Information Technology Reference
In-Depth Information
particular style; (ii) one must ensure that there is no overlap between styles; (iii) one
cannot use exclusively the most representative images of each style, otherwise the
tasks may become trivial and, therefore, useless.
The first problem can be partially solved by using a relevant external source for
the images. Unfortunately, the only published digital sets of artistic images we are
aware of are those provided by Directmedia/The Yorck Project publications. How-
ever, the quality of the collections is far from perfect (they include black and white
versions of some images, frames, detailed images of parts of other artworks, etc.).
One can also resort to online databases of paintings. The collection “Oil paintings by
Western masters” contains 46,000 images and can be found in the peer-to-peer net-
work. The
Worldimages
website (
http://worldimages.sjsu.edu/kiosk/artstyles.htm
),
the website
http://www.zeno.org
,
developed by the creators of “The Yorck Project”,
and online museum websites are also good sources of images.
Wallraven et al. (
2008
) analysed the perceptual foundations of the traditional cat-
egorisation of images into art styles, finding supporting evidence. They concluded
that style identification was predominantly a vision problem and not merely a his-
torical or cultural artefact.
Wallraven et al. (
2009
) presented an experiment that analysed the capacity of a
group of non-experts in art to categorise a set of artworks in styles. One of the met-
rics they analysed is the artist consistency, which was higher if paintings of the same
painter were put in the same cluster. In one experiment, they obtained an average
artist consistency of 0.65. The conclusions were that “experts were able to reliably
group unfamiliar paintings of many artists into meaningful categories”. In the same
paper, the authors employed a set of low-level measures (Fourier analysis, colour
features, Gist, etc.) and a k-means algorithm to categorise the artworks into styles.
They concluded that low-level features were not adequate to artistic style classifica-
tion: “the fact that neither texture, nor colour-based, scale-sensitive or complexity
measures correlate at any dimension casts doubt on whether another [low level]
measure will do much better” (Wallraven et al.
2008
).
Marchenko et al. (
2005
), based on the colour theory of Itten (
1973
), characterised
regions of the image in terms of “artistic colour concepts”, while Yan and Jin (
2005
)
used several colour spaces to gather information with the aim of retrieving and clas-
sifying oil paintings.
There are several papers in the content-based image retrieval literature that pro-
pose image classification based on the “type” of image, distinguishing professional
photos from amateur ones, e.g. (Tong et al.
2004
); or photos from: (i) paintings
(Cutzu et al.
2003
), (ii) computer graphics (Athitsos et al.
1997
), (iii) computer-
generated images (Lyu and Farid
2005
). These tasks result in an interesting test field
for AJS, creating the opportunity of using AJSs in image classification tasks that are
far from aesthetics. These works can also provide tools (e.g., features, classification
methods, etc.) of interest to the creative computer community, in particular to those
researchers involved in artistic tasks.
