1.2 Acquisition Probability
Suppose the correct matching point is (*, j*). In practical
systems, often it is the target point. So, it will be called
‘target point’ in this paper. A good image matching system
should make the matching point as close as possible to the
target point. But because of the influence of noise, they
generally do not coincide even if there exist no geometrical
distortions between the two images. Therefore, we define
acquisition probability P, as the probability where the image
matching point coincides with the target point. For the
MAD algorithm, we have
Py = P{f(3,3) > f(&",5"),V(5,5) € G, (i #7) N (J # m
2
2. EVALUATION OF ACQUISITION
PROBABILITY
2.1 Johnson's Evaluation
Different image correlation models result in different eval-
uation approaches and results. In [2], based on the image
model of pixel-independency, the following evaluation of ac-
quisition probability for the MAD algorithm is derived:
5 1 eo (= — 0)" pi
Pers Le ro uem (3)
and
1 "eo (21 = F1)”
Pole) ren La A
where K = N,_,, N, is the number of all points in the
searching area G; zo, cà represent the mean and variance of
f(i*,j*) = z' respectively (see [1]); similarly, z1, o2 repre-
sent the mean and variance in mismatching point(s, 7) and
fla,ÿ)= 71, x = 7° = Tp:
(3) also can be expressed as
P, i Í^ ioni y X
BE vel. 202
P{f(i, 5) > a, V(hi)E G,li Ai*)N(i#j")}dx (5)
}dz1 (4)
In contrast to [2], we use the expression f(i,7) > 2’ instead
of f(i,j) > z' in (5). In case of equality, at least two mini-
mum points in the searching area G appear and the matching
point can not be determined.
2.2 The Acquisition Probability Based on the
Pixel-Correlation Model
In real images, neighbour pixels generally correlated[4]. There-
fore, the evaluation of acquisition probability by (3) is not
accurate. In order to get a more accurate evaluation, an
image correlation model must be used. In [1], based on
the image correlation model proposed in Reference [4], the
probability density distribution function in a single-valley
area containing the minimum point has been given. Now,
we will give the probability density distribution function in
the whole searching area G. In order to simplify the analy-
sis, we transfer the 2-D searching area into a 1-D sequence
by scanning. So, we have
f(g) = f(6,5) (6)
here g = (¢ — 1)m/ + j; (¢,7) is a mismatching point in the
searching area(m/ x n/).
376
For every f(g), it is satisified with the following Gauss dis-
tribution:
f(g) ^ N (25,02) (7)
where
= 2 jlo}, |],
Tg = TET = mpi c eer = nt *e2]
2 eii 5-7] (8)
2r 2 lg —1 1o 73"
3 = a - DBeia - esp(- Pa P1. Ea ly 4 og
(9)
For random sequence {f(g9)}(¢ = 1,2, ...,N,_,), suppose
their joint probability distribution is the N, ,-dimensional
joint Gauss distribution
PO Ya Ur) =
1 1 T =
aaa zum a FF v -Y»
with
Kz N.i
Y- (91,92, 9k)"
Y= (51.7.7)
2 2
Sz d pg pepe
2 2 2
kc Rau CK
(10)
of, are the covariances of f(k) and f(g):
ot, = 0,0, (11)
Peek Na
So we have
P(f(53) » 2, V(5j)eG, (iz i)n(G # j*)}
Too roo Too
j / wf Dini; ya; - V) dyidya dy,
T r-Tro
Too
Il
+Xo /x+To
1 e d 2
f rare ~ PF Tr - Py (12)
with W = (w,w,...,w),w=2 + To = 2’
According to (5), the acquisition probability for the MAD
algorithm is:
Tow v1 (z — zo)?
P = ex „DL X
: -oo V2T00 pi 208 }
+00 1
ya (2m )K/2(det Y) X
ezp[- (Y -Yyy'(v-YyaYde (13)
W,Y,Y and Y, can be determined by (8) ~ (11) when the
reference image's variance eZ, the signal to noise ratio SNR,
the correlation length Az, A, and m,n, M, N are known. So,
the acquisition probability for the MAD algorithm can be
calculated from (13).
2.3 The Evaluation of Acquisition Probability
for the MAD Algorithm
In image matching, usually N4 is very big. For example,
when the reference image is of size 64 x 64 and the sensed im-
age is of size 32x32, N,—1 = (64—32+1)(64—32+1) = 1088.
= YD jus kM* ba A I — RA a FA
ty
e
In
Se
So
No
red
app
Inc
wit.
ent
pho
hav
Fig
pro:
qua
ima
gra]
mat
ence
Syst
area
get
SNF