Académique Documents
Professionnel Documents
Culture Documents
3 MARCH 1999
534
INVITED SURVEY PAPER Surveys on Image Processing TechnologiesAlgorithms, Sensors and Applications
SUMMARY In this paper, the authors present general views A similar means of problem setting was employed
of computer vision and image processing based on optimization. in other elds of engineering at that time [2]. The
Relaxation and regularization in both broad and narrow senses
method of Isomichi [3], [4], which was presented in the
are used in various elds and problems of computer vision and
image processing, and they are currently being combined with same period, formulates problems to t more general
general-purpose optimization algorithms. The principle and case functions of one-dimensional signals in quantied em-
examples of relaxation and regularization are discussed; the ap- pirical data. This method denes the problem explic-
plication of optimization to shape description that is a particu- itly, i.e., the solution is a function with which the sum of
larly important problem in the eld is described; and the use of
a genetic algorithm (GA) as a method of optimization is intro-
the error between the tting function and the observed
duced. data, and the function that evaluates the appropriate-
key words: computer vision, image processing, optimization, ness of the tting function, becomes the minimum.
relaxation, regularization, snakes, genetic algorithm Methods to minimize the linear sum of indices with
dierent dimensions in this manner are widely used in
1. Introduction the processing of pattern data, with which input and
output are ambiguous. However, CVIP problems are
The eld of applied research in computer vision and im- often incompatible with computer calculations, since
age processing (hereafter abbreviated as CVIP) is ex- the function to be tted is not obtained by analysis or is
panding from industrial applications under controlled too complicated. Thus, attempts have been made in the
lighting to vision systems which people use in their en- eld of CVIP to obtain quasi-optimal solutions, which
vironment. What is required in such expansion is a are coherent overall, in the framework of relaxation.
robust method that is resistant to noise and environ- To conduct numerical analysis of large-scale simul-
mental changes. Application of optimization may pro- taneous equations, a series of methods called either the
vide a breakthrough for CVIP. Specically, this is an iteration method or the relaxation method is widely
approach to nd out the most probable solution from used. With these methods, the optimal solution is con-
images. verged in the nal stage by iterating calculations to
Conventionally, CVIP problems, such as extrac- correct an approximate solution that is obtained at the
tion of curved lines, have been resolved by applying a start. Although the methodology of relaxation has not
method to determine parameters by the least squares been dened explicitly, its concept is applied to vari-
method or some other means by giving appropriate ous problems. The denition of relaxation that is most
functions, such as polynomials and spline functions. widely used in the eld of CVIP is the general name
The method of Otsu [1] which was presented in 1979, of methods with which a local constraint is transmitted
contemplates optimal function tting problems, includ- to the totality by the iteration of parallel operations to
ing the simplicity of the functions to be tted. With minimize (or maximize) a certain index dened to the
this method, a square norm of a quadratic dierential of totality.
a function is used as the appropriateness of the func- Examples of the application of relaxation in a
tion to be tted. This method is considered to have broad sense include Waltzs ltering [5] and the relax-
made a great step forward from the previous stage of ation labeling [6] of Rosenfeld, et al., both of which
problem setting, where the functions were selected by were presented in the 1970s. Thereafter, attempts were
humans and only the parameter setting was resolved made to solve simultaneous equations by the (original)
numerically. relaxation method in some elds of computer vision
dealing with early vision. In all of these attempts,
Manuscript received January 2, 1999. the simultaneous equations are resolved in a calculus
The author is with Electrotechnical Laboratory, of variations framework by formulating the early vision
Tsukuba-shi, 3058568 Japan. as optimization problems. This trend has become more
The author is with Hiroshima City University,
pronounced since Poggio, et al. [7] explained the early
Hiroshima-shi, 7313194 Japan.
The author is with the Graduate School of Information vision by the regularization theory and proposed an in-
Science, Nara Institute of Science and Technology, Ikoma- tegrated framework.
shi, 6300101 Japan. In the eld of neural computing, a neural network
SAKAUE et al: OPTIMIZATION APPROACHES IN COMPUTER VISION AND IMAGE PROCESSING
535
variations which deals with the extremal problem of a When C is the minimum, the following is obtained.
function that represents a certain index (e.g., a kind of
energy) that needs to be minimized (or maximized) in C
= 0 (i = 1, , m; j = 1, , n) (5)
the entire image. Such a problem can be resolved using ui,j
either the direct method or the Euler equation. When
From formulas (4) and (5), C is the quadratic form of
the Euler equation, which is a necessary condition, is
ui,j ; thus, formula (5) is a simultaneous linear equation
used, it becomes a problem of resolving simultaneous
with m n unknowns. Furthermore, since ui,j appears
equations obtained by discretizing this partial dier-
only in terms ci,j , ci+1,j , ci1,j , ci,j+1 , ci,j1 in C, for-
ential equation per pixel (or datum point). Since the
mula (5) can be modied as follows.
coecient matrix of such simultaneous equations often
becomes large-scale and non-dense, relaxation is gener- ui+2,j + ui,j+2 + ui2,j + ui,j2
ally used as a method for numerically solving the prob- +2(ui+1,j+1 + ui1,j+1 + ui+1,j1 + ui1,j1 )
lems of CVIP. In this paper, the framework of calculus
of variations + simultaneous equations + relaxation in 8(ui+1,j + ui1,j + ui,j1 + ui,j+1 ) + 20ui,j
the eld of CVIP is called deterministic relaxation. =0 (6)
Deterministic relaxation is described below using
Briggs [11] modied the formula as shown below, and
some concrete examples (Fig. 1). Now, a problem to in-
solved it by the Jacobian method.
terpolate sporadic altitudinal data obtained by obser-
vation to a curved surface (e.g., a ground surface) based (k+1)
ui,j
1 (k)
on the principle of minimum curvature is assumed [11].
(k) (k) (k)
Here, the curved surface, u(x, y), to be obtained is con- = 8 ui+1,j + ui1,j + ui,j1 + ui,j+1
sidered to be a thin elastic plate xed by the observed 20
(k) (k) (k) (k)
datum points, and interpolation is conducted by ob- 2 ui+1,j+1 +ui1,j+1 +ui+1,j1 +ui1,j1
taining a curved surface that minimizes the energy of
(k) (k) (k) (k)
deection expressed by the formula below. ui+2,j + ui,j+2 + ui2,j + ui,j2 (7)
IEICE TRANS. INF. & SYST., VOL.E82D, NO.3 MARCH 1999
536
Here, ui,j at the observation points are xed. The early 3. The state is shifted from x to x at a probability of
value is obtained by linear interpolation from the ob- min {1, exp { (E(x ) E(x)) /Tk }}, and the new
servation points. Another method of solution is em- state is considered as x.
ployed in the former half of Ref. [12]. A solution can be 4. Steps 2 and 3 are iterated until reaching the equi-
obtained by both of these methods by iterating local librium state.
parallel operations. Formula (7) is in fact a discrete ex- 5. Steps 2 to 4 are iterated while gradually decreas-
pression of the Euler equation in formula (2). After all, ing the temperature (Tk Tk+1 ) until the value of
the Euler equation was solved by these methods with the energy becomes almost invariable in the equi-
empirical data as the boundary condition. librium state.
2.1.2 Stochastic Relaxation Many stages of CVIP require some kind of label-
ing operations to pixels, such as edge detection, seg-
When the energy function is not in the quadratic form, mentation, and stereo matching. When the Markov
the problem becomes non-linear simultaneous equa- Random Field (MRF) can be hypothesized as a prob-
tions. Although a formula for iterative operations can ability model of image, i.e., when the state of the re-
be derived in this case, the convergence to the optimal spective pixels depends only upon their neighborhood,
solution that minimizes the energy is not guaranteed the energy of the entire image is the sum of local ener-
by iterative operations starting with an arbitrary ini- gies. Accordingly, the energy variations of the totality
tial value when several solutions exist (i.e., the origi- E = E(x) E(x ), which determines the state vari-
nal energy function is non-convex and has local min- ations of the totality in Step 3 above, are equivalent to
ima). The same applies to the process of energy min- the energy variations in the neighborhood of a pixel in
imization by the Hopeld-type neural network. Thus, attention; thus, an annealing process is established by
the means of selecting the appropriate initial value and local parallel operations.
avoiding local minima need to be contrived. To counter A CVIP method with which the state variations
such non-convex problems, a method to use multiple are iterated stochastically by conducting local paral-
scales [13] is employed in the framework of determinis- lel operations, is called stochastic relaxation [10], [16].
tic relaxation, with which processing is conducted grad- Stochastic relaxation has hitherto been used for prob-
ually from coarse scales to ne scales. lems of image restoration, segmentation, binocular
On the other hand, simulated annealing [14], [15] stereo, etc.
utilizes analogies from statistical thermodynamics to Geman, et al. [10] applied stochastic relaxation to
search for the global minimum of non-convex energy the problems of image restoration. They prevented the
functions. This algorithm was proposed as a discrete sections of discontinuity in the region boundary from
combinatory optimization method, which makes the becoming excessively smooth by preparing a stochastic
state of a system undergo a stochastic transition by es- variable for the existence of region boundary, which is
tablishing a stochastic process, with which the state x called a line process, in addition to the stochastic vari-
appears by a probability expressed by formula (8) when able for the gray levels. This contrivance is used for
x (which represents the combination of labels; i.e., the other applied problems of CVIP [17][20].
state of the system) has energy E(x). Despite their capacity for dealing with various
types of problems, stochastic methods have a disadvan-
1
p(x) = exp {E(x)/T } (8) tage in that a huge amount of operations is required to
Z obtain solutions. It is thus desirable to use determin-
Here, Z is a constant for normalization, while T is a istic relaxation, if possible. However, there are not so
parameter corresponding to the temperature in the many classes (types) of problems that can be solved by
statistical mechanics. With simulated annealing, the deterministic methods at present, and stochastic meth-
system converges to one of the states of energy mini- ods are the only alternative means. Thus, a massively
mum by the probability of 1 when the initial value of parallel computer that can adequately handle such op-
the temperature parameter T is suciently high, and erations should be developed.
when T0 by decreasing the temperature continuously.
The basic steps of annealing based on the proba- 2.1.3 Regularization Method
bility process of Metropolis, which is often used, are as
follows: In the eld of CVIP, many problems of early vision are
1. A suciently high initial temperature T0 is deter- formulated as optimization problems [21][27].
mined, as is the initial state of the system. This Problems of early vision can be considered as prob-
state is considered as x (its energy is E(x)). lems to restore the descriptions of an original three-
2. One freedom of the system is selected at random; dimensional world from images created by projecting
its value is varied at random; and this temporary the data of the three-dimensional world onto a two-
state is considered as x (its energy is E(x )). dimensional world. However, problems are often essen-
SAKAUE et al: OPTIMIZATION APPROACHES IN COMPUTER VISION AND IMAGE PROCESSING
537
tially indeterminable, because the data of the three- There are several other methods proposed for es-
dimensional world that need to be restored are con- timating a regularization parameter, with which a pa-
tained in the two-dimensional image in a degenerated rameter is estimated as a distributed constant based on
shape. Such problems are called ill-posed problems, the hypothesis that the errors between the observed val-
as mentioned by Hadamard. ues and estimated values are dispersed uniformly [30]
A well-posed problem satises the following condi- [33]. Such methods are eective for processing the
tions: (1) Its solution exists; (2) its solution is unique; neighborhood of the boundary of discontinuity, partic-
and (3) its solution depends continuously on the initial ularly when dealing with a solution space that contains
data. Should it fail to meet any of these conditions, discontinuity.
it is called an ill-posed problem. A method available
for giving a mathematically stable solution to an ill- 2.1.4 Countermeasures for Discontinuity
posed problem is regularization [28]. Poggio, et al. [7]
proposed that the problems of early vision be discussed Methods using a stabilizing functional may not be ap-
uniformly in terms of the theory of regularization. propriate when the solution to be obtained is not a
Problems that can be solved by regularization are smooth function. To counter such problems, various
those with a possible solution space that is much greater studies have been conducted to establish methods to
than input. Specically, regularization is applicable use functions in the quadratic form as the base [1]
when possible solutions of problems exist innitely, as [4]. Other methods being studied include methods to
in the case of a problem to restore a three-dimensional use the spline function [34], [35], GNC and other meth-
structure from image data that are projected two- ods to realize weak continuity constraint [34], [36][39],
dimensionally. Regularization is a method for obtaining and methods to use a multi-valued function [40]. De-
a mathematically stable solution from such a solution spite its great disadvantage of being time-consuming,
space. the stochastic relaxation with the line process, that is
With standard regularization, when data that re- mentioned in Sect. 2.1.2, may also be eective because
ceive a linear operation A from unknown z are observed, it is applicable to problems that cannot be solved by
as expressed below, deterministic methods.
y = Az (9) Terzopoulos introduced a controlled continuity
constraint [34], [41]. This constraint is expressed by a
an inverse problem that estimates z from data y is for- generalized spline with several degrees based on conti-
mulated as an optimization problem that minimizes the nuity control functions. To solve problems of 3D surface
objective function (called energy) as follows: reconstruction, a functional that expresses a controlled
E = ||Az y||2 + ||P z||2 (10) continuity constraint was adopted as follows:
2
The rst term in formula (10), ||Az y|| , expresses 1
S (u) = (x, y) (x, y)(u2xx + 2u2xy + u2yy )
the dierence from the observation data, and is called a 2
penalty functional. The second term, ||P z||2 , is called +{1 (x, y)}(u2x + u2y ) dxdy (11)
a stabilizing functional, which expresses general con-
straint conditions (e.g., smoothness) for solutions in the Here, (x, y) and (x, y) are continuity control func-
real world. is a regularization parameter that deter- tions (0 1, 0 1). This functional is a
mines the proportions of these two functionals. This weighted linear combination of an energy functional of
formulation is designed to seek a solution that satis- a thin plate and an energy functional of a thin mem-
es the constraint conditions fairly well with few con- brane, and can control the smoothness properties of
tradictions to observation data. Formula (10) can be the constraint by adjusting the control functions at ar-
minimized in the framework of calculus of variations. bitrary points.
With standard regularization, the norm of a quadratic The operations of basic control are as follows: i)
form is adopted, and P is a linear operator. (x, y) and (x, y) are set to be non-zero in all non-
How to determine the regularization parameter continuous points (x, y); ii) the value of (x, y) is set to
is a matter of great concern. A method to estimate a be close to 0 when the discontinuity of direction needs
regularization parameter, taking advantage of the er- to be established; and iii) the value of (x, y) is set to
ror from the amount obtained by observation, was pro- be close to 0 at the discontinuous points of depth.
posed by Amano, et al. [29], in which a function that If the values of (x, y) and (x, y) are known in
minimizes the dierence from the sample is presented advance, formula (11) is a quadratic form, and can
as a stabilizing function. Although some samples need be solved in the framework of standard regularization.
to be prepared in advance, this method enables the To detect discontinuity in advance, the so-called edge-
utilization of errors in observed values by presenting detection method can be used. However, during the op-
several samples, and estimates the regularization pa- erations of 3D surface reconstruction, Terzopoulos [41]
rameter based on these errors. found that the discontinuity of depth at a point where
IEICE TRANS. INF. & SYST., VOL.E82D, NO.3 MARCH 1999
538
the moment of deection of a thin plate was 0, and dis- explicitly state or estimate which one of the two curved
covered the discontinuity of direction at the maximum surfaces the empirical data belong to.
point. A similar approach is described in Ref. [35]. If
(x, y) and (x, y) are incorporated to calculus of vari- 2.2 Examples of the Application of Regularization
ations as unknown values, it is quite troublesome be-
cause dual minimization processes appear. 2.2.1 Reconstruction of a Three-Dimensional Curved
Grimson, et al. [42] proposed a method for esti- Surface
mating the discontinuity for use in problems to restore
a curved surface based on discrete depth data. Accord- Here, a curved surface reconstruction problem that es-
ing to this restoration method, the boundary of discon- timates the original curved surface is established [12],
tinuity is estimated by making the error distribution [41], assuming that data d is given as a result of apply-
uniform, using the distribution of errors between the ing a sampling operation S to a given curved surface u.
given depth data and the tted curved surface. This This is identical to a problem of obtaining elevations
concept is the same as that in the guideline for esti- densely from sporadic data in a digital terrain map
mating a regularization parameter. (DTM). Although the values of empirical data were
Based on an idea that partial discontinuity is plau- stored in the case of the problem of interpolation in-
sible but imposes a penalty, Blake [43], [44] gave the troduced in Sect. 2.1.1, the tolerance obtained from the
name weak constraint to a constraint condition that data of observation points needs to be given in actual
minimizes the summation of variations in continuous operations as shown in Fig. 2, due to the inaccuracy of
sections and penalties in discontinuous sections. This actual empirical data. With the number of pixels of
problem is non-convex. GNC (graduated non-convexity m n, the so-called quadratic variation is often used as
algorithm) is proposed as a method to solve this prob- the constraint of smoothness. As with formula (1), this
lem. With this method, a sequence of functions that is a thin plate spline, with which continuity is given to
falls in-between the original objective function F (non- the normal vector. Formula (10) can be modied as
convex) and the convex envelope F (convex) of F is follows:
created, and local optimal solutions are obtained se- n m 2 2
quentially from F by the hill-climbing method to grad- 2 u
E= Si,j (ui,j di,j ) +
ually approach the solution at F . j=1 i=1 x2
Shizawa [40] succeeded in using two functions as 2 2 2 2
a constraint by expanding standard regularization to u u
+2 + dxdy (15)
multi-valued functions. This method is applicable to xy y 2
a problem to restore several overlapped surfaces called
transparency. Regularization that constrains a curved To obtain a curved surface ui,j (1 i m, 1 j n)
surface y = f (x) is expressed as a minimization prob- that minimizes the above formula, simultaneous equa-
lem of the following function. tions with m n unknowns as shown below are solved
by calculus of variations:
N
2 E
E (1) [f ] = y(i) f (x(i) ) + ||Sf (x)||2 (12) = 0 (1 i m, 1 j n) (16)
i=1 ui,j
When f1 (x) and f2 (x) are constraint functions, the con- If the partial dierentiations, such as 2 /x2 , in for-
straint condition can be expressed as follows: mula (15) are substituted to a dierence as follows (h
is the grid space):
(y f1 (x))(y f2 (x))
= y 2 (f1 (x) + f2 (x))y + f1 (x)f2 (x) 2u ui+1,j 2ui,j + ui1,j
= (17)
=0 (13) x2 h2
Formula (16) becomes linear simultaneous equations of
If F (x) = f1 (x)f2 (x), G(x) = (f1 (x)+f2 (x)), regular-
the values of elevations at all points in image as fol-
ization with the two functions as constraint conditions
lows, which can easily be solved [12].
becomes a minimization problem of the following func-
tion.
N
2
E (2) [F, G] = F (x(i) ) + G(x(i) )y(i) + (y(i) )2
i=1
E 25 cameras with parallel lines of sight is examined. With
= 2Si,j (ui,j di,j ) + ui,j
ui,j h4 this system, the points within the three-dimensional
8 space infallibly appear on the same horizontal scan-
4 (ui+1,j + ui1,j + ui,j+1 + ui,j1 ) ning line (epipolar line) in the images on both left and
h
right. It is assumed that the brightness of correspond-
3
+ 4 (ui+2,j + ui2,j + ui,j+2 + ui,j2 ) ing points is similar, and that the depth of the target
2h scene varies continuously. When the brightness of the
1 left and right images is expressed as L(x, y) and R(x, y),
+ 4 (ui+2,j+2 + ui+2,j2
4h respectively, a stereo matching problem, which obtains
the points (x + d(x, y), y) in the left image (that corre-
+ui2,j+2 + ui2,j2 )} spond to the respective points (x, y) in the right image)
is formulated as a problem to obtain a disparity func-
=0 (18) tion d(x, y) that minimizes the following functional in
the framework of regularization.
Although it is nothing more than obtaining a spline
2
curved surface, this method is greatly advantageous E= {L(x + d(x, y), y) R(x, y)}
in that the perspective is quite favorable when adding
other constraint conditions because the energy mini- + dx (x, y)2 + dy (x, y)2 dxdy (20)
mization is explicitly expressed by a formula using the
framework of regularization. Muraki, et al. [45] ob- Since this functional is not in a quadratic form in
tained a favorable result (Fig. 3) by adding the con- regard to the function d(x, y) to be obtained, the prob-
straint that the variations of elevation become 0 on the lem is how to avoid local minima. There are two meth-
contour lines of a map, as expressed below: ods available to counter this problem, i.e., a method to
use deterministic relaxation by multiple scales based on
2
u u an Euler equation [46],
x(x, y) + y(x, y) dxdy (19)
x y
{L(x + d(x, y), y) R(x, y)} Lx (x + d(x, y), y)
to formula (15), assuming that contour lines are par- {dxx (x, y) + dyy (x, y)} = 0 (21)
tially obtained from a map. Here, (x(x, y), y(x, y))
is the direction cosine of contour lines at a point (x, y) and a method to use stochastic relaxation by anneal-
in the image. ing [20].
When discontinuous regularization proposed by
2.2.2 Binocular Stereo Terzopoulos [41] is applied simply to an image with high
texture, a boundary of discontinuity can easily be de-
The standard stereo camera system that consists of tected as in Refs. [47] and [48]. Boult, et al. [49] pro-
posed a method to roughly t a curved surface to the
extracted characteristic points by using a function t-
ting, and to interpolate their intervals. The scale of the
function to be tted here is discussed in Ref. [50].
It + U T I = 0 (22)
to consecutively estimate regularization parameters, so Here, Eint is the internal energy that expresses the
that the distribution of errors between the observed smoothness of contours, called the energy that is con-
values and estimated values become uniform. cerned in the shape, to which the sum of energy that is
SAKAUE et al: OPTIMIZATION APPROACHES IN COMPUTER VISION AND IMAGE PROCESSING
541
mosomes are constructed by the values of the coordi- ble of conducting local optimization, and GA are com-
nates at three points in the case of circle detection. bined (Fig. 6) [92]. In the three-dimensional shape re-
The idea is that a more global circle can be detected construction from gray-scale images [93], GA is used to
by crossing over two circles, i.e., one made by giving minimize the error between a shadow image in the esti-
three points in the image and the other by giving dif- mated three-dimensional shape, under the given light-
ferent three points. In the polygonal approximation of ing and picturing conditions, and a shadow image that
a closed curve [91], whether or not the respective points is obtained by observation.
that divide the closed curve into equal N sections are In both cases, the estimated two-dimensional data
considered as the vertexes of a polynomial polygon is (image, disparity distribution, and depth distribution)
expressed by 1 and 0, respectively, and coded as the are expressed as chromosomes without coding, and
N -bit chromosomes. the cross-over is conducted by two-dimensional region
reshuing. The problem can be solved only if the two-
4.2.4 Image Reconstruction dimensional data generated by genetic operations meet
the constraint conditions particular to the problem and
Several attempts have been made to use GA as an op- can be evaluated quantitatively. Specically, the recon-
timization method in the application of regularization struction can be solved as a forward problem not as an
to CVIP. inverse problem. This may be the most advantageous
In the stereo matching of binocular stereoscopic vi- point in using GA for image reconstruction.
sion the disparity distribution where the gray scale level
becomes identical in the corresponding points on the 5. Conclusion
left and right is obtained. Here, the two-dimensional
active contour model, Active Net [67], which is capa- In this paper, the authors presented general views of
image processing and computer vision based on opti-
mization. It should be noted that diverse approaches,
too numerous to be covered in this paper, have been
taken. Nonetheless, a big question that remains to be
answered is: Why can the visual function of nding
the most probable solution from an image be done
by humans quickly, accurately, and naturally, but is so
dicult for a machine?
References
PAMI, vol.6, no.6, pp.721741, 1985. [35] D. Lee and T. Pavlidis, One-dimensional regularization
[11] L.C. Briggs, Machine contouring using minimum curva- with discontinuities, Trans. PAMI, vol.10, no.6, pp.822
ture, Geophysics, vol.39, no.1, pp.3948, 1974. 829, 1988.
[12] W.E.L. Grimson, An implementation of a computational [36] A. Blake and A. Zisserman, Some properties of weak con-
theory of visual surface interpolation, CVGIP, vol.22, tinuity constraints and the GNC algorithm, Proc. IEEE
pp.3969, 1983. CVPR 86, pp.656667, 1986.
[13] A.P. Witkin, D. Terzopoulos, and M. Kass, Signal match- [37] S.Z. Li, Reconstruction without discontinuities, 3rd.
ing through scale space, Int. Journal Computer Vision, ICCV, pp.709712, 1990.
vol.1, no.2, pp.133144, 1987. [38] R.L. Stevenson, B.E. Schmitz, and E.J. Delp, Disconti-
[14] S. Kirkpatrick, C.D. Gelatt, Jr., and M.P. Vecchi, Op- nuity preserving regularization of inverse visual problems,
timization by simulated annealing, Science, vol.220, IEEE Trans. Syst. Man. & Cybern., vol.24, no.3, pp.455
no.4598, pp.671680, 1983. 469, 1994.
[15] V. Granville, M. Krivanek, and J.P. Rasson, Simulated [39] C. Schnorr, Unique reconstruction of piecewise smooth
annealing: A proof of convergence, IEEE Trans. PAMI, images by minimizing strictly convex nonquadratic func-
vol.16, pp.652656, 1994. tionals, Journal Mathematical Imaging and Vision, vol.4,
[16] M.S. Landy, A brief survey of knowledge aggregation pp.189198, 1994.
methods, Proc. 8th ICPR, pp.248252, 1986. [40] M. Shizawa, Extension of standard regularization the-
[17] D. Lee and G.W. Wasilkowski, Discontinuity detection and ory into multi-valued functionsReconstruction of smooth
thresholdingA stochastic approach, Proc. IEEE CVPR multiple surfaces, IEICE Trans., vol.J76-D-II, no.6,
91, pp.208214, 1991. pp.11461156, 1994 (in Japanese).
[18] D. Keren and M. Werman, Probabilistic analysis of regu- [41] D. Terzopoulos, The computation of visible-surface repre-
larization, IEEE Trans. PAMI, vol.15, no.10, pp.982995, sentations, IEEE Trans. PAMI, vol.10, no.4, pp.417438,
1993. 1988.
[19] D. Keren and M. Werman, A Bayesian framework for reg- [42] W.E.L. Grimson and T. Pavlidis, Discontinuity detection
ularization, ICPR 94, pp.7276, 1994. for visual surface reconstruction, CVGIP, vol.30, pp.316
[20] S.T. Barnard, Stochastic stereo matching over scale, Int. 330, 1985.
Journal Computer Vision, vol.3, no.1, pp.1732, 1989. [43] A. Blake, The least-disturbance principle and weak con-
[21] B.K.P. Horn and B.G. Shunk, Determining optical ow, straints, Pattern Recognition Letters, vol.1, nos.56,
AI, vol.17, pp.185203, 1981. pp.393399, 1983.
[22] K. Ikeuchi and B.K.P. Horn, Numerical shape from shad- [44] A. Blake, Comparison of the eciency of deterministic
ing and occluding boundaries, AI, vol.17, pp.141184, and stochastic algorithms for visual reconstruction, IEEE
1981. Trans. PAMI, vol.11, no.1, pp.212, 1989.
[23] D. Marr and T. Poggio, A theory of human stereo vision, [45] S. Muraki, N. Yokoya, and K. Yamamoto, 3D surface re-
Proc. Royal Soc. London Ser. B 204, pp.301328, 1979. construction from contour line image by a regularization
[24] A. Blake, Relaxation labelingThe principle of least dis- method, Proc. SPIE, vol.1395, Close-Range Photogram-
turbance, Pattern Recognition Letters, vol.1, nos.56, metry Meets Machine Vision, pp.226233, Sept. 1990.
pp.385391, 1983. [46] N. Yokoya, Surface reconstruction directly from binocu-
[25] D. Terzopoulos, Multilevel computational processes for lar stereo images by multiscale-multistage regularization,
visual surface reconstruction, CVGIP, vol.24, pp.5296, ICPR 92, I, pp.642646, 1992.
1983. [47] R. March, A regularization model for stereo vision with
[26] T. Pajdla and V. Hlavac, Surface discontinuities in range controlled continuity, Pattern Recognition Letters, vol.10,
images, 4th. ICCV, pp.524528, 1993. pp.259263, 1989,
[27] T. Poggio and C. Koch, Ill-posed problems in early vision: [48] R. March, Computation of stereo disparity using regular-
From computational theory to analogue networks, Proc. ization, Pattern Recognition Letters, vol.8, pp.181187,
Royal Soc. London Ser. B 226, pp.303323, 1985. 1988,
[28] A.N. Tikhonov and V.Y. Arsenin, Solutions of Ill-Posed [49] T.E. Boult and L. Chen, Analysis of two new stereo algo-
Problems, Winston, Washington DC, 1977. rithms, Proc. IEEE CVPR 88, pp.177182, 1988.
[29] A. Amano, Y. Sakaguchi, M. Minoh, and K. Ikeda, Snakes [50] T.E. Boult, What is regular in regularization?, 3rd.
using a sample contour model, Proc. 1st ACCV, pp.538 ICCV, pp.644651, 1990.
541, 1993. [51] M.A. Snyder, The mathematical foundations of smooth-
[30] R. Hummel and R. Moniot, Solving ill-conditioned prob- ness constraints: A new class of coupled constraints, IUW,
lems by minimizing equation error, 1st ICCV, pp.527533, pp.154161, 1990.
1987. [52] M.A. Snyder, On the mathematical foundations of
[31] D. Keren and M. Werman, Variations on regularization, smoothness constraints for the deterimination of optical
ICPR 90, pp.9398, 1990. ow and for surface reconstruction, IEEE Trans. PAMI,
[32] A.M. Thompson, J.C. Brown, J.W. Kay, and D.M. vol.13, no.11, pp.11051114, 1991.
Titterington, A study of methods of choosing the smooth- [53] M.J. Black and P. Anandan, A framework for the robust
ing parameter in image restoration by regularization, estimation of optical ow, 4th. ICCV, pp.231236, 1993.
IEEE Trans. PAMI, vol.13, no.4, pp.326339, 1991. [54] S. Raghavan, N. Gupta, and L. Kanal, Computing
[33] N. Mukawa, Vector eld reconstruction using variance- discontinuity-preserved image ow, ICPR 92, A, pp.764
based resolution control and its application to optic ow es- 767, 1992.
timation, IEICE Trans., vol.J78-D-II, no.7, pp.10281038, [55] C. Schnorr, Computation of discontinuous optical ow by
July 1995 (in Japanese) (published in English in Systems domain decomposition and shape optimization, Int. Jour-
and Computers in Japan, vol.27, no.13, 1996). nal Computer Vision, vol.8, no.2, pp.153165, 1992.
[34] D. Terzopoulos, Regularization of inverse visual problems [56] C. Schnorr, Determining optical ow for irregular domains
involving discontinuities, IEEE Trans. PAMI, vol.8, no.4, by minimizing quadratic functionals of a certain class, Int.
pp.413424, 1986. Journal Computer Vision, vol.6, no.1, pp.2538, 1991.
IEICE TRANS. INF. & SYST., VOL.E82D, NO.3 MARCH 1999
546
[57] D. Geiger and A. Yuille, A common framework for image [79] S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum,
segmentation, Int. Journal Computer Vision, vol.6, no.3, and A. Yezzi, Gradient ows and geometric active contour
pp.227243, 1991. Models, 5th ICCV, pp.810815, 1995.
[58] T.E. Boult and M. Lerner, Energy-based segmentation of [80] L.D. Cohen and R. Kimmel, Global minimum for active
very sparse range surfaces, IUW, pp.565572, 1990. contour models: A minimum path approach, Proc. IEEE
[59] G. Chen and H. Yang, Edge detection by regularized cu- CVPR 96, pp.666673, 1996.
bic B-spline tting, IEEE Trans. Syst., Man. & Cybern., [81] G. Sapiro, Vector-valued active contours, Proc. IEEE
vol.25, no.4, pp.636643, 1995. CVPR 96, pp.680685, 1996.
[60] M. Gokmen and C.C. Li, Edge detection and surface recon- [82] D.E. Goldberg, Genetic ALgorithms in Search, Optimiza-
struction using rened regularization, IEEE Trans. PAMI, tion, and Machine Learning, Addison-Wesley, 1989.
vol.15, pp.492499, 1993. [83] A.M. Gillies, An image processing computer which learns
[61] N. Yokoya, M. Kaneta, and K. Yamamoto, Recovery of by example, SPIE, vol.155, pp.120126, 1978.
superquadric primitives from a range image using simulated [84] K. Yamamoto, K. Sakaue, H. Matsubara, and K.
annealing, ICPR92, (I), pp.168172, 1992. Yamagishi, Miracle-IV: Multiple image recognition sys-
[62] S. Muraki, Volumetric shape description of range data us- tem aiming concept learningIntelligent vision, ICPR 88,
ing blobby model, ACM Computer Graphics, vol.25, no.4, pp.818821, 1988.
pp.227235, 1991. [85] H. Matsubara, K. Sakaue, and K. Yamamoto, Learning
[63] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: Active a Visual Model and an Image Processing Strategy from a
contour models, Int. Journal Computer Vision, vol.1, no.3, Series of Silhouette Images on MIRACLE-IV, Symbolic
pp.321331, 1988. Visual Learning, K. Ikeuchi and M. Veloso, eds., Oxford,
[64] A.L. Yuille, P.W. Hallinan, and D.S. Cohen, Feature ex- pp.171191, 1997.
traction from faces using deformable templates, Int. Jour- [86] I. Yoda, Automatic generation of directional erosion and
nal Computer Vision, vol.8, no.2, pp.99111, 1992. dilation sequence by genetic algorithms, Proc. ACCV 95,
[65] S.R. Gunn and M.S. Nixon, Snake head boundary extrac- II, pp.103107, 1995.
tion using global and local energy minimisation, ICPR 96, [87] H. Iba, H. deGaris, and T. Sato, A bug-based search strat-
pp.581585, 1996. egy for problem solving, ETL-TR92-24, Electrotechnical
[66] R. Chung and C. Ho, Using 2D active contour models for Laboratory, 1992.
3D reconstruction from serial sections, ICPR 96, pp.849 [88] T. Tagao, T. Agui, and H. Nagahashi, Pattern matching
853, 1996. of binary shapes using a genetic method, IEICE Trans.,
[67] K. Sakaue and K. Yamamoto, Active net model and its ap- vol.J76-D-II, no.3, pp.557565, March 1993 (in Japanese).
plication to region extraction, The Journal of the Institute [89] T. Nagao, T. Takeshi, and H. Nagahashi, Extraction of
of Television Engineers of Japan, vol.45, no.10, pp.1155 straight lines using a genetic algorithm, IEICE Trans.,
1163, 1991 (in Japanese). vol.J75-D-II, no.4, pp.832834, April 1992 (in Japanese).
[68] S. Araki, N. Yokoya, and H. Takemura, Tracking of mul- [90] G. Roth and M.D. Levine, A genetic algorithm for primi-
tiple moving objects using split-and-merge contour models tive extraction, Proc. of the 4th International Conference
based on crossing detection, Proc. Vision Interface 97, on Genetic Algorithm, pp.487494, 1991.
pp.6572, Kelowna, British Columbia, Canada, 1997. [91] T. Yamagishi and T. Tomikawa, Polygonal approximation
[69] C. Davatzikos and J.L. Prince, Convexity analysis of ac- of closed curve by GA, IEICE Trans., vol.J76-D-II, no.4,
tive contour problems, Proc. IEEE CVPR 96, pp.674679, pp.917919, April 1993 (in Japanese).
1996. [92] K. Sakaue, Stereo matching by the combination of ge-
[70] N. Ueda and K. Mase, Tracking moving contours us- netic algorithm and active net, IEICE Trans., vol.J77-D-II,
ing energy-minimizing elastic contour models, ECCV 92, no.11, pp.22392246, Nov. 1994 (in Japanese). (published
pp.453457, 1992. in English in Systems and Computers in Japan, vol.27, no.1,
[71] M. Hashimoto, H. Kinoshita, and Y. Sakai, An object pp.4048, Scripta Technica, Inc., 1996.)
extraction method using sampled active contour model, [93] H. Saito and N. Tsunashima, Estimation of 3-D paramet-
IEICE Trans., vol.J77-D-II, no.11, pp.21712178, 1994 (in ric models from shading image using genetic algorithms,
Japanese). ICPR 94, A, pp.668670, 1994.
[72] B. Olstad and A.H. Torp, Encoding of a priori information
in active contour models, IEEE Trans. PAMI, vol.18, no.9,
pp.863872, 1996.
[73] T. CcInerney and D. Terzopoulos, Topologically adaptable
snakes, 5th ICCV, pp.840845, 1995.
[74] A.A. Amini, S. Tehrani, and T.E. Weymouth, Using dy-
namic programming for minimizing the energy of active
contours in the presence of hard constraints, 2nd ICCV, Katsuhiko Sakaue received the
pp.9599, 1988. B.E., M.E., and Ph.D. degrees all in elec-
[75] A.A. Amini, T. Weymouth, and R. Jain, Using dynamic tronic engineering from University of To-
programming for solving variational problems in vision, kyo, Tokyo, Japan, in 1976, 1978 and
Trans. PAMI, vol.12, no.9, pp.855867, 1990. 1981, respectively. In 1981, he joined the
[76] D.J. Williams and M. Shah, A fast algorithm for active Electrotechnical Laboratory, Ministry of
contours and curvature estimation, CVGIP, vol.55, no.1, International Trade and Industry, and en-
pp.1426, 1992. gaged in researches in image processing
[77] V. Caselles, R. Kimmel, and G. Sapiro, Geodesic active and computer vision. He received the En-
contours, 5th ICCV, pp.694699, 1995. couragement Prize in 1979 from IEICE,
[78] R. Malladi, J.A. Sethian, and B.C. Vemuri, Shape mod- and the Paper Award in 1985 from Infor-
eling with front propagation: A level set approach, IEEE mation Processing Society of Japan. He is a member of IEEE,
Trans. PAMI, vol.17, pp.158175, 1995. IPSJ, ITE.
SAKAUE et al: OPTIMIZATION APPROACHES IN COMPUTER VISION AND IMAGE PROCESSING
547