3 Boundaries
What constitutes a boundary? Since the human visual system
is so highly sophisticated, this seems to be a trivial question.
However, a large problem in boundary detection is obtaining a
suitable definition of a boundary which is generally applicable.
Definitions can be given in two domains: (1) image domain,
and (2) task domain. Within the image domain the most
important feature that may represent a boundary is the edge:
An edge is an intensity discontinuity in the image.
In course of time an abundance of algorithms have been de-
veloped, designed to trace local intensity discontinuities in
the image. We treat some of them in sections 4 and 5.
The objects one wants to extract, depend on the task do-
main at hand. For example, for a large scale base map of
an urban area one wants other objects than for a national
road data base. This observation results in the important
understanding that a perfect outlining of relevant objects in
aerial and space imagery can not be established by bottom-up
approaches alone. We define a boundary as:
A boundary is a (closed) outlining of an object
which is relevant for the task at hand.
A common preassumption is that abrupt intensity changes in
the image correspond to meaningful object boundaries in the
scene. This may be valid for highly restricted scenes, such
as present in industrial environments, but for unrestricted
scenes, a boundary may be visible in the image as abrupt in-
tensity changes, but this is neither a necessary nor a sufficient
condition. Not all intensity changes correspond with relevant
object outlinings. They may be, for example, due to noise,
texture and shadows. Furthermore, boundaries may not show
up as intensity changes, due to low contrast or occlusion. To
resolve this problem a priori information about the objects
relevant to the task domain is necessary. This information
can concern radiometric and geometric properties.
The exploration of radiometric properties is carried out regu-
larly by the remote sensing community for multispectral aerial
and space imagery by statistical pattern recognition tech-
niques using as features (functions of) the grey values of the
different spectral bands. However, pixel-based classification
is highly prone to error, resulting in the wish of developing
object-based classification techniques, which require, at turn,
reliable delineation of objects.
Geometric properties may concern generic or specific aspects.
Generic geometric aspects concern descriptions of general ap-
pearances of boundaries, such as: (1) smooth continuation,
i.e. nearby edge pixels will point approximately in the same
direction, (2) thinness, e.g. edges should be one pixel thick,
which criterion is often necessary for further processing, and
(3) connectivity, e.g. boundaries will be closed. A classical
mechanism to incorporate generic geometric information is
by probabilistic relaxation (e.g. Schachter et al., 1977; Pe-
leg, 1980; Prager, 1980; Hancock & Kittler, 1990; Duncan
& Birkholzer, 1992). For edge detection purposes on images
of unrestricted scenes this mechanism shows several severe
drawbacks.
Specific geometric information concerns more detailed de-
scriptions about the shape and possibly size of the bound-
aries, for example, the objects of interest have circular or
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
rectangular boundaries. The exploration of specific geomet-
ric information has been subject of extensive research. For
example the Hough transform (for a good survey with many
references, see lllingworth & Kittler (1988)), boundary de-
lineation by dynamic programming ( e.g. Montanari, 1971;
Martelli, 1976; Elliott & Srinivasan, 1981; Gerbrands, 1988;
Lemmens et al., 1990), and perceptual grouping (e.g. Witkin
& Tenenbaum, 1986; Khan & Gillies, 1992; Lu & Agger-
wal, 1992; Mohan & Nevatia, 1992; Price & Huertas, 1992;
Lin et al, 1995) are well-known approaches. Lemmens et al.
(1990) suggested to use dynamic programming for road de-
tection. Within the photogrammetric community De Gunst
et al. (1991), and Grün & Li (1994) explored the approach
for this purpose.
We focus in the sequel on local edge detection schemes, be-
cause they form the fundamental stage in any boundary de-
lineation task. We loosely divide them into two broad classes:
(1) monadic schemes and (2) plural schemes. Monadic
schemes base the decision whether an edge is present or not,
directly on the responses of the operators, without further ex-
amination of additional responses, while plural schemes carry
out such an evaluation in some form.
4 Monadic Local Edge Detection
A local edge detector is an operator of small spatial extent
that traces changes in the image function to classify each
pixel as edge or non-edge according to some decision rule. No
a priori information about the scene structure or contextual
information is employed. Monadic methods can be generally
classified into one of four categories: (1) Differentiation, (2)
Surface fitting, (3) Template matching, and (4) Curvature
determination. Other approaches have been proposed, such
as those based on morphological operators (cf. Lee et al.
1987). We do not treat them here.
4.1 Discrete Differentiation
Abrupt intensity changes are traceable by computing the
partial derivatives in two orthogonal directions: gx =
09/02; gu = 9g/9y), usually along the grid lines. The small-
est step possible on a sampled space is the sample interval
Az. By definition the pixel size is unity: Az = 1. Hence the
discrete derivative in z-direction of the 2-D discrete function
g becomes:
g; — limas—i[g(j-- Az) - g()]/A» — 9(3 1) 9G) (1)
Accordingly one can define gy, the first derivative in y-
direction. (For the Roberts operator g; and gy become:
gs — g(i 3--1)—g(1--1, 3); gy — 9(5 3) (1-1, 3 1).) The
edges located by the gradient components according Eq.(1)
are situated at the grid lines yielding interpixel (crack) edges.
The computation can be done by convolving g with the masks
[1 1] and [71 1]7, where T' denotes the transpose. These
above masks are not symmetrically positioned. To evade this
we may choose the central derivative:
ga = limaz—o[9(j + FAx) — 9(j — 5A2]/Az
which results in: g2 = #[g(7 +1) — g(3 — 1)]. Accordingly gy
is defined. The computation of g; and gy can now be done
by convolving g with the masks [—1 0 1] and [-1 0 1],
respectively. The orientation of the edge is defined by:
0 — arctan gy/g« -- 1/27, and the edge strength (magnitude)
by: M, = 4/92 + g?. Alternative definitions of edge strength
that bypass squaring and square rooting reducing computa-
436
tion tim
and the
magnitu
depend
grid. F.
tor may
ideal st:
operato
degrees
M3 is d
is more
to temr
(section
occupy
of the |
ally in t
distanc«
ing all €
look-up
rived fr
version
entiatio
directio
gradien
nent: [-
sign. |
operato
The So
small si
backgrc
lie close
grey va
propert
Since a
tensity
proache
ceeds t
tected |
The ea
tor, we!
much r
a featu
be at le
feature
of the
edges à
ues are
The dis
Gaussiz
remain
image |
4.2 S
In an a
edge fe
model
functio
The lo
tions, t
to the
norm, |
Prewit
idea- a|