are showed in figure 1. In practical application, we can choose a 
different directional wavelet spatial filter g a according to the 
required by equation (11). 
are n-dimensional feature vectors formed by the high- 
frequency details of different frequency bands (direction) and 
the low-frequency information 
(a) (b) (c) 
(d) (e) 
Figure 1.Gaussian directional wavelet extracting the high- 
frequency features of 4 directions. 
where (a)The standard test image 
(b) High-frequency features of 30° 
(c) High-frequency features of 60° 
(d) High-frequency features of 120° 
(e) High-frequency features of 150° 
3. SOM IMAGE SEGMENTATION 
3.1 Self-organizing map 
The model of Fig.3 , introduced by Kohonen , displays the 
structure of self-organizing map. A self-organizing image 
segmentation based on competitive, cooperative learning and 
adaptive adjusting is one of unsupervised classified neural 
networks. A self-organizing map transforms an incoming image 
of arbitrary dimension into a two-dimension discrete map. A 
self-organizing map is characterized by the learning process 
which has no use for the prior information about the correct 
classes of input patterns. The network classifies them by 
statistical characteristics of the input pattern and the output 
neurons of the network compete among themselves to be 
activated or fired , with the result that only one output neuron , 
or one neuron per group, is on at one time in the learning 
process of network. The excited synaptic weight connects input 
pattern layer with competition layer and lateral inhibition 
connections exist between layers. 
Two-dimensional array of postsynaptic neurons . 
Figure 2. Kohonen model 
2.2 Remote sensing images feature vectors constructing 
To construct images feature vectors, select 
TC 
a = k — (k = 0,1,2 ...,m is a fixed 
m 
value, 0° < OL < 360°) , By equation (12) according to 
wavelet transform we can get the form 
x = { A r’ D °» ! a = a„a 2 ,a 3 ,...a n _\ } ]S/SJ 
3.1.1 Competitive Process : Let m denote the dimension of 
the input space. Let an input pattem(eigenvector) selected from 
the input space be denoted by 
X = ( Xj, x 2 , x 3 
The synaptic weight vector of each neuron in the network has 
the same dimension as the input space. Let the synaptic weight 
vector of neuron j be denoted by 
Wj — j %j25 x J3 > xj m ) , y — (1,2,3 • • •, /) 
Where 1 is the total number of neurons in the network. Winning 
neuron is mathematically equivalent to minimizing the 
Euclidean distance between X and Wj 
In the equation, A^ is the low-frequency features information 
when the scale is j. Vectors 
i(x) = arg min{ jX - Wj || } 
(13) 
X jj 9 ^2 j+i I ^ CZ l ,CC 2 ,Ct 3 ,..-Ci n _\ } \<j<j 
According to Eq.(4), i(x) is the subject of attention because we 
want the identity of neuron i. The particular neuron i that 
satisfies this condition is called the best-matching or winning 
neuron for the input eigenvector X.