Boundary Segmentation

Detection of Corner, Inflection and Transition Points

Kazuhide SUGIMOTO Fumiaki TOMITA

Real World Computing Partnership Electrotechnical Laboratory

Tsukuba Mitsui Bldg., 1-6-1 Takezono, 1-1-4 Umezono,
Tsukuba, Ibaraki 305, Japan Tsukuba, Ibaraki 305, Japan
[email protected] t omit [email protected]

Abstract where the change of the curvature is small. Even

For future intelligent man-machine system with vi- if possible, they need delicate threshold value for
sion, it is necessary to visualize the results of shape segmentation[l2]. Methods for polygonal approxima-
and motion analysis of observed objects in the images. tion are not good enough for analysis and interpre-
A s for object recognition, there are at least three steps. tation of planar curved figures[l3][14]. In this paper,
The first is to detect edges which correspond to the we propose a method for segmentation of boundaries
boundaries of objects (image segmentation). The sec- which does not need delicate threshold value for seg-
ond is to segment each boundary into simple line or mentation by adapting it to the details of a feature
curve segments (feature extraction). The third is to automatically. Furthermore, this method can detect
match those features between the data and the model. all kinds of feature points listed above and calculate
This paper presents a new method for the second step: the curvature and the normal vector at each point on
boundary segmentation. It can detect not only corners the boundary with good accuracy.
but inflection points on which the sign of the curvature
changes and transitional points on which a line and a 2 Feature points for segmentation
curve connect smoothly without any delicate threshold. Since line sections on the boundary are key fea-
It also calculates the curvature and the normal vector ture for shape recognition and description, it is im-
at each point on the boundary with good accuracy. The portant to detect transitional points with good accu-
features extracted by the proposed method are useful for racy. As for curve sections it is difficult to determine
both machine vision and visualization. the best suited section for quadratic curve approxima-
tion. If the error value for curve approximation is used
for segmentation, a boundary is often segmented into
1 Introduction too many short sections at the point even where the
Segmentation of boundaries in image is one of the sign of the curvature does not change and curvature
crucial steps in shape recognition and description. The changes smoothly. This over segmentation should be
goal of a boundary segmentation algorithm is to par- avoided because it increases the number of segments
tition a boundary of the regions into the line or curve on a boundary and also becomes the cause of the diffi-
edge segments. These segments can be used as units culty in finding key features. Since it is easy to detect
for correspondence in edge based stereo[l] and units end points and branching points using the connectivity
for feature matching process between known physi- number, the other feature points of which the connec-
cal properties and observed object features[2]. Fur- tivity number is two should be talk about and they
thermore, line and curve segments are good unit for are defined as follows.
qualitative and hierarchical definition and symbolic
description and classification of objects[3]. However, corner point: a point at which the direction of the
many of the popular techniques for boundary seg- normal vector changes abruptly
mentation make use of local k-curvature[4] which is
sensitive to noise, so they need smoothing process inflection point: a point at which the sign of the
using such as Gaussian filtering which is commonly curvature changes from positive(negative) to neg-
used also t o describe the hierarchical structure of ative(positive).
images[5][6][7][8],and zero-crossin of this smoothed
curvature to detect only corners[9]~10][11].Almost of transition point: a point at which the curvature
them are impossible to detect various feature points changes from zero to non-zero (a point at which
on boundaries, such as inflection points on which two curves that have the same convexity prop-
the sign of the curvature changes and transitional erty and different curvature connect smoothly (so
points on which line and curve connect smoothly, called smooth join) is not treated as transition
since zero-crossing cannot be applied t o the points point.)

13
0-8186-5875494$3.00Q 1994 IEEE
3 Calculation of curvature and normal e Calculate Ej to every neighbor point Pj(i - IC 5
vector +
j 5 i I C ) which is used for straight line approxi-
mation.
3.1 Directional boundaries
There are two types of directional boundaries which 0 Let Le; be the maximum error value of Ej (i - IC 5
surround the region on the right, one is the peripheral j5 i + IC).
boundary the other is the hole boundary as shown in
figure 1. After this, the order of the edge points on How to calculate Cei is as follows.
the boundary for tracking depends on this direction.
Let 0 be the center of the approximated circle
CP..
Let Ij be the intersection of Cpi and the line
of region 2 which passes through both points Pj and 0.

U region 1

Figure 1: Directional boundaries

Let E , be the distance between Pj and I j .
Calculate Ej to every neighbor point Pj(i - IC 5
j 5 i + IC) which is used for the circle approxima-
3.2 Approximation of circle and straight tion.
line
First, to every point, both straight line I and circle Let Cei be the maximum error value of Ej (i - IC 5
Cp, are approximated using a set of neigtbor points j5 i + IC).
within the distance K from the noticed point Pi by the
least squares method to calculate the local temporal Let E, be the smaller error value Lei or Cei at the
curvature and the temporal normal vector as shown in noticed point Pi.
figure 2. Here, K is the threshold value. 3.4 Temporal curvature and normal vec-
tor
Next, the local temporal curvature and the tempo-
approximated line using p i neighbors ral normal vector are calculated at every point Pi. If
1Pi : a x + b y + c = O Lei 5 C e ; , then the curvature is zero and the direc-
I tion of the normal vector differs $(radian) from the
one of the boundary and outward from the region. If
Boundary C e , < Le,, then the curvature is $. Here, T is the
radius of the circle C p i . The direction of the normal
vector is equal to the one of the vector from 0 to Pi.
The sign of the curvature depends on which side of
error region 0 is. If 0 is inside of the region then the sign
El of curvature is plus (plus means convex), if outside,
minus (concave). These processes are applied to every
point Pi on the same boundary and calculate the local
l/r:curvature temporal curvatures and the temporal normal vectors.
3.5 Renewal of curvature and normal vec-
tor
However, E, becomes big near the corner and the
noisy point. Let the line and the circle which are
\
approximated circle using pi neighbors
approximated to the noticed point P, using a set of
neighbor points be 1 and c respectively as shown in
figure 3(a). Both of the temporal normal vector and
c p i : x2+ y2+ ax + by + c = 0 the temporal curvature sometimes can be unexpected
value in this case, whether the line is selected or not.
Figure 2: Approximation of circle and straight To solve this problem, the lines and the circles on the
line, and calculation of curvature and normal backward point Pb and forward point P f which are the
vector distance of K away from the noticed point P, in the
same figure are used for comparison. Then select the
3.3 Error calculation line or the circle which has the minimum error value
Then, calculate the error value Lei for line approxi- Emin from lines or circles at Pb and P f t o calculate
mation and C e ; for circle approximation at each point the accurate curvature and the normal vector, and let
Pi. How to calculate Lei is as follows. Emin be the error value E , at the noticed point P,.
Furthermore, when the point Pb is selected, Emin is
e Let Ej be the distance between Pj and the foot of compared with E, of all points between P, and Pa. At
the perpendicular dropped from point Pj to l p , . every point where Ei is bigger than Emin, the direction

14
of the normal vector and the curvature is calculated curvature as nearly equal to zero at the points where
again by using the line or circle equation on Pa. When the change of the curvature is small (see position 150
the point Pf is selected, this renewal process should t o 500 in the same figure). And zero-crossing of this
be applied to every point between P, and Pf in the
same way.
normal vector using circle equation
-
ec
smoothed curvature can be used only for detection of
the corner points.
Figure 5 shows the curvatures on the boundary in
figure 7(a) by our new method, temporal (with thick
normal vector using line equation
\ - 1
e1 line) and after renewal (with dotted line).
Figure 6 shows the direction of the normal vectors
on the boundary in the same figure, before (with thick
line) and after renewal (with dotted line).

approximated circle C using

Pc neighbors
(a):Error near the corner point

0'
(b):Error near the noisy point Dlrretlon of the Normnl
Ask
Figure 3: Renewal of curvature and normal vec- r
c..mol-
tor 3w

250

1M

130

1.w
030

OW

-050

I.M

-110

-2.M

-
.2so

-9M
I I I I I
OM 2" 4" Mow b"

Figure 6: Renewal of normal vector

4 Segmentation point detection

4.1 Corner point
Corner points are detected as the point where
Figure 4: k-curvature the direction of the normal vector changes abruptly.
When the difference is bigger than the threshold, the
Figure 4 shows the usual local k-curvatures on the point can be detected as the corner point.
boundary in figure 7(a), original (with dotted line) and 4.2 Transition point
smoothed curvature by Gaussian filtering ( U = 10) Candidates of transition points are detected at the
1with thick line). There are some problems that the
ormer is sensitive to noise and the latter calculate
point where the curvature changes from zero to non-
zero. However these points can be observed near the

15
noisy points. So it is necessary to select only the Figure 7 shows the result of segmentation of bound-
point where the line section and the curve section aries. 0 , A and x denote the corner point, the in-
smoothly join from these candidates. If both sections flection point and the transition point respectively.
have enough length (longer than the threshold K ) and 4.4 Noisy point
the point is not the corner point, it can be detected as
the transition point. Points where the error value for line or circle ap-
proximation becomes big should be treated as noisy
4.3 Inflection point point. However, when the default value K is longer
than the length of the line or curve section that should
Inflection points are detected at the point where be segmented, error value for approximation becomes
sign of the curvature changes from positive(negative) big also at each point in such section and desirable
t o negative(positive). The point where two curve sec- result cannot be obtained.
tions, one is convex and the other is concave, smoothly
join can be detected as the inflection point. However, 5 Adaptation of threshold value K
there can be the case that there exists short line sec- The bigger the threshold value K becomes, the ro-
tion which is shorter than the threshold K between buster to noise the calculation accuracy of the cur-
these two curves. In such case, the midpoint of this vature and the normal vector is. So the initial value
short line section becomes the inflection point. of K should be set as big as possible to the length of
the section that should be segmented. However, big K
cannot be applied to the short section. When K is big,
poits in short section are treated as noisy point. It is
necessary only for such points to change the thresh-
old value K smaller (ex. half size) and calculate the
curvature and the normal vector again using renewed
K to detect the feature points. While there is noisy
point and K is bigger than the minimum length of
the section that should be segmented, this process is
repeated.
Figure 8 shows the result of segmentation of bound-
aries (the size of image is 512x512). U , A, x and 0
denote corner point, inflection point, transition point
and noisy point respectively. Figure 8(a) shows the
result of detection of these feature points using 20 as
the threshold value K without adaptaion. Figure 8(b)
shows the one with adaptation of K using 20 as the
initial value and changed to 5 (i of the initial value)
automatically so that the hopeful result is acquired.
6 Conclusions
In this paper, we presented a method that can de-
tect the various feature points (corner, inflection and
transition) and calculate the curvature and the normal
vector at each point on the boundary with good accu-
racy, and does not require the delicate threshold values
for segmentation and interactive renewal of them. It
takes only five threshold values, the first one is for the
difference of the angle at the corner point(defau1t value
is fixed to 0.3 radian), the second one is the thresh-
old of radius for the linearity(defau1t value is fixed to
256), the third one is the error value for approxima-
tion (default value is fixed to 1 pixel), the fourth one
is the minimum length of the section that should be
segmented and the rest is adaptive K (the initial value
is fixed to 20). These threshold values except K need
no delicate adjastment for input images. In this way,
adaptive boundary segmentation which needs no del-
icate threshold value for segmentation is realized.
This method is very important for many kinds of
higher-level post-processing and may be useful to vi-
(b) sualize many characteristics (types of feature points,
0:Corner point, A:Inflection point, x:Traneition point curvature and direction of normal at each point, etc)
and can be the one connection between machine vision
Figure 7: Result of segmentation and visualization.

16
References
[l] G.G.Medioni , R.Nevatia, “Segment-based stereo
matching,” PTOC.of Image Understanding Work-
shop, pp. 128-136, 1983.
[2] R.B.Fishler, “Model invocation for three dimen-
sional scene Understanding,” PTOC.of lUth In-
ternational Joint Conference on Artificial Intelli-
gence, pp. 805-807, 1987.
[3] F.Tomita , M.Koizumi, “A step toward generic
object recognition,” Proc. of 1 l t h International
Conference on Pattern Recognition, pp. 632-636,
1992.
[4] Rosenfe1d.A. and A.C.KAK, Digital picture pro-
cessing, New York : Academic Press, 1976.
[5] H.Asada M.Brady, “The Curvature Primal
Sketch,” +AMI-8, NO.l pp. 2-13, 1986.
[6] A.P.Witkin, “Scale space filtering,” Proc. of 8th
International Joint Conference on Artificial In-
telligence, pp. 1019-1022, 1983. (a) K=20
[7] A.L.Yuille , T.Poggio, “Scaling theorems for zero
crossings,” PAMI-8, NO.l pp. 15-25, 1986.
[8] J.Babaud , A.P.Witkin , M.Baudin
and R.O.Duda, “Uniqueness of the Gaussian ker-
nel for scale space filtering,” PAMI-8, NO.l pp.
26-33, 1986.
[9] J.Q.Fang , T.S.Huang, “A corner finding algo-
rithm for image analysis and registration,” Proc.
of AAAI Conf., pp. 46-49, 1982.
[lo] 0.A.Zuniga ,R.Haralick, “Corner detection using
the facet model,” Proc. of IEEE CVPR., pp. 30-
37, 1983.
[ll] R.Mehrotra , S.Nichani, “Corner Detection,”
Pattern Recognition, Vo1.23, No.11 pp. 1223-1233,
1990.
[12] C.H.Teh , R.T.Chin, “On the detection of dom-
inant points on digital curves,” PAMI-11, N 0 . 8
pp. 859-872, 1990.
[13] T.Pavlidis, “Algorithms for shape analysis and
waveforms,” PAMI-2, N 0 . 7 pp. 301-312, 1980. (b) K from 20 to 5
0:Corner point, &Inflection point, x:Transition point,
[14] J.L.Dunham, “Optimum uniform piecewise linear @:Noisy point
approximation of planar curves,” PAMI-8, NO.l
pp. 66-75, 1986. Figure 8: Adaptation of K

17