Information Technology Reference
In-Depth Information
11.3.2.4 Image Classification Based on Online Evaluation
We used the dataset provided by Datta et al. (
2006
) that was analysed in Sect.
11.2.2
.
The database contains 832 images with an aesthetic rating
≥
5.8 and 760 images with
a rating
4.2. However, when we carried out our experiment, some of the images
used by Datta were no longer available at
photo.net
, which means that our image
set is slightly smaller. We were able to download 656 images with a rating of 4.2 or
less, and 757 images with a rating of 5.8 or more.
We conducted 50 runs, each with different training and validation sets, randomly
created with 80 % and 20 % of the images, respectively. The success rate in the
validation set was 77.22 %, which was higher than the ones reported in the original
paper (Datta et al.
2006
) but lower than the one obtained by Wong and Low (
2009
),
using 10 % of the images in each set.
≤
11.3.2.5 Integration in an Image Generation System
A previous version of the AJS described here was used in conjunction with a genetic
programming evolutionary art tool. The main goal of this experiment, reported by
Machado et al. (
2007
), was to develop an approach that promoted stylistic change
from one evolutionary run to the next. The AJS assigns fitness to the evolved images,
guiding the evolutionary engine.
The AJS is trained by exposing it to a set of positive examples made up of art-
works of famous artists, and to a set of negative examples made up of images gen-
erated randomly by the system. The goal is twofold: (i) evolving images that relate
with the aesthetic reference provided by the positive examples, which can be con-
sidered an inspiring set; (ii) evolving images that are novel relative to the imagery
typically produced by the system. Thus, more than trying to replicate a given style,
the goal is to break from the traditional style of the evolutionary art tool. Once novel
imagery is found (i.e. when the evolutionary engine is able to find images that the
AJS fails to classify as being created by it), these images are added to the negative
set of examples, the AJS is re-trained and a new evolutionary run begins. This pro-
cess is iteratively repeated and, by this means, a permanent search for novelty and
deviation from the previously explored paths is enforced.
Next, the genetic programming engine and the AJS performed 11 consecutive
iterations (Machado et al.
2007
). In each iteration, the evolutionary engine was able
to find images that were misclassified by the AJS. Adding this set of examples to
the dataset forced the AJS to find new ways to discriminate between paintings and
the images created by the evolutionary art tool. The evolutionary engine and the
AJS performed well across all iterations. The success rate of the AJS for validation
set images was above 98 % in all iterations. The evolutionary engine was also al-
ways able to find novel styles that provoked misclassification errors. In Fig.
11.3
we
present some examples of images created in the 1st and 11th iteration.
Overall, the results indicate that the internal coherency of each run is high, in
the sense that runs converge to imagery of a distinctive and uniform style. The style
