Autopoiesis and Image Processing: Detection of Structure and

Organization in Images

Mario K6ppen I, Javier Ruiz-del-Solar2

i Dept. of Pattern Recognition, Fraunhofer IPK-Berlin,
Pascalstr. 8-9, 10587 Berlin, Germany.
Email: mario.koeppen@ipk.fhg.de.
2 Dept. of Electrical Eng., Universidad de Chile
Casilla 412-3, 6513027 Santiago, Chile.
Email: j.ruizdelsolar@computer.org.

Abstract

The theory of Autopoiesis describes what the living systems are and not what they do.
Instead of investigating the behavior of systems exhibiting autonomy and the concrete
implementation of this autonomy (i.e. the system structure), the study addresses the
reason why such behavior is exhibited (i.e. the abstract system organization). This article
explores the use of autopoietic concepts in the field of Image Processing. Two different
approaches are presented. The first approach assumes that the organization of an image is
represented only by its grayvalue distribution. In order to identify autopoietic
organization inside an image's pixel distribution, the steady state Xor-operation is
identified as the only valid approach for an autopoietic processing of images. The effect
of its application on images is explored and discussed. The second approach makes use of
a second space, the A-space, as the autopoietic-processing domain. This allows for the
formulation of adaptable recognition tasks. Based on this second approach, the concept of
autopoiesis as a tool for the analysis of textures is explored.

Keywords: Autopoiesis, Steady State Image Processing, Auto-projective Operators,

Texture Analysis, Texture Retrieval Systems, Autopoietic-Agents

1 Introduction

The theory of Autopoiesis, developed by the Chilean biologists Humberto Maturana and
Francisco Varela, attempts to give an integrated characterization of the nature of a living
system, which is framed purely with respect to the system in and of itself. The term
autopoiesis was coined some twenty-five years ago by combining the Greek auto (self-)
and poiesis (creation; production). The concept of autopoiesis is defined as [Varela,
1979, p. 13]:

'An autopoietic system is organized (defined as a unity) as a network of

processes of production (transformation and destruction) of components that
produce the components that:
 443

1. through their interactions and transformations continuously regenerate and

realize the network o f processes (relations) that produced them; and

2. constitute it (the machine) as a concrete unity in the space #1 which they [the
components.] exist by specifying the topological domain of its realization as
such a network. '

In other words an autopoietic system produces its own components in addition to

conserving its organization. A network of local transformations produces elements, which
maintain a boundary. This boundary captures the domain, in which the local
transformations take place. In this context, l/~ can be defined as autopoietie organization
realized in a physical space. The autopoietic theory describes what the living systems are
and not what they do. Instead of investigating the behavior of systems exhibiting
autonomy and the concrete implementation of this autonomy (i.e. the system structure),
the study addresses the reason why such behavior is exhibited (i.e. the abstract system
organization). A complete material concerning the autopoietic theory (tutorials, study
plan, bibliography, Internet links, etc.) can be found in the Internet site Autopoiesis and
Enaction: The Observer Web [Whitaker, 1996].
The autopoietic theory has been applied in diverse fields such as biology,
sociology, psychology, epistemology, software engineering, artificial intelligence and
artificial life. In this context, this article tries to explore the use of autopoietic concepts in
the field of Image Processing.
Two different approaches will be presented. These approaches differ in their
interpretation of the domain of the processes. The first approach, presented in section 2,
assumes that the domain of an image is represented only by its grayvalue distribution. In
order to identify autopoietic organization inside an image's pixel distribution, the steady
state Xor-operation is identified as the only valid approach for an autopoietic processing
of images. The effect of its application on images is explored. The second approach makes
use of a second space, the A-space, as autopoietic processing domain. This allows the
formulation of adaptable recognition tasks. Based on this second approach, the concept of
autopoiesis as a tool for the analysis of textures is explored in section 3. As a concrete
example, a Texture Retrieval System based on the use of an autopoietic-agent is presented.
Finally, in section 4 some conclusions are given.

2 Recognition of image structures by using auto-projective operators

In a first attempt, the question arises whether images by themselves preserve some kind
of autopoietic organization. Because of images generally are considered as static
representations of real-world objects, but autopoiesis is constituted by a network of
dynamic transformations, the image must be processed by suitable operators in order to
reveal possible organizational principles. Thereby, the original image appears to be like a
"frozen" state of its intrinsic dynamical processes. Two approaches are possible from
 444

now on: the first approach assumes no relation between these dynamics and the real-
world objects pictured in the image, in contrary to the second approach, which names the
kind of features of real-world objects from which the dynamics are driven.

This section is concerned with the first approach i.e., image dynamics are restricted to the
distribution of colors or grayvalues in the image. No reference is given to the pictured real-
world objects. In order to "melt" the image due to a possible intrinsic dynamic, an
operator is searched with the following two essential properties:

1. The operator should be applied point-wise. Normally, image operators like the
Laplacian, the Sobel or the median operator are applied to all image pixels, at once. But, as
it was mentioned in the introduction, autopoietic systems constitute the domain of its
realization. For an effective search of these domains, the size of them can not be
predicted. Hence, the application domain of the operator must be balanced between pure
local analysis (a pixel and its neighborhood) and global analysis (all image pixels). Pre-
defined image operators do not offer such a choice. By repetitively applying the local
operator point-wise, the effect of a local operation is spread out over the image and more
complicated patterns of interaction are possible. As a reminder for a similar procedure in
genetic algorithms, we refer to this manner of image operator application as steady state
image processing.

2. The requirements of autopoietic organization detection include the requirements for

rejecting this hypothesis as well. The operators we are interested in, must be able to
recover the original image, at least to some degree, or with a probability much greater than
zero. If the image dynamics appears to be not autopoietic, the process must be able to
fail, i.e. to leave the image as it was at the beginning of the processing.

Speaking more formally, let | be the image operation, which is applied point-wise. A
sequence of points p(T) is generated randomly, where p(T) is the point chosen at time
step T. Let g(p) be the grayvalue of point p in the image. In this article, all points are
equally possible in the random sequence. Non-adjacent points in the sequence could be
neighbors in the image. Assume p(Ti) and p(T2) are such a pair of points with T~<'1"2.
Then, while applying the operator | to the point p(Tt) and its neighbors, p(T2) is also
affected. But later, at step/'2, the application of the operator onto the modified p(T2) also
affectsp(Tt). Our demand is to have a non-zero probability of reproducing p(Ti)'s original
value by this procedure. This demand can be fulfilled using the Xor-operation. This can
be verified by considering the following three properties of the Xor-operation:

Commutativity: a|174
Associativity: (a@b)|174174
Auto-projection: (a|174
 445

The third property explains the fundamental role of the Xor-operation in data coding, as
well as for sprite algorithms in computer animations. It can be easily shown that the Xor-
operation and its negate are the only binary operations fulfilling all of these three
properties. Hence, for detecting organized autopoietic structures in an image's grayvalue
distribution, it is necessary to apply the Xor-operation in a steady state manner. The
resulting algorithm is as simple as in the following:

Repeat:

1. Take an arbitrary pixelpl and choose one of its neighborsp2.

2. Replace g(Pz) with g(Pz) | g(Pz).

Alternatively, operations similar to the Xor-operation can be designed based on the

theory of periodic Abelian groups (Xor gives a 2-periodic Abelian Group). The
periodicity ensures the possible reproduction of its operands. However, the larger the
period, the fewer the reproduction probability.

2.1 Diseussion

For a further understanding of the effect of this operation consider figure 1. There, a face
image (a), the result of the repetitive application of the above algorithm after 1000
"generations" (b), and a dilated version of the second image (c), are shown. A white
circular contour around the phong-pattern on the forehead can be seen. This phong-
pattern is a result of the lighting conditions during image acquisition. Phong-patterns are a
major problem for facial recognition tasks. By using the Xor-operator, they can be easily
detected. But where comes this circle around the phong-pattern from?
To understand this, consider the effect of the same procedure onto a gradient
image (figure 2). Also, the original image, the image after 1000 steps and its dilated version
are shown. A white line appears in the middle of the image, around the grayvalue 128, but
not exactly in this position. The explanation of this effect reminds of another famous role
of the Xor-operation, as a benchmark for neural networks. The Xor-operation can not be
linearly separable. A neural network needs a hidden layer to learn the Xor-operation. The
gradient image helps to give an imagination of this fact. If pl has grayvalue 0, then g(Pl) |
g(P2) gives g(Pz), i.e. for low grayvalues, the Xor-operator tends to be the identity
transformation. Ifg(pl) is the maximum grayvalue (255 in our case), g(Pl) | g(P2) gives
the inverse of g(Pz), i.e. for high grayvalues it tends to be the inverting transformation.
But there is no linear descent from identity to inverse! Hence, we must have a non-linear
anomaly between these two extremes. The white line represents this anomaly.
Grayvalues around 128 tend to complete each other to the maximum grayvalue 255. The
white line appears to be a boundary between a gradual descent from the maximum to the
minimum grayvalue. This way boundary exchange processes can be identified; i.e. the
boundary must be a closed one to prevent the Xor-operations from pocketing it. Hence,
 446

the Xor-operation is able to detect the gradient-descending organization of grayvalue

distributions in an image, no matter how this descent is structured (line-likeness,
circularity, linearity ascending or faster). But also, the white line is a boundary separating
intcrior from outside.

Figure I. A face image (a), tile result of the repetitive application of the above algorithm after 1000
"generations" (b), and a dilated version of the second image (c).

Figure 2. A gradient image (a), tile result of the repetitive application of the above algorithm after 1000
"generations" (b), and a dilated version of the second image (c).

Two summarize the foregoing discussion: For the detection of the autopoietic
organization of a grayvalue distribution, or better, the actual grayvalue distribution as a
"frozen" state of a possible autopoietic organization, the Xor-operator must be applied in
a steady state manner, i.e. on a sequence of randomly chosen image points. Only the Xor-
operator has the property of auto-projection, which ensures a much greater than zero
probability of regenerating the original image. This is true as long as binary numbers
represents grayvalues. The application of the Xor-operator onto images yields phong-like
structures, which prove to be the only organizational issues of intensity ordering in an
autopoietic manner.
These first results are encouraging enough to continue this work. It has been
shown, that the search for autopoietic organization in grayvalue distributions of images
reveals new structural properties of them, which are hardly to find by mean o f other
image processing operations. Further research on the Xor-operator should explore the role
of the probability distribution for the random sequence of pixel positions. Also, other
 447

ordering relations in the image than the conventional intensity ordering should offer new
application tasks for the Xor-operator.

3 Autopoiesis and Texture Analysis

3.1 Textures and Autopoiesis

Texture perception plays an important role in human vision. It is used to detect and
distinguish objects, to infer surface orientation and perspective, and to determine shape in
3D scenes. Even though texture is an intuitive concept, there is no universally accepted
detinition for it. Despite this fact we can say that textures are homogeneous visual
patterns that we perceive in natural or synthetic images. They are made of local
micropatterns, repeated somehow, producing the sensation of uniformity [Ruiz-del-Solar,
1997]. It is important to point out, that textures can not be characterized only by their
structure because the same texture, viewed under different conditions, is perceived as
having different structures.
In the fi:~mework of the theory of autopoiesis, Maturana and Varela make a
complementary definition of the concepts of organization and structure of a system. The
organization of a system defines its identity as a unity, while the structure determines
only an instance of the system organization. In other words, the organization of a system
defines its invariant characteristics. The concept of autopoiesis captures the key idea that
living systems are systems that self maintain their organization (see introduction). In the
context of texture analysis, the systems to be analyzed are the textures. As it was
established, the concept of organization must be used to characterize a system and in our
case to characterize a texture. For this reason, in this section the concept of autopoiesis is
explored as a tool for texture identification, which corresponds to an important task in the
field of texture analysis. The analogy between the process of autopoietic organization (i.e.
lil~) in a chemical medium and the process of texture identification is used.
Before to apply the concept of autopoiesis as a tool for texture identification a
computational model of autopoiesis must be defined. Varela et al. developed the first
model that was capable of supporting autopoietic organization [Varela et al., 1974].
Recently, McMuilin developed the SCL model, which corresponds to an improvment of
the model, presented by Varela [McMullin, 1996a and 1997b]. The SCL model from
McMullin is used here.

3.2 The SCL Model

SCL involves three different chemical elements (or particles): Substrate (S), Catalyst (K)
and Link (L). These particles move in random walks in a discrete, two-dimensional space.
In this space, each position is occupied by a single particle, or is empty. Empty positions
are managed by introducing a fourth class of particles: a Hole (H). SCL supports six
distinct reactions among particles [McMullin, 1997b]:
 448

1. Production:
K+S+S ........ > K+L+H

2. Disintegration:
L ........ > S+S

3. Bonding:
Adjacent L panicles bond into indefinitely long chains

4. Bond decay:
Individual bonds can decay, breaking a chain

5. Absorption:
L+S ........ > L*

6. Emission:
L* ........ > L+S

The autopoietic organization is produced, when a chain of L dements forms a

boundary, which defines a concrete unity in the space. Of course, this boundary must be
continuously regenerated (see introduction). The L elements are produced only in the
prcscnce of a catalyst (Production reaction). For this reason, we can say that in this
model, an autopoietic organization is produced only in the presence of a catalyst.

3.3 The modified SCL Model

The original SCL model was modified to allow the identification of textures, by
introducing the idea of a texture-dependent catalyst. That means, a catalyst that is tuned
with a defined texture and that produced an autopoietic organization only in this texture.
To implement this idea an autopoietic image A(i,j) is defined for each texture image T(i,j).
Each pixel of A(i,j) has a corresponding position in T(ij) and is represented by 2 bits
(enough for represent four particles). A T-Space is associated with the texture images
TO'j) and a A-Space is associated with the autopoietic images A(i,j) (see figure 3). The
reactions defined by the SCL model, that is the possible autopoietic organization, take
place in the A-Space, but by taking into account information from the T-Space (textures).
T-Space A-Space

K
.s2

Figure 3. The A-Space, where the autopoietic organization is created, and the T-Space, where convolution
between the texture and the Gabor-Filter is performed, are shown.
 449

Gabor-Filters are able to characterize textures by decomposing them into different

orientations and frequencies (scales) [Ruiz-del-Solar, 1997]. Here, a Gabor-Filter is
associated with tile catalyst, to allow it (the catalyst) to be tuned with a particular
texture. The Gabor-Filter interacts directly with the textures in the T-Space (convolution
operation) and the result of this interaction is used to modulate the reactions in the A-
Space.
From all the reactions defined by the SCL model only the Production reaction was
modified, because it is the only one where the catalyst operates and the L particles are
created. The new Production reaction is defined by:

Production:

K+SI+S 2 ........ > K+L+H

Cl=Nl*Gk
C2=N2*Gk

if(Cl>TH and C2>TH) {

if(Ct> C2) {
St ---> L
$2 ---> H
}
else {
Si ---> H
$2 ---> L

where Gk is the Gabor-Filter associated with the catalyst K; N/ and N2 are the
neighborhood, in the T-Space, of St and $2, respectively (see figure 3); Ct and C2 are the
results of the convolution (performed in the T-Space); and TH is a threshold value. If in
the A-Space of a given texture a chain of elements form a boundary, after an interaction
time, then tile catalyst K has identified the texture (in its T-Space) as corresponding to the
class of textures characterized by the Gabor-Filter Gx.

3.4 An autopoietic-agent for Texture Retrieval

To illustrate the idea of texture identification by using a computational model of

autopoiesis, a system for retrieval of textures in image databases is proposed (see figure
4). The system is based in the use of an autopoietic-agent (the texture-dependent catalyst
described in section 3.3), which is generated by using the texture description contained in
the query. This autopoietic-agent is tuned with this description, which means it can
interact (produce autopoietic organization) only with the texture that corresponds to this
description. The autopoietic-agent is sent to every texture of the database and allowed to
interact with the substrate particles of the A-Space of the textures. After an interaction
time, the texture, where an autopoietic organization was produced (in its A-Space), is
retrieved.
 450

query
~ autoagent
poeitci- ~ "9 ~

Texturei Textures'
9 Database

Figure 4. ProposedTextureRetrievalSystem(A3G:AutomaticAutopoietie-AgentGenerator;TA2T:
Textural Autopoietie-AgentTester).

4 Conclusions

The use of autopoietic concepts in the field of Image Processing was explored. Two
different approaches were presented. The first approach, presented in section 2, assumes
that the organization of an image is represented only by its grayvalue distribution. In
ordcr to identify autopoietic organization inside an image's pixel distribution, the steady
state A'or-operation was identified as the only valid approach for an autopoietic
processing of images. The application of the Xor-operator onto images yields phong-like
structures, which prove to be the only organizational issues of intensity ordering in an
autopoietic manner. These first results are encouraging enough to continue this work. It
was shown that the search for autopoietic organization in grayvalue distributions of
images reveals new structural properties of them, which are hardly to find by means of
other image processing operations. Further research on the Xor-operator should explore
the role of the probability distribution for the random sequence of pixel positions. Also,
other ordering relations in the image than the conventional intensity ordering should offer
new application tasks for the Xor-operator.
The second approach, presented in section 3, makes use of a second space, the A-
space, as autopoietic processing domain. This allows the formulation of adaptable
recognition tasks. Based on this second approach, the concept of autopoiesis as a tool for
thc analysis of textures was explored. The SCL model, a computational model of
autopoiesis, was modified to allow the identification of textures, by introducing the idea
of a texture-dependent catalyst. As a demonstrating example, a Texture Retrieval System
based on the use of an autopoietic-agent, the texture-dependent catalyst, was presented.
Further research must be performed to apply this concept in the solution of real-world
problems.

References

McMullin, B. (1997a). Computational Autopoiesis: The original algorithm. Working

Paper 97-01-001, Santa Fe Institute, Santa Fe, NM 87501, USA.
http ://www. santafe, edu/sfi/publications/Working-Papers/97-01-001 /
 451

McMullin, B. (1997b). SCL: An artificial chemistry in Swarm. Working Paper 97-01-002,

Santa Fe Institute, Santa Fe, NM 87501, USA.
http://www.santafe.edu/sfi/publications/Working-Papers/97-01-002/

Ruiz-del-Solar, J. (1997). TEXSOM: A new Architecture for Texture Segmentation. Proc.

of the Workshop on Self-Organizing Maps - WSOM 97, pp. 227-232, June 4-6, Espoo,
Finland.

Varela, F.J. (1979). Principles of Biological Autonomy, New York: Elsevier (North
t [olland).

Varela, F.J., Maturana, H.R., and Uribe, R. (1974). Autopoiesis: The organization of
living systems, its characterization and a model. BioSystems 5: 187-196.

Whitaker, R. (1996). Autopoiesis and Enaction: The Observer Web.

http://www.informatik.umu.se/~rwhit/AT.html