Professional Documents
Culture Documents
2, MARCHIAPRIL 1993
-
, “Application of describing functions in the transient response
analysis of a three-term controller,” IEEE Trans. Syst.. Man, Cybern., 1) fF(S) 5 PF(2).
pp. 603-606 this issue.
A. Gelb and W. E. Van der Velde, Multiple-Input Describing Functions 2) fF(S) 5 P,(X)
Vague Sets
Wen-Lung Gau and Daniel J. Buehrer
11. VAGUESETS
on whether s does or does not belong to I’, our knowledge about s [0.2,0.8] can be interpreted as “the vote for a resolution is 2 in
is very exact and the theory reverts back to that of ordinary sets (Le., favor, 2 against, and 6 abstentions.”
sets with two-valued characteristic functions). The number in favor for [0.4,0.7] is greater than for [0.2,0.7], but
A vague set is in R1 is illustrated in Fig. 1. the linear ordering gives a “truth” to [0.4,0.7] which is less than
When X is continuous, a vague set 1- can be written as the “truth” of [0.2,0.8]. Hence, [0.4,0.7] and [0.2,0.8] are not really
comparable, and this definition is not true in the real world. For this
reason, some restrictions must be added, making the containment
relation only a partial ordering. Therefore, Definition 4 is used. Now,
When X is discrete, a vague set I’ can be written as a comparable example using Definition 4 is as follows: [0.4,0.7] 5
[ O S , 0.81, since
n
[0.4,0.7] can be interpreted as “the vote for a resolution is 4 in
I*= C
t=1
[t\ (st), 1 - fx ( s t ) ] / s L . .r, E -Y. (4) favor, 3 against, and 3 abstentions.”
[0.5,0.8] can be interpreted as “the vote for a resolution is 5 in
For example, assume that X = [I,2 , . . ..lo]. SMALL is a vague favor, 2 against, and 3 abstentions.”
Furthermore, the number in favor for [0.5,0.8] is greater than
set of X defined by
for [0.4,0.7] and the number opposed for [0.5,0.8] is less than
small = [l,1]/1 + [0.9,1]/2 + [0.6.0.8]/3 for [0.4,0.7]. Hence [0.4,0.7] and [ O S , 0.81 are comparable and
Definition 4 is true in the real world.
+
[0.3,0.5]/4 + [0.1.0.2]/5. Definition 5: Two vague sets A and B are equal, written as
A = B , if and only if rl C_ B , and B A; that is
We start with several definitions involving vague sets that are
obvious extensions of the corresponding definitions for ordinary sets t 4 = f B (9)
and fuzzy sets. 1-fA =1-fR. (10)
Definition 2: A vague set is empty if and only it its truth-
membership and false-membership functions are identically zero Definition 6-Union: The union of two vague sets A and B with
on X . respective truth-membership and false-membership functions t A , f4,
Definition 3: The complement of a vague set A is denoted by A’ t g and fB is a vague set C , written as C = A U B , whose truth-
and is defined by membership and false-membership functions are related to those of
A and B by
tA’(s) =f4(s) (5)
1-f4t(s)=l-f4(.r). (6) tc = m a x ( t 4 , t ~ ) (11)
1 - fr = max( 1 - f 1.1 - f ~ =) 1 - nlin(fA, fs). (12)
r ) 1 - f 4 / ( s=)
In the sequel, instead of writing t A i ( s )= f ~ ( and
1 - t 4 ( s ) for all s in s,we shall write more simply f 4 , = f A , A more intuitively appealing way of defining the union is the
1-fA‘ = 1-f4. following: The union of two vague sets A and B is the smallest
612 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 23, NO. 2, MARCHIAPRIL 1993
fdx), lfdx) Convex vague set Non-convex vague set t4(1.2), o f = 1 - f 4 ( 1 1 j 5 1 - f 4 ( S 2 j , then I.:, E re and
A Xa.1+ ( 1 - X j.r2 E r, by the convexity of r,.
1$4&1+Cl-Xzr2, Hence
1 -
t4(X.rl + (1 - X j s z ) 2 at = t I ( s 1 ) = niin(fA(.rl).t4(s2)j
1-fA(~.?1+(1-XjSZj >of=1-f4(Xl)
= niin(1 - f A ( S l ) , 1 - f 4 ( 2 2 ) ) .
A* Conversely, if =I is convex in the sense of the second definition
and c ) ~= t l(sl),o f = 1 - fa(sl),then may be regarded as
12
11
the set of all points for which t A ( s z ) 2 t 4 ( . r l j r 1 - f~(sz)2
Fig 3 C o n v e x and non-convex v a g u e sets in E' 1 - f 4 ( . r 1 ) . In virtue of (16) and (17), every point of the form
Asl + (1- Xj.rp, 0 5 X 5 1, is also in r, and hence rnis a
convex set. Q.E.D.
Property 6-Identity: -4U0 = .-I. -4n-Y = -4, where 0 = [O,O],
Theorem 5: If A and B are convex, so is their intersection.
and I = [l.11
Proof: Let C = A n B , then
Proof:
1) t4"R = Inax(t.\.tR)
1- f 4 U ~= niax(1 - f z . 1 - f B j = 1 - niin(f4.fRj
u ti ZUB)t = niin(f 1. f~ j = t l l n B /
1 - f( l U B ) i = 1 - mas(f 4 . t B j = inin( 1 - t 1 . 1 - t R j =
f 1'"B'.
2) f i n R = niiii(t I . f R j
1 - f m B = niiii(1 - f z . 1 - f B j = 1 - n i a s ( f 4 . f B j
u ti InB)' = n l a ~ ( f ,f. R ) = t I'"B'
l - f ( I n ~ )=l 1 - n i i i i ( t 1 . t ~ )= n i a x ( l - t q . 1 - t g j =
f I'UB'. Q.E.D.
Property 9-~nvolution: ( ~ ' j ' = .-I
Proof: t = f i . 1 - f 4 f = 1 - t I r u t( = t 1.1-
f ( 4 y = 1- f l . Q.E.D.
Definition 9StrongZy Convex: A vague set A is strongly convex
if for any two distinct points s1 and s 2 , and any X in the open
IV. CONVEXITY
interval (0,l)
We assume for concreteness that .Y is a real Euclidean space E".
Definition 8: Convexity: A vague set -4 is convex if and only if + (1 - X).rzj > r n i n ( t ~ ( r l j . t s ( z 2 ) ) (18)
t / L ( h 1
for all and , r ~in X and all X in [0, I]. This is illustrated in Fig. 3 from ordinary sets, fuzzy sets, and interval-valued sets. The notions of
for 7) = 1. inclusion, union, intersection, complement, relation, and composition
To show the equivalence between the above definitions, note that if have been extended to vague sets. Finally, convex vague sets and
-4 is convex in the sense of the first definition, and nf = f 4 ( s1) 5 strongly convex sets have been introduced.
614 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 23, NO. 2, MARCHIAPRIL 1993
The major advantage of vague sets over fuzzy sets is that vague I. INTRODUCTION
sets separate the positive and negative evidence for membership of Few problems in computational vision have drawn more attention
an element in the set. We not only have an estimate of how likely it than the representation of surfaces. A desirable property of any
is that an element is in the set, but we also have an lower and upper surface representation is that it be invariant with respect to rigid trans-
bound on this likelihood. This lower/upper bound can be used to formations of the surface in 3-space, otherwise one cannot guarantee
perform constraint propagation. For instance, constraint propagation the integrity of the representation under changes in viewpoint. This
can be used to detect inconsistencies in assignments of intervals to paper considers the computation of invariant surface representations
Boolean expressions involving the sets. given 3-D depth information from single and multiple views. Our
We are currently using constraint propagation to generalize the approach makes use of an elastically deformable model that offers the
methods of first-order automatic theorem proving so as to find a geometric flexibility to satisfy a plethora of multimodal and multiview
fixed-point minimal-sized possibility interval for each predicate in constraints. We specialize this model to assume an initial cylindrical
the input clauses, or to find inconsistencies, if they exist [7]. shape as in Terzopoulos et al. [21]. We dub the process of surface
fitting with elastically deformable surface as dynamic surface fitting.
REFERENCES In dynamic surface fitting, all the data are transformed into externally
L. A. Zadeh, “Fuzzy sets,” Inform. Control, vol. 8, pp. 338-353, 1965. applied forces. Constraints from single as well as multiple views are
D. Dubois and H. Prade, Fuzzy Sets andsystems: Theory andilpplications. applied as external forces to mold the elastic tube model into a shape
New York: Academic, 1980. consistent with the data.
G. J. Klir and T. A. Folger, Fuzzy Sets, Uncertainty, and Information. Surface reconstruction typically precedes the symbolic represen-
Englewood Cliffs, NJ: Prentice Hall, 1988.
I. Grattan-Guiness,“Fuzzy membership mapped onto intervals and many-
tation of surfaces. The reconstruction process integrates raw shape
valued quantities,” Zeitschr. Math. Logik. and Grundlagen d. Math., data from multiple visual sources. Using this information it generates
Bd. 22, S, pp. 149-160, 1975. a dense, coherent surface. Several variational reconstruction methods
C. Ponsard, “Hierarchie des places centrales et graphes +-flous,” Environ. have been proposed. An efficient noninvariant method was developed
Plann. A 9, pp. 1233-1252, 1977.
by Terzopoulos [19]. Blake [2] pointed out that the noninvariance
J. R. Quinlan, “Inferno: A cautious approach to uncertain inference,”
Computer J., vol. 26, no. 3, pp. 255-268, 1983. was inherent to the functional used in the variational principle
D. J. Buehrer, “Vague logic: A first-order extension of interval probability formulation, which quantifies the small deflection bending energy of
theory,” in Proc. IEEE Workshop Imprecise and Approximate Computa- a thin flexible plate. He proposed an invariant functional involving
tion, Dec. 1, 1992, Phoenix, AZ, pp. 83-87. the principal curvatures (minimum and maximum normal curvatures)
of the surface. This is known as the large deflection bending energy
of a thin plate.
A problem with Blake’s approach is that the large deflection energy
is a nonquadratic functional that, moreover, is nonconvex. The result
Intrinsic Parameters for Surface Representation is a difficult, nonlinear system of equations with multiple suboptimal
Using Deformable Models solutions. Blake and Zisserman [3] use a deterministic method for
finding the global optimum, which they claim compares favorably
Baba C. Vemuri and Ravikanth Malladi against Monte Carlo techniques such as simulated annealing [ 111.
Like ofher methods for solving large nonlinear systems, however,
the computational complexity of their algorithm is equivalent to the
Abstruct- Three-dimensional viewpoint invariance is an important solution of many linear systems of comparable size. By contrast,
requirement on the representation of surfaces for recognition tasks. noninvariant reconstruction as in [ 191 uses a quadratic functional,
Parameterized surfaces possess this desirable property. In general, the
parameters in a parametric surface representation can be arbitrarily and finding the unique minimum requires the solution of a single
defined. A canonical, intrinsic parameterization provides a consistent, linear system.
invariant form for describing surfaces. Our goal here is to define and In this paper, we develop a viewpoint invariant surface repre-
construct such a parameterization. A new technique for achieving this sentation scheme whose computational complexity is closer to that
goal is presented by using an elastically deformable model. The salient
features of our method are that it provides a unified and general of noninvariant reconstruction. To achieve our goal of invariance,
framework for reparameterization of a surface and easily allows for however, we must dissociate the surface description from the viewer
incorporation of multiview data sets. The canonical parameterization of centered coordinate frame. This can be done easily by employing an
the surface is defined in terms of the surface lines of curvature. Depth intrinsically parameterized surface energy functional in the variational
constraints are first imposed as an external force field on the deformable
model that molds itself to be consistent with the data. Principal vectors reconstruction. We adopt an intrinsic parameterization for the surface
computed from this conformed model surface are then imposed as a that depends purely upon the surface shape and is thereby invariant
force field on the parameter curves of the model. The parameter curves with respect to rigid motion. We express the resultant surface in a
deform to become tangential to the principal vectors thereby yielding an parametric form, where the parameters are curves in predetermined
invariant surface parameterized by the lines of curvature. Extension of
the canonical parametric grid to multiple views is then demonstrated by
directions along the surface [24]-[26], [28]. To obtain a unique
incorporating depth and curvature constraints from multiple views. parameterization, we choose these directions to be the principal
directions, which correspond to the minimum and maximum normal
Manuscript received April 21, 1991; revised July 18, 1992. This work was curvatures of the surface. These directions by Euler’s theorem [9]
supported in part by National Science Foundation under the contract No. EET- are mutually orthogonal. Our final representation is therefore an
8810751.
The author is with the Department of Computer & Information Sciences, orthogonally gridded surface where the grid coordinates are along the
University of Florida, Gainesville, FL 32611. integral curves of the principal vector fields. We call this a canonical
IEEE Log Number 9206200. representation. Orthogonal grids are useful in applications involving
1 - -