Classification
of
Painting styles
To try our algorithm on another possible application, we took 35
pictures (about 600x800 pixels size) drawn by 5 different painters,
trying to identify the painters by their drawing style. In the
experiments we have carried out, each of the painters was classified
correctly, and the obtained clusters were stable. We used windows of
128x128 and 64x64 pixels, overlapping and not overlapping, with
different
wavelet filters, and always converged to the same result: Figure 11.
The obtained clusters differentiate between impressionism style of Van
Gogh, characterized by a large amount of bold brush-strokes, classic
style of Rembrandt, with its soft lines and smooth textures, landscape
reproductions of Shishkin, rich of small scrupulous details, cubism
of Picasso and marine landscapes of Aivazovsky.
It should be noted that in this example the correct classification was
obtained only when we took sufficiently fine segmentation of the input
images (into about 20-25 segments). In order to choose a segmentation
with an appropriate number of segments for the subsequent
classification
step, we looked at the inverse rate distortion function D(R), as
described in the paper (Figure 12).
The points where the decrease of the distortion slows down relative to
the increase in the mutual information correspond to stable
segmentation
solutions. The vertical grid lines in the graph correspond to the
increments of the number of segments in the partition by one. By
looking
at the minima of the second derivative of the function, we identify
stabilization points in the segmentation solution. In the example
shown, one of the evident (actually most evident) local minima
corresponds to segmentation into 24 segments, which is actually the
point where the classification results become correct and stabilize.
When taking segmentations with a greater number of segments we always
converged to the same correct classification of the input images. At
the
same time, when classifying images using segmentation into less than 24
segments, a small number of misclassifications were present.
When using large (256x256 and more) windows, at all the segmentation
resolutions in subsequent classification some of the pictures were
misclassified (for example, some of the Picasso paintings were mixed up
with those of Aivazovsky). As well, we got poor results when used one
big window of the whole image size, which is equivalent to classifying
the images without segmenting them first. These results justify the
advantage of image segmentation as a preprocessing step for
classification.