ne
Ent
nce
the
kes
ich
kle
the
the
ent
rst
tio
mal
nal
ent
ent
lily
ing
class.
For definiteness sake, let a Doppler
frequency generated by remote sensi &
surface with a heterodyne-reception
infrared coherent laser radar be the
parameter to be estimated. Segmentation
will be considered to be aimed at
discerning patterns of moving objects of
the scene.
PROBLEM FORMALIZATION
Given: the Doppler image specified in the
form of a M«N frequenoy matrix F=f js a
similar intensity matrix A={a } and
image segment types numbered from O to
R-1. Assume that each pixel (T,y) can
take an arbitrary state Boy"
corresponding to one of the segment type
Nos. Segmentation is aimed at generating
the image Q consisting of the subset Q,
Uu Q,- Q and Q, n Q, = at V t £ J, (2)
where Q,7C(7,9):8,, 7 L,3,L, 4(0,...,B-1).
Ihe segmentation is implemented according
to a postertort probability maximum
criterion
p(FIQ,,~,Q)P(Q,,-,Q) 6)
P(Q,,,Q, |F)= ^ max,
p(F)
where P(Q,,...,Q, |F) - is the probability
of Q,,....0Q, regions presence in image on
condition that F image is observed;
p(F|Q,,...,Q,) - is joint probability
density of all pixels Doppler frequenoies
on condition that image is partitioned
into regions Q, SQ; P(Q,,...,0,) - is
a probability of ::Q
presence in : p(F) - is
unconditional j probability density
of all pixels Doppler frequencies.
regions
Consider the cofactor P(Q,,...,Q,) in
(3). To desoribe laser radar images, use
can be made of Markovian random-field
models (Kelly,1988; Besag,1974; Derin,
1986; Hanson, 1982; Therrien, 1986) where
each region is desoribed by iis own
stationary random process and the
transition from one image region to
another is modeled by a Markovian
process.
Let us use 3, to designate a set of
imu)
states of eight pixels adjacent to (r,y)
and N to symbolize a set of states of
(zy)
all the pixels without (I,y). Now the
Markovian model satisfies
Pre POS,
CN
(zy) FE
eu)
— Te t 1 B
where PS 19 (ay)? is the (r,y) pixel
S state probability on condition that
885
neighbouring pixels have S ay) states,
corresponding to specified image
partitioning.
It is known that the Markovian model
satisfies the Gibbs distribution (Derin,
1986), which ean be written as
1 1
P(Q ss ell )= — € — V Q 5)
J i 5 exp { : 5 of } (
Q
where C is the pixel set termed the
clique which consists either of
individual pixels or of their Oups,
satisfying the condition that if (1,J)ec
and (k,4)€0 for (1,1)#(h,1), then 11,7)
and (k,1) are adjacent pixels. C is the
set of oliques belonging to different
types. Vo(Q,...,Q,) is the function
depending only on pixels of type "c"
cliques, intended for the specified
fragmentation of the image and termed the
potential function. T is the constant. Bo
is the normalization factor.
To solve the above problems, it is
expedient to determine clique types in
accordance with Fig.1.
EBB
# [cul
| M | Æ | type a type b type c
type d type e
Fig.1. Clique Types for 8-Connection
Neighbourhood of Pixel (r,y).
À potential function for the Type 1
clique consisting of two pixels (r,y,)
and (TY) can be specified in the form
; p.f s. y. uv
(6)
where - is the parameter corresponding
to ihe Type I clique. For
individual-pixel cliques, the potential
function can be defined as
V,(æ.w) = à, if Bu (T)
where a, is the parameter asscoiated with
the Type 1 clique. Then the potential
unction for the type "¢" clique, v_(Q),
specified all over the image Q will be a
sum of potential functions (6) or (7) for
the entire image.
The normalization factor HB shall be
0
selected on the basis of the condition