) = T (r):
The atlas principle can be associated with a commutative semigroup de-
scribing the way in which the scales combine in a composition of scale-space
operators. We dene T (s)T (t) = T (t)T (s) = T (s+_ t) where +_ is a binary oper-
ation on T that is associative and commutative, i.e. (T ; +) _ is a commutative
semigroup. In this paper we only consider a special class of scale semigroups
(T ; +).
_
Denition 4 (+ -scale semigroup) The binary operator + : T ! T de-
ned by s + t = (s + t ) 1 forms a commutative semigroup (T ; + ) for
0 <   1.
For  = 1 we obtain the standard additive semigroup (T ; +), whereas  =
1 leads to the supremal semigroup (T ; _). Within the scale-space context
this choice for a scale semigroup has particular advantages. It is easy to see
that s  t ) s + r  t + r for any xed r, meaning that the ordering of
scale values is preserved when we perform a subsequent observation at scale r.
In morphological terms we say that + is an increasing operator in both its
arguments; in semigroup theory the semigroups (T ; +) _ that have this property
are called linearly ordered. Furthermore it is easily shown that for nite 
(i.e.  < 1) any observation at scale t following an observation at scale s will
increase the scale, i.e. s; t < s + t. For the supremum semigroup we only have
s; t  s +1 t showing that the scale does not decrease by making an observation
at any nite scale.
Now we are ready to give a formal denition of a scale-space. In our ap-
proach the starting assumption is that the scale-space is a semi-group of opera-
tors on the image space L under composition, compatible with a given linearly
ordered semi-group +_ on the range of scale values T and invariant with respect
to a given scaling S .
Denition 5 (Scale-space) Let (T ; +_ ; ) be a linearly ordered semigroup and
let S be a scaling on the image space L: The family fT (t)gt>0 of operators on
L is called an (S; +)_ scale-space if:
T (t)T (s) = T (t+_ s); s; t > 0;
T (t)S (t) = S (t)T (1); t > 0:
An alternative denition of a scale-space operator follows easily from the above
denition:
Denition 6 (Scale-space construction) Let (T ; +_ ; ) be a linearly ordered
semigroup, let be an image operator on L; and let S be a scaling on L: The
4 REIN VAN DEN BOOMGAARD AND HENK HEIJMANS
family fT (t)gt>0 dened with T (t) = S (t) S (t),1 , denes a (S; +)
_ scale-space
_
if T (t)T (s) = T (t+s); s; t > 0:
In the beginning of this section we dened that anamorphic scalings are equal
up to an order preserving reparameterization of the scale. It can be shown [10]
that anamorphic scalings lead to anamorphic scale-spaces.

3. Convex analysis and mathematical morphology

The morphological operators that we will be looking at, are the classical struc-
tural erosions and dilations. The set L = F un(R2; R ) is a complete lattice
under the pointwise ordering. Every translation invariant adjunction on L is
of the form ("b ; b ) with:
V
"b (f )(x) = h2R2 [f (x , h) + b(h)] (1)
W
b (f )(x) = h2R2 [f (x + h) , b(h)] : (2)
In these expressions, the function b is called the structuring function1 . We
will assume throughout this paper that a structuring function b is lower semi
continuous and convex with b > ,1. The expression for the erosion "b is a
well-known operation in convex function analysis, where it is known under the
name inmal convolution and denoted by f  b. The dilation b will be denoted
as f
b. Our choice for the inmum convolution as the erosion leads to some
properties that may look a bit awkward to the seasoned morphologist. For
instance it is the erosion that is associative and commutative in this approach,
i.e. (f  b)  b0 = f  (b  b0) and f  b = b  f whereas in the classical approach
based on the Minkowski operations it is the dilation that has these properties.
This section presents a brief overview of some basic results in convex analysis
that play a role in the sequel; refer to [11] for a comprehensive account.
Denition 7 (Convex sets and convex functions) A set X  Rd is con-
vex if tx + (1 , t)y 2 X when x; y 2 X and 0  t  1. A function f : Rd ! R
is convex if f (tx + (1 , t)y)  tf (x) + (1 , t)f (y); for x; y 2 Rd and 0  t  1.
It can be easily shown that the inmal convolution of two convex functions
yields a convex function (see [11]). Convexity of sets and of functions are
compatible notions; let IX be the indicator function of a convex set X given by
IX (x) = 0 for x 2 X and IX (x) = 1 for x 2= X; then IX is a convex function.
Denition 8 ( Subpolynomial functions) A convex function f is called sub-
polynomial of degree k where k  1 if f (tx) = tk f (x); x 2 Rd ; t > 0: The family
of convex subpolynomial functions of degree k is denoted as SP (k).
Special cases worth mentioning are the subpolynomial functions of degree 1
that are called sublinear and the subpolynomial functions of degree 1 being
1 The classical structural erosion is dened as ( )( ) = V
f a f a x
h2R2 [ ( + ) , ( )]
f x h a h

showing that  =  where ( ) = , (, ).

f b f b; b x b x
 MORPHOLOGICAL SCALE-SPACE OPERATORS 5
indicator functions of convex sets. An example of a sublinear function is the
support function HB of a set B dened for x 2 Rd as HB (x) = supy2B hx; yi.
It is not dicult to show that O 2 B is equivalent with HB (x)  0 for all
x 2 Rd. Not only is the support function sublinear, but it can also be shown
that any sublinear function is the support function of some closed convex set
B . For a disk B = fx j kxk  1g the support function is the corresponding
norm HB (x) = kxk.
In convex function analysis the function conjugate being a dual represen-
tation of a (convex) function plays an important role. The conjugate function
is also known as the Fenchel conjugate, the Young-Fenchel conjugate or the
Legendre transform.
Denition 9 (Conjugate function) The conjugate f  of a function f : Rd !
R is dened by
f  ( ) = sup fhx;  i , f (x)g
x2Rd
Conjugation in the context of mathematical morphology has been emphasized
by Dorst and van den Boomgaard [12] and by Maragos [13], who call this
operation the morphological slope transform. Refer to [14] for a systematic
treatment of the slope transform in the complete lattice framework. To a
certain extent the slope transform plays a role in mathematical morphology
which is somewhat comparable to the role of the Fourier transform in linear
signal processing. This is mainly because of the following result.
Proposition 10 Let f and g be convex functions, then (f  g) = f  + g:
Indicator functions and support functions are linked through conjugation: if B
is a nonempty closed convex set, then IB = HB and HB = IB .
For a number k 2 [1; 1] we dene its reciprocal k through the relation:
1=k + 1=k = 1.
Proposition 11 A function f is subpolynomial of degree k if and only if its
conjugate f  is subpolynomial of degree k .
If k = 1 then f is sublinear and therefore we know that f = HB for some closed
convex set B . Because the conjugate of a support function is the indicator
function of the same set we have f  = HB = IB 2 SP (1); i.e. the conjugate
of a sublinear function is an subpolynomial function of degree 1:

4. Morphological scale-space operators

In this section we are exclusively conserved with the class of scale-space op-
erators on L = F un(Rd; R ) induced by the morphological structural erosion
"b f = f  b. Without loss of generality we may concentrate on the erosions
only, as it can be shown[10] that in case ("; ) is an adjunction and T" denes
a scale-space then T denes a scale-space as well.
6 REIN VAN DEN BOOMGAARD AND HENK HEIJMANS
In this paper we will restrict ourselves to scalings of the form S  (t) as
dened in section 2. A straightforward calculation shows that the scaling of
the images completely `carries over' to the scaling of the structuring function
that is used:
T" (t)f = f  S  (t)b:
For T" (t) to be a scale-space operator we should have that T" (t)T" (s) = T" (s+_ t).
Using associativity of the erosion we obtain a semigroup requirement on the
family of scaled structuring functions S (t)b with respect to the erosion opera-
tor: S  (t)b  S  (s)b = S  (s+_ t)b: Taking conjugates on both sides we arrive
at (S  (t)b) + (S  (s)b) = (S  (s+_ t)b) . The conjugate of a scaled function
(S  (t)b)( ) equals t1, b (t2,1  ) thus:
t1, b (t2,1  ) + s1, b (s2,1  ) = (s+_ t)1, b ((s+_ t)2,1  ): (3)
Erosion scale-space analysis thus boils down to the analysis of the above equa-
tion. Without any assumption on b (other then b, and thus b being convex)
we can notice the special case for  = 2 , then: t 2 b ( )+ s 21 b ( ) = (s+_ t) 12 b ( ):
1 1 

Regardless of b we thus have that:

Proposition 12 The erosion "(f ) = f  b with b convex induces an (S 21 ; + 21 )
scale-space.
The scaling S 21 (t)f is anamorphic with the umbral
p scaling U (t)f , where U (t)f =
tf (
=t) with t > 0; more precisely S 12 (t) = U ( t). Notice
, that regardless

of the
structuring function b we have that U (t)f  b = U (t) f  U ( 1t )b showing the
umbral scale invariance of the erosion. It is a well-known result from classical
morphology that umbral scales add up in a sequence of erosions (dilations) if
and only if the structuring function is convex. From this point of view the
above proposition tells us little news.
The algebraic form of the equation governing erosion scale-space (eq.(3))
suggests that choosing a subpolynomial function may lead to other types of
erosion scale-spaces. Assuming that b 2 SP (k ) we may rewrite eq. (3):
  
t1,+k (2,1) b( ) + s1,+k (2,1) b ( ) = (s+_ t)1,+k (2,1) b ( ): (4)
It should be noted that the case k = 1 (i.e. b 2 SP (1)) has to be treated
separately as the above equation only makes sense for k < 1; i.e. k > 1. First
we concentrate on nite k . Equation (4) shows that:
Proposition 13 The erosion "(f ) = f  b with b 2 SP (k) for k > 1 induces
an (S  ; + ) scale-space if  = 1 ,  + k (2 , 1):
A well-known example in this class of erosion induced scale-spaces are associ-
ated with the quadratic structuring functions bQ (x) = hQx; xi where Q is a
positive denite diagonal matrix. There k = k = 2 and thus "(f ) = f  bQ
induces an (S  ; +3,1 ) scale-space in case  > 13 . The parabolic morphological
scale-space is illustrated in gure 1.
 MORPHOLOGICAL SCALE-SPACE OPERATORS 7

Fig. 1. Erosion scale-spaces. The top row shows the parabolic scale space, the second
row shows a at disk scale-space. On the left in both rows the `zero scale image' is shown.
f
The four other images on the top row show the erosion  1 ( ) where is the parabolic
f S t b b
function ( ) = k k2 for the rst row and = B with a disk of radius 1 for the second
b x x b I B
row.

Another example is the case k = 1; i.e. k = 1 corresponding with b = IB

for some convex set B . Then we have  = 1 ,  + 2 , 1 = . Thus erosions
using at convex structuring functions induce scale-spaces irrespective of the 
parameter dening the balance between the spatial scaling and the grey value
scaling. Figure 1 also shows a morphological scale-space induced by a at
erosion using a disk shaped structuring element.
The special case that k = 1 is extensively studied in [10], here we only
present an important case. The case that k = 1 corresponds with a convex
indicator function b = IB where B is a convex set. In case B contains the
origin we may rewrite equation (3), leaving out the nite multiplications of the
functions b = IB and get IB (t2,1  ) + IB (s2,1  ) = IB ((s+_ t)2,1  ). Note
that IB (x) = IB= (x); which allows us to rewrite the above into It1,2 B +
Is1,2 B = I(s+_ t)1,2 B : Because IX + IY = IX \Y we nd It1,2 B + Is1,2 B =
I(t1,2 ^s1,2 )B . For  > 1=2 we have t1,2 ^ s1,2 = (t _ s)1,2 which leads
to s+_ t = s _ t: In case b = IB ; b = HB i.e. b is the support function of a
convex set B containing the origin. Summarizing:
Proposition 14 The erosion "(f ) = f  b with b = HB where B is convex set
containing the origin, induces an (S  ; _) scale-space if  > 1=2.
The scale-space operator in this case is T (t)f = f S  (t)b: Note that S  (t)b(x) =
t1, b(x=t ) = t1,2 b(x): Because  > 1=2 we have that t1,2 is a decreasing
function in t: The erosion f  HB resembles the band-pass lter known from
linear signal processing. Taking the conjugate (or slope transform) we obtain
(f  HB ) = f  + HB = f  + IB which equals f  ( ) if  2 B and 1 if  2= B:
Thus the erosion lters out all slopes that lie outside B .
8 REIN VAN DEN BOOMGAARD AND HENK HEIJMANS
5. Conclusions
In this paper we proposed an algebraic construction technique for morphological
scale-spaces induced by erosions (inmal convolutions) using convex structuring
functions. This technique is also valid to analyze the morphological scale-spaces
induced by openings and closings and even to analyze linear scale-spaces based
on convolutions. A comprehensive account can be found in [10].
In this paper we have presented a short overview of some results from convex
function theory. Analysis of linear scale-space operators makes heavily use of
the Fourier transform. Morphological operators allow for a similar change
in representation to facilitate the analysis of scale-space operators: the slope
transform or conjugate as it is well-known in convex function theory. Using
results from convex function theory we have presented some of the important
classes of morphological erosion scale-spaces.
All structuring functions that have been used in literature to construct mor-
phological scale-spaces, ranging from at convex structuring functions (`sets')
to the parabolic and quadratic structuring functions being the morphological
equivalent of the Gaussian function, are member of the class of subpolynomial
functions that we have proposed in this paper.

References
1. A. Witkin. Scale space ltering. pages 1019{1022, Palo Alto, California, 1983. Morgan-
Kaufmann.
2. J.J. Koenderink. The structure of images. Biological Cybernetics, (50):363{370, 1984.
3. L. M. J. Florack. Image Structure, volume 10 of Computational Imaging and Vision
Series. Kluwer Academic Publishers, Dordrecht, 1997.
4. J. Weickert, S. Ishikawa, and A. Imiya. On the history of Gaussian scale-space ax-
iomatics. In J. Sporring, M. Nielsen, L. Florack, and P. Johansen, editors, Gaussian
scale-space theory, pages 45{58. Kluwer, Dordrecht, 1997.
5. G. Matheron. Random Sets and Integral Geometry. John Wiley & Sons, New York,
1975.
6. P.T. jackway. Morphological scale-space. In Proceedings 11th IAPR International Con-
ference on Pattern Recognition, pages 252{255. IEEE Computer Society Press, 1992.
7. R. van den Boomgaard. Mathematical Morphology: Extensions towards Computer
Vision. PhD thesis, University of Amsterdam, 1992.
8. R. van den Boomgaard and A. Smeulders. The morphological structure of images:
the dierential equations of morphological scale space. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 16:1101{1113, 1994.
9. Luis Alvarez, Frederic Guichard, Pierre-Louis Lions, and Jean-Michel Morel. Axioms
and fundamental equations of image processing. (multiscale analysis and p.d.e.). C. R.
Acad. Sci., Paris, Ser. I, 315, No.2., 1992.
10. H.J.A.M. Heijmans and R. v.d. Boomgaard. Algebraic construction of linear and mor-
phological scale-spaces. in preparation, 1999.
11. J.-B. Hiriart-Urruty and C. Lemarechal. Convex Analysis and Minimization Algo-
rithms. Parts I and II. Springer, Berlin, 1993.
12. L. Dorst and R. van den Boomgaard. Morphological signal processing and the slope
transform. Signal Processing, 38:79{98, 1994.
13. P. Maragos. Morphological systems: slope transforms and max-min dierence and
dierential equations. Signal Processing, 38:57{77, 1994.
14. H. J. A. M. Heijmans and P. Maragos. Lattice calculus of the morphological slope
transform. Signal Processing, 59(1):17{42, 1997.