A Technique for Computerised Brushwork Analysis
Dmitry Murashov
1
, Alexey Веrezin
2
and Ekaterina Ivanova
2
1
Dorodnicyn Computing Centre of RAS, Vavilov st. 40, 119333, Moscow, Russian Federation
2
State Historical Museum, Red Square, 1, 109012, Moscow, Russian Federation
Keywords: Attribution of Paintings, Images of Paintings, Brushwork, Image Ridges, Structure Tensor.
Abstract: In this work, the problem of computer-assisted attribution of fine-art paintings based on image analysis
methods is considered. A technique for comparing artistic styles is proposed. Textural features represented
by histograms of brushstroke ridge orientation and local neighborhood orientation are used in this work to
characterize painter's artistic style. The procedures for feature extraction are developed and the parameters
are chosen. The paintings are compared using three informative fragments segmented in a particular image.
Selected image fragments are compared by information-theoretical dissimilarity measure. The technique is
tested on images of portraits created in 17-19th centuries. The preliminary results of the experiments
showed that the difference between portraits painted by the same artist is substantially smaller than one
between portraits painted by different authors. The proposed technique may be used as a part of
technological description of fine art paintings for attribution. The unsolved problems are pointed out and the
directions of further research are outlined.
1 INTRODUCTION
The paper is devoted to the problem of developing
image analysis techniques for computer-assisted
attribution of fine-art paintings. In the glossary of
the National Gallery, attribution is defined as an
assessment of who was responsible for creating a
particular work. Sometimes the term "attribution" is
interpreted more widely and includes also an
assessment of art school, time, country, etc.
(Obukhov, 1959). One of the trends in attribution
today is related to the analysis of digital images of
paintings and called as “Computer-assisted
Connoisseurship” (Stone, 2010). The idea of
applying image analysis in attribution is that to
compare images of authentic and studied paintings
by features characterizing individuality of an artist.
This idea is based on the concepts of Giovanni
Morelli, who laid foundations of the method for
comparative analysis in fine arts (Morelli, 1900),
and his followers (Berenson, 1903), (Ignatova,
1994). With the individuality of artist the experts
associate features of brushstrokes that are forming a
painted surface.
In the studies related to the tasks of attribution,
the feature sets that allow capturing appearance of
individual painting techniques, or unconscious
rhythm that distinguishes manner of an artist, are
extracted from the images. When developing
methods for attribution of paintings, the researchers
analyze geometry and texture of individual strokes,
configurations of groups of strokes, spatial
frequencies, and other properties of the paint layer
texture. As the tools, they use a variety of methods
for image analysis created in the last few decades. In
many papers the low-level texture features are
extracted from coefficients of wavelet transform and
values of Gabor filters responses.
In publications of different research groups, two
main approaches to the task have been proposed.
The first is based on the exhaustive comparison of
square image fragments of the researched images
(Johnson, 2008). The features are usually derived
from the coefficients of the orthogonal transforms
(in particular, wavelet transform). This approach is
of high computational complexity and is sensitive to
conditions of image acquisition and hardware
parameters (Polatkan, 2009).
The second approach provides features computed
from the segmented brushstrokes (Lettner, 2005),
(Shahram, 2008), (Li, 2012). But one can find too
few paintings with a sufficient number of
distinguishable brushstrokes that can be successfully
segmented in automatic mode, or even manually. It
221
Murashov D., Berezin A. and Ivanova E..
A Technique for Computerised Brushwork Analysis.
DOI: 10.5220/0005362302210226
In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 221-226
ISBN: 978-989-758-091-8
Copyright
c
 2015 SCITEPRESS (Science and Technology Publications, Lda.)

can be done, for example, in images of paintings by
van Gogh, P.J. Pollock and some others. Therefore,
it is preferable to use features that can be computed
directly from images, but not from the segmented
brushstrokes.
In this paper, a technique for comparing artistic
styles is proposed. Following recommendations of
art experts (Ignatova, 1994), we propose to use for
comparing paintings a group of brushstrokes that
form homotypic details of these paintings. In
(Sablatnig, 1998) for attribution of portrait
miniatures, the homotypic details of the human faces
were compared. The fragments were segmented
using geometric model of face. In this work, we use
images of the homotypic informative face details:
forehead, nose, and cheek. It should be noticed that
the size of paintings and, respectively, of images in
current research is much larger than in (Sablatnig,
1998), and selected image samples differ from those
in (Sablatnig, 1998). Three types of informative face
fragments used in this work are shown in Figure 1.
In the current research we use features that capture
directions of artist's brush. Procedures for feature
extraction are developed. Selected image fragments
are compared using information-theoretical
dissimilarity measure based on Kullback-Leibler
divergence. The technique is tested on 11 images of
portraits created in 17th-19th centuries. The
preliminary results of the experiments showed that
the difference between portraits painted by the same
artist is substantially smaller than one between
portraits painted by different authors, and these
groups of paintings can be separated.
The paper is organized in the following way. In
the next section the formal problem formulation is
given. In sections 5 and 6 we describe the analyzed
images of paintings and textural features capturing
artistic manner. Feature extraction procedures are
proposed. Then we introduce a technique for
comparing images of portraits represented by
extracted features. In two last sections we present
the results of computing experiments and make
conclusions.
2 PROBLEM FORMULATION
The problem is formulated as follows. Let
j
U
be
images of paintings by
J
authors,
1,2,...,
j
J=
;
2
:UR R→
. Let
2
:,
i
j
uRRΩ→ Ω⊂
be an
informative sample of type
i
taken from image
j
U
,
1,2,..., Ii =
. Fragment
i
j
u
is characterized by a
feature vector
12
, ,... ,...,
T
iiiisiS
jjjj j
xx x x


=


x
,
()
is i
j
sj
x
u γ =
,
2
:
s
R
RR γ ×→
,
1, 2,...,
s
S=
.
The difference between two samples
i
j
u
and
i
k
u
of
type
i
of images
j
U
and
k
U
we define as
()
2
1
(,)
S
iisis
jk j k
s
Ddxx
=
=

,
(1)
where
(,)
is is
j
k
dx x
is a measure of difference
between the features of type
s
extracted from these
two image samples. The difference
j
k
D
between the
images
j
U
and
k
U
is calculated from differences
between corresponding informative samples as
follows:
()
2
1
I
i
jk jk
i
DD
=
=

.
(2)
Let
l
U be an image with unknown attribution,
1lJ=+
. It is necessary to find image
m
U (and
the author of the painting) providing minimum of
distance
ml
D .
In the next section, the properties of the analyzed
images are described.
3 IMAGE DATA
The data used in the research are the image
fragments of artworks painted in 17 – 19th centuries
by different authors. The images are fixed by a
digital camera. The size of the images is about
4270x2850 pixels. The distortions conditioned by
acquisition process are compensated and images are
uniformly oriented. The size of informative
fragments varies from 990x814 to 1800x1000 pixels.
Resolution of fragments is about 200 dots per cm
that corresponds to the quality of the data used in the
analogous studies. For example, Johnson et al.
(Johnson, 2008), Polatkan et al. (Polatkan, 2009),
and Li et al. (Li, 2012) analyzed images obtained at
resolution of 196 dots per inch.
Some of the paintings have retouched and
repainted areas. The features should be extracted
only from areas with original brushwork. Thereby,
retouched and repainted areas should be excluded
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
222

during feature extraction. A technique developed in
(Murashov, 2014) is used for localizing damaged
paint layer areas.
In the next section, feature description of
informative samples of images of paintings is given.
(a) (b) (c)
Figure 1: Informative fragments of portraits: (a)
“forehead”; (b) “nose”; (c) “cheek”.
4 BRUSHWORK FEATURES
Taking into account complexity of brushstroke
segmentation, it is preferable to use image features
that do not need segmentation of a single
brushstroke. The features will be extracted only
from the areas containing maximum information
about the artistic manner of the painter. In this work,
histograms of brushstroke ridge orientation and local
neighborhood orientation are considered as the
features of a brushstroke group that describe the
individual artistic manner, specific to a particular
detail of a painting.
Image ridges are localized by modified technique
described in (Eberly, 1996). A set of points forming
a ridge of an object in grayscale image is extended
by including parabolic and umbilic points of a
grayscale image relief.
Let the grayscale image relief be a function
22
(,)
f
CRR∈ . It is assumed that 0Df ≠ ,
()
,
T
xy
Df f f= . We denote NDfDf= ,
TDf Df
⊥
=
, and
()
,
T
yx
Df f f
⊥
=− , where N is
a normal and
T is a tangent to level lines
(isophotes) of image
f
. The following expression
based on Hessian describes the local properties of
function
f
:
22
22
1
,
TT
TT
g
NDfN NDfT
k
Df
TDfN TDfT
μ
μ
−=






(3)
where
()
2T
kTDfDfT=−
is an isophote
curvature;
2
()
T
TDfDfN μ =−
is a gradient
flowline curvature;
2
()
T
g
NDfDfN=−
is a
measure of gradient variation along the flowlines. At
ridge points of
f
the following conditions are
taking place:
0
μ
= and
}
{
max 0,kg> . We also
consider the points of
f
where one or both
eigenvalues of matrix (3) are zero. For obtaining
ridge orientation histogram, orientation angle of the
inertia axes of ridge connected components is
calculated (Jähne, 2005):
1,1
2,0 0,2
2
1
arctan ,
2
μ
θ
μμ
=
−
(4)
where
,ij
μ
are the elements of inertia tensor:
2,0 1,1
1,1 0, 2
.
μμ
μμ
−
Μ=
−






For computing direction histogram, length (in
pixels) of ridge connected components is taken into
account.
Another feature describing local orientation of
painting texture is based on the notion of structure
tensor, or the second moment matrix at a point
x
weighted by a window function:
2
() ( ( )( ( )) ( )
T
f
pR
x
Df p Df p w x p dp μ
∈
=−

,
(5)
where
()wx p− is a window Gaussian function
(Lindeberg, 1994). The angle of simple
neighborhood orientation
ϕ
is determined as:
1,1
2,0 0,2
2
1
arctan ,
22
f
ff
μ
π
ϕ
μμ
=
−
+
(6)
where
,
f
ij
μ
are the components of the structure
tensor (5). The proposed feature description of
artistic manner is illustrated in Figure 2. In Figure 2
(a) – (c), the images of a human face detail "cheek",
taken from three portraits, are shown. The portraits
(a) and (b) are created by F. Rokotov, portrait (c) is
painted by another artist. In Figure 2 (d) – (f),
corresponding ridge orientation histograms are
given, and in Figure 2 (g) – (i), simple neighborhood
orientation histograms, obtained from images (a) –
(c), are depicted.
The procedures for computing described above
features are developed. For obtaining histogram of
orientation angles of grayscale image ridges, the
following operations are performed: (a) image
rotation and scaling; (b) extension of image dynamic
range; (c) creating a mask of informative fragment;
(d) creating masks of craquelure and damages; (e)
ATechniqueforComputerisedBrushworkAnalysis
223

combining masks; (f) image masking; (g) Gaussian
blurring; (h) localizing ridges of grayscale image
relief; (i) defragmenting obtained image ridges; (j)
filtering connected components of ridges by size; (k)
computing orientation angle values and building a
histogram. For extracting the second feature, firstly
the image is downsampled with a factor equal to 2.
Then the operations (a)-(g) listed above are
performed, and histogram of simple neighborhood
orientation is obtained. The procedure for creating a
craquelure mask includes operations of “black top-
hat”, adaptive thresholding, interactive selection of
connected components, morphological opening and
dilation. The next section deals with description of a
difference measure that is used for comparing
images of details of paintings.
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 2: Illustration of a brushwork feature description:
(a) – (c) are the images of a human face detail "cheek"; (d)
– (f) are the ridge orientation histograms obtained from
images (a) – (c); (g) – (i) are the corresponding simple
neighborhood orientation histograms.
5 MEASURE OF IMAGE
DIFFERENCE
For comparing fragments of artworks, statistical
tests (Li, 2012), cluster analysis, and classification
techniques (Johnson, 2008), (Polatkan, 2009) are
used. In this paper, the image samples are compared
using information-theoretical measure of difference,
because this measure fits the features represented by
distributions. The measure is constructed on the
basis of Kullback-Leibler divergence as follows
(Escolano, 2009):
()
() ()
1
()log ()log
2() ()
,
is is
jk
HH
pq
pq
qp
dx x
ϕϕ
ϕϕ
ϕϕ
ϕϕ
ΦΦ
ΦΦ
ΦΦ
∈∈
=
−







(7)
where ()p
ϕ
Φ
and ()q
ϕ
Φ
are the probabilities of the
event, when orientation angle values in samples
i
j
u
and
i
k
u
are equal to
ϕ
;
H
is the alphabet of random
variable
Φ representing the orientation angle. The
measure
()
,
is is
jk
dx x
is non-negative and symmetric.
Firstly, according to the procedure proposed above,
we create craquelure masks for specified portrait
regions (forehead, nose, and cheek, see Figure 1). We
apply created masks to image patches at the step of
computing features. Using the developed feature
extraction procedures, we obtain histograms of
orientation angles defined by expressions (4) and (6).
After that, using measure of difference (7) and
expression (1), the distances
i
jk
D
between fragments
of type
i in portraits j and
k
are computed. At the
next step, as defined by expression (2), we aggregate
computed distances
i
jk
D
between corresponding
portrait regions into the values of distances
j
k
D
between portraits
j and
k
.
6 PRELIMINARY RESULTS
Computing experiment was carried out for choosing
parameters of feature extraction procedures and
testing the proposed features for applicability to
attribution tasks. In the experiment we use images of
three portraits by F. Rokotov and eight portraits by
other artists dated to 17-19th centuries.
For choosing parameters of feature extraction
procedures we computed distances
()
2
1
(, )
I
sisis
jk j k
i
Ddxx
=
=

between images represented by a particular feature
s
at different values of parameters. The first parameter
under consideration is the lower bound
b
of the size
of ridge connected components. Distances
s
jk
D
between portraits for values of
b
equal to 8 and 12
pixels were computed. Histograms of distances
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
224

between images of paintings, represented by
orientation angle of ridge connected components at
8b = and 12b = , are shown in Figure 3. Here, the
distances between three portraits by F. Rokotov are
denoted as "own" and distances between portraits by
different artists are denoted as "alien".
Figure 3: Histograms of distances between images of
paintings, represented by orientation angle of ridge
connected components at
8b = and 12b = ; f is a
frequency of a distance value
D .
The second parameter is the size of the window
function
w
in (5). Distances
s
jk
D
between portraits
for values of window function size
a
equal to 3, 5,
7, and 11 pixels were computed. Averages and
standard deviations found from histograms of
distances between "own" and "alien" images of
paintings represented by angle of simple
neighborhood orientation
ϕ at different values of
a
are presented in Figure 4. The point at
3a =
corresponds to noise components of painting's
texture. Other values of window function size
produce distances that are revealing different
components of brushwork texture characterized by
different spatial frequencies.
The results of comparing images using features
(3) and (5), extracted from the homotypic regions
(forehead, nose, and cheek, see Figure 1) of three
portraits by F. Rokotov and eight portraits by other
artists are obtained and presented in Figure 5. Here,
a histogram of distances
jk
D
between paintings by
F. Rokotov is denoted as "own". Histogram of
distances between paintings created by different
artists is denoted as "alien". It follows from Figure 5
that distance between paintings created by the same
artist is less than the distance between any two
paintings created by different authors. This result fits
the following description of brushworks made by the
art experts that show the difference of the styles. The
portraits by F. Rokotov are characterized by an
arrangement of strait brushstrokes of various length,
producing zigzag patterns. In the female portrait by
V. Tropinin the embossed brushstrokes are weakly
expressed and variously oriented. The portrait by P.
D. de Rij is characterized by the accurate
brushstrokes that creating the form of face details. In
the portrait by A. Kharlamov, short, narrow-meshed,
variously oriented, and intersecting brushstrokes
create illusion of three-dimensionality. V.Lik's style
of brushing is considered to be rather chaotic. The
peculiarities of the artistic style of I. Ligotsky are
conditioned by a uniform texture without any
diversity of paint layer relief. Dense long
brushstrokes reflect the structure of facial muscles.
In the male portrait by F. Riss some of the short
embossed strokes are visible in bright white regions
of paint layer. The brushstrokes are mainly long and
waved, exactly and firmly follow face details.
Figure 4: Averages and standard deviations found from
histograms of distances between "own" and "alien"
paintings at different values of window function size
a
.
Figure 5: Histograms of distances between paintings
created by the same artist ("own") and different artists
("alien").
So, “own” and “alien” paintings, represented by
feature description that we proposed in this work,
can be separated, and a threshold value for
attribution decision can be chosen.
0
0.1
0.2
0.3
0.4
35711
Avg own Avg alien
Std own Std alien
D
a
0
2
4
6
8
10
12
14
16
0 0.1 0.14 0.18 0.3 0.5 0.7 0.9
f
D
Own
Alien
ATechniqueforComputerisedBrushworkAnalysis
225

7 CONCLUSIONS
The problem of computer-assisted attribution of
fine-art paintings based on image analysis methods
is considered. A feature description of artistic
brushwork based on textural characteristics of
paintings is proposed. Selected image features in the
form of orientation angle distributions give
quantitative description of painter's artistic style and
provide suitable accuracy of features computation.
Feature evaluation does not require segmentation of
a single brushstroke and is not sensitive to image
acquisition conditions as opposed to the
conventional techniques. The results of computing
experiments showed the efficiency of the proposed
features for comparing artistic styles. The proposed
feature set may be used as a part of technological
description of fine art paintings for restoration and
attribution. The future research will be aimed at the
problems, remained untouched in the current
research. Firstly, as the art experts pay attention to
the geometry of the brushes used by the artists, the
research will be aimed on the frequency analysis of
the images. Secondly, the results of the experiment
presented above, showed that different values of the
feature extraction procedures provide capturing
different components of the painting's texture. So,
another aim of the research will be concerned with
multiscale feature description of paintings. Next, in
the presented research we tested the proposed
feature space and procedures on a limited number of
paintings. In particular, we invoked in experiments
only three paintings created by the same artist (F.
Rokotov) and eight portraits by other artists. In order
to obtain more accurate experimental results, we are
planning to extend our image dataset. It is important
to represent each artist in the dataset by a few
artworks. And at last, it is necessary to develop a
procedure for making decisions on similarity of
brushwork techniques and artistic styles.
ACKNOWLEDGEMENTS
This work is supported by the RFBR grants No 12-
07-00668 and No 15-07-09324.
REFERENCES
Obukhov, G., 1959. A brief glossary of fine art. Moscow,
The soviet painter. (In Russian).
Stone, D., Stork, D., 2010. Computer-assisted
Connoisseurship: The Interdisciplinary Science of
Computer Vision and Image Analysis in the Study of
Art, IP4AI3, Museum of Modern Art, New York, May
27, pp. 9-10.
Morelli, G., 1900. Italian Painters. Critical Studies of
Their Works. John Murray, London.
Berenson. B., 1903. The Study and Criticism of Italian Art.
George Bell and sons, London.
Ignatova, N, 1994. Analysis of oil painting textures. In
Fundamentals of Oil painting Examination. The
guidelines, Moscow, Grabar restoration Centre, vol. 1.
Johnson, C. R., Hendriks, E., Berezhnoy, I.J., Brevdo, E.,
Hughes, S.M., Daubechies, I., Li, J., Postma, E.,
Wang, J.Z., 2008. Image processing for artist
identification (Computerized analysis of Vincent van
Gogh's painting brushstrokes). Signal Processing
Magazine, IEEE. Vol. 25, No 4. - P. 37-48.
Polatkan, G., Jafarpour, S., Brasoveanu, A., Hughes, S.,
Daubechies, I., 2009. Detection of forgery in paintings
using supervised learning. ICIP2009, IEEE. P. 2921-
2924.
Lettner, M., M. Sablatnig, M., 2005. Texture Analysis for
Stroke Classification in Infrared Reflectogramms. H.
Kalviainen et al. Eds., SCIA 2005, Springer-Verlag
Berlin-Heidelberg, LNCS. Vol. 3540. - P. 459-469.
Shahram, M., Stork, D.G., Donoho, D., 2008. Recovering
layers of brush strokes through statistical analysis of
color and shape: an application to van Gogh's Self
portrait with grey felt hat. Computer Image Analysis in
the Study of Art. D. Stork, J. Coddington, Eds.: Proc.
of the SPIE. Vol. 6810. - P. 68100D-1 - 68100D-8.
Li, J., Yao, L., Hendriks, E., Wang, J. Z., 2012. Rhythmic
brushstrokes distinguish van Gogh from his
contemporaries: findings via automated brushstroke
extraction. IEEE TPAMI. Vol. 34, No 6. - P. 1159-
1176.
Sablatnig, R., Kammerer, P., Zolda, E., 1998. Structural
Analysis of Paintings Based on Brush Strokes. Proc.
of SPIE Scientic Detection of Fakery in Art. SPIE.
Vol.. 3315. – P. 87–98.
Murashov, D., 2014. Localization of differences between
multimodal images on the basis of an information-
theoretical measure, Pattern Recognition and Image
Analysis. Springer, Vol. 24, No 1. - P. 133-143.
Escolano, F., Suau, P., Bonev, B., 2009. Information
Theory in Computer Vision and Pattern Recognition. –
London, Springer Verlag.
Eberly, D., 1996. Ridges in Image and Data Analysis.
Klewer Academic Publishers.
Lindeberg, T., 1994. Scale-space Theory in Computer
Vision. Kluwer Academic Publishers.
Jähne, B. 2005. Digital Image Processing. 6th ed.
Springer.
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
226