DeepAI

A Class of LULU Operators on Multi-Dimensional Arrays

The LULU operators for sequences are extended to multi-dimensional arrays via the morphological concept of connection in a way which preserves their essential properties, e.g. they are separators and form a four element fully ordered semi-group. The power of the operators is demonstrated by deriving a total variation preserving discrete pulse decomposition of images.

• 3 publications
• 1 publication
03/03/2010

Properties of the Discrete Pulse Transform for Multi-Dimensional Arrays

This report presents properties of the Discrete Pulse Transform on multi...
07/06/2017

Algorithms for zero-dimensional ideals using linear recurrent sequences

Inspired by Faugère and Mou's sparse FGLM algorithm, we show how using l...
04/29/2019

Partially Ordered Statistics Demapping for Multi-Dimensional Modulation Formats

We propose a very low-complexity and high-performance algorithm for soft...
05/01/2020

Multi-dimensional Arrays with Levels

We explore a data structure that generalises rectangular multi-dimension...
03/21/2013

Multi-dimensional sparse structured signal approximation using split Bregman iterations

The paper focuses on the sparse approximation of signals using overcompl...
07/13/2016

Deleting and Testing Forbidden Patterns in Multi-Dimensional Arrays

Understanding the local behaviour of structured multi-dimensional data i...
06/22/2018

On the Design of Multi-Dimensional Compactly Supported Parseval Framelets with Directional Characteristics

In this paper, we propose a new method for the construction of multi-dim...

1 Introduction

The LULU operators and the associated Discrete Pulse Transform developed during the last two decades or so are an important contribution to the theory of the nonlinear multi-resolution analysis of sequences. The basics of the theory as well as the most significant results until 2005 are published in the monograph [13]. For more recent developments and applications see [1], [4], [7], [8], [14]. Central to the theory is the concept of separator. This concept is defined in [13] only for operators on sequences due to the context of the book. However, it is meaningful in more general settings. In fact, some of the axioms have been used earlier, e.g. see [18], for functions on arbitrary domains. We will give the definition of separator for operators on real functions defined on a domain with a group structure.

Let a be an abelian group. Denote by

the vector lattice of all real functions defined on

with respect to the usual point-wise defined addition, scalar multiplication and partial order. For every the operator given by

 Ea(f)(x)=f(x+a), x∈Ω,

is called a shift operator.

Definition 1

An operator is called a separator if

 (i) P∘Ea=Ea∘P, a∈Ω; (ii) P(f+c)=P(f)+c, f,c∈A(Ω), c - constant function; (iii) P(αf)=αP(f), α∈R, α≥0, f∈A(Ω); (iv) P∘P=P;(Idempotence) (v) (id−P)∘(id−P)=id−P.(Co-idempotence)

Here denotes the identity operator and the operator is defined in terms of the point-wise linear operations for the operators on , that is, . The first two axioms in Definition 1 and partially the third one were first introduced as required properties of nonlinear smoothers by Mallows, [9]. Rohwer further made the concept of a smoother more precise by using the properties (i)–(iii) as a definition of this concept. The axiom (iv) is an essential requirement for what is called a morphological filter, [18], [19], [21]. In fact, a morphological filter is exactly a syntone operator which satisfies (iv). Let us recall that an operator is called syntone if

 f≤g⟹P(f)≤P(g) ,  f,g∈A(Ω).

The co-idempotence axiom (v) in Definition 1 was introduced by Rohwer in [13], where it is also shown that it is an essential requirement for operators extracting signal from a sequence.

The LULU theory was developed for sequences, that is, the case . Given a bi-infinite sequence and the basic LULU operators and are defined as follows

 (Lnξ)i=max{min{ξi−n,...,ξi},...,min{ξi,...,ξi+n}}, i∈Z. (1) (Unξ)i=min{max{ξi−n,...,ξi},...,max{ξi,...,ξi+n}}, i∈Z. (2)

It is shown in [13] that for every the operators and as well as their compositions are syntone separators. Hence they are an appropriate tool for signal extraction. Furthermore, these operators form the so called strong LULU semi-group. This a four element semi-group with respect to composition, see Table 1, which is fully ordered with respect to the usual point-wise defined order

 P≤Q⟺P(f)≤Q(f), f∈A(Ω). (3)

We have

 Ln≤Un∘Ln≤Ln∘Un≤Un. (4)

Let us recall that, according to the well known theorem of Matheron [10], in general, two ordered morphological operators generate a six element semi-group which is only partially ordered.

The power of the LULU operators as separators is further demonstrated by their Total Variation Preservation property. Let be the set of sequences with bounded variation, that is,

 BV(Z)={ξ∈A(Z):∑i∈Z|ξi−ξi+1|<∞}.

Total Variation of a sequence is given by .

Definition 2

An operator is called total variation preserving if

 TV(ξ)=TV(P(ξ))+TV((id−P)(ξ)) ,  ξ∈BV(Z). (5)

We should note that since is a semi-norm on we always have

 TV(ξ)≤TV(P(ξ))+TV((id−P)(ξ)).

Hence, the significance of the equality (5) is that the decomposition does not create additional total variation. In particular, this property is very important for the application of the LULU operators to discrete pulse decompositions of sequences.

The aim of this paper is to generalize the LULU operators to functions on in such a way that their essential properties are preserved. In Section 2 the definitions of the basic operators and on are derived by using a strengthened form of the morphological concept of connection. Then we show that indeed these operators replicate the properties of the LULU operators for sequence. More precisely, we prove that: (i) they are separators (Section 2); (ii) their smoothing effect can be described in a similar way to the -monotonicity of sequences (Section 3); (iii) they generate a four element fully ordered semi-group (Section 4). The developed theory can be applied to many problems of Image Analysis and it is the intention of the authors to research such applications in the future. However, as an illustration and demonstration of the power of this approach we apply the newly defined operators to deriving a total variation preserving discrete pulse decomposition of images. Noise removal and partial reconstructions are discussed in Section 6.

2 The basic operators Ln and Un.

The definition of the operators and for sequences involves maxima and minima over sets of consecutive terms, thus, making an essential use of the fact that is totally ordered. Since , , is only partially ordered the concept of ‘consecutive’ does not make sense in this setting. Instead, we use the morphological concept of set connection, [19].

Definition 3

Let be an arbitrary non-empty set. A family of subsets of is called a connected class or a connection on if
(i)
(ii) for all
(iii) for any family

 ⋂i∈ICi≠∅⟹⋃i∈ICi∈C

This definition generalizes the topological concept of connectivity to arbitrary sets including discrete sets like . If a set belongs to a connection then is called connected.

It is clear from Definition 3 that a connection on does not necessarily contain sets of every size. For example, and are connections on but neither of them contain sets of finite size other than 0 and 1. In the definition of the operators and we need sets of every size. We assume that the set is equipped with a connection which satisfies the following three conditions

 ∙ Zd∈C (6) ∙ For any a∈Zd,Ea(C)∈C % whenever C∈C (8) ∙ If V⊊W, V,W∈C, then there exists x∈W∖V such that V∪{x}∈C

The aim of the conditions (6)–(8) is to define a connection which is sufficiently rich in connected sets. This is demonstrated by the following property, which is obtained via iterative application of the property (8):

 ∙ Let V⊊W, V,W∈C. For every k∈N such that card(V)

As usual, is the number of the elements in the set , that is, the size of . For we have . Given a point and we denote by the set of all connected sets of size , which contain point , that is,

 Nn(x)={V∈C:x∈V, card(V)=n+1}. (10)

Now the operators and are defined on as follows.

Definition 4

Let and . Then

 Ln(f)(x) = maxV∈Nn(x)miny∈Vf(y), x∈Zd, (11) Un(f)(x) = minV∈Nn(x)maxy∈Vf(y), x∈Zd. (12)

Let us first see that Definition 4 generalizes the definition of and for sequences. Suppose and let be the connection on generated by the pairs of consecutive numbers. Then all connected sets on are sequences of consecutive integers and for any we have

 Nn(i)={{i−n,i−n+1,...,i},{i−n+1,i−n+2,...,i+1},...,{i,i+1,...,i+n}}

Hence for an arbitrary sequence considered as a function on the formulas (11) and (12) are reduced to (1) and (2), respectively.

Theorem 5

(Order Properties)

Proof. We will only prove the inequalities involving since those involving are proved similarly.

a) Let . For every and we have

 miny∈Vf(y)≤f(x).

Hence

 Ln(f)(x)=maxV∈Nn(x)miny∈Vf(y)≤f(x), x∈Zd.

Therefore, , , which implies .

b) Let . For any and , we have . Therefore

 Ln(f)(x)=maxV∈Nn(p)miny∈Vf(y)≤maxV∈Nn(p)miny∈Vg(y)=Ln(g)(x), x∈Zd.

c) Let . It follows from (2) that for every and there exists a set such that . Therefore

 miny∈Vf(y)≤miny∈Wf(y)≤maxS∈Nn1(x)miny∈Sf(y)=Ln1(f)(x).

Hence

 Ln2(f)(x)=maxV∈Nn2(x)miny∈Vf(y)≤Ln1(f)(x),x∈Zd.

Theorem 6

For any the operators and are separators.

Proof. We will only verify the conditions (i)–(v) in Definition 1 for since is dealt with in a similar manner.

(i) Let and . Using the property (8), for every we have

 Nn(x+a)=a+Nn(x)={a+V:V∈Nn(x)}

Therefore,

 Ea(Ln(f))(x) = Ln(f)(x+a)=maxV∈Nn(x+a)miny∈Vf(y) = maxV∈Nn(x)miny∈a+Vf(y) = maxV∈Nn(x)miny∈Vf(y+a) = maxV∈Nn(x)miny∈VEa(f)(y), x∈Zd

(ii) Let , where is a constant function with a value of . Then for every we have

 Ln(f+c)(x) = maxV∈Nn(x)miny∈V(f+c)(y) = maxV∈Nn(x)miny∈V(f(y)+θ) = (maxV∈Nn(x)miny∈Vf(y))+θ = Ln(f)(x)+c(x)

(iii) Let and , . For every we have

 Ln(αf)(x)=maxV∈Nn(x)minq∈V(αf)(y)=α(maxV∈Nn(x)minq∈Vf(y))=αLn(f)(x).

(iv) The inequality

 Ln∘Ln≤Ln

is an immediate consequence of Theorem 5. Then it is sufficient to prove the inverse inequality. Let and . We have

 Ln(Ln(f))(x)=maxW∈Nn(x)miny∈WmaxV∈Nn(y)minz∈Vf(z). (13)

But implies . Therefore for every and we have

 maxV∈Nn(y)minz∈Vf(z)≥minz∈Wf(z).

Using that the right hand side is independent of we further obtain

 miny∈WmaxV∈Nn(y)minz∈V≥minz∈Wf(z), W∈Nn(x).

Then it follows from the representation (13) that

 Ln(Ln(f))(x)≥maxW∈Nn(x)minz∈Wf(z)=Ln(f)(x).

(v) The co-idempotence of the operator is equivalent to . The inequality is an easy consequence of Theorem 5. Hence, for the co-idempotence of it remains to show that . Assume the opposite. Namely, there exists a function and such that . Using the definition of this inequality implies that there exists such that for every we have , or equivalently

 f(y)>Ln(f)(y), y∈V. (14)

Let be such that . Then for every we have

 Ln(f)(y)=maxW∈Nn(y)mint∈Wf(t)≥mint∈Vf(t)≥f(z). (15)

Taking in (14) and (15) we obtain a contradiction which completes the proof.

3 The operators Ln and Un as smoothers

Similar to their counterparts for sequences the operators and defined in Section 2 smooth the input function by removing sharp peaks (the application of ) and deep pits (the application of ). The smoothing effect of these operations is made more precise by using the concepts of a local maximum set and a local minimum set given below.

Definition 7

Let . A point is called adjacent to if . The set of all points adjacent to is denoted by , that is,

An equivalent formulation of the property (8) of the connection is as follows:

Definition 8

A connected subset of is called a local maximum set of if

Similarly is a local minimum set if

The next four theorems deal with different aspects of the application of and to functions in . Their cumulative effect will be discussed at the end of the section. All theorems contain statements a) and b). Due to the similarity we present only the proofs of a).

Theorem 9

Let and . Then we have
a) if and only if there exists a local maximum set such that and ;
b) if and only if there exists local minimum set such that and .

Proof. a) Implication to the left. Suppose that there exists a local maximum set , . Consider an arbitrary and let . Then, since the size of is larger than the size of we have . Furthermore, by (16) we have . Let . Since , we have that but Then using also that is a local maximum set we obtain

 miny∈Wf(y)≤f(z)

Since the set is arbitrary, this inequality implies that .

Implication to the right. Suppose . Let be the largest (in terms of ) connected set containing such that

 f(y)≥f(x), y∈V. (17)

The set is obviously unique and can be constructed as , where is the morphological point connected opening generated by , see [19] or [20], and . We have , , because otherwise (17) is satisfied on the larger connected set . Therefore

Hence is a local maximum set.

Assume that . It follows from (2) that there exists such that . Then

 Ln(f)(x)=maxV∈Nn(x)miny∈Vf(y)≥miny∈Wf(y)≥miny∈Vf(y)=f(x).

This contradicts the assumption . Therefore, .

Theorem 10

Let . Then
a) the size of any local maximum set of the function is larger than ;
b) the size of any local minimum set of the function is larger than .

Proof. a) Assume the opposite, that is, there exists a local maximum set of such that . By Theorem 9 we have that

 Ln(Ln(f))(x)

Since is idempotent, see Theorem 6, this implies the impossible inequality , which completes the proof.

Theorem 11

Let and let .
a) If then ;
b) If then .

Proof. a) For any the set is connected and of size larger than . Therefore, by (2), for every there exists such that . Then, using also the given inequality, for every and we have

 minz∈Wf(z)=minz∈W∪Vf(z)≤minz∈Syf(z)≤Ln(f)(y).

Hence

 Ln(f)(x)=maxW∈Nn(x)minz∈Wf(z)≤miny∈VLn(f)(y).

Theorem 12

Let and .
a) If is a local minimum set of then there exists a local minimum set of such that .
b) If is a local maximum set of then there exists a local maximum set of such that .

Proof. a) Let be a local minimum set of . Then

Let be such that and let

 Y={y∈V:f(y)

An easy application of Theorem 11 shows that . Let and let be the largest (with respect to inclusion) connected subset of which contains . As in the proof of Theorem 9, the set can be obtained through . For every we have . Therefore is a local minimum set of .

Theorems 912 provide the following characterization of the effect of the operators and of a function :

• The application of () removes local maximum (minimum) sets of size smaller or equal to .

• The operator () does not affect the local minimum (maximum) sets in the sense that such sets may be affected only as a result of the removal of local maximum (minimum) sets. However, no new local minimum sets are created where there were none. This does not exclude the possibility that the action of () may enlarge existing local maximum (minimum) sets or join two or more local maximum (minimum) sets of into one local maximum (minimum) set of ().

• () if and only if does not have local maximum (minimum) sets of size or less;

Furthermore, as an immediate consequence of Theorem 10 and Theorem 12 we obtain the following corollary.

Corollary 13

For every the functions and have neither local maximum sets nor local minimum sets of size or less. Furthermore,

 (Ln∘Un)(f)=(Un∘Ln)(f)=f

if and only if does not have local maximum sets or local minimum sets of size less than or equal to .

We should remark that in the one dimensional setting, the sequences without local maximum sets or local minimum sets of size less than or equal to are exactly the so-called -monotone sequences. Hence Corollary 13 generalizes the respective results in the LULU theory of sequences, [13, Theorem 3.3].

4 The LULU semi-group

In this section we consider the operators , and their compositions. The main result is that , , and form a semi-group with respect to composition with a composition table as given in Table 1. Furthermore, the semi-group is totaly ordered as in (4) with respect to the point-wise defined partial order (3).

Theorem 14

The operators and are idempotent, that is,

 Ln∘Un∘Ln∘Un=Ln∘Un , (18) Un∘Ln∘Un∘Ln=Un∘Ln . (19)

Proof. Using the order properties in Theorem 5 and the idempotence of and , see Theorem 6, we have

 Ln∘Un∘Ln∘Un≤Ln∘Un∘id∘Un=Ln∘Un∘Un=Ln∘Un Ln∘Un∘Ln∘Un≥Ln∘id∘Ln∘Un=Ln∘Ln∘Un=Ln∘Un

which implies (18). The equality (19) is proved similarly.

Theorem 15

For any we have

 Ln∘Un∘Ln=Un∘Ln. (20)

Proof. It follows from Theorem 5 that

 Ln∘Un∘Ln≤id∘Un∘Ln=Un∘Ln. (21)

Assume that (20) is violated. In view of (21), this means that there exists and such that

 Ln(Un(Ln(f)))(z)

It follows from Theorem 9 that there exists and such that is a local maximum set for . Then, by Theorem 12, there exists such that is a local maximum set of the function . We have . However, does not have any local maximum sets of size less than or equal to , see Theorem 10. This contradiction completes the proof.

As in the case of sequences, the key result for the set

 {Ln, Un, Ln∘Un, Un∘Ln} (22)

to be closed under composition is the equality in Theorem 15. Now one can easily derive the rest of the formulas for the compositions of the operators in this set. The composition table is indeed as given in Table 1. Furthermore, Theorem 15 implies the total order on the set (22) as in (4). Indeed, we have

 Ln=id∘Ln≤Un∘Ln=Ln∘Un∘Ln≤Ln∘Un∘id=Ln∘Un≤id∘Un=Un

Therefore, the operators and for functions on generate via composition a semi-group with exactly the same algebraic and order structure as the semi-group generated by the operators and for sequences.

5 Discrete pulse transform of images

In this section we apply the LULU operators defined and investigated in the preceding sections to derive a discrete pulse decomposition of images. A grayscale image is given through a function on a rectangular domain , the value of being the luminosity at the respective pixel. For the theoretical study it is more convenient to assume that the functions are defined on the whole space . To this end one can, for example, define on the set as a constant, e.g. 0. Hence we consider the set .

Appropriate connections for images are defined through a relation on reflecting what we consider neighbors of a pixel in the given context. Figure 1 gives some examples of the the neighbors of the pixel .

We call a set connected if for any two pixels there exists a set of pixels such that each pixel is neighbor to the next one, is neighbor to and is neighbor to . We assume that the neighbor relation on is such that

 ∙ r is reflexive, symmetric and shift invariant (23) ∙ ((i,j),(i±1,j))∈r and ((i,j),(i,j±1))∈r,% for all i,j∈Z. (24)

The conditions (23)–(24) ensure that the set of connected set defined through this relation is a connection in terms of Definition 3 and satisfies the conditions (6)–(8). Hence we can apply the operators and discussed in the preceding sections to functions on . Similar to the case of sequences we obtain a decomposition of a function by applying iteratively the operators with increasing from 1 to . This can be done in different ways depending on sequencing of the and . Since this section is intended as a demonstration rather than presenting a comprehensive discrete pulse transform theory, we will take one particular case when follows . Define the operators , , by and . Then for any and we have

 f=(id−U1∘L1)(f)+((id−U2∘L2)∘F1)(f)+((id−U3∘L3)∘F2)(f) + ... +((id−Um∘Lm)∘Fm−1)(f)+Fm(f) (25)
Definition 16

A function is called a pulse if there exist a connected set and a real number such that

The set is called support of the pulse and is denoted by .

Figure 2 gives an example of a pulse. It should be remarked that the support of a pulse may generally have any shape, the only restriction being that it is connected.

The usefulness of the representation (25) of a function is in the fact that all terms are sums of pulses as stated in the next theorem.

Theorem 17

Let .

• For every the function is a sum of discrete pulses with disjoint support, that is, there exist and discrete pulses , , such that

 ((id−Un∘Ln)∘Fn−1)(f)=γ(n)∑s=1ϕns (26)

and

 supp(ϕns1)∩supp(ϕns2)=∅ for s1≠s2. (27)
• Let be such that , and . Then

 supp(ϕn1s1)∩supp(ϕn2s2)≠∅⟹supp(ϕn1s1)⊂supp(ϕn2s2) (28)

Proof. a) Denote . We have

 ((id−Un∘Ln)∘Fn−1)(f)=(id−Ln)(g)+(id−Un)(Ln(g)), (29)

where the first term in the sum on the right hand side is nonnegative while the second one is nonpositive. Let