Professional Documents
Culture Documents
Component-Trees and Multivalued Images: Structural Properties
Component-Trees and Multivalued Images: Structural Properties
Component-Trees and Multivalued Images: Structural Properties
DOI 10.1007/s10851-013-0438-3
Abstract Component-trees model the structure of grey- the design of methods devoted to process or analyse grey-
level images by considering their binary level-sets obtained level images, based on hypotheses related to the topology
from successive thresholdings. They also enable to define (connectedness) and the specific intensity (local extrema) of
anti-extensive filtering procedures for such images. In or- structures of interest. Based on these properties, component-
der to extend this image processing approach to any (grey- trees have been involved in several image processing appli-
level or multivalued) images, both the notion of component- cations, especially for filtering and segmentation.
tree, and its associated filtering framework, have to be gen- The success of component-trees in the field of grey-level
eralised. In this article we deal with the generalisation of image processing, together with the increasing need for ap-
the component-tree structure. We define a new data struc- plications involving multivalued images, motivate their ex-
ture, the component-graph, which extends the notion of tension to the case of such images, which do not take their
component-tree to images taking their values in any (par- values in totally ordered sets, but in any (partially or totally)
tially or totally) ordered set. The component-graphs are de- ordered ones. In particular, this work takes place in the con-
clined in three variants, of increasing richness and size, text of the extension of mathematical morphology to multi-
whose structural properties are studied. valued images.
After a preliminary study of the relations between compo-
Keywords Mathematical morphology · Component-tree · nent-trees and multivalued images from a methodological
Multivalued images · Anti-extensive filtering · point of view [1], a generalisation of component-trees to
Component-graph such images has been initiated in [2]. The present work
develops this framework. In particular, this article deals
with data structure issues (algorithmic issues will be con-
1 Introduction sidered in a further article). We describe a new data struc-
ture, the component-graph, which extends the notion of
component-tree to images taking values in any ordered set.
The component-tree is a data structure which models some
The component-graphs are declined in three variants of in-
characteristics of grey-level images by considering their bi-
creasing richness and size. Their structural properties are
nary level-sets obtained from successive thresholding op-
studied, in particular under various hypotheses related to
erations. Component-trees are particularly well-suited for
frequent image value spaces.
This article is organised as follows. Section 2 describes
the context of this work. Section 3 provides notations. Sec-
N. Passat
CReSTIC, Université de Reims Champagne-Ardenne, Reims, tion 4 gives a formal definition of the classical notion of
France component-tree. Sections 5–7 constitute the contribution of
e-mail: nicolas.passat@univ-reims.fr the article. Section 5 defines the notion of component-graph.
Section 6 establishes the structural links between different
B. Naegel ()
ICube, CNRS, Université de Strasbourg, Strasbourg, France variants of this notion. Section 7 studies the influence of
e-mail: b.naegel@unistra.fr specific image value spaces, thus leading to structural sim-
J Math Imaging Vis
plifications. Section 8 summarises the principal properties [29–33]. These approaches present the advantage of embed-
of component-graphs. Section 9 concludes the article. ding multivalued images into simpler frameworks which au-
thorise to process them similarly to grey-level ones, enabling
in particular to reduce the algorithmic complexity induced
2 Related Works by partially ordered sets. Unfortunately, they may also po-
tentially hide the information intrinsically carried by these—
2.1 Component-Trees more complex but richer—partially ordered value spaces.
In the present work, we deal with the general case of
any (partially or totally) ordered value spaces, without at-
Initially proposed in the field of statistics [3, 4], the compo-
tempting to modify the order, then providing a contribution
nent-tree (also known as dendrone [5, 6], confinement tree
to the extension of mathematical morphology to multivalued
[7] or max-tree [8]) has been (re)defined in the framework
images, and more especially to the one of connected filter-
of mathematical morphology (see, e.g., [9, Chap. 7], or [10])
ing (see also [34, 35] for other recent contributions on this
and involved in the development of morphological operators
topic).
[8, 11].
From a methodological point of view, some efforts
have been conducted to enable the efficient computation of
3 Notations
component-trees [8, 12–14]. From an applicative point of
view, component-trees have been involved in the develop-
The inclusion (resp. strict inclusion) relation on sets is
ment of several image processing and analysis techniques.
noted ⊆ (resp. ⊂). The cardinality of a set X is noted |X|.
Most of them are devoted to filtering or segmentation [8,
The power set of a set X is noted 2X . If P ⊆ 2X is a partition
15–18]. Other applications have also been considered, for
of X, we write X = P.
instance, image registration [7, 19], image retrieval [20, 21],
A function F from a set X to a set Y is noted F : X → Y ,
image classification [22], interactive visualisation [23], mul-
and the set of all the functions from X to Y is noted Y X . If
tithresholding [24] or document binarisation [25].
X ⊆ X and Y ⊆ Y , we note F (X ) = {F (x) | x ∈ X } and
In the field of filtering and segmentation, the proposed
F −1 (Y ) = {x ∈ X | F (x) ∈ Y }. If F is a bijection, we also
methods have been designed to detect some structures of in-
note F −1 : Y → X its associated inverse function.
terest by using information modelled by attributes [11, 26],
Let be a (binary) relation on a set X. The restriction of
and stored at each node of the tree. These attributes are cho-
to a subset Y ⊆ X will generally still be noted (except
sen according to hypotheses related to the applicative con- if a new notation is introduced).
text. The subtree obtained by pruning the component-tree of We say that is an equivalence relation if is reflexive,
the image, with respect to these attributes, can then be used transitive and symmetric. For any x ∈ X, the equivalence
to reconstruct a binary (segmentation) or grey-level (filter- class of x with respect to is noted [x] . The set of all
ing) result. these equivalence classes is noted X/.
We say that is an order relation (and that (X, ) is
2.2 Mathematical Morphology and Multivalued Images an ordered set) if is reflexive, transitive and antisymmet-
ric. Moreover, we say that is a total (resp. partial) or-
Mathematical morphology has been first defined on binary der relation (and that (X, ) is a totally (resp. partially)
images, and then on grey-level ones (see [9] for a recent state ordered set), if is total (resp. partial) (i.e., if ∀x, y ∈ X,
of the art on mathematical morphology). Its extension to (x y) ∨ (y x) (resp. if ∃x, y ∈ X, (x
y) ∧ (y
x))).
multivalued (e.g., colour, multispectral, label) images is an For any symbol further used to denote an order relation
important task, motivated by potential applications in mul- (⊆, ≤, , etc.), the inverse symbol (⊇, ≥, , etc.) denotes
tiple areas. Several contributions have been devoted to this the associated dual order, while the symbol without lower
specific purpose (a whole state of the art is beyond the scope bar (⊂, <, , etc.) denotes the associated strict order.
of this article; see, e.g., [27] for a recent survey). The Hasse diagram of an ordered set (X, ) is the couple
In general, the spaces in which such images take their (X, ≺) where ≺ is the cover relation associated to , defined
values are not canonically equipped with total orders (by for all x, y ∈ X by x ≺ y iff x < y and there is no z ∈ X such
opposition to the case of grey-level images), but with partial that x < z < y.
ones. Several strategies have been considered to deal with If (X, ) is an ordered set and x ∈ X, we note x ↑ = {y ∈
this issue. Except in few works (see, e.g., [28]), they intend X | y x} and x ↓ = {y ∈ X | y x}, namely the sets of
to split these value spaces into several totally ordered ones the elements greater and lower than x, respectively. If Y ⊆
(marginal processing), or to define ad hoc total order rela- X, the sets of all the maximal and minimal elements of Y
` a
tions on them (vectorial processing), with several variants are noted Y and Y , respectively. The supremum and
J Math Imaging Vis
the infimum of Y are noted (when they exist) Y and In the illustrations of Sects. 4–7, the sets Ω will be finite
Y , respectively (we will note and for ⊆ and subsets of R2 equipped with the usual arc-based connectiv-
⊆
, respectively). The maximum and the minimum of Y ity.
b c
are noted (when they exist) Y and Y , respectively.
If Y is defined as {x | p(x)} where p is a Boolean predicate, 4.2 Images
b b
we will sometimes note p(x) x, instead of Y ; the same
c
remark holds for , , , , , . Let Ω be a nonempty finite set. Let V be a nonempty fi-
(Note that the symbol ≤ will be used to denote two dis- nite set equipped with an order relation . We assume that
tinct orders: the standard order on Z and the pointwise order (V , ) admits a minimum, noted ⊥. An image is a function
on functions; the context of use allows the reader to unam- I : Ω → V . The sets Ω and V are called the support and
biguously associate the correct semantics to each occurrence the value space of I , respectively. For any x ∈ Ω, I (x) ∈ V
of the symbol.) is the value of I at x. Without loss of generality, we assume
that I −1 ({⊥})
= ∅. If (V , ) is a totally (resp. partially) or-
dered set, we say that I is a grey-level (resp. a multivalued)
4 Component-Trees image.
Let X ⊆ Ω and v ∈ V . The thresholding function λv is
4.1 Connectivity defined by
λv : V Ω → 2 Ω
Intuitively (and informally), the notion of connectivity on a (1)
I → x ∈ Ω | I (x) v
set Ω allows to decide whether it is possible to go from a
point (or a subset) of Ω to another one while always re- The cylinder function C(X,v) is defined by
maining in Ω. If this property is verified, we say that Ω is
connected. Several (similar, and sometimes equivalent [36]) C(X,v) : Ω → V
From the order relation defined on V , and the inclusion Definition 1 (Component-graph(s)) Let I : Ω → V be an
relation ⊆ on 2Ω , we can define the order relation on Θ image. The Θ- (resp. Θ̇-, resp. Θ̈-)component-graph of I
as follows is the Hasse diagram G = (Θ, ) (resp. Ġ = (Θ̇, ), ˙ resp.
¨
G̈ = (Θ̈, )) of the ordered set (Θ, ) (resp. (Θ̇, ), resp.
(X1 ⊂ X2 )∨ (Θ̈, )). For the sake of concision, the term Θ̊-component-
(X1 , v1 ) (X2 , v2 ) ⇔ (9)
((X1 = X2 ) ∧ (v2 v1 )) graph and the notation G̊ = (Θ̊, )˚ will sometimes be used
This order relation on the valued connected components, to denote the three kinds of component-graphs. The ele-
previously introduced in [46], can be seen as the analogue of ments of Θ̊ are called Θ̊-nodes (or simply, nodes); the el-
ements of ˚ are called Θ̊-edges (or simply, edges); (Ω, ⊥)
the (order) inclusion relation on the connected components. a
In particular, for any (X1 , v1 ), (X2 , v2 ) ∈ Θ, it verifies the is called the root; the elements of Θ̊ are called the leaves
following properties of the Θ̊-component-graph.
(X1 ∩ X2
= ∅) ∧ (v2 v1 ) ⇒ (X1 ⊆ X2 ) (10) An example of component-graph is illustrated in Fig. 2.
From Formula (3), the reconstruction of I from its valued
(X1 , v1 ) (X2 , v2 ) ⇒ (v1
< v2 ) (11) connected components is given by
5.2 Component-Graphs ≤
j
I= CK (17)
In first approximation, the component-graph G of an image K∈Θ̊
I : Ω → V is defined as the Hasse diagram of the ordered
set (Θ, ). However, three variants of component-graphs
can relevantly be considered by defining two additional sub- 6 General Remarks on the Structure
sets Θ̇, Θ̈ ⊆ Θ of valued connected components (the use- of Component-Graphs
fulness of which will be justified in the sequel)
In this section, we study some links existing between the
h different structures of Θ̊-component-graphs and the Hasse
Θ̇ = {X} × v | X ∈ C λv (I ) (12) diagram (V , ≺) of (V , ).
X∈Ψ
≤
j 6.1 Links Between (Θ, ) and (V , ≺)
Θ̈ = (Ω, ⊥) ∪ Θ ⊆ Θ I = CK (13)
K∈Θ
The Θ-component-graph of I : Ω → V locally inherits
from the structure of (V , ≺). More precisely, for any leave
Broadly speaking, Θ gathers all the valued connected com- a
ponents induced by I ; Θ̇ gathers the valued connected com- K = (X, v) ∈ Θ, the ordered set (K ↑ , ) is “similar” to
ponents of maximal values for any connected components; the ordered set (v ↓ , ).
and Θ̈ gathers the valued connected components associated This similarity is first expressed by the fact that each
to the cylinders functions which are sup/max-generators of I valued connected component of K ↑ can be associated to a
(see Formula (3)). unique value of v ↓ and vice versa.
We note (resp. , ˙ resp. )
¨ the cover relation associated a
to the order relation on Θ (resp. to the restriction of to Property 1 Let K = (X, v) ∈ Θ. The function
Θ̇, resp. to the restriction of to Θ̈). From these definitions,
σ : K ↑ → v↓
we derive that (18)
(Y, u) → u
Θ̈ ⊆ Θ̇ ⊆ Θ (14)
is a bijection between K ↑ and v ↓ .
and
Proof The fact that σ ((Y, u)) ∈ v ↓ for any (Y, u) ∈ K ↑ de-
a
j j j rives from the fact that K ∈ Θ (this is generally not true
Θ= Θ̇ = Θ̈ = (Ω, ⊥) (15) otherwise). The injectivity and surjectivity of σ then derive
from Properties (P1) and (P2).
i i i
Θ= Θ̇ = Θ̈ (16)
This similarity is also expressed by the fact that the re-
We then have the following definition for the three vari- lation between two values of v ↓ is preserved between their
ants of component-graphs. associated valued connected components in K ↑ .
J Math Imaging Vis
Fig. 2 (a) A multivalued image I : Ω → V with Ω ⊂ R2 , and V = value σ (K) = v (see Property 1). The comparison between these four
{a, b, c, d, e, f, g, h, i, j }. (b) The Hasse diagram of the ordered set curves and the three curves of (b) illustrates the “flattening” of (K ↑ , )
(V , ). For the sake of readability, each value of V is associated to by comparison to (σ (K)↓ , ) (see Property 3). (n) The Θ̇-component-
an arbitrary colour. The (Hasse diagrams of the) ordered sets (f ↓ , ), graph of I . The nodes I, J, K, L, R and S, in Ġ are associated to the
(i ↓ , ), (j ↓ , ) are identified by the closed green, cyan and orange
sets of nodes {D, I}, {E, J}, {F, K}, {G, H, L}, {O, R} and {M, N,
curves, respectively. (c)–(l) Thresholded images λv (I ) for v ∈ V . (m)
The Θ-component-graph of I . The letters (A–S) in nodes correspond Q, S} in G, respectively (see Property 4). (o) The Θ̈-component-graph
to the associated connected components (c)–(l). For any one of the four of I . The nodes C and J, which are surrounded by {K, L} and {R, S},
a
leaves K = (X, v) ∈ Θ, the (Hasse diagram of the) ordered set respectively, belong to Θ̇ but not to Θ̈ (see Property 6) (Color figure
(K ↑ , ) is identified by a closed curve of colour corresponding to the online)
a
Property 2 Let K = (X, v) ∈ Θ. The function σ −1 : In general, this homomorphism is not an isomorphism.
v ↓ → K ↑ induces a homomorphism from (v ↓ , ) to In particular, it is possible that (X1 , v1 ) (X2 , v2 ) while
(K ↑ , ), i.e., for any K1 = σ −1 (v1 ), K2 = σ −1 (v2 ) ∈ K ↑ , v1
v2 . This property derives from the definition of (For-
we have mula (9)), in which ⊆ is considered prioritarily to .
From a theoretical point of view, the ordered set (K ↑ , )
is then richer than (v ↓ , ). (In other words, the cardinality
(v1 v2 ) ⇒ (K1 K2 ) (19)
of is higher than the cardinality of .) However, from a
practical point of view (K ↑ , ) is (most of the time) less
Proof The result derives from Property (P2). rich than (v ↓ , ≺).
J Math Imaging Vis
a
Property 3 Let K = (X, v) ∈ Θ. Let v1 , v2 ∈ v ↓ and Broadly speaking, all the edges between two nodes of
K1 = σ −1 (v1 ), K2 = σ −1 (v2 ) ∈ K ↑ . Let us suppose that (Θ, ) associated to a same connected component disappear
K1 K2 while v2
v1 . Let v3 ∈ v ↓ and K3 = σ −1 (v3 ) ∈ ˙ while any edge between two nodes of (Θ, ) as-
in (Θ̇, ),
K ↑ . We have sociated to distinct connected components leads to edges be-
˙ respectively associated
tween all pairs of nodes of (Θ̇, )
(v3 ≺ v1 ) ⇒ (v3 ≺ v2 ) ⇒ K1
K3 (20) to these two distinct connected components. These links
between Θ- and Θ̇-component-graphs are exemplified in
(v2 ≺ v3 ) ⇒ (v1 ≺ v3 ) ⇒ K3
K2 (21)
Fig. 2(m, n).
Proof If v3 ≺ v1 , v2 , we have v3 v1 , v2 and then In this example, for any K ∈ Θ̇, we have [K]∼Θ̇ = {K}.
K1 , K2 K3 . By hypothesis, we then have K1 K2 K3 . However, this is not necessarily true in general. Neverthe-
By definition of it comes K1
K3 , and Formula (20) less, since the existence of an edge between K1 and K2 in
then holds. A similar reasoning can be applied for For- ˙ implies the existence of a similar edge between any
(Θ̇, )
mula (21). K1 and K2 in [K1 ]∼Θ̇ and [K2 ]∼Θ̇ , respectively, we can un-
Fig. 3 (a) A multivalued image I : Ω → V with Ω ⊂ R2 , and (b) The Θ̇-component-graph of I . Each node of the graph is a val-
V = {a, b, c, d, e, f, g, h, i, j, k, l, m}. (d) The Hasse diagram of the ued connected component made of a connected component (A–F) of
ordered set (V , ). For the sake of readability, each value of V is a thresholded image depicted in (e)–(j), and a value corresponding to
associated to an arbitrary colour. (e)–(j) Thresholded images λv (I ) the associated colour in (d). (c) The “compact” Θ̇-component-graph
for v ∈ V : (e) λa (I ); (f) λb (I ) = λc (I ) = λd (I ) = λe (I ) = λf (I ); ˙ of I , which has the same structure as (Ψ, ⊆) (see Prop-
(Θ̇/∼Θ̇ , )
(g) λg (I ) = λi (I ) = λj (I ) = λk (I ); (h) λh (I ); (i) λl (I ); (j) λm (I ). erty 5) (Color figure online)
then, there exists x ∈ X such that I (x) = v. The a fact that 7.1 (V , ) is a Lower-Piecewise Lattice
K ∈ Θ̈ then derives from Formula (13). If K ∈
/ Θ̇ and
X = (X ,v )∈K ↓ \{K} X , then for each x ∈ X, there ex- An ordered set (X, ) is an upper- (resp. lower-)semilattice
ists K ∈ K ↓ \ {K} b such that v < CK (x). Consequently,
if for any x, y ∈ X, {x, y} (also noted x ∨ y) (resp.
we have CK (x) < ≤ C
K ∈K ↓ \{K} K
(x), and then, from For- {x, y} (also noted x ∧ y)) exists. It is a lattice if it is
mula (13), K ∈
/ Θ̈. both an upper- and a lower-semilattice.
An ordered set (X, ) is an upper- (resp. lower-)piecewi-
The links between Θ̇- and Θ̈-component-trees are exem- se lattice if for any x ∈ X, the ordered set (x ↑ , ) (resp.
plified in Fig. 2(n, o). (x ↓ , )) is a lattice. Note that an upper- (resp. lower-)
Note finally that there is no straightforward way to lo- semilattice is an upper- (resp. lower-)piecewise lattice, but
cally characterise the edges of Θ̈ from the ones of Θ̇. In- the converse is not true in general.
deed, the existence (resp. non-existence) of an edge between In this section, we assume that (V , ) is a lower-
two nodes of Θ̈ depends on the non-existence (resp. exis- piecewise lattice.
tence) of a transitive path composed of successive Θ̈-edges
between these nodes. 7.1.1 Structure of the Θ̊-Component-Graphs
7 About the Influence of (V , ) on Component-Graphs The Θ- component-graphs inherit from the structure of
(V , ).
In this section, we study how certain order relations
defined on V influence the structural properties of the Property 7 Let (V , ) be a lower-piecewise lattice. Then
component-graphs. In particular, we consider (from the
most general to the most specific) three kinds of ordered (Θ, ) is an upper-piecewise lattice. (29)
sets, frequent in image applications: lower-piecewise lattices a
(Sect. 7.1), lower-piecewise totally ordered sets (Sect. 7.2), Proof Let K = (X, v) ∈ Θ. It derives from Property 2
and totally ordered sets (Sect. 7.3). that (K ↑ , ) is a lattice. Since for any x ∈ X where (X, )
J Math Imaging Vis
is a lattice, (x ↑ , ) is still a lattice, (Θ, ) is an upper- Proof From Formula (32), for each K ∈ Θ̇, we have
piecewise lattice. [K]∼Θ̇ = {K}. The result then follows from Property 5.
As a corollary, we have the following property, related to Moreover, by extending the relation (and then ) from
the structure of the equivalence classes of ∼Θ . Θ to Θ/∼Θ as follows
k
k
Property 8 Let (V , ) be a lower-piecewise lattice. Let [K1 ]∼Θ [K2 ]∼Θ ⇔ [K1 ]∼Θ [K2 ]∼Θ (34)
K ∈ Θ, then
we have the following property.
[K]∼Θ , is a lower-semilattice. (30)
Property 11 If (V , ) is a lower-piecewise lattice, then
a
Proof Let K = (X, v). Let K = (Y, u) ∈ Θ such that
Y ⊆ X. From Property 7, (K ↑ , ) is a lattice. Moreover (Θ/∼Θ , ) is isomorphic to (Θ̇, ). (35)
we have [K]∼Θ ⊆ K ↑ . As (V , ) is a lower-piecewise
lattice, (u↓ , ) is a lattice. Let (X, v1 ), (X, v2 ) ∈ [K]∼Θ . Proof The proof derives from the equality between any
c
We have X ⊆ λv1 ∨ v2 (I ), and then, from Property (P3), K ∈ Θ̇ and [K]∼Θ .
X ∈ C[λv1 ∨ v2 (I )]. Consequently, we have (X, v1 ∨ v2 ) ∈
[K]∼Θ , and the result follows. This identification is completed by the fact that for all
K ∈ Θ̇, ([K]∼Θ , ) is a lower-semilattice (Property 8), and
by the following property.
It then derives from this property that the Θ̇-component-
graphs also inherit from the structure of (V , ). Property 12 Let (V , ) be a lower-piecewise lattice. For
all K1 , K2 ∈ Θ (with [K1 ]∼Θ
= [K2 ]∼Θ ) we have
Property 9 Let (V , ) be a lower-piecewise lattice. Then
k
k
(Θ̇, ) is an upper-piecewise lattice. (31) (K1 K2 ) ⇒ ˙
[K1 ]∼Θ [K2 ]∼Θ (36)
Proof The result derives from the fact that (Θ, ) is an Proof The result derives from Property 4.
upper-piecewise lattice and that ([K]∼Θ , ) admits a mini-
mum for any K ∈ Θ. Broadly speaking, all the edges between two nodes of
(Θ, ) associated to a same connected component disappear
˙ while any edge between two nodes of (Θ, ) as-
in (Θ̇, ),
Note that, in general, (Θ̈, ) is not an upper-piecewise
sociated to distinct connected components lead to edges be-
lattice. These properties are exemplified in Fig. 4.
tween the nodes of (Θ̇, )˙ respectively associated to these
two distinct connected components. These links between
7.1.2 Links Between Θ- and Θ̇-Component-Graphs Θ- and Θ̇-component-trees are exemplified in Fig. 4(r, s).
Under the current hypotheses (see Property 8), Formula (12) 7.2 (V , ) is a Lower-Piecewise Totally Ordered Set
can be rewritten as
We say that an order relation on a set X is upper-
j (resp. lower-)piecewise total (and that (X, ) is an upper-
Θ̇ = X, v X ∈ Ψ (32) (resp. lower-)piecewise totally ordered set) if for any x ∈ X,
X∈C [λv (I )] (x ↑ , ) (resp. (x ↓ , )) is a totally ordered set.
In this section, we assume that (V , ) is a lower-
This formula leads to the following result. piecewise totally ordered set. Note that (V , ) is then also a
lower-piecewise lattice.
Property 10 Let (V , ) be a lower-piecewise lattice. By
considering the bijection between Θ̇ and Ψ which asso- 7.2.1 Structure of the Θ̊-Component-Graphs
ciates any node (X, v) to X,
Under the current hypotheses, the result of Property 2 is
(Θ̇, ) is isomorphic to (Ψ, ⊆). (33) strengthened.
J Math Imaging Vis
Property 13 Let K = (X, v) ∈ Θ. By considering the bijec- Property 14 Let (V , ) be a lower-piecewise totally or-
tion σ : K ↑ → v ↓ , dered set. Then
↑
K , is isomorphic to v ↓ , . (37)
(Θ̊, ) is an upper-piecewise totally ordered set. (38)
Proof The result derives from the totality of on v ↓ .
Proof The result derives from Property 13 and the fact
The Θ̊-component-graphs then inherit from the structure that any subset of a totally ordered set is itself totally or-
of (V , ). dered.
J Math Imaging Vis
k
Property 17 Let (V , ) be a totally ordered set. Then
(K1 K2 ) ⇒ ˙ K2
[K1 ]∼Θ (40)
(Θ̇, ) = (Θ̈, ) (42)
J Math Imaging Vis
Proof The result derives from Property 6 and Proper- Proof Formula (43) derives from Properties 10 and 17. For-
ties (P2) and (P3). mula (44) is a rewriting of Formulae (7) and (17).
Property 18 Let I : Ω → V be a grey-level image, and T Table 1 describes the component-graphs/tree defined by the
its component-tree. Then different ordered sets induced by Θ̊ and Ψ . Tables 2 and 3
summarise the isomorphism relations between them, and the
Ġ, G̈ are isomorphic to T (43) nature of their order. These properties are classified accord-
ing to the nature of the ordered set (V , ).
and
8.2 Space Complexity
≤
j ≤
j
I= C(X,m(X)) = CK (44) From an algorithmic point of view (and, a fortiori, from an
X∈Ψ K∈Θ̊ applicative one), the size of the Θ̊-component-graph G̊ of
J Math Imaging Vis
an image I : Ω → V strongly conditions the ability to pro- 4. Hartigan, J.A.: Statistical theory in clustering. J. Classif. 2(1), 63–
cess this image via a filtering framework relying on G̊. From 76 (1985)
5. Hanusse, P., Guillataud, P.: Sémantique des images par analyse
a theoretical point of view, we have the following upper- dendronique. In: RFIA, Proceedings, pp. 577–588 (1991)
bounds for the number of nodes of G̊ 6. Chen, L., Berry, M.W., Hargrove, W.W.: Using dendronal sig-
natures for feature extraction and retrieval. Int. J. Imaging Syst.
|Θ| = O |Φ|.|V | (45) Technol. 11(4), 243–253 (2000)
7. Mattes, J., Demongeot, J.: Efficient algorithms to implement the
|Θ̇| = O |Φ|.|V | (46) confinement tree. In: DGCI, Proceedings. Lecture Notes in Com-
puter Science, vol. 1953, pp. 392–405. Springer, Berlin (2000)
|Θ̈| = O |Φ| (47) 8. Salembier, P., Oliveras, A., Garrido, L.: Anti-extensive connected
operators for image and sequence processing. IEEE Trans. Image
where Φ ⊆ 2Ω is the set of all the flat zones of I , with Process. 7(4), 555–570 (1998)
thus |Φ| ≤ |Ω| (and generally, |Φ| |Ω|). From a prac- 9. Najman, L., Talbot, H. (eds.): Mathematical Morphology: from
Theory to Applications. ISTE/J. Wiley, New York (2010)
tical point of view, for real images, the actual number of 10. Salembier, P., Wilkinson, M.H.F.: Connected operators: a re-
nodes in the Θ̊-component-graph of an image I : Ω → V view of region-based morphological image processing techniques.
will be generally (much) lower than these bounds, since it IEEE Signal Process. Mag. 26(6), 136–157 (2009)
11. Breen, E.J., Jones, R.: Attribute openings, thinnings, and granu-
will depend on |Φ|, |V |, but also on the structure of the im-
lometries. Comput. Vis. Image Underst. 64(3), 377–389 (1996)
age itself, and in particular the number of its maxima and 12. Najman, L., Couprie, M.: Building the component tree in quasi-
their “height”, with respect to (V , ). In particular, we will linear time. IEEE Trans. Image Process. 15(11), 3531–3539
have (2006)
13. Berger, C., Géraud, T., Levillain, R., Widynski, N., Baillard, A.,
|Θ̈| ≤ |Θ̇| ≤ |Θ| ≤ σ (K)↓ (48)
Bertin, E.: Effective component tree computation with application
to pattern recognition in astronomical imaging. In: ICIP, Proceed-
a
K∈ Θ ings, pp. 41–44 (2007)
14. Wilkinson, M.H.F., Gao, H., Hesselink, W.H., Jonker, J.E., Mei-
jster, A.: Concurrent computation of attribute filters on shared
memory parallel machines. IEEE Trans. Pattern Anal. Mach. In-
9 Conclusion tell. 30(10), 1800–1813 (2008)
15. Jones, R.: Connected filtering and segmentation using component
The notion of component-graph has been proposed as an ex- trees. Comput. Vis. Image Underst. 75(3), 215–228 (1999)
16. Dokládal, P., Bloch, I., Couprie, M., Ruijters, D., Urtasun, R., Gar-
tension of the notion of component-tree for images taking nero, L.: Topologically controlled segmentation of 3D magnetic
their values in any (partially or totally) ordered sets. Some resonance images of the head by using morphological operators.
structural properties of three variants of component-graphs Pattern Recognit. 36(10), 2463–2478 (2003)
have been investigated, in particular for value spaces which 17. Ouzounis, G.K., Wilkinson, M.H.F.: Mask-based second-
generation connectivity and attribute filters. IEEE Trans. Pattern
are likely to appear in real images (e.g., piecewise totally Anal. Mach. Intell. 29(6), 990–1004 (2007)
ordered sets or lattices). 18. Passat, N., Naegel, B., Rousseau, F., Koob, M., Dietemann,
In further works, we will propose to generalise the stan- J.L.: Interactive segmentation based on component-trees. Pattern
dard antiextensive grey-level image filtering framework Recognit. 44(10–11), 2539–2554 (2011)
19. Mattes, J., Richard, M., Demongeot, J.: Tree representation for im-
based on component-trees [8, 15], thus relaxing the con- age matching and object recognition. In: DGCI, Proceedings. Lec-
straints linked to total orderings on image values. In par- ture Notes in Computer Science, vol. 1568, pp. 392–405. Springer,
ticular, the component-graphs(s) construction and the non- Berlin (1999)
20. Mosorov, V.: A main stem concept for image matching. Pattern
trivial issue of image reconstruction will be discussed [47].
Recognit. Lett. 26(8), 1105–1117 (2005)
21. Alajlan, N., Kamel, M.S., Freeman, G.H.: Geometry-based im-
Acknowledgements The research leading to these results has re- age retrieval in binary image databases. IEEE Trans. Pattern Anal.
ceived funding from the French Agence Nationale de la Recherche Mach. Intell. 30(6), 1003–1013 (2008)
(Grant Agreement ANR-10-BLAN-0205). 22. Urbach, E.R., Roerdink, J.B.T.M., Wilkinson, M.H.F.: Connected
shape-size pattern spectra for rotation and scale-invariant classi-
fication of gray-scale images. IEEE Trans. Pattern Anal. Mach.
Intell. 29(2), 272–285 (2007)
References 23. Westenberg, M.A., Roerdink, J.B.T.M., Wilkinson, M.H.F.: Vol-
umetric attribute filtering and interactive visualization using the
1. Naegel, B., Passat, N.: Component-trees and multi-value im- max-tree representation. IEEE Trans. Image Process. 16(12),
ages: a comparative study. In: ISMM, Proceedings. Lecture Notes 2943–2952 (2007)
in Computer Science, vol. 5720, pp. 261–271. Springer, Berlin 24. Menotti, D., Najman, L., de Albuquerque Araújo, A.: 1D compo-
(2009) nent tree in linear time and space and its application to gray-level
2. Passat, N., Naegel, B.: An extension of component-trees to partial image multithresholding. In: ISMM, Proceedings, INPE, vol. 1,
orders. In: ICIP, Proceedings, pp. 3981–3984 (2009) pp. 437–448 (2007)
3. Wishart, D.: Mode analysis: a generalization of the nearest neigh- 25. Naegel, B., Wendling, L.: A document binarization method based
bor. In: Numerical Taxonomy, pp. 282–319. Academic Press, San on connected operators. Pattern Recognit. Lett. 31(11), 1251–
Diego (1969) 1259 (2010)
J Math Imaging Vis
26. Urbach, E.R., Boersma, N.J., Wilkinson, M.H.F.: Vector attribute 44. Braga-Neto, U., Goutsias, J.: Connectivity on complete lattices:
filters. In: ISMM, Proceedings. Computational Imaging and Vi- new results. Comput. Vis. Image Underst. 85(1), 22–53 (2002)
sion, vol. 30, pp. 95–104. Springer, Berlin (2005) 45. Ronse, C.: Idempotent block splitting on partial partitions, I: Iso-
27. Aptoula, E., Lefèvre, S.: A comparative study on multivariate tone operators. Order 28(2), 273–306 (2011)
mathematical morphology. Pattern Recognit. 40(11), 2914–2929 46. Caselles, V., Monasse, P.: Geometric Description of Images as To-
(2007) pographic Maps. Lecture Notes in Mathematics, vol. 29. Springer,
28. Ronse, C., Agnus, V.: Morphology on label images: flat-type op- Berlin (2010)
erators and connections. J. Math. Imaging Vis. 22(2), 283–307 47. Naegel, B., Passat, N.: Toward connected filtering based on
(2005) component-graphs. In: ISMM, Proceedings. Lecture Notes in
29. Barnett, V.: The ordering of multivariate data. J. R. Stat. Soc., Ser. Computer Science, vol. 7883, pp. 350–361. Springer, Berlin
A, Stat. Soc. 139(3), 318–354 (1976) (2013)
30. Goutsias, J., Heijmans, H.J.A.M., Sivakumar, K.: Morphologi-
cal operators for image sequences. Comput. Vis. Image Underst.
62(3), 326–346 (1995) Nicolas Passat was born in 1978.
31. Talbot, H., Evans, C., Jones, R.: Complete ordering and multivari- He studied computer science at the
ate mathematical morphology. In: ISMM, Proceedings, pp. 27–34. Université d’Orléans and the Uni-
Kluwer Academic, Norwell (1998) versité Strasbourg 1 (M.Sc., 2002;
32. Aptoula, E., Lefèvre, S.: On lexicographical ordering in multivari- Ph.D., 2005, Habilitation, 2011),
ate mathematical morphology. Pattern Recognit. Lett. 29(2), 109– specialising in image analysis. In
118 (2008) 2006, he worked at the Univer-
33. Angulo, J.: Geometric algebra colour image representations and sité Paris-Est Marne-la-Vallée, as a
derived total orderings for morphological operators—Part I: post-doctoral fellow. Between 2006
Colour quaternions. J. Vis. Commun. Image Represent. 21(1), 33– and 2012, he worked at the Univer-
48 (2010) sité de Strasbourg, as an assistant
34. Gimenez, D., Evans, A.N.: An evaluation of area morphology professor. Since 2012, he has been
scale-spaces for colour images. Comput. Vis. Image Underst. professor at the Université de Reims
110(1), 32–42 (2008) Champagne-Ardenne. His scientific
35. Soille, P.: Constrained connectivity for hierarchical image parti- interests include mathematical mor-
tioning and simplification. IEEE Trans. Pattern Anal. Mach. Intell. phology, discrete topology, medical imaging and remote sensing.
30(7), 1132–1145 (2008)
36. Mazo, L., Passat, N., Couprie, M., Ronse, C.: Paths, homotopy Benoît Naegel was born in 1978.
and reduction in digital images. Acta Appl. Math. 113(2), 167– He studied computer science at
193 (2011) the Université Strasbourg 1 (M.Sc.,
37. Maunder, C.R.F.: Algebraic Topology. Dover, New York (1996) 2000; Ph.D., 2004), specialising in
38. Stong, R.E.: Finite topological spaces. Trans. Am. Math. Soc. image analysis. In 2005, he worked
123(25), 325–340 (1966) at the University of Applied Sci-
39. Rosenfeld, A.: Connectivity in digital pictures. J. Assoc. Comput. ences of Geneva as a research as-
Mach. 17(1), 146–160 (1970) sistant. Between 2007 and 2011, he
40. Kong, T.Y., Rosenfeld, A.: Digital topology: introduction and sur- worked at the Université Nancy 1, as
vey. Comput. Vis. Graph. Image Process. 48(3), 357–393 (1989) an assistant professor. Since 2011,
41. Kovalevsky, V.A.: Finite topology as applied to image analysis. he has been assistant professor at the
Comput. Vis. Graph. Image Process. 46(2), 141–161 (1989) Université de Strasbourg. His sci-
42. Ronse, C.: Set-theoretical algebraic approaches to connectivity in entific interests include mathemati-
continuous or digital spaces. J. Math. Imaging Vis. 8(1), 41–58 cal morphology, document analysis,
(1998) medical imaging and remote sens-
43. Serra, J.: Connectivity on complete lattices. J. Math. Imaging Vis. ing.
9(3), 231–251 (1998)