Information Technology Reference
In-Depth Information
The sets were generated at random and care was taken to ensure that the two
patterns resulting from an item with three images were integrated into the same
set. Thus, it was guaranteed that the correct image was not simultaneously used for
training and testing the neural network.
Considering the 20 experiments carried out, the global success rate in the test
sets was 74.49 %. Which corresponds to a percentage of 71.67 % correct answers in
the Design Judgment Test.
4
The result is similar to the maximum success rate pre-
viously achieved with the heuristic AJS (73.3 %) by adjusting the parameters. This
reinforces the conclusion that it is possible to capture some of the aesthetic princi-
ples considered by Maitland Graves in the DJT. They also show that it is possible to
learn principles of aesthetic order based on a relatively small set of examples. The
fact that the approach was not able to achieve the maximum score in the DJT has
two, non exclusive, explanations: (i) the features are unable to capture some of the
aesthetic principles required to obtain a maximum score in the DJT; (ii) the set of
training examples is not sufficient to allow the correct learning of these principles.
Although the results obtained by the system are higher than the human averages
reported in the previously mentioned studies, these results are not comparable. In
addition to the issues we mentioned when analysing the results of the heuristic based
classifier, the nature of the task is different herein: humans do not make their choices
based on a list of correct choices for other items of the test.
11.3.2.3 Author Identification Experiments
In Machado et al. (
2004
) we presented the results obtained by a previous version
of our AJS in an author identification task. The image dataset was made up of 98
paintings from Goya, 153 from Monet, 93 from Gauguin, 122 from Van Gogh, 81
from Kandinsky, and 255 from Picasso. Although the system obtained high success
rates (above 90 %), further experiments revealed that the reduced number of images
and their nature made the classification task easier than expected.
Taking into account the dataset limitations mentioned in Sect.
11.2.2
, we created
a dataset composed of images from three prolific painters, from chronologically
consecutive artistic movements:
Claude-Oscar Monet (Impressionism, mid 19th century). It consists of 336 im-
ages, most of them landscapes and portraits.
Vincent van Gogh (Post-Impressionism, late 19th century): a total number of
1046 well-known images from his work, including landscapes, portraits, self-
portraits and still lifes.
Pablo Picasso (Cubism and Surrealism, early 20th century): a total of 540 images
belonging to different stages were used, ranging from the Blue Period to the author's
surrealist stage.
We avoided using greyscale images and images with insufficient resolution.
Some of the images (12 from Picasso and 8 from Van Gogh) included the frames
4
Some of the test items are triads, hence the lower percentage.
