Analysing Superimposed Oriented Patterns
Ingo Stuke, Til Aach Erhardt Barth, Cicero Mota
Institute for Signal Processing,
University of Liibeck,
Ratzeburger Allee 160,
D-23538 Liibeck, Germany
{ stuke,aach} @isip.uni-1uebeck.de
Abstract
E.stiniation af local orientation in inlages is often posed
as the task [ifjnding the niiiiimrmi iariance axis in a lo-
cal neighborhood. The .solution ir given as the eigenvector
belonging to the siimller eigenvalue afa 2 x 2 tensor: Ide-
ully, the tensar is rank-dejcient, i.e.. the sniuller eigenvulue
i,s rem. A large niininial eigenidue .si,qtial.s the presence of
niare than one locul orientation. We describe a frumenwrk
for estimating such superhiposed orientations. Oar aiiuly-
sis of superiniposed orientations is based on the eigensys-
tern anu/?sis qf a .suitublj extended tensor We show how
IO efficiently c a r 0 out the eigensystem analysis using ten-
sor inwriants. Unlike in the single orientation case, the
eijiensystem analysis does not directly yield the orienta-
tioris, rather; if provides .so-called niixed orientation parani-
eter.s. We tlierejbre show how to decanipuse the mixed ori-
en~ufion paraineter.~info the individual orientatians. There,
iri trim, ullow to separate the superinipsed patterns.
1. Introduction
Analysis of local orientation in an image is an essen-
tial step in, e.g., directional filtering [I, 21, directional in-
terpolation [81, feature extraction, tracking, motion estima-
tion [12, 11, 9, 51 and pattern analysis [7,10, 51. Multiple
orientations appear in non-opaque imagery like X-ray, ul-
trasound, and Computer Tomography and in image features
like comers, crossings and bifurcations. Here we present an
approach for estimating multiple orientations for additively
overlaid patterns. The solution is based on earlier results
for multiple transparent motions [13, 121. Orientation esti-
mation for two superimposed patterns is expressed as the
Work supponed by Deutsche F0ctchungs:emeinschaft under Ba
I I7617-2
0-7803-8387-7/04/$20.00 0 2004 IEEE
Institute for Neuro- and Bioinformatics,
University of Lubeck
Ratzeburger AlIee 160,
D-23538 Liibeck, Germany
{barth,mota} @inb.uni-luebeck.de
eigensystem analysis of a generalized structure tensor. Un-
like in the single-orientation case, the tensor concept now
leads to a nonlinear problem that is here linearized by intro-
ducing so-called mixed-orientation paranieters. The mixed-
orientation parameters form a unique descriptor of double-
orientation neighborhoods, but do not provide the orienta-
tions explicitly. Decomposing the mixed-orientation param-
eters into the individual orientations is solved by seeking
the intersections of a line and a circle. Furthermore, we de-
rive a hierarchical algorithm which allows to deal with sin-
gle and superimposed orientations.
2. Single orientation estimation
A two-dimensional image f(x),x = (z, y) is oriented
in a region R if
f(x) = f(x + U) (1)
+
V x , x U E R. The vector U = (cos@,sine)= describes
the orientation of f(x)in terms of the angle 0, which is con-
ventionally restricted to lie in the interval (--71/2,11/2]. In
order to estimate the local orientation, we follow the differ-
ential approach introduced in [6, 5 , 10, 43, which finds the
minimum gray level variance axis within the local neigh-
borhood R. Let
denote the derivative operator with respect to the vector
( ~ 0 ~ 4 , s i n @Unless
) ~ . f(x) is locally constant, it has an
unique orientation 0 at x if and only if its directional deriva-
tive in the direction of 0 vanishes, i.e.
4@)f(x)= cOs(e)f&) + sin(e)f,(x) (3)
= UTVf(X) = 0
133
In practice, local orientation is estimated within a local and
neighborhood R by minimizing the residual
df = (f=Z,fzy,fgy)T. (1 1)
2
€(U) =
J, (uTVf(x)) d R = U ~ J Is.1.
UuTu, =1
(4)
The vector c is the result of the concatenation of two direc-
tional derivative operators, and thus combines both orienta-
tions 8 and y.We will refer to c as the mixed-orientation
with respect to U. The constraint uTu = 1 excludes the triv- vector. Let F ( w ) denote the Fourier transform o f f (x)over
ial solution U = 0. The symmetric 2 x %-tensorJ I is Cl. We then obtain in the Fourier domain
o ( 8 ) ~ ( 7 ) f ( x=) 0 -4 ( u T w ) ( v T w ) F (=
~ )0, (12)
where w is the frequency vector. The above equation im-
plies that the local spectrum F ( w ) must be zero except for
Minimization of the composite criterion
the lines uT w = 0 and vTw = 0. The non-zero spectral
E(u) = u T J l u + X(1- U ~ U ) (6) components at uT w = 0 correspond to the Fourier trans-
form of gl (x)and those: at vT w = 0 to the transform of
with X being an arbitrary constant, is equivalent to g2(x). Consequently, for w = 0 the frequency components
of both layers are superimposed. Eq. (8) and Eq. (12) ensure
J ~ =uX u s.t. uTu = 1. (7) that Eq. (9) is a n e c e s s q and sufficientcondition for the ex-
The solution U is then the eigenvector of J, correspond- istence of a double orientation as long as Eq. (9) is evalu-
ing to its smallest eigenvalue X2. The minimum residual is ated within a region R. As in the single-orientation case, the
equal to this smallest eigenvalue, i.e. €(U) = X2. Eq. (4) residual error €2 of Eq. (‘2) within R is defined as
evidently assumes that only a single orientation is present
E ~ ( c )= c ~ J ~ c (13)
within a.Compliance with this assumption is indicated by
a low value of X2, and a high value for the larger eigen- where
value XI, Indeed, if the single-orientation hypothesis is ide-
ally met. the residual X2 is zero, and hence, rank(J1) = 1. J2 = l ( d f ) ( d f ) ’ d a
Violation of the single-orientation hypothesis is indicated
by a large residual X2. This in turn indicates that more than
one oriented structure is present in R. In the following, we
will show how multiple superimposed orientations can be
locally estimated.
[
=J, fZJY,
fA
fzzfz,
fzzfzy
fzy
fzyfyf,,
fz*fy,
fiyf,,
ftY 1
Minimizing €2 with respect to c under the constraint cTc =
dQ.
(14)

1, we obtain c as the eigenvector of the 3 x 3 tensor J2

3. Estimation of two overlaid orientations
We model a double-oriented pattern by f (x)= g1 (x)
g2(x). Since the two components g1 and g? are assumed to
be ideally oriented in the directions 8 and y. respectively,
they satisfy the following equations
T T
n.(O)g1 = U Vg1 = 0 a ( y ) g 2 = v vg2 = 0
where U = (cos 8, sin and v = (cos y,sin T ) ~ There-
.
fore, the composite image f(x) satisfies (for multiple mo-
tions the analog equations have been derived in [13])
a(@(r)f =
a a a factor R, i.e.
( c o s ( 6 ’ ) ~+sin(O)-)(cos(y)- ax
ay
= cTdf = 0
where
c = (cos(8)cos(y),sin(8 + y ) , s i n ( 8 ) ~ i n ( y ) ) ~(IO)
that corresponds to the smallest eigenvalue AB. The mini-
mum residual error is then equal to AB. Confidence for the
+ double-orientation hypothesis is high if XB is small and the
other eigenvalues are large. The resulting mixed-orientation
vector c is an unambiguous descriptor of double orienta-
tion neighborhoods, which could be used as, e.g., tracking
feature. It does, however, not yet explicitly provide the ori-
(8) entations.
3.1. Decomposing the mixed-orientationvector
The eigenvector analysis of the generalized tensor Jz in
Eq. (13) defines the vector c only up to an unknown scaling
c = R (cos(8) cos(y), tin(0 + T), sin(0) (15)
= ( U , b, c ) ~ .
(9)
Though R is unknown, it complies with the constraint
cTc = 1. i.e., U’ + +
b2 c2 = 1.We first observe that
a ~ c = R c o s ( 8 f y ) ,b = R s i n ( O + y ) (16)

134
 +
These expressions cannot be inverted to find 0 y and 0 -
y because, even though 0 and y are restricted to lie within
(-7r/2,7r/2], we have - R < 0 f y 5 R . The inversion of
the cosine and sine functions is, however, ambiguous in this
interval.
An alternative solution, similar to solutions derived for
multiple motions, consists of seeking the mots of a poly-
nomial constructed from the components a , 6 , and c of the Figure 1. Geometry of the separation.
mixed-orientation vector, see [12,31. We use
b
-
a
-
- sin(0 7) = tan(@) tan(r),
+

cos(0) cos(y)
+ (17) +
where m, n, and T are normalized to satisfy m2 n2 = 1.
From b.(21) and the additional constraint cTc = 1 (see
C sin(0) sin(y)
- = = t a n ( 0 ) . tan(?) above) we can draw the following two conclusions: (i)
a cos(8) cos(y) 1 - 2ac is always positive and (ii) the points P and Q are
as coefficients to construct the polynomial given as the intersections of the line defined by Eq.(21) and
the unit circle. Note that the vector el = (m, n) is orthogo-
b
P ( z ) = z2 - - z
a
+ -cn = ( z - tan0) ( z - tan?) (18)
nal and e 2 = (-n, m ) is parallel to the line. Since T is the
distance between the line and the origin, the points P and
Q are found by the linear combination
We could then determine tan(0) and tan(?) as roots of
P ( ; ) , The angles could therefore be unambiguously ob-
tained within (-7r/2, ~ / 2 ] .We would, however, have to P,Q=Tel+sez (23)
deal with the singularity a1 a = 0 in cases where 0 = ~ / 2
or 7 = a/2. where s = m. Figure 1 depicts the geometric inter-
pretation of the separation procedure. Finally, P and Q de-
Starting from the above result, we now show a solu-
termine the orientations.
tion that avoids these difficulties. We propose a decompo-
sition method that determines the auxiliary vectors P =
(cos(28),sin(28)) and Q = (cos(27).sin(27)) as the in-
4. Confidence measures
tersections of a line with the unit circle. To do so, we first
replace the variable z of P ( z )by one of the roots of P ( z ) ,
As we have seen, orientation estimation is ideally only
here denoted by y/x where the ratio is either defined by
possible if the suucture tensors J1or J2have a single zero
x = cos0 and y = sin0 in case of P or x = cosy and
eigenvalue. As shown in [ 121, the above confidence crite-
I = sin 7 for Q. Eq. (18) then becomes equivalent to
J
rion can be expressed in terms of the tensor-invariants with-
cx2 - bxy + ay2 = 0. (1%
out explicit evaluation of the eigenvalues. Such a confidence
criterion is especially useful for 2 x 2 and 3 x 3 matrices be-
Now we observe that cause its evaluation based on the minors of the tensor is sim-
pler and faster. The tensor J1 has the invariants:
1 1
22 : cos20 = E + xcos2e, (20)
K1 = d e t J l = X I X1
= (f,”,(f;) - ( f Z f d 2 (24)
ry = sin0cose =
I .
- sin20 HI = traceJ1 = XI + X2 = (f,“)+ (fi),
2
where (f,“) = Jnf2dQ. and so on. Clearly, H1 = 0 if
Note that analog equations are valid for y. Therefore, the
and only if AI = A2 = 0. This can only occur in homoge-
two points P = (cos 20, sin 20) and Q = (cos 27. sin 2 7 )
neous regions where no orientations can be estimated. In
lie on the line
case of single orientations the confidence criterion trans-
mX +nY -T = 0, (21) lates to HI S 0 and K I = 0. As shown in [12], these in-
variants obey K 1 5 H:. Equality holds for isotropic re-
where X and Y are the first and second components of P gions only, i.e., HI = *‘
A = 0. Thus the confidence cri-
and Q, respectively. The parameters m , n, T are terion becomes K1 5 EIH: with a confidence parameter
€1.
In regions where the single-orientation model is rejected
by the above confidence criterion, we test for the double-

135
orientation model. For tensor J2 we use the invariants
IC, = det(J2) = X1X2X3 = (f&)Af33
- (frzfry)Af32 (fzzfyy)nf31+
SZ = (xlX2 + xlX3 + x2x3)
= All1 + A122 + A l 3 3 ,
where AI,, are the minors obtained from Jz by eliminat-
ing row 4 - i and column 4 - j. In analogy to the single-
orientation case, we obtain the confidence criterion K;
e?$. As shown in [12], for any matrix having a single zero
eigenvalue the eigenvector corresponding to the zero eigen-
value can be expressed in terms of the minors. For J1 this
eigenvector is
For J2 there are three possibilities to compute the eigenvec-
tor corresponding to the zero eigenvalue:
c; = R , ( ~ I ~ ~ , - . ~ I , ~ =
: A( ~I ., ,I, )b~~ , e ; i) ~=. 1 , 2 : 3 (27)
where the scaling factors R, are not needed in our analysis,
hut can be detennined by (a, ci)' + +
b: = R? according
toEq.(16).
5. Layer separation
For some applications it is useful to separate the super-
imposed pattems using the double-orientation information.
Let SI(.) and gZ(x) be oriented in directions O and y. re-
spectively. The directional derivative a(O)f(x)nulls out
g l ( x ) (see Eq. (S)), and yields in the Fourier domain
n(O)f(x) = a(O)g2(x)o-.uTuG2(u) =: Gz(w). (28)
The pattern g2(x) is thus recovered up to a convolution. At
frequencies where uT U # 0 we obtain Cz(w) by
The frequencies where uT U = 0 correspond to the support
of Gl(w) and therefore
G2(wT U ) = 0.
Since both lines (determined by U and v respectively) pass
through the origin, it is impossible to separate the frequency
components at the origin. Accordingly, the mean values of
the patterns cannot he recovered.
(25)
<
Figure 2. Two overlaid patterns.
6. Results
Figure 2 shows two sinusoidal patterns that are superim-
posed: one is vertically 'oriented and the other is a section
of a circle. We added Gaussian white noise with an SNR
of 25dB. The estimated orientation vectors are depicted in
the right image (15-fold sub-sampling). For each pattern
we performed an error analysis by comparing the estimates
with the known ground truth. For the vertical oriented pat-
tern the mean squared angular error is 0.28 degrees and its
standard deviation 0.29 degrees. For the circular pattern, the
errors are estimated relative to the tangent, and we observed
an MSE of 1.48 and a STD of 0.94 degrees. The parame-
ters €1 and €2 are set to 0. I and 0.4 respectively, and the size
of the integration area Cl was 7 x 7 pixels. The top left im-
age in Figure 3 depicts a fabric, which was taken from the
MIT-VisTex database. The estimated orientations are shown
top right for every tenth pixel. Even though the image does
not satisfy the superposition assumption, the estimated ori-
entations correspond well to the orientations in the texture.
The parameters € 1 and i : are ~ set to 0.2 and 0.6, respec-
tively, and the size of the integration area is 21 x 21 pix-
els. The second example of figure 3 shows a small part of
an X-ray image with a blood vessel crossing. In this case the
superposition is approximately additive. The estimated ori-
entations are depicted on the right image. Note that in the
overlapping region the two orientations are quite accurately
detected. For the non-overlaid vessel parts the single ori-
entation is also accurately estimated. We used a integration
filter size of 9 x 9 pixels and the parmeters = 0.6 and
cZ = 0.6. In all three examples we used a Gaussian pre-filter
with size 3 x 3 pixels and standard deviation of one pixel.
(30) Derivatives were computed by finite differences. Figure 4
shows an example for layer separation. The first row de-
picts the two original layers (top left and middle) and their
overlay (top right). In the second row we show the sepa-
rated layers. We were able to recover the layers except for
their mean values.
136

Figure 3. Results for fabric with two opaque
orientations (top) and an X-rayed blood ves-
sel crossing (bottom).
Figure 4. Layer Separation
7. Conclusions
We presented methods for the detection, estimation, and
separation of two locally overlaid orientations in images.
A key element is a generalized structure tensor that is de-
fined for any number of orientations. We have used this ten-
sor to first detect the number of orientations based on con-
fidence measures that do not involve an eigenvalue analy-
sis. Instead, these measures can be fast and efficiently cal-
culated by the invariants of the generalized structure ten-
sor, which in turn are calculated in terms of the tensor
137
minors. The overlaid orientations are unambiguously de-
scribed by a mixed-orientation vector, which is given as
the tensor's eigenvector corresponding to the smallest (ide-
ally zero) eigenvalue. As for the confidence measures, we
have shown that the mixed-orientation vector can be calcu-
lated from the tensor minors without an explicit eigensys-
tem analysis. It is a powerful feature of local image con-
tent, and uniquely decomposable into the individual orien-
tations. Finally, we have used the orientation parameters to
separate the overlaid image patterns. Our results, for real
and synthetic images, illustrate the potential of our frame-
work. The orientation parameters are quite accurately re-
covered under noisy conditions and a wide range of relative
orientations. Test patterns were well separated into the orig-
inal layers. Further examples illustrate the benefit for tex-
ture analysis and vessel segmentation.
References
[ I I T. Aach, D. Kunz. Anisotropic spectral magnitude estima-
tion filtersfor noise reduction and image enhancement. Pmc.
IEEE h a Con$ bn. Proc.:335-8, Lausanne, Sept. 1619,
1996.
121 T. Aach, D. Kunz. A lapped directional transform for spec-
tral image analysis and its application to restoration and en-
hancement. Signal Processing, 802347-64, 2000.
131 E. Barth, 1. Stuke, C. Mota. Analysis of motion and cun'a-
ture in image sequences. Pmc. IEEE Sourhwesr S y n p . I n t .
Anal. lnrerp.:20610, Santa Fe, NM, Apr. 7-9, 2002.
141 I. Bigiin, G. H. Granlund. Optimal orientation detection of
linear symmetry. lEEE Isr Irirl. Cont: Conzp. Ws.:433-8,
London; June 1987.
..
151 J. Biciin.
e . G. H. Granlund. I. Wiklund. Multidimensional ori-
entation estimation with.application to texture analysis and
optical flow. lEEE T P A M I , 13(8):775-90, 1991.
161 S. Di Zenzo. A note on the gradient ofa multi-image. Comp.
K.s.Graph. Im.Pmr.,33:116-25, Jan. 1986.
171 W. T.Freeman, E. H. Adelson. The design and use of steer-
ablefilters. IEEE TPAMI, 13(9):891-906, 1991.
181 J. Hladuvka, E. Griiller. Direction-driven shape-based inter-
polation of volume data. T. Ertl et.al. (eds.), Pmc. vis. Model.
visualiz.:l13-20, Stuttgart, Nov. 21-23,2001. Aka GmbH.
191 B. Jahne. Image sequence analysis in environmental and life
sciences. B. Michaelis, G. Krell (eds), Parr. Recog., 25rh
DAGM Sw?ip., kcr. N o r ~ sComp. Scie., vol. 2781:608-17,
Magdeburg, Sept. 10-12.2003. Springer.
1101 M. Kass. W. Witkin. Analyzing oriented patterns. Camp,
vis. Graph. Im. Pmc., 37:362-85, 1987.
1 I I I R. Mester. Some steps towards a unified motion estimation
procedure. In Pmc. 45rh IEEE MidWesr Synip. Circ. Sysr.,
Tulsa. OK, Aug. 47.2002.
1121 C. Mota, 1. Stuke, E. Barth. Analytic solutions for multiple
motions. In Pnic. IEEE In:. ConJ Im. Proc., vol. 11: 917-20.
Thessaloniki. Oct. 7-10, 2001.
I131 M. Shizawa, K. Mase. Simultaneous multiple optical flow es-
timation. In IEEE Con$ Comp. vis. Parr. Recog.. vol. 1:274-
8, Atlantic City, NJ, June 1990.