The final section will present the result of
applying the algorithms to the radar image in
Figure 1.
Figure 1. Original Radar Image of an Airfield.
PROCEDURE
EDGE PRESERVING SMOOTHING -> EDGE
DETECTION (MAGNITUDE AND DIRECTION)
—> RELAXATION FOR EDGE REINFORCEMENT
-> THINNING -> CONNECTED COMPONENTS
-» REGION PROPERTY CALCULATIONS -»
EXTRACT CONNECTED COMPONENTS OF THE
AIRFIELD -> BORDER FOLLOWING ->
ELIMINATE PIXELS NOT ON THE OUTER-
MOST BORDERS -> GENERATE FREEMAN CHAIN
CODES -> POLYGON APPROXIMATION
Figure 2. Procedure Used for Extracting the
Airfield
Edge Preserving Smoothing
The purpose of an edge preserving smoothing algo-
rithm is to eliminate noise and to preserve edges
from degradation. The variation of the gray tone
in a neighborhood around each pixel is used to
determine the direction that is most homogeneous.
Smoothing is then performed in this direction.
The particular approach to edge preserving smooth-
ing used in this research consisted of analyzing
the gray tone variations within each 5- by 5-
pixel area in the image. For each 5- by 5-pixel
area, nine geometric figures are formed using the
center pixel. Four of the geometric figures are
pentagons. Four of the geometric figures are
hexagons. One of the geometric figures is a
Square. Each of the four pentagon figures is
formed by using the center pixel and one of the
outermost edges of the 5- by 5-pixel area. Each
of the hexagon figures is formed by using the cen-
ter pixel and one of the outermost corners of the
>- by 5-pixel area. The 3- by 3-pixel square is
formed using the center pixel and its first near-
est neighbors. The pixels associated with each
of the geometric figures are used to compute the
mean and variance of the gray tone for each fig-
ure. The pentagon and hexagon figures each have
7 pixels associated with them. The Square has
9 pixels associated with it. A list of nine means
809
and nine variances is generated from all of the
computations involving the nine geometric figures.
The gray tone value of the center pixel is replac-
ed by the particular mean gray value that is asso-
ciated with the smallest variance. The theory
behind this edge preserving technique was devel-
oped by Nagao and Matsuyama (Nagao and Matsuyama,
1980). The algorithm can also be used in an iter-
ative manner, that is, the output of one smoothing
operation can be used as the input to another.
Edge Detection (Magnitude and Direction)
After edge preserving smoothing has been performed,
an edge detection operator is used to enhance
edges and to compute the direction of each edge.
The edge detection operator used was the Sobel
operator. This operator consists of two 3 by 3
masks. The masks are applied to each pixel to
calculate a magnitude image and a directional
image. The magnitude image is the edge enhanced
image. The directional image contains the dir-
ection of the edge at each pixel. The direction
of an edge is defined as the angle between the
edge and the x-axis. The x-axis extends along the
top row of the image with the origin at the pixel
in the upper left-hand corner. The y-axis extends
downward along the first column of the image. The
magnitude image is computed by taking the square
root of the sum of the squares of the result of
applying the two masks at each pixel. The dir-
ectional image is calculated by taking the inverse
tangent of the ratio of the results of applying
the two masks. Because the direction of an edge
has a 180 degree ambiguity, a convention must be
established to eliminate this ambiguity and estab-
lish a fixed direction for each edge. The conven-
tion used in this research was that the edge dir-
ection was taken in such a way that the darker
side is always on the left when facing in the
direction of the edge.
Relaxation for Edge Reinforcement
The result of applying the Sobel edge operator
yields an image in which some edges are defined
very well, some edges are poorly defined, and some
edges have holes in them. In addition, some large
responses are obtained where there are no edges.
These errors occur because of noise in the origi-
nal image and also because the Sobel edge detector
is not perfect. The purpose of the relaxation
calculations is to enhance edges by increasing the
gray tone value of the pixels that are really on
edges, and to decrease the gray tone value of the
pixels that are not on edges. Initially, the
magnitude and direction of the edge at each pixel
are obtained from the edge detection operation.
The magnitude at each pixel location is divided by
the maximum of the magnitudes over the entire
image in order to define the probability of an
edge at each pixel. The location of each pixel
will be designated by the quantity (i,j), where i
represents the row dimension and j represents the
column dimension. The relaxation process for edge
reinforcement consists of defining a new edge
probability and a new edge angle at each (i,j) in
terms of the old ones at (i,j) and its neighbors.
The neighbors used in this research were the first
and second nearest neighbors. This definition of
neighbors is not a restriction on the basic techni-
que. As explained by Schachter, et al. (Schachter,
et al., 1977) the number of neighbors used in the
relaxation calculations is arbitrary. However, if
more neighbors are used, the result will be longer
computation times. The calculation of the new
EE
ucc
=