Pattern Recognition 47 (2014) 623–633

Contents lists available at ScienceDirect

Pattern Recognition
journal homepage: www.elsevier.com/locate/pr

Unsupervised segmentation and approximation of digital

curves with rate-distortion curve modeling
Alexander Kolesnikov a,n, Tuomo Kauranne b
a
Arbonaut Ltd., Kaislakatu 2, FI-80130 Joensuu, Finland
b
Lappeenranta University of Technology, PO. Box 20, FI-53851 Lappeenranta, Finland

art ic l e i nf o a b s t r a c t

Article history: This paper considers the problem of unsupervised segmentation and approximation of digital curves and
Received 31 December 2012 trajectories with a set of geometrical primitives (model functions). An algorithm is proposed based on a
Received in revised form parameterized model of the Rate–Distortion curve. The multiplicative cost function is then derived from
4 August 2013
the model. By analyzing the minimum of the cost function, a solution is defined that produces the best
Accepted 8 September 2013
possible balance between the number of segments and the approximation error. The proposed algorithm
Available online 16 September 2013
was tested for polygonal approximation and multi-model approximation (circular arcs and line segments
Keywords: for digital curves, and polynomials for trajectory). The algorithm demonstrated its efficiency in com-
Shape parisons with known methods with a heuristic cost function. The proposed method can additionally be
Graphical model
used for segmentation and approximation of signals and time series.
Piecewise linear approximation
& 2013 Elsevier Ltd. All rights reserved.
Curve fitting

1. Introduction (2) Min-ε problem: With a given number of approximating line

Polygonal approximation of digital curves is used in image
processing, computer graphics, pattern recognition, data retrieval,
CAD, GIS, shape analysis and encoding. In all above applications,
it is used to reduce the amount of data that need to be processed,
stored and transmitted. In addition to polygonal approximation,
signals, curves and trajectories can be approximated with non-
linear functions like arcs, polynomials and splines [1–7]. There are
two main approaches [8] to the problem of shape representation
with a polygonal curve: (1) identification of the dominant points
as vertices of the approximating polygon, or (2) approximation of
the segments of the input curve with a continuous sequence of
line segments. The former approach is based on analysis of the
curvature or other features to identify the corner points of the
curve [9–13]. The approach has the drawback that the algorithms
are sensitive to contour noise and often include input or hidden
parameters. When using the latter approach, the problem of the
approximation of digital curves is usually formulated in one of the
following two forms [7,14]:
(1) Min-# problem: With a given constraint on the approxima-
tion error, approximate a curve with the minimum number of line
segments (the description length).
n
Corresponding author. Tel.: þ 358 468 104 853.
E-mail addresses: alexander.kolesnikov@arbonaut.com (A. Kolesnikov).
tuomo.kauranne@lut.fi (T. Kauranne).
segments M (description length R), approximate the input curve
with a minimum approximation error.
The solution to the problem depends on the error measure in
use. For an error measure of L1 (the maximum deviation) the
solution can be found with the following heuristic [15–20] or
optimal algorithms [21–24]. With L2-norm (Integral Square Error,
ISE), the problem can be solved with heuristic [20,25] or optimal
algorithms [26–31].
Thus, an approximation can be found if the number of seg-
ments M (the description length R) or the error bound ε0 is known.
However, before an approximation can be formed, we need to
know the most suitable value of the input parameter and how
many segments are required to represent the curve adequately.
Therefore, a method is needed that determines the input para-
meter of the approximation algorithm for a concrete curve.
A possible approach to the problem is to find a balance
between the number of segments M and the approximation error
E by introducing a cost function that incorporates both M and E.
For an additive error measure L2, the Lagrange multiplier method
can be used. The Lagrange multiplier algorithm searches for a
solution with a minimal value of the additive cost function
C ¼E2 þ λM [2,4,5,32], where E2 is the Integral Square Error. The
trade-off between the number of segments M and the Integral
Square Error E2 is controlled by means of the user-defined
heuristic Lagrange multiplier λ.
In [9], a multiplicative criterion Figure of Merit (FOM) was
introduced for evaluating solutions that have been obtained by
using heuristic algorithms for polygonal approximation: FOM¼E2M.

0031-3203/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.patcog.2013.09.002
624 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
Later, this criterion was presented as an “optimization error” [10],
“weighted sum of square error” [11] and “compromise ratio” [13].
However, Rosin [33] has demonstrated that the two terms, M and
E2, in the FOM are not balanced – the heuristic criterion gives
solutions with too many linear segments. In the modified criterion
FOM-n (n¼2 or 3) proposed in [12], the number of segments M was
penalized by raising it to a power n that would have the effect of
reducing this bias: FOMn ¼ E2 M n :
pffiffiffiffiffi [17], as the “relative
The criterion FOM-2 was proposed earlier
error”, in the following form: Er ¼ M E2 , but this publication
remained unnoticed. Although originally created as a criterion for
evaluating heuristic algorithms [8,12,17,25], the criterion FOM-n
has been used for determining the number of linear segments
[20,34–36]. However, it is still not clear which value of the
parameter n is more appropriate, n ¼2 or n ¼3.
For approximation with the error measure L1, the multiplica-
tive criterion “weighted maximum error”, WE1 was introduced in
[11]: WE1 ¼ E1 M: The modified Figure of Merit (MFOM-3) was
proposed in [19]. The criterion incorporates the criterion FOM-2
and the maximum deviation E1 with the number of segments M
as follows:
MFOM-2 ¼ MFOM-2E1 M ¼ E2 E1 M 3
In the algorithm [19], the set of polygonal approximating
curves was sequentially constructed for input contour with a
heuristic algorithm for increasing error tolerance for maximum
deviation until the first local minimum of the MFOM-3 criterion is
reached.
In this paper, we approach the problem from another angle.
Based on the fractal properties of curves, we first construct a
Fig. 1. General scheme of the full-search DP algorithm for polygonal approximation for error measures L2 (left) and L1 (right) [26]. M.
parameterized model of the approximation error as a function of
the number of segments M. Then, we derive a multiplicative cost
function for approximation with error measures L2 and L1. The
minimum of the cost function gives us the optimal number and
type of approximating segments.
The paper is organized as follows. Section 2 discusses the
problem of segmentation and approximation of curves. In
Section 3, we introduce a parameterized model of the Rate–
Distortion. Section 4 presents a criterion that gives a solution to
the problem in question. Section 5 sets out the results of our
experiments for test curves. Section 6 presents the conclusions.
2. Rate–Distortion curve
An open planar N-vertex curve P is defined as ordered set of
points: P¼ {p(1), …, p(N)}, where p(n) ¼(x(n), y(n)). The polygonal
curve P is approximated by another polygonal curve Q¼{q(1), …, q
(Mþ 1)} with M linear segments, where q(m)¼p(im).
The approximation error with the measure L1 for a curve
segment {p(im), …, p(im þ 1)} is defined as the maximum of the
Hausdorff distance dH between points p(n) of the curve segment
ð1Þ
and the corresponding approximating linear segment S(im, im þ 1)
defined by the points p(im) and p(im þ 1). Maximum deviation for
the curve P is defined as follows:
 
δðMÞ  E1 ðMÞ ¼ max max fdH ðpðnÞ; Sðim ; im þ 1 ÞÞg : ð2Þ
1rmrM im r n r im þ 1
The approximation error with the measure L2 for a curve
segment {p(im), …, p(im þ 1)} is defined as the sum of squared
 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
Euclidean distances dE between a point p(n) of the curve segment
and the approximation line L(im, im þ 1) defined by the points p(im)
and p(im þ 1). The total approximation error E2(M) for the curve P is
defined as the sum of the approximation errors for all the
segments
M M im þ 1
2
E2 ðMÞ ¼ ∑ e2 ðim ; im þ 1 Þ ¼ ∑ ∑ dE ðpðnÞ; Lðim ; im þ 1 ÞÞ:
m¼1 m ¼ 1 n ¼ im
Approximation error with L2–norm is also characterized by Root-
Mean-Square-Error s:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sðMÞ ¼ E2 ðMÞ=N:
In the general case, we have a set of K approximation models
(functions) Φ¼ {φ1, …, φK}, where the model functions are
described by the corresponding number of parameters {λ1, …,
λK} [7]. For two-model approximation with line segments (φ1) and
circular arcs (φ2), the description length of the models are defined
by the corresponding number of parameters to represent the
model function: λ1 ¼ 2 and λ2 ¼3. For a polynomial model of order
q we thus have λk ¼q.
The input curve P is divided into segments S(im, im þ 1), and each
segment is approximated by a function φk from the set Φ. The
description length, r(im, im þ 1; φk), for the curve segment S(im, im þ 1)
is defined as a description length λk of the selected model φk. for
the segment: r(im, im þ 1; φk)¼ λk. The total description length R for
Fig. 2. Approximation of the curve Great Britain with L2-norm: (a) the Rate–Distortion curve in linear scale with its model; and (b) the Rate–Distortion curve in log–log scale
with its model.
Fig. 3. Approximation of the curve Great Britain with L1-norm: (a) the Rate–Distortion curve in linear scale with its model; (b) the Rate–Distortion curve in log–log scale
with its model: D1 ¼ 1.09 and a¼ 0.92.
625
the curve is obtained from the sum of the description lengths of all
segments
M
R ¼ ∑ rðSðim ; im þ 1 ; φk ÞÞ: ð5Þ
m¼1
The functions E1(m) and E2(m) are called Rate–Distortion curves
ð3Þ for the error norms L1 and L2, correspondingly.
Solution to the min-ε problem can be found with a Dynamic
Programming (DP) algorithm for the k-link shortest path in the
weighted acyclic graph of size M  N (state or search space) [26–
30]. For multi-model approximation, a set of solutions can be
calculated with the optimal DP algorithm [7] for the description
ð4Þ length in the range [R1, R2].
The cost function C(n,m) is defined as the minimum approx-
imation error for the n-vertex curve {p1, p2,…, pn} approximated by
m line segments. Thus, the function C(N, m) for m ¼[1, M] gives the
Rate–Distortion curve. A 2D array of parent states A(n,m) is also
needed to construct the curve partition into segments. The general
scheme of the full-search DP algorithm for polygonal approxima-
tion is given in Fig. 1. Modification of the algorithm for multi-
model approximation is given in [7].
Approximation error Ep(j,n) for curve segment {pj, …, pn} can be
calculated in O(N) and O(1) time for measures L1 and L2,
correspondingly [26]. The errors for all segments with end point
pn are calculated beforehand and stored in an array of size N. The
total time complexity of the DP algorithm for L1-norm is cubic:
T1 ¼O(N3), i.e., defined by the time for calculation of approximation
626 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633

errors for O(N2) curve segments. For L2-norm, the complexity is error ε^ p on the number of segments m
defined by complexity of the DP algorithm for the shortest k-link
ε^ p ðmÞ ¼ C p m  1=Dp : ð8Þ
path: T2 ¼ O(MN2) [26]. The space complexity of the algorithms is O
(MN), which is defined by the size of the search graph (size of the The parameter Dp is evaluated from the corresponding log-linear
arrays C and A). regression model for m in the range [M1, M2] as follows:
( )
M2
Dp ¼ arg min ∑ ðlog ε^ p ðmÞ  log εp ðmÞÞ2 : ð9Þ
3. Model of Rate-Distortion curve m ¼ M1

This gives:
To construct model of the Rate–Distortion curve, we consider
the fractal dimension of curves. There are two main practical 2
ðM 2  M 1 þ 1Þ∑M M2
M 1 lg ðmÞ  ð∑M 1 lgðmÞÞ
2 2
methods to find the fractal dimension for planar curves: the box- Dp ¼ :
counting method [37] and the method with calculation of the ðM 2  M 1 þ 1Þ∑M M2 M2
M 1 lgðmÞlgðεp ðmÞÞ  ∑M 1 lgðmÞ∑M 1 lgðεp ðmÞÞ
2

curve length [38–40]. The fractal dimension H is defined with the
box-counting dimension method as follows: the curve P is covered
with a regular grid with cell size ε  ε and the number n(ε) of the
cells (boxes) that cover the curve is calculated. Then, the fractal
dimension H is defined as follows:
log nðεÞ
H ¼ lim : ð6Þ
ε-0 log ε
Thus, the number n(ε) of covering boxes of size ε  ε for the curve
can be estimated by the following power law: nðεÞ  Cε  H .
Following this approach, we cover the curve with m variable-
length rectangles of width wp, where the width depends on the
error measure: w1  ε1 ¼δ(m) and w2  ε2 ¼s(m) for the measures
L1 and L2, correspondingly. Thus, the assumption is that the
number of such rectangles m is described by the following
function of the error εp [41]:
D
mðεp Þ  εp p ;
where the parameter Dp depends on the error measure. Relation (7)
is inverted to get the dependence of the estimated approximation
ð10Þ
Thus, a model is constructed that describes “on average” the
behavior of the Rate–Distortion curve for m in the range [M1, M2].
The parameter Dp for a curve is not necessarily equal to the fractal
dimension H of the curve [41]. Indeed, the number n(ε) of covering
boxes of size ε  ε for a straight line segment of length L is as
follows: n(ε)  L/ε. Thus, for the straight line, the fractal dimension
H is equal to 1. On the other hand, only one rectangle of width wp
is needed to cover the line segment, which gives a value of
parameter Dp for the straight line: Dp ¼ 0.
An example of the Rate–Distortion (R–D) curve for the test
contour Great Britain is given in Figs. 2 and 3, the parameter of the
R–D curve model is D2 ¼ 1.10 and D1 ¼1.09, correspondingly.
4. Parameterized criterion pFOMp
ð7Þ
From (8), it follows that the function I RDC ¼ ε^ p ðmÞm1=Dp is a
constant for the model of the Rate–Distortion curve in the range
[M1, M2]. We can derive the criterion Fp(m) directly from the

Fig. 4. The test data set: (a) Curve Great Britain, N ¼ 10,909; (b) Curve #1, N ¼ 1953; (c) Curve #2, N ¼ 2661 [19]; (d) Curve #3, N¼ 460 [12]; (e) Curve #4, N ¼ 428 [12]; (f) Curve
#5, N ¼ 1159 [19]; (g) Curve #6, N ¼1042 [36]; (h) Curve #7, N ¼2141 [19]; (i) Curve #8, N ¼ 1029, (j) Curve #9, N ¼939; (k) Trajectory Mario 8, N ¼ 1861 [42]; and (l) 2D point
set Shark, N ¼ 449 [43].
 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633 627

model invariant IRDC to evaluate solutions as follows: F p ðmÞ ¼ In the case of multi-model approximation, the total description
εp ðmÞm1=Dp . However, to keep consistency with the criteria FOM-n length R of the curve is used instead of the number of segments m.
and WE1 [12,11], the parameterized Figure of Merit pFOMp is Then the criterion pFOMp should be written as follows:
defined in the following way: (
( E2 Ra ; a ¼ 2D2 1 ; p ¼ 2;
E2 ma ; a ¼ 2D2 1 ; p ¼ 2; pFOM p ¼ 1 ð13Þ
pFOM p ¼ 1 ð11Þ E1 Ra ; a ¼ D1 ; p ¼ 1:
E1 ma ; a ¼ D1 ; p ¼ 1:
The algorithm for unsupervised approximation therefore consists
The global minimum of the criterion pFOMp at the point M*
of three steps:
gives a solution with the best balance between the error Ep and the
number of segments m in terms of the multiplicative cost function
Step 1: Calculate approximation errors Ep with the correspond-
M* ¼ arg min fEp ðmÞma g ð12Þ ing optimal algorithm and find the model parameter a;

Fig. 5. Rate–Distortion curves with models for approximation with L2-norm for: (a) Curves #1–6 (polygonal approximation); (b) Curves #7–9 (multi-model approximation
with line segments and circular arcs); and (c) Trajectory Mario 8 (multi-model approximation with polynomials).

Fig. 6. Rate–Distortion curves for approximation with L1-norm for: (a) Curves #1–6 (polygonal approximation); and (b) Curves #7–9 (multi-model approximation with line
segments and circular arcs).
628 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
Step 2: Find the global minimum of the criterion pFOMp by the
number of segments m (or the description length R);
Step 3: Approximate the input curve with the optimal algo-
rithm by using the found number of approximating segments
M* (description length R*).
The heuristic criterion FOM-n was proposed in [12] to over-
come the drawback of the criterion FOM by penalizing the number
of segments m. In the criterion FOM-n, an integer value of the
parameter n should be guessed, whereas with the approach
proposed in this paper, the parameter a is obtained from the
Rate–Distortion curve. The criterion WE1 is a particular case of the
criterion pFOM1 for a¼ 1.
The total time complexity of the algorithm is cubic. This is
defined by the complexity of the optimal DP algorithm for
polygonal and multi-model approximation of curves.
5. Results and discussions
The proposed algorithm was tested on a HP EliteBook with
Microsoft Visual Studio 2010. The results of the experiment are set
out for the curve Great Britain, nine digitized curves [5,22,24], the
trajectory Mario 8 [42] and the 2D point set Shark [43] (see Fig. 4).
The set of the test shapes included smooth and noisy curves.
Fig. 7. The test curve Great Britain: (a) the criteria FOM-2, FOM-3, and pFOM2: M* ¼22 for pFOM2; and (b) results of the optimal polygonal approximation with L2-norm for
M* ¼ 22.
Fig. 8. The test curve Great Britain: (a) the criteria WE1, MFOM-3, and pFOM1: M* ¼ 34 for pFOM1; and (b) results of the optimal polygonal approximation with L1-norm for
M* ¼ 34.
The proposed algorithm was tested with single- and multi-
model approximation. The former case comprised polygonal
approximation of the curve Great Britain and test curves #1–6
with L2 and L1-norms. The 2D point set Shark has been approxi-
mated by polygonal curve with the Optimal Transport algorithm
[43]. In the latter case, the set of approximating models included
line segments and circular arcs for test curves #7–9. The test
trajectory Mario 8 was approximated with a set of polynomials of
the 1st and 2nd order, where the order of polynomials for the
segments was defined with the optimal algorithm [7].
The results of experiments have demonstrated that the pro-
posed parametric model describes the behavior of Rate–Distortion
curves well enough for most of the test curves, especially for the
coastline of Great Britain, the 2D point set Shark, the trajectory
Mario 8 and the curves #1, #2, #4, #7 (see Figs. 2, 3, 5, 6 and 11).
For artificial objects, like curves #3, #7 and #8, the R–D curve
deviates significantly from its model (see Figs. 5 and 6). However,
we are more interested in deviating R–D curves because in such a
case, the criterion pFOMp has a distinct global minimum. More-
over, a larger deviation of the R–D curve from its model, makes it
easier to find an adequate number of segments.
Parameter a of the model depends on the interval of the
numbers of segments [M1, M2] (Description Lengths [R1, R2]) used
for construction of the Rate–Distortion curve. In other words, we
should define the Region of Interest (ROI) for the approximation
 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
errors [εL, εU]. It is unreasonable to take too small a lower bound,
because the criterion pFOMp has a degenerate solution: pFOMp ¼0,
when Ep ¼0. On the other hand, the upper bound should be much
less than the geometrical size of the shape. For test curves #1–9,
we have s0 ¼[0.3, 8.0], and δ0 ¼ [0.3, 8.0]. For the test trajectory, the
range of approximation errors is s0 ¼[3.0, 12.0].
Test curves #1–6 are sorted by the model parameter a for
L2-norm in ascending order, where the parameter a is varied from
2.15 to 3.66 (see Fig. 5). The parameter a can be treated as a
measure of curve smoothness. For smooth curves, the approxima-
tion error decreases faster with increasing number of segments
than for noisy ones, as can be seen by comparing the parameter a
for two similar curves with different levels of noise, where a ¼2.15
(noisy curve #1) and a ¼3.02 (smooth curve #5).
We have evaluated performance of the criterion pFOMp in
comparison with some other multiplicative criteria for the curve
Great Britain, the curves #1–6, and 2D point set Shark. The
criterion pFOM2 was compared with the criteria FOM-2 and
FOM-3 [12] (see Figs. 7, 9, 11 and 12). The criterion pFOM1 was
then compared with the criteria WE1 and MFOM-3 [11,19] (see
Figs. 8 and 10). The criteria are scaled to the range [0,1] to
emphasize their minimums in the range under study.
With the proposed method, the number of segments is evaluated
from the Rate–Distortion curve. The resulting number of segments
depends on the error measure in use. The approximation error with
Fig. 9. Criteria FOM-2, FOM-3 and pFOM2 for approximation with L2-norm of test curves #1–9.
629
measure L1 is more sensitive to outliers than those with the additive
measure L2. That is why the number of segments evaluated with
the measure L1 is usually bigger than that with the measure L2
(see Figs. 13 and 14). Appropriate choice of error measure depends on
the purpose of curve segmentation.
The criteria FOM-2 or FOM-3 give the same results as the
criterion pFOM2 if a E2 (curve #1 and Great Britain) or a E3 (curve
#5), but which criterion to be used has to be guessed in every
particular case (see Figs. 7 and 13). For smooth shapes, both
criteria overestimate the number of segments M. For curve #6
with a6 ¼3.66, the proposed method gives M* ¼20 compared with
M* ¼47 and M* ¼34 for the criteria FOM-2 and FOM-3, respectively
(see Fig. 9). For well-structured artificial contours like curves #7
and 8, the heuristic criteria also give good results. However, in
total, the proposed criterion pFOMp gives more adequate results
than the heuristic criteria under study. In the criterion pFOMp,
sensitivity to the error Ep and the number of segments M is
balanced by the model parameter a: the criterion is less sensitive
to the number of segments for noisy curves and more sensitive for
noisy curves
The proposed unsupervised algorithm provides adequate shape
representation with a smaller number of segments than the
heuristic algorithms [12,35] (see Figs. 13–15). For example, for
curve #3, the proposed method gives M* ¼40 and M* ¼ 42 for L2
and L1-norm, respectively, against M ¼61 for the heuristic
630 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633

Fig. 10. Criteria FOM-2, WE1, MFOM-3 and pFOM1 for approximation with L1-norm of test curves #1–9.

Fig. 11. Test 2D point set Shark, N ¼ 449: (a) Rate–distortion curve with its model: a¼ 2.26; (b) Criteria FOM-2, FOM-3, and pFOM2; and (c) result of approximation with the
found number of the edges: M* ¼30 for pFOM2.
 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633 631

Fig. 12. (a) Criteria FOM-2, FOM-3 and pFOM2 for the test trajectory Mario 8; and result of polynomial approximation of the trajectory with found parameters, R¼ 222,
s ¼4.33: (b) in the plane (Time, X), and (c) in the plane (Time, Y).

Fig. 13. Results of approximation for error measure L2: (1) Curve #1, M* ¼ 27; (2) Curve #2, M* ¼ 30; (3) Curve #3, M* ¼ 40; (4) Curve #4, M* ¼ 35; (5) Curve #5, M* ¼ 34;
(6) Curve #6, M* ¼ 22; (7) Curve #7, Mlines ¼8 and Marcs ¼ 3, R¼ 25; (8) Curve #8, Mlines ¼8 and Marcs ¼ 4, R¼ 28; and (9) Curve #9, Mlines ¼3 and Marcs ¼13, R ¼44.

algorithm method in [12]. In [35], the first minimum of the
criterion FOM-2 has been used to evaluate the number of seg-
ments for optimal polygonal approximation of the curve #6 with
L2-norm: M ¼47 (see Fig. 15). The proposed method for this curve
gives a more adequate value: M* ¼ 22 (see Fig. 13).
The proposed method is not very sensitive to the presence of
noise: quite adequate results are obtained for smooth and noisy
versions of the same shape: M1* ¼27 for curve 1 and M5*¼ 35 for
curve 5 (L2-norm), M1* ¼ 34 and M5* ¼38 (L1-norm), respectively.
However, the corresponding approximation errors differ greatly
for two curves: s1 ¼1.40 and s5 ¼0.63 (L2-norm); and δ1 ¼3.2 and
δ5 ¼1.6 (L1-norm).
Due to the simplicity of the algorithm, the processing time is
acceptable even for a few thousand vertices. For example, the
processing time for approximation of 2661-vertex curve #3 is 1 s
and 30 s for measures L2 and L1, correspondingly. The processing
time for the 10,909-vertex curve Great Britain for measures L2 and
L1 is 10 s and 30 min, correspondingly. The processing time for
Full Search Dynamic Programming algorithm with error measure
L1 is too high for practical applications for N  10,000 vertices.
In such a case, it is recommended to reduce the number of points
in the input curve to a reasonable value [18] or to use a faster
heuristic or near-optimal algorithm for approximation to calculate
the Rate–Distortion curve.
632 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
Fig. 14. Results of approximation for the error measure L1: (1) Curve #1, M* ¼ 34; (2) Curve #2, M* ¼ 34; (3) Curve #3, M* ¼ 42; (4) Curve #4, M* ¼ 38; (5) Curve #5, M* ¼ 32;
(6) Curve #6, M* ¼ 23; (7) Curve #7, Mlines ¼9 and Marcs ¼ 3, R¼ 27; (8) Curve #8, Mlines ¼ 8 and Marcs ¼ 4, R ¼28; and (9) Curve# 9, Mlines ¼ 4 and Marcs ¼ 12, R ¼44.
Fig. 15. Examples of polygonal approximation with heuristic algorithms: (1) Curve #3, M ¼61 [12]; (2) Curve #4, M ¼ 46 [12];and (3) Curve #6, M ¼ 47 [36].
The proposed approach with Rate–Distortion curve modeling
has been used to solve a similar problem in a different area, i.e.
determining the number of clusters in multidimensional data
[44].
6. Conclusions
This paper considered the problem of optimal representation of
digital curves with a set of geometric primitives. The proposed
method is based on an analysis of the approximation error as a
function of the number of segments (description length). The
parameterized model of the Rate–Distortion curve was presented
for polygonal and multi-model approximation for error measures
L2 and L1. The model parameter is a characteristic of the curve
smoothness – a bigger value of the parameter means a smoother
curve. By basing our algorithm on the constructed model, a
multiplicative cost function was derived for determining the
number of segments. The minimum of this criterion gives a
solution with the best balance between the number of segments
and the approximation error.
The proposed algorithm demonstrated good results for contour
data when compared with heuristic multiplicative criteria tested
for unsupervised approximation. The known heuristic criteria can
be treated as particular cases of the presented criterion. The
proposed algorithm has potential for use in computer graphics,
CAD, digital cartography, spatial information systems, and shape
analysis, modeling and comparison.
Conflict of interest
None declared.
Acknowledgment
We would like to thank Fernando de Goes et al. for providing
experimental data for the 2D test point set Shark.
References
[1] R. Nygaard, A.K. Katsaggelos, Rate–Distortion optimal signal compression
using second order polynomial approximation, in: Proceedings of Interna-
tional Conference on Acoustics, Speech, and Signal Processing, ICASSP, 2001,
vol. 4, pp. 2617–2620.
[2] R. Mann, A.D. Jepson, T. El-Maraghi, Trajectory segmentation using dynamic
programming, in: Proceedings of International Conference on Pattern Recog-
nition, ICPR, 2002, vol. 1, pp. 331–334.
[3] L. Yin, Y. Yajie, L. Wenyin, Online segmentation of freehand stroke by dynamic
programming, in: Proceedings of International Conference on Document
Analysis and Recognition, ICDAR, 2005, vol. 1, pp. 197–201.
[4] E. Bodansky, A. Gribov, Approximation of a polyline with a sequence of
geometric primitives, in: Proceedings of International Conference on Image
Analysis and Recognition, ICIAR, LNCS, Springer, Heidelberg, 2006, vol. 4142,
pp. 468–478.
[5] N. Miller, R. Mann, Detecting hand-ball events in video sequences, in:
Proceedings of Canadian Conference on Computer and Robot Vision, CRV,
2008, pp. 139–146.
[6] F. Tortorella, R. Patraccone, M. Molinara, A dynamic programming approach
for segmenting digital planar curves into line segments and circular arcs, in:
Proceedings of IAPR International Conference on Pattern Recognition, ICPR,
2008, pp. 1–4.
[7] A. Kolesnikov, Segmentation and multi-model approximation of digital curves,
Pattern Recognition Letters 33 (9) (2012) 1171–1179.
[8] F. Arrebola, F. Sandoval, Corner detection and curve segmentation by multi-
resolution chain-code linking, Pattern Recognition 38 (2005) 1596–1614.
[9] D. Sarkar, A simple algorithm for detection of significant vertices for polygonal
approximation of chain-coded curves, Pattern Recognition Letters 14 (1993)
959–964.
[10] J.M. Iñesta, M. Buendia, M.A. Sarti, Reliable polygonal approximation of
imaged real objects through dominant point detection, Pattern Recognition
31 (1998) 685–697.
[11] W.-Y. Wu, An adaptive method for detecting dominant points, Pattern
Recognition 36 (2003) 2231–2237.
 A. Kolesnikov, T. Kauranne / Pattern Recognition 47 (2014) 623–633
[12] M. Marji, P. Siy, Polygonal representation of digital planar curves through
dominant point detection – a nonparametric algorithm, Pattern Recognition
37 (2004) 2113–2130.
[13] R. Dinesh, D.S Guru, Finite automata inspired model for dominant point
detection: a non-parameter approach, in: Proceedings International Confer-
ence on Computing: Theory and Application, ICTA, 2007, pp. 579–583.
[14] H. Imai, M. Iri, Polygonal approximations of a curve (formulations and
algorithms), in: G.T Toussaint (Ed.), Computational Morphology, North-
Holland, Amsterdam, 1988, pp. 71–86.
[15] D.H. Douglas, T.K. Peucker, Algorithm for the reduction of the number of
points required to represent a line or its caricature, The Canadian Cartogra-
pher 10 (2) (1973) 112–122.
[16] J. Sklansky, V Gonzalez, Fast polygonal approximation of digitized curves,
Pattern Recognition 12 (1980) 327–331.
[17] A. Held, K. Abe, C. Arcelli, Towards a hierarchical contour description via
dominant point detection, IEEE Transactions on Systems, Man, and Cyber-
netics 24 (1994) 942–949.
[18] A. Pikaz, I. Dinstein, An algorithm for polygonal approximation based on
iterative point elimination, Pattern Recognition Letters 16 (1995) 557–563.
[19] T.P. Nguyen, I. Debled-Rennesson, Parameter-free method for polygonal
representation of noisy curves, in: Proceedings of International Workshop
on Combinatorial Image Analysis, IWCIA, 2009, pp. 65–78.
[20] M.T. Parvez, S.A. Mahmoud, Polygonal approximation of digital planar curves
through adaptive optimizations, Pattern Recognition Letters 31 (2010)
1997–2005.
[21] G. Papakonstantinou, Optimal polygonal approximation of digital curves,
Signal processing 8 (1985) 131–135.
[22] J. Dunham, Optimum uniform piecewise linear approximation of planar
curves, IEEE Transactions on Pattern Analysis and Machine Intelligence 8 (1)
(1986) 67–75.
[23] H. Imai, M. Iri, Polygonal approximations of a curve (formulations and
algorithms), in: G.T Toussaint (Ed.), Computational Morphology, North-
Holland, Amsterdam, 1988, pp. 71–86.
[24] W.S. Chan, F. Chin, On approximation of polygonal curves with minimum
number of line segments or minimum error, Int. J, Computational Geometry
and Applications 6 (1996) 59–77.
[25] M. Shearer, J.J. Zou, Detection of dominant points based on noise suppression
and error minimization, in: Proceedings of International Conference on
Information, Technology and Applications, ICITA, 2005, vol. 1, pp.772–775.
[26] J.C. Perez, E. Vidal, Optimum polygonal approximation of digitized curves,
Pattern Recognition Letters 15 (1994) 743–750.
Alexander Kolesnikov received a M.Sc. degree in Physics and a Ph.D. degree in Computer Science. He used to be a Senior Researcher at the Research Institute in Novosibirsk
(Russia) and the University of Eastern Finland. His main research interests are signal and image processing and compression.
Tuomo Kauranne is Associate Professor in Applied Mathematics at Lappeenranta University of Technology. His research interests include algorithms for geospatial data
processing and their application in modeling the environment and climate change.
633
[27] C.-C. Tseng, C.-J. Juan, H.-C. Chang, J.-F Lin, An optimal line segment extraction
algorithm for online chinese character recognition using dynamic program-
ming, Pattern Recognition Letters 19 (1998) 953–961.
[28] M. Salotti, An efficient algorithm for the optimal polygonal approximation of
digitized curves, Pattern Recognition Letters 22 (2) (2001) 215–221.
[29] M. Salotti, Optimal polygonal approximation of digitized curves using the sum
of square deviations criterion, Pattern Recognition 35 (2002) 435–443.
[30] A. Kolesnikov, P. Fränti, Data reduction of large vector graphics, Pattern
Recognition 38 (3) (2005) 381–394.
[31] A. Kolesnikov, ISE-bounded polygonal approximation of digital curves, Pattern
Recognition Letters 33 (10) (2012) 1329–1337.
[32] A. Gribov, E. Bodansky, A new method of polyline approximation, in:
Proceedings of International Conference on Structural, Syntactic and Pattern
Recognition, LNCS, 2004, vol. 3138, pp. 504–511.
[33] P.L. Rosin, Techniques for assessing polygonal approximation of curves, IEEE
Transactions on Pattern Analysis and Machine Intelligence 19 (1997) 659–666.
[34] M. Marji, P. Siy, Polygonal representation of digital planar curves through
dominant point detection – a nonparametric algorithm, Pattern Recognition
37 (2004) 2113–2130.
[35] A. Carmona-Poyato, N.L. Fernandez-Garcıa, R. Medina-Carnicer, F.J. Madrid-
Cuevas, Dominant point detection: a new proposal, Image and Vision
Computing 23 (2005) 1226–1236.
[36] A. Carmona-Poyato, F.J. Madrid-Cuevas, R. Medina-Carnicer, R. Mundoz-
Salinas, A new measurement for assessing polygonal approximation of curves,
Pattern Recognition 44 (1) (2011) 45–54.
[37] B. Mandelbrot, How long is the coast of Britain? Statistical self-similarity and
fractional dimension, Science 156 (3775) (1967) 636–638.
[38] F. Normant, C. Tricot, Method for evaluating the fractal dimension of curves
using convex hulls, Physical Review A 43 (12) (1991) 6518–6524.
[39] K. Falconer, Fractal Geometry, Wiley, New York, 2003.
[40] F.M. Barnsley, Fractals Everywhere, second ed., Morgan Kaufmann, 1993.
[41] J. Jimenez, J.L. Navalon, Some experiments in image vectorisation, IBM Journal
of Research and Development 26 (1982) 724–734.
[42] E. Keogh, UCR Time Series Datasets at 〈http://www.cs.ucr.edu/  eamonn/
time_series_data/〉.
[43] F. de Goes, D. Cohen-Steiner, P. Alliez, M. Desbrun, An optimal transport
approach to robust reconstruction and simplification of 2D shapes, Computer
Graphics Forum 30 (5) (2011) 1593–1602.
[44] A. Kolesnikov, E. Trichina, Determining the number of clusters with Rate–
Distortion curve modeling, in: Proceedings of International Conference on
Image Analysis and Recognition, ICIAR, 2012, vol. 1, pp. 43–50.