Artistic Style Characterization of Vincent Van Gogh’s Paintings using
Extracted Features from Visible Brush Strokes
Tieta Putri, Ramakrishnan Mukundan and Kourosh Neshatian
Department of Computer Science, University of Canterbury, Christchurch, New Zealand
Keywords:
Stylometry, Style Characterization, Feature Extraction, Painting Classification.
Abstract:
This paper outlines important methods used for brush stroke region extraction for quantifying artistic style
of Vincent Van Gogh’s paintings. After performing the region extraction, stroke-related features such as
colour and texture features are extracted from the visible brush stroke regions. We then test the features by
performing a binary classification between painters from different art movements and painters from the same
art movement.
1 INTRODUCTION
Identifying artistic styles in digital paintings have
been of great interest for researchers in the field of
Computer Vision. It has many applications such as for
cultural heritage preservation (Putri and Arymurthy,
2010), differentiating art movement period (Johnson.
et al., 2008), building a style-based image retrieval
system (Lombardi et al., 2004) and forgery detection
(Rosseau, 1968). There are many factors that de-
termine an artistic style. Such factors are the brush
stroke characteristics and colour palette used by the
artist and the way objects are drawn. From those fac-
tors, brush stroke characteristics contribute the most
to an artistic style (Zang et al., 2013). For instance,
Pointillist-style consists of small, elliptical and re-
peated brush strokes that are put together in such way
that it will form the object when a viewer looks at it
from a certain distance (see Fig. 1).
The existence of many painting styles with each of
them having several unique brush stroke characteris-
Figure 1: Example of computer-generated Pointillist ren-
dering (Putri, 2012).
tics has motivated considerable research into brush
stroke analysis, which can be done mathematically
and statistically with the aid of stylometry. The aim
of stylometry is to quantify artistic styles with a se-
ries of extracted features from the digitized artworks
(Hughes et al., 2010). An image will represented as
a string of features of statistical, texture, colour or
shape, which will be analyzed using machine learning
techniques. In this paper, we use stroke-based sty-
lometry as we characterize various paintings by exa-
mining the statistical properties of the visible brush
strokes.
This work presents important image processing
methods for extracting visible brush strokes from a set
V of digital paintings by Vincent Van Gogh. Using ex-
tracted brush strokes, we describe feature extraction
methods based on their texture and shape. The ex-
tracted features can then be compiled into a feature set
S which serves as the quantified brush strokes proper-
ties. Since brush strokes appearances are closely re-
lated to the artistic style of the painting itself (Strass-
man, 1986), S can be seen as the style representation
of V.
This paper is organized as follows: In Section 2,
we describe some related work in artistic style charac-
terization and brush stroke extraction. Section 3 gives
a detailed description of the datasets and methods
used in our work. Then, in Section 4, we provide
some results and discussions. Finally, Section 5 con-
cludes this paper and outlines our future research di-
rections.
378
Putri, T., Mukundan, R. and Neshatian, K.
Artistic Style Characterization of Vincent Van Gogh’s Paintings using Extracted Features from Visible Brush Strokes.
DOI: 10.5220/0006188303780385
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 378-385
ISBN: 978-989-758-222-6
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

2 RELATED WORKS
2.1 Artistic Style Characterization
In a stylistic painterly non-photorealistic (NPR) sys-
tem, the characterization of artistic style is necessary
for capturing, representing, and remapping a parti-
cular artistic style to an input image. Every digi-
tized paintings can be seen as a composition of two
components: the style and the content (Gatys et al.,
2015). Artistic style characterization process extracts
the style component of digitized paintings as a set of
features. The features are then used by the NPR sys-
tem as a heuristic in the painterly rendering process.
Research done by Hughes et al. (2010) investi-
gated the characterization of artworks done by the
Flemish painter Pieter Bruegel the Elder using sparse
coding analysis. The aim of the research was to dis-
tinguish the authentic Bruegel paintings from the im-
itations by determining their similarity of the sparse
model. The sparse model attempts to describe the
image space by training a set of orthogonal basis func-
tions that effectively represent the space. Sparse co-
ding is proven to be an effective method for feature
modelling in drawings and in other two-dimensional
media due to the sparseness of the artworks’ statistical
structures that are considered to give a high contribu-
tion to the perception of similarity.
Sener et al. (2012) extracted various features for
identifying children’s book illustrators. From illustra-
tion samples by authors Alex Scheffler, Debi Gliori,
Dr. Seuss and Korky Paul, features such as 4x4x4
bin RGB histograms, gist (Oliva and Torralba, 2001),
colour dense SIFT (Lowe, 2004) and gradient his-
tograms are extracted. Support Vector Machine with
various kernels are then used for classification. From
their experimentation, it was found that these features
are useful for distinguishing one artist’s style from
another.
The extension of the work of Sener et al. (2012)
by Vieira et al. (2015) uses a set of 93 different fea-
tures extracted from various digital paintings by 12
artists. Among those features are image energy and
entropy along with their statistical properties. Rele-
vant features were selected by measuring the cluster
dispersion using scatter matrices. Image energy and
entropy are proven to be more representative of style
than any other colour-based features. This research
successfully identifies the correlation between several
Baroque painters based on their works.
2.2 Brush Stroke Extraction
Brush strokes are the medium used by painters to
communicate what they want to convey in their pain-
tings. The way they are drawn can also provide some
information related to the painter, for instance the
painter’s art movement and his/her emotional state
(Callen, 1982). Because of this, brush stroke extrac-
tion has an important role in the area of digital pain-
ting analysis since brush strokes contain a lot of in-
formation that can be used as features to represent a
painting.
Li et al. (2012) described a brush stroke extraction
method for distinguishing Van Gogh’s paintings from
his contemporaries. Their method was used for dis-
tinguishing Van Gogh’s paintings from two different
periods, which are Paris and Arles-St.Remy period.
Their work consists of developing statistical frame-
work for the assessment of the distinction level of
different painting categories, brush stroke extraction
algorithm, and numerical features for brush stroke
characterization. They used the EDISON edge de-
tection algorithm developed by Meer and Georgescu
(2001). After edges are detected, edge linking algo-
rithm and enclosing operation are performed in or-
der to close the gaps between edge segments. Then,
the processed edges are extracted using the connected
component labelling. Finally, brush stroke condi-
tions are defined as: the brush skeleton not severely
branched; the ratio of broadness to length is within
the range of [0.05, 1.0]; and the ratio of the brush size
to two times length times width span is within [0.5,
2.0]. The brush skeleton is produced by the thinning
operation of the extracted connected components.
Johnson. et al. (2008) did a mathematical analy-
sis for the classification of Van Gogh paintings. They
examined high resolution grayscale scans of 101 pain-
tings, which consist of: 82 paintings by Van Gogh, 6
paintings by other painters and 13 others which are
loosely classified to be Van Gogh or non-Van Gogh
by art experts. In their research, they combined two
kinds of features that are extracted from the paintings,
which are texture-based feature obtained by wavelets
and stroke-based geometric features obtained by edge
detection. They argue that it is extremely challenging
to locate strokes accurately from grayscale images in
a fully automated manner.
Berezhnoy et al. (2009) elaborated a method
called as prevailing orientation extraction technique
(POET). This method focuses on brush stroke tex-
ture orientation extraction for segmenting individual
brush strokes in Van Gogh’s painting. The method
consists of two stages: the filtering stage and the
orientation extraction stage. In the filtering stage, a
Artistic Style Characterization of Vincent Van Gogh’s Paintings using Extracted Features from Visible Brush Strokes
379

rotation invariant circular filter with good response
for band-passing is applied. The orientation extrac-
tion stage extracted the principal orientation of brush
strokes from the filtered images. The filtered images
were transformed into binary images using multilevel
thresholding before the orientations were extracted.
The evaluation of POET is based on the cross com-
parison between the judgments of POET and human
subjects.
3 CHARACTERIZING ARTISTIC
STYLE
3.1 Brush Stroke Extraction
In this paper, the term brush stroke refers to the trails
of paints in the canvas that are produced by the painter
during the painting process. A brush stroke can be
placed on a canvas using other tools that are not ne-
cessarily in the form of brushes. The brush strokes
that we are interested in are the ones who are visible
enough to the viewers (i.e. not concealed behind other
brush strokes). We employed three methods for brush
stroke extraction as detailed below.
3.1.1 Iterative Brush Region Extraction
The method by Putri and Mukundan (2015) employed
colour-based method that extracts brush area with the
assumption of colour homogeneity inside the brush
regions. The extraction is done by using a brush tem-
plate called blob, which is a circular region that will
move through the input image and detect regions with
uniform colour. The formal definition of a blob of
radius R with centre at pixel location P
0
is given by:
B
R
(P
0
) = {P ∈ I | kP − P
0
k ≤ R} (1)
The colour constraint for the blob defined in Eq.
(1) is the subset:
S
R
(P
0
) = {P ∈ B
R
(P
0
) | kv(P) − v(P
0
)k < ∆E} (2)
with the condition of:
#S
R
(P
0
) > 0.9(#B
R
(P
0
)) (3)
The definitions of variables in the above equations are
listed in Table 1.
The extraction is done sequentially from left-
to-right and top-to-bottom with the largest possible
brush stroke radius R. The region that is detected to
be inside the blob, which satisfies the conditions given
in Eq. (1), (2) and (3), is considered to be a part of a
brush stroke.
Table 1: Notations for Eq. (1), (2), and (3).
B
R
Set of pixels with a radius of R referred as a blob
S
R
Subset of B
R
P
0
Pixel location of the centre of B
R
P ∈ I Pixel in an image I
v(P) Colour component of P
kx − yk Distance value between two object x and y
∆E Error threshold for colour comparison
#A Number of elements in set A
3.1.2 Texture Boundary Detection
This method detects brush strokes by identifying dif-
ferent textures in the painting image. A brush stroke
can have a very different texture from other neigh-
bouring brush strokes. This happens due to various
factors, such as artist preferences, paint concentra-
tion, stroke orientation, and so forth. This method
used image entropy to measure the randomness of the
pixel information stored in every visible brush stroke.
In this method, local entropy value of the neighbour-
hood is calculated for each pixel in the extracted visi-
ble brush strokes. To obtain the visible brush strokes
area from the image I, we perform these steps:
1. Convert I to grayscale image I
gray
.
2. Adjust the contrast of I
gray
using histogram equ-
alization.
3. Perform binary thresholding with Otsu’s method
(Otsu, 1975) to cluster the area with visible brush
strokes.
4. The visible brush strokes is the area in I
gray
that
have the binary value 1. This area is called as
I
grayvis
.
After obtaining I
grayvis
, we then extract the entropy
value for each of its pixels. The entropy value in a
pixel P is given as follows:
E
P
= −
∑
(n
P
log
2
n
P
) (4)
where n
P
is histogram counts of the neighbourhood of
P. In this method, an 8x8 neighbourhood is used.
3.1.3 Gabor Filters
In this method, a filter bank of Gabor filters with vari-
ous scales and rotations is applied in order to evaluate
the distribution of intensity level. Gabor filters are
proven to be a robust method for analysing oil pain-
ting images which have textured brush strokes (Putri
and Arymurthy, 2010).
The two-dimensional Gabor filter is defined as:
g
λ,θ,σ,φ
(s,t) = e
−(
s
0
2
σ
2
s
+
t
0
2
σ
2
t
)
cos(
s
0
λ
+ φ) (5)
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
380

From Eq. (5), a filter response of signal f is defined
as:
R
λ,θ,σ,φ
(x, y) =
ZZ
W
f (x − s, y − t)g
λ,θ,σ,φ
(s,t)ds dt
(6)
For every pixel in the painting, the Gabor energy is
defined as:
e
λ,θ
(x, y) =
q
R
λ,θ,1,0
(x, y)
2
+ R
λ,θ,1,
π
2
(x, y)
2
(7)
The definitions of variables in the above equations are
listed in Table 2.
The Gabor energies given in Eq. (7) are computed
for λ
i
, i = 1, ..., 6 and θ
i
=
iπ
8
, i = 0, ..., 7. Each pair of
Gabor filters with the combination of λ
i
and θ
i
detects
the image intensity transition via convolution. Every
convolution will produce energy values for each pixel.
The total energy from every convolution is the num-
ber of contours (or light-dark transition), thus will de-
tect regions with different textures (Johnson. et al.,
2008).
Table 2: Notations for Eq. (5), (6), and (7).
g
λ,θ,σ,φ
Gabor filter with parameters λ, θ, σ, and φ
(s,t) Two-dimensional position of the impulse
λ Filter scale, also known as spatial frequency
θ Filter orientation
σ
s
and σ
t
Standard deviation of circular Gaussian envelope
σ
s
= 1 and σ
t
= 1
φ Phase offset of the filter response
φ = 0 for the real component
φ =
π
2
for the imaginary component
s
0
s cos θ + t sin θ
t
0
s sin θ + t cos θ
R
λ,θ,σ,φ
Gabor filter response with parameters λ, θ, σ, and φ
f Signal which response is to be calculated using R
λ,θ,σ,φ
W Filter window
e
λ,θ
Gabor energy on scale λ and orientation θ
3.2 Feature Extraction
From the three aforementioned methods, we extract
features that are related to the texture and shape of
the brush strokes from every painting patch from the
dataset. Shape features are extracted from the re-
sults of iterative brush region extraction and texture
boundary detection, while Gabor energy features are
extracted from the results of Gabor filters method.
The dataset and the generation of patches will be ex-
plained further in Subsection 3.3.
The shape features consist of the region properties
of a brush stroke, such as:
1. Major axis length: The length of major axis of
the ellipse that has the same normalized second
central moments as the region.
2. Minor axis length: The length of minor axis of
the ellipse that has the same normalized second
central moments as the region.
3. Eccentricity: The eccentricity of the ellipse that
has the same normalized second central moments
as the region.
4. Perimeter: The distance around the region boun-
dary.
5. Orientation: The angle between the x-axis and the
major axis.
For every detected brush stroke, the shape features
mentioned above along with their mean and standard
deviation are computed, giving 10 features for every
patch.
The Gabor energies given in Eq. (7) are calculated
for every pixel in the patches with 6 different scales
and 8 different orientations. For each patch, the mean
and standard deviation of the energies for each scale
and orientation are obtained, thus giving us 96 fea-
tures for every patch.
Consequently, each patch in the corpora can be
seen as a row of data which consists of a total of
106 features. The features then got selected to im-
prove the accuracy and decrease the training time of
the classification. The feature selection is done by
Weka’s AttributeSelection
1
filter which evaluates sub-
set of features by examining each of their individual
ability to predict the correct class.
3.3 The Datasets
3.3.1 The Van Gogh Corpus
Table 3: The Van Gogh Corpus.
Title Resolution
Le Moulin de la Galette 3840x3082
Self-Portrait with Grey Felt Hat 2606x3163
Self-Portrait with Straw Hat 2452x3068
Cabbages and Onions 3840x2975
A Pair of Leather Clogs 3840x3034
The Garden of Saint-Paul Hospital 3840x3039
Landscape at Twilight 3507x1719
Tree Roots 3840x1879
View of Auvers 3840x3694
Wheat Fields 3840x3153
Wheat Field under Thunderclouds 3840x1885
Wheat Field with Crows 3840x1939
In our work, we choose several paintings by Vin-
cent van Gogh from his different art periods. Van
1
The AttributeSelection filter can be found on Weka Ex-
plorer GUI in Weka → Filters → Supervised → Attribute
→ AttributeSelection.
Artistic Style Characterization of Vincent Van Gogh’s Paintings using Extracted Features from Visible Brush Strokes
381

Gogh’s works are chosen due to their distinguish-
able brush stroke characteristics which are bold, wide,
repetitive and have the ability to convey objects with
a certain level of abstraction such that the object
appears to be fleeting in the viewer’s eyes (Callen,
1982). The corpus is made of 12 paintings with high-
resolution obtained from The Van Gogh Museum via
Google Art Project. Table 3 provides the complete list
of paintings in the corpus.
3.3.2 The Testing Corpora
Two kinds of testing corpora are used to validate the
extracted features proposed in this work. Both cor-
pora consist of images by painters other than Van
Gogh. Those corpora are called as the Rembrandt cor-
pus and the Impressionists corpus.
The Rembrandt corpus consist of the works by
Rembrandt Harmenszoon van Rijn, a Realist painter
whose brush stroke properties are very different to
Van Gogh’s. Table 4 provides the complete list of
paintings in this corpus. This dataset is used for a
classification benchmark to validate the representabi-
lity of features extracted from the Van Gogh corpus.
The Impressionists corpus consist of the works
of other Impressionists such as Claude Monet, Paul
C
´
ezanne and Auguste Renoir. The complete list
of paintings in this corpus can be seen in Table 5.
Van Gogh has a unique way to create Impressionistic
brush stroke, thus differentiating his works from other
Impressionists can be a useful performance measure-
ment for his brush stroke analysis.
In all corpora, each painting is divided into a
set of 500x500 patches. After dividing the images
into patches, any remaining blocks that are less than
500x500 pixels are omitted. Each painting consists of
approximately 30-50 patches.
We use the L*a*b (CIELAB) colour space since it
can simulate colour perception in a way that is close
to the human visual system (Reinhard et al., 2008).
Since all the images in all corpora are in RGB format,
we convert all of the pixel values of them to CIELAB
space before we process them to extract features.
3.4 Artistic Style Characterization
Pipeline
The artistic style characterization in this work is done
by these following processes:
1. Divide each painting in the Van Gogh corpus into
500x500 patches.
2. Extract visible brush strokes from every patches
using the three proposed brush stroke extraction
methods. The visible brush strokes are brush
Table 4: The Rembrandt Corpus.
Title Resolution
Parable of the Hidden Treasure 3703x2864
The Entombment of Christ 2024x1604
Judas Returning the Thirty Pieces of Silver 2048x1585
The Apostle Paul in Prison 3168x3727
David with the Head of Goliath before Saul 2048x1412
Balaam and the Ass 2252x3000
Man in a Gorget and a Cap 2358x3208
The Spectacle-Pedlar 2793x3284
The Operation 1410x1724
The Abduction of Europa 3000x2342
The Raising of Lazarus 4113x4905
The Parable of Rich Fool 2998x2228
Tobit Accusing Anna of Stealing the Kid 2058x2724
Family Portrait 3000x2264
Descent from the Cross 2789x3840
The Stoning of Saint Stephen 2024x1458
Table 5: The Impressionists Corpus.
Title Painter Resolution
At the Water’s Edge Paul C
´
ezanne 4000x3146
Bazille and Camille Claude Monet 2975x4000
Flowers in a Rococo Vase Paul C
´
ezanne 2452x3068
Oarsmen at Chatou Auguste Renoir 4000x3255
Sainte-Adresse Claude Monet 4000x2786
The Japanese Footbridge Claude Monet 4000x3219
Woman with a Parasol Claude Monet 3220x4000
strokes that have identifiable form, i.e. the obvi-
ous brush strokes that are not located behind any
other brush strokes.
3. Extract features f
1
, ..., f
n
from the visible brush
strokes. For every patch p, the features are then
grouped into a set S
p
= { f
1
, ..., f
n
}.
4. Repeat process number 1-3 for the testing cor-
pora. The obtained feature set is T
p
= {g
1
, ..., g
n
}.
5. Do a classification-based test from both S
p
and
T
p
. The feature set obtained from the Van Gogh
corpus is then tested by a classification-based test
with the testing feature sets obtained from the tes-
ting corpora.
From the result of process number 5, if S
p
are se-
parable for every Van Gogh painting patch p then the
features f
1
, ..., f
n
are considered representative for
quantifying Van Gogh’s painting style.
4 EXPERIMENTAL RESULTS
4.1 Brush Stroke Extraction
The extraction results can be seen in Fig. 2. Tex-
ture boundary detection and Gabor filter give good
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
382

a
b
c
d
Figure 2: (a) The input image and the brush extraction re-
sult using: (b) iterative brush region extraction, (c) texture
boundary detection and (d) Gabor filters.
result in capturing visible brush strokes due to their
ability to detect regions based on the texture within
them. Iterative brush region extraction is fast in the
implementation but unsuccessful in detecting small
textured brush strokes that are placed in the same area.
For instance, it misses smaller strokes in the sky area
of Fig. 2b.
4.2 Distinguishing Van Gogh from other
Painters
The features from the extracted brush strokes are then
classified according to their respective class. There
are three classes, which are Van Gogh (VG), other Im-
pressionists (NVG) and Rembrandt (R). The purpose
of the classification is to validate the representability
of the extracted features, i.e. how effective are the
features in quantifying the artistic style of Van Gogh.
The classifications are done in Weka in two ways:
first as binary classifications between the classes VG
and R, and second as binary classifications between
the classes VG and NVG. The experiments are done
using 10-fold cross validation (CV) with 70/30 per-
centage split, in which 70% of the data are used for
training and the rest 30% are used for testing. Multi-
layer Perceptron (MLP) and J48 classifiers are used to
their reliability and capability to learn the data adap-
tively (Su et al., 1996) (Bhargava et al., 2013). The
configurations for the classifiers are given in Table 6
and 7.
In Table 8, it can be seen that the results of the
classification between the works of Van Gogh and
Rembrandt are high in accuracy and F-measure value.
This is expected since those two painters come from
different art movement periods thus have a very dif-
ferent style to depict an object. While as an Impres-
sionist, Van Gogh tends to use bold strokes to convey
only the most essential part of the objects; Rembrandt
use very tiny strokes to depict his objects as real as
possible, thus making him a Realist.
Table 6: Weka J48 Classifier Configurations.
Parameters Value
Binary splits False
Option for collapsing the tree True
Pruning confidence 0.25
Option for making split point actual value False
Minimum number of instances 2
Number of folds for reduced error pruning 3
Option for reduced error pruning False
Seed for random data shuffling 1
Option to perform subtree raising True
The classification results between Van Gogh and
his fellow Impressionists show less accuracy levels
than the results between Van Gogh and Rembrandt,
but are still satisfactory. Being under the same in-
Artistic Style Characterization of Vincent Van Gogh’s Paintings using Extracted Features from Visible Brush Strokes
383

Table 8: The Classification Results of Van Gogh’s Brush Stroke Features.
Class Classifier Testing Mode Accuracy F-Measure
VG & R MLP 10-fold CV 99.56% 0.996
VG & R MLP 30/70 99.53% 0.995
VG & R J48 10-fold CV 97.79% 0.978
VG & R J48 30/70 98.57% 0.986
VG & NVG MLP 10-fold CV 97.74% 0.977
VG & NVG MLP 30/70 split 98.76% 0.988
VG & NVG J48 10-fold CV 87.57% 0.876
VG & NVG J48 30/70 87.58% 0.875
Table 7: Weka MLP Classifier Configurations.
Parameters Value
Option to autocreate the network connections True
Option to allow learning rate decay False
Learning rate for the backpropagation 0.3
Momentum rate for the backpropagation 0.2
Option to filter nominal to binary True
Option to normalize attributes True
Option to normalize numeric class True
Number of epochs 500
Threshold for number of consecutive errors 20
Percentage of validation set 0
Value to seed the random number generator 0
fluence of Impressionism art movement, the painters
in the Impressionists class have similar style to Van
Gogh in terms of object representation. Nevertheless,
Van Gogh has different way to portray light and use
curvy and expressive brush strokes to create visual
effect that will make his paintings more engaging to
the viewers. This makes him as a Neo-Impressionist
painter who employs the techniques of the Impres-
sionists in unconventional ways (Callen, 1982).
5 CONCLUSION AND FUTURE
WORK
5.1 Conclusion
In this paper, three methods are used for quantifying
artistic style by analysing the visible brush strokes.
Those three methods are iterative brush region ex-
traction, texture boundary detection and Gabor filters.
Then, based on their shape and texture, the properties
of extracted strokes are encapsulated in a set of fea-
tures. The features are then classified to test their re-
presentability of quantifying a particular artistic style.
The experiment results show that the proposed me-
thods give satisfactory results in producing features
that are able to differentiate the works by Van Gogh
from the works by another Impressionists painters.
5.2 Future Work
The immediate extension of this work is NPR
parametrization based on the extracted features for
painterly stroke-based rendering of photograph im-
ages. The parameters will be used for guiding the
digital brush which will be modelled as a group of co-
ordinated particles that travels across the digital can-
vas. After the rendering results are produced, their
aesthetics will then be assessed using Convolutional
Neural Network to eliminate the biases that will be
introduced by artist-based assessment.
Another possible extension of this work is to build
a robust style-based painting retrieval system which
can be used for retrieving artworks based on their
artistic style, painter or art movement era. This sys-
tem will be beneficial for art education and apprecia-
tion in museums or art galleries.
ACKNOWLEDGEMENTS
Tieta Putri would like to thank the Indonesia Endow-
ment Fund for Education for their financial support
during her Ph.D. study at University of Canterbury.
REFERENCES
Berezhnoy, I., Postma, E., and Herik, H. V. D. (2009). Auto-
matic extraction of brushstroke orientation from pain-
tings. In Machine Vision and Applications, volume 20,
1, pages 1–9.
Bhargava, N., Sharma, G., Bhargava, R., and Mathuria, M.
(2013). Decision tree analysis on j48 algorithm for
data mining. In International Journal of Advanced Re-
search in Computer Science and Software Engineer-
ing, volume 3, 6.
Callen, A. (1982). Techniques of the Impressionists. QED
Publishing/Chartwell Books.
Gatys, L., Ecker, A., and Bethge, M. (2015). A neural algo-
rithm of artistic style. In Nature Communications.
Hughes, J., Graham, D., and Rockmore, D. (2010). Quan-
tification of artistic style through sparse coding analy-
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
384

sis in the drawings of pieter bruegel the elder. In Pro-
ceedings of the National Academy of Sciences, volume
107, 4, pages 1279–1283.
Johnson., C. J., Hendriks, E., Berezhnoy, I., Brevdo, E.,
Hughes, S., Daubechies, I., Li, J., Postma, E., and
Wang, J. (2008). Image processing for artist identi-
fication. In IEEE signal processing magazine, volume
25, 4, pages 37–48. IEEE.
Li, J., Yao, L., Hendriks, E., and Wang, Y. (2012). Rhythmic
brushstrokes distinguish van gogh from his contempo-
raries: findings via automated brushstroke extraction.
In IEEE Trans. on Pattern Analysis and Machine In-
telligence, volume 34, 6, pages 1159–1176. IEEE.
Lombardi, T., Cha, S., and Tappert, C. (2004). A graph-
ical user interface for a fine-art painting image re-
trieval system. In ACM SIGMM international work-
shop on Multimedia information retrieval, pages 107–
112. ACM.
Lowe, D. (2004). Distinctive image features from scale-
invariant keypoints. In International Journal of Com-
puter Vision, volume 60, 2, pages 91–110.
Meer, P. and Georgescu, B. (2001). Edge detection with em-
bedded confidence. In IEEE Trans. on Pattern Anal-
ysis and Machine Intelligence, volume 23, 12, pages
1351–1365. IEEE.
Oliva, A. and Torralba, A. (2001). Modeling the shape of
the scene: A holistic representation of the spatial en-
velope. In International Journal of Computer Vision,
volume 42, 3, pages 145–175.
Otsu, N. (1975). A threshold selection method from gray-
level histograms. Automatica, 11(285-296):23–27.
Putri, T. (2012). Non-photorealistic rendering of pointillist-
inspired images by an evolutionary process. Master’s
thesis, RMIT University.
Putri, T. and Arymurthy, A. (2010). Image feature ex-
traction and recognition of abstractionism and real-
ism style of indonesian paintings. In IEEE Advances
in Computing, Control and Telecommunication Tech-
nologies, pages 149–152. IEEE.
Putri, T. and Mukundan, R. (2015). Iterative brush path
extraction algorithm for aiding flock brush simulation
of stroke-based painterly rendering. In International
Conference on Evolutionary and Biologically Inspired
Music and Art, pages 152–162. Springer.
Reinhard, E., Khan, E., Akyuz, A., and Johnson, G. (2008).
Colour Imaging: Fundamentals and Applications.
AK Peters Ltd., Natick, MA.
Rosseau, T. (1968). The stylistic detection of forgeries. In
The Metropolitan Museum of Art Bulletin, volume 26,
6, pages 247–252. The Metropolitan Museum of Art.
Sener, F., Samet, N., and Sahin, P. (2012). Identification
of illustrators. In European Conference on Computer
Vision, pages 589–597. Springer.
Strassman, S. (1986). Hairy brush. In ACM Com-
puter Graphics and Interactive Techniques Computer
Graphics, volume 20, 4, pages 225–232.
Su, M., Jean, W., and Chang, H. (1996). A static hand
gesture recognition system using a composite neural
network. In Fifth IEEE International Conference on
Fuzzy Systems, pages 786–792. IEEE.
Vieira, V., Fabbri, R., Sbrissa, D., da Fontoura Costa, L.,
and Travieso, G. (2015). A quantitative approach to
painting styles. In Physica A: Statistical Mechanics
and its Applications, volume 417, pages 110–129.
Zang, Y., Huang, H., and Li, C. (2013). Stroke style analysis
for painterly rendering. In Journal of Computer Sci-
ence and Technology, volume 28, 5, pages 762–775.
IEEE.
Artistic Style Characterization of Vincent Van Gogh’s Paintings using Extracted Features from Visible Brush Strokes
385