ROAD NETWORK DETECTION BY GROWING NEURON GAS 
Arpad Barsi 
Department of Photogrammetry and GeoinformaticsBudapest University of Technology and 
EconomicsBudapest, Hungary- barsi@eik.bme.hu 
Commission III Working Group III/5 
KEY WORDS: neural network, image analysis, self-organization 
ABSTRACT: 
The growing neuron gas (GNG) algorithm is an excellent self-organization tool, which efficiently combines the graph and neural 
network techniques. The gas formation starts with two connected neurons and during the training both the neurons (nodes) and their 
connections are iteratively added.The most relevant advantage of this technology is that it forms the final graph structure considering 
the input data points regarding the control parameters, but without the strict requirement of any prior hypothesis of the 
graph. Although the graph nodes can represent data points of any arbitrary number of dimensions, in this specific application they are 
taken as two-dimensional ones. The data points derived by simple image processing operations, like thresholding the intensity values 
and by other similar low-level segmentation techniques. The algorithm is fast and can handle even larger set of data points.The paper 
gives an overview about the main self-organizing and unsupervised neural network techniques. It’s followed by the description of the 
growing neuron gas algorithm, and then its application in road network detection is presented. The illustration of the proposed 
method with aerial and satellite imagery also contains accuracy and performance analysis, of course in comparison with other 
detection methodologies. 
1. INTRODUCTION 
Neural networks are currently accepted analyzing tools in data 
processing, also in digital image analysis. The unsupervised 
networks have some basic automatism which prognoses rapid 
spreading. The self-organization is a special type in the 
unsupervised network group. 
The paper will give a short overview about the three main types 
of self-organizing neural techniques, a presentation of their 
simple algorithmic backgrounds, then the description of the test 
areas and the performed experiments are followed, and it has 
been finished with the results and conclusion. The main point of 
the paper is the most flexible self-organizing methodology. 
Because both aerial and satellite imagery is used in the tests, the 
presented techniques can be taken as generally suitable ones for 
remote sensing image analysis. 
2. BASICS OF SELF-ORGANIZING NEURAL 
NETWORKS 
2.1. Self-organizing maps (SOM) 
The self-organizing (feature) map technique was developed by 
T. Kohonen as a special case of unsupervised learning technique 
[Kohonen2001], The network has a neuron lattice; in most cases 
it is a regular rectangular or hexagonal grid. The learning rule is 
very simple: it is the winner-takes-all rule. The neurons of the 
network have weights, stored in a vector m. The input data 
points - also organized into a vector, called x - are shown to the 
network, where a winner is firstly selected: 
c = argmin|x-m.||] (1) 
i 
Then the weights of the winner and its neighboring neurons 
have been increased from epoch t+1 based on epoch t: 
m j (t +1) - m j (/) ■+ h ci (t ) • [x - m j (t )] (2) 
where considering the neighborhood N c {t) 
f«(0 ieN c{ { ) 
[O otherwise 
(3) 
Usually a{t) is a simple, e.g. linear decreasing function. 
The training has two phases: a most raw “ordering” and fine 
“tuning”. The algorithm is the same for both phases, only the 
parameter sets are differing. 
The application of SOM is widely spread from analyzing speech 
to images [Kohonen2001], Data mining takes SOM as a basic 
technique, too [Kohonen2001], 
2.2. Self-organizing neuron graph (SONG) 
The self-organizing neuron graph technique is a generalization 
of SOM’s by introducing graphs for describing the neuron 
structure. The basic learning rule has been kept, only the graph 
feature must be considered, therefore the processing is started 
by a generalized neighborhood analysis, which is used in 
propagating the weight update for the winner’s neighborhood. 
The neighborhood is taken by generalized adjacency (A) of 
graphs [Barsi2003a]: 
. /\ \a{t) A H <a(/) 
M')= „ " r (4) 
0 otherwise