Information Technology Reference
In-Depth Information
vision, such as Datta et al. (
2006
;
2008
), Ke et al. (
2006
), Cutzu et al. (
2003
). One
of the advantages of this kind of systems is their potential use to perform different
tasks, and to be adapted to different aesthetic preferences. Classification tasks are
particularly useful for validation purposes since they tend to be objective and allow
a direct comparison of the results obtained by several systems (provided that they
are applied to the same datasets).
Relatively few attempts have been made in the visual arts field to integrate eval-
uation skills into an image generation system. Neufeld et al. (
2007
) presented a
genetic programming engine generating non-photorealistic filters by means of a
fitness function based on Ralph's bell curve distribution of colour gradient. This
model was implemented by carrying out an empirical evaluation of hundreds of
artworks. Their paper contains examples of some of the non-photorealistic filters
created.
Kowaliw et al. (
2009
) compared biomorphs generated in three different ways:
at random, through interactive evolution, and through evolution guided by a set of
image metrics used in content based image retrieval. They compared the results of
the three methods taking into account a model of creativity explained in Dorin and
Korb (
2009
), coming to the conclusion that automatic methods gave rise to results
comparable to those obtained by interactive evolution.
Baluja et al. (
1994
) used an artificial neural network trained with a set of im-
ages generated by user-guided evolution. Once trained, the artificial neural network
was used to guide the evolutionary process by assigning fitness to individuals. Al-
though the approach is inspiring, the authors consider the results somewhat disap-
pointing.
Saunders (
2001
) used a similar approach, proposing the use of a Self Organising
Map artificial neural network for the purpose of evolving images with a sufficient
degree of novelty. This approach is restricted to the novelty aspects of artworks.
Svangård and Nordin (
2004
) made use of complexity estimates so as to model the
user's preferences, implying that this scheme may be used for fitness assignment.
The authors introduced some experiments in which they used sets of two randomly
generated images, and compared, for each pair, the system's choices with those
made by the user. Depending on the methodology used, the success rates ranged
between 34 % and 75 %. Obviously, a result of 35 % is very low for a binary clas-
sification task. No example of the images considered was presented, which makes
it impossible to evaluate the difficulty of the task and, as such, the appropriateness
of the methodologies that obtained the highest averages. Additional information on
the combination of AJSs in image generation systems can be found in Chap. 10 in
this volume.
Although the integration of AJSs in image generation systems is an important
goal, having autonomous, self-sufficient AJSs presents several advantages:
•
It allows one to assess the performance of the AJSs independently, providing a
method for comparing them. This allows a more precise assessment of the AJS
abilities than possible when comparing AJSs integrated with image generation
systems, since the strengths and weaknesses of the image generation systems
may mask those of the AJS;
