THE EVALUATION OF ACQUISITION
PROBABILITY IN IMAGE MATCHING!
"E | Ding Mingyue?
imee | sut ait ;
SPOT | Institute of Pattern Recognition and Artificial Intelligence
| Huazhong University of Science and Technology
| Wuhan, Hubei 430074, P.R.China
| ABSTRACT:
ries | In modern navigation and guidance systems, image matching is often used as an efficient approach
ay : | to increase the registration accuracy. Acquisition probability of image matching is one of the most
| important parameters in registration accuracy analysis of image matching. It represents the correctness
| of the position estimated by the navigation and guidance system with respect to the real position in
975) | flight. For example, in missile homing guidance, it is the probability of hitting a target. So, it is
et | the main basis for designing a navigation and guidance system. In image matching, Mean Absolute
rche | Difference (MAD) is one of the most often used algorithms. It has a lot of advantages such as high
hier | registration accuracy, high noise robustness and can be easily realized by hardware etc. In this paper,
| first, the acquisition probability for the MAD algorithm is derived based on the image pixel-correlation
z model. Then, in order to evaluate the value of acquisition probability for the MAD algorithm, an
ns", approximation formula is given. Finally, the experiments with different optical aerial photographs
| and infrared remoted sensing photographs have been conducted on a IBM-PC microcomputer system
and a S575 image processing system . By the experimental comparion to the evaluation of Johnson
Et | it is demonstrated that the evaluation of acquisition probability for the MAD algorithm proposed in
Ces | this paper is more accurate and close to the real acquisition probability.
| Key Words : accuracy, image matching, navigation and registration.
the
once
|
|
and |
| 1. DEFINITION M
iple | M-m+i
| 1.1 Image Matching Point m s
Suppose S and R represent sensed image (m X n) and ref- (4,j
| erence image (M x N) respectively . The purpose of image | Be; N
sis matching is to determine the position (ig, jo) where the ref- N-n+i
erence subimage is most similar to the sensed image S by $ G-miij-n-4li
translating the reference subimage R; ; in the searching area
cat G, as shown in Fig. 1. The position (o, jo) is called match- R
the ing point between image S and image R. For the MAD Fig. 1 The matching area of reference image
| algorithm, the similarity between two images is measured
| with the mean absolute difference function f(i,j): The smaller the MAD value, the more similar the two im-
ies "m m ages. Therefore, for the MAD algorithm, the image match-
ive f(üj3)s + > >. |Ri+k-1,j+g-1)- S(k,g) | (1) ing point is the minimum of f(i, j); it can be mathmatically
> MN =19=1 expressed as:
| where 0<i<M-n+10<j<N+n-1 fio, jo) = min_ f(i, 7)
iode
la |
ées |
éal
The work is supported by the Alexander von Humboldt Foundation in Germany
Currently on leave at the Institute for Robotics and Process Control, Technical University of Braunschweig, W-3300
Braunschweig, Germany
