The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
2) Generate a random number in every subinterval. 
3) With the assistant of the Yl number’s random com 
bination in step.2 n m individuals can be got. Among 
the n m individuals N ones will be randomly se 
lected as the initial population. 
This initialized approach has do great favor to the group’s 
quality increase. According to the method, similar individual 
won’t be chosen and the population has larger probability to in 
clude optimal solution. With a proper initial result the algo 
rithm’s search efficiency and the probability of convergence to 
global optimum are improved. 
2.4 Fitness Function 
Fitness function reflects individual’s adaptable ability to envi 
ronment, it is the controller of evolution trend of genetic algo 
rithm. During the process of designing genetic algorithm fitness 
function only has to meet non-negative and comparable condi 
tions. Before introducing our own fitness function, we define an 
new variation to describe target’s character target character 
istics value. It can be calculated following formula (4): 
Ratx-±-xP if R& <0.95 
Ovssr-l r Rct-k +L*df 
0.95x-j±rxP others 
(7) 
Where, k x > 0,k 2 >0 and k x + k 2 = 1. t is the current evo 
lution time, T is the maximum evolution time, P is an adjust 
able parameter which is used to limited the scale of crossover. 
dif presents the population difference. 
2.5.2 Custom Crossover Approach: According to natural 
evolution theory, the combination of individuals who have simi 
lar genotypes will greatly increase the probability of pathogenic 
generation. In order to simulate this natural phenomenon, we 
add an classification step in crossover. According to the indi 
vidual’s difference the population is divided into two spaces 
male and female using K-means clustering analysis method, the 
difference between two individuals is measured by formula(9), 
where X t is the vector form of individual X t . 
&(a) 
£ e Ao)/f w 
(4) 
dtf(X¡,Xj) 
if №1*, 
others 
(9) 
Where, Q presents a local area, its central point a is point- 
target. / (a) is the gray value of central point a , is the local 
area point’s gray value . In order to reflect the difference be 
tween target and background we use the gray value contrast of 
target and background, the accumulation step can partially re 
duce the noise effect. Exponent function used here acts as an 
amplifying lens to amplify the difference between background 
and target. Objective function can be defined as: 
N 
t a &(X) = — S(a)) ( 5 ) 
i=1 
This classification method doesn’t base on the individual’s 
fitness value, the two parent space only represents two kinds of 
individuals which own different characters. Therefore excellent 
individuals are distributed randomly in the two species. As we 
do crossover operation the parent are selected separately from 
male and female space. This crossover approach has great ad 
vantages in mining potential high-fitness individual and en 
hancing the algorithm’s search capability. 
So as to enhance the algorithm’s detecting ability, the formu 
las we choose in crossover step is a combination approach of 
mathematical method and heuristic method. The formal one is 
good at generating linear combination offspring of two com 
plementary parental individuals, yet heuristic crossover can 
generate the parent’s linear extrapolation according to individ 
ual’s fitness. Their formula is as follows: 
Where, N is the size of population, is the output value of 
the i th sample’s morphology operation, is the i th sam 
ple’s normalized target characteristics value. Here our fitness 
function can be defined as the reciprocal of objective function: 
fit(X) ~ 
(6) 
2.5 Crossover Operator 
2.5.1 Adaptive Crossover Probability: In this paper, the 
crossover probability is based on the population difference and 
evolution time. The crossover probability Cross is defined as 
follows 
ÍY l =X^+r(X i -Xj) if = rX, + (1 -r)Xj 
\y 2 =X, {Y 1 =(\-r)X i +rX ] 
(10) 
Where r is an random number in(0,1). If fit {X) > fit (X) do 
as formula(a), otherwise formula(b) should be used. 
If offspring meet one of the following two conditions, the 
current crossover operation is invalid, try again. 
1) f, f are beyond the solution space; 
2) The better one’s fitness in f, Y 2 is smaller the 
worse in X t , Xj. 
The better one in parents and filial generation will be selected. 
This crossover operation ensures that excellent genes have been 
inherited by offspring. So a sub-optimal individual in parental 
generation and a sub-optimal individual in the next generation 
will be rejected when choosing the next generation. As a result 
of this operation, the population size will be stable, poor indi- 
1300