2. FEATURE DETECTING LAYER
In the common feature detecting problem, we always
extract out the edges of the image in binary. The
edges represent the discontinuties of the grey level
in the image. But we loss so many important informa-
tion in it such as the contrast of the edges to the
background, the average value in a region with its
neighbour. The feature we extracted out here is some-
what different from the conventional feature. The
feature of every interest point is a group of six
values ín sequence such as: exist edge or not, the
mean value with iis eight neighbour points, mean
square invariance with its eight neighbour points,
medium value with ita eight neighbour points......
The purpose of this is to describe the feature of
interest points in every detail while not deseribe
it in a binary value only to represent whether there
is edge existed or not.
Here, we attribute the feature detecting process as
a recognition process compared with a group of atan-
dard templete. Thia feature detecting layer is form-
ed by a BP network. The structure of the network is
shown in Fig. 3.
Fig. 3. Feature-detecting layer
The input node number : 9 Xx (k=0...8)
The hidden node number : 18 Yj (j=0..17)
The output node number : 8 Zi (i=0...8)
The calculation is done in parallel in the same
layer but consequently from top to bottom between
layers. The input of the network ia the grey value
in one 3x3 window.
The output of the hidden layer is:
Qf Em te 01; m
170 1... „17
The output of the output-layer is :
Zi = fil fni 59-94 (a)
where fi is a non-linear function
fi (ap) = : (8)
1 + e7(9 i- 94)
The network is trained according to the Back Propag-
ation algorithm. The training step of the network is
&s followed :
418
step 1 :
Initiate Wikj » Waji
small non-sero value .
step 2 :
Input the templete image Xp and the expected out-
put value Dj. (Di is got from the standard output
from the templete image )
step 8 :
Shift the window along scanning line and ealeulate
the output of the output- layer Zi.
step 4 :
Adjust the weights and the threshold of the the
neiwork according the followed rule.
The weights and thresholds of the second layer:
915, 03g randomly with
Waji(t+1) = Way (6) + noôg; * Yjyi
* 9 * (Wgjy1(0 - Wgjyi (t-D ) (4)
9;(6+1) 7 940) - » » 69; » Caj (8)
where 511 = Zi * (1-Zi) + (D4-24),
Cai is constant, 1 = 0...6 , j = 0...17,
The weights and the thresholds of the first layer:
Vikj ($81) 7 Wigj (€) * n 9» 54j * Ip
+ 06» Wlggj(t - Vg (t-0) ) (6)
9j (t1) = 0j (8) - » * 9j(0 » C1 (7
where 81; 7» Yi * 5» Ya]
C1j is constant, j = 0...17, k= 0....8
Step 6 :
Compare Z;; with Dj;, if [Zij-Dijl < z, go to
step 2; otherwise the training process ends.
In the actual application, we take 2-0. 1.
After the training process is completed, the network
can be applied to detect the feature of the interest
point. The feature of one interest point is denoted
a8 Fj, j, where | means the position of the interest
point in the Image. j means the type of the feature.
Here, j is from 0 to 6.
3. PATTERN BECOGNITION LAYER
It ie not an easy task to find the corresponding
points between the left image and the right image,
especially when a number of interest points oceurs
in one image but does not oecur in another image.
Therefore, only a number of interest points in left
image may find corresponding interest points in
right image and vice versa. Each interest point in
the matching process should satisfy the uniqueness
constraint.
In this paper, we are supposed that the left image
and the right image has been rectifyed after rela-
tive orientation, so that the search of corresponding
interest points can be done alone the corresponding
epipolar line. The epipolar lines are parallel to
each other. So we mateh the corresponding interest
points in one dimension. The structure of the conti-
nual edges is reflect in the mutual restriction bet-
ween the adjacent epipolar line. That is to say, if
continual edges occur across two epipolar lines, the
