Page 29 - AC-2-2

Arts & Communication Identification of Pollock Art

authentic paintings by Pollock. Previous cases showed that That transformation ensures that the colors analyzed by
paintings claimed to be authentic Pollock’s paintings were the way they are perceived by the human eye, rather than
often not provided with strong evidence of authenticity, or by triplets of RGB values that follow the way computer
were shown to use painting materials that only existed after hardware works, but do not always reflect the way colors
the alleged date of creation. As the results of this paper are perceived by the human brain.
23
show, these paintings also have substantial differences In addition to the features in the WND-CHARM
from Jackson Pollock’s paintings. feature set, UDAT also uses the Gini coefficient, which
37
A dataset of drip paintings not created by Pollock was was adopted from the field of economy to measure the
37
also used in a previous study. However, these paintings distribution of light in an image. The Gini coefficient is
23
were not presented by their owners as authentic Pollock measured as the area under the curve of the histogram
art. These paintings were created to provide drip paintings of pixel intensities, which reflects the inequality of the
inspired by Jackson Pollock and were used for decorative intensities. Another visual measurement that was added is
purposes. They were created with no attempt to fully the image entropy. The image entropy e is computed using
mimic Pollock’s style or claim that these paintings were Equation I.
authentic art created by Pollock himself. In fact, even a e = −1 Σ P logP (I)
non-expert could identify that these paintings are different i i i
from the authentic work of Jackson Pollock, and no claim Where P is the number of pixels in the histogram
i
of authenticity was made, suggested, or implied. The new bin i divided by the total number of pixels in the image.
dataset, however, contains paintings that were claimed Image entropy is a simple measurement that reflects the
by its owners to be authentic Jackson Pollock’s art. The consistency of the pixel intensity values in the image. In the
paintings were all normalized by their size to 850k pixels, UDAT implementation, 32 bins are used. To extract more
such that the aspect ratio of each of the original images information from the images, the features described above
was preserved, and therefore, no distortion could affect the are computed from the original pixel values of the image, as
visual content. well as from the Fourier transform, Chebyshev transform,
Wavelet (Symlet 5, level 1) transform, color transform,
32
38
2.2. Analysis method hue transform, and combinations of these transforms.
The WND-CHARM feature set is described in details and
The image analysis tool used to analyze the images is experimental results in Shamir Tarakhovsky, Shamir et al.,
UDAT, 24,25 which is a multi-purpose data analysis tool Shamir et al. 26-28
that can also perform comprehensive analysis of images.
Its image analysis is based on the WND-CHARM feature Once the numerical image content descriptors are
set, 26-28 which includes a comprehensive set of numerical computed, the values can be compared to identify
image content descriptors that reflect many aspects of differences between different sets of paintings. Numerical
the visual content. The comprehensive nature of the set values can also be used to classify the paintings using
of numerical image content descriptors and its ability to machine learning and identify similarities between the
measure a broad range of visual cues allow it to handle the paintings or groups of paintings. UDAT is an image
complex nature of visual art. 2,13,28,29 analysis tool that is not necessarily an image classifier. It
therefore uses instance-based machine learning to identify
In summary, the WND-CHARM feature set reflects and profile patterns of similarities between artistic styles
an image with a set of 2885 different numerical visual represented by groups of paintings. 25,28,29
content descriptors. These descriptors include texture
descriptors such as Gabor, Tamura, and Haralick, Instance-based methods can also be used for
30
31
32
color features, edge statistics, objects statistics, Radon classification by determining the predicted class according
transform, distribution of pixel intensity values, to the label of the training instance with the shortest
33
28
multi-scale histograms, Chebyshev statistics, Zernike distance to the test painting. In fact, instance-based
34
polynomials, and fractal features. These features are methods are some of the earliest methods of supervised
35
36
39
described in detail in Shamir Tarakhovsky, Shamir et al., machine learning and are still extremely popular and
Shamir et al. 26-28 To handle the variety of colors in the commonly used.
paintings, colors are estimated by applying a fuzzy logic An important aspect of the method is its ability to
analysis on the Hue Saturation Value (HSV) domain. identify specific differences between different sets of
32
This analysis converts the colors used in the paintings and paintings. In that sense, UDAT is different from image
represented as triplets of numerical values into the human classification methods that are based on deep neural
perception of the colors, determined by fuzzy logic rules. network. Such neural networks use rules that are complex
Volume 2 Issue 2 (2024) 3 doi: 10.36922/ac.1628