Hu Xiangyun
AN APPROACH OF SEMIAUTOMATED ROAD EXTRACTION FROM AERIAL IMAGE BASED
ON TEMPLATE MATCHING AND NEURAL NETWORK
Xiangyun HU, Zuxun ZHANG, Jianqing ZHANG
Wuhan Technique University of Surveying and Mapping, China
National Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing
hxy @iris4 wtusm.edu.cn
KEY WORDS: Road Extraction, Semiautomatic Extraction, Template Matching, Binary Template,
Correlation Coefficient, Optimization, Artificial Neural Network, Hopfield Model.
ABSTRACT
In this paper, we propose a semiautomatic road extraction scheme that is based on template matching and optimization
by Hopfield neural network. In the semiautomatic way, a road is extracted automatically after a series seed points have
been given coarsely by the operator through a convenient interactive image-graphics interface. Attending to accuracy,
robustness, speed and interactivity, we use a binary profile template as the local gray model to speed up the template
matching and build a Hopfield neural network to select the ‘best road way’ form the candidates gotten from template
matching. The template is generated by 'darkness-brightness-darkness' local road feature so it is mainly aim at
extraction of ‘light ribbon like road’. The Hopfield model is built according to the geometric and gray constraint of road
on aerial image. Even there is serious noise, the algorithm extracts road well. The algorithm can extract the road of
which width is from a few pixels to more than 100 pixels. This paper describes the principle and steps of the approach.
Some experimental results and discussions about semiautomatic road extraction are also given.
1 INTRODUCTION
Automatic or semiautomatic extraction of roads from digital images has been a hot research subject for resent years
because road is important geo-information. Unfortunately, there is not any practical system that can automatically and
robustly extract road from different scales and complexity images. The reason is, from the viewpoint of image
understanding, that the extraction is not only low level feature extraction, but also a recognition problem that needs to
use ‘high level knowledge’, whereas how to handle and use knowledge of the scene is an unsolved problem. There are
many embarrassments on overcoming noise or uncertainty and modeling the road feature from kinds of images. So,
from the viewpoint of building a practical system, semiautomatic extraction has gotten special attention. In the other
words, the so-called strategy of human-machine cooperation can conquer the ‘bottleneck’ of recognition and relieve the
boring and laborious digitize work, and compare to fully manual operation, accurately position and delineate the road
feature.
Some algorithms of semiautomatic road extraction have been proposed for resent years. Such as dynamic programming
(Gruen A, Li H H, 1995) and active contour (snakes) (Trinder, 1995, Gruen A, Li H H1996), they are all optimization
problems based on geometric and photometric constraint of road feature. In this paper, we present an algorithm of
extraction based on the optimization by Hopfield neural network that faces to practical production of digital map from
images. The next section provides an overview of our methods and then the steps of the algorithm and experimental
results are given. Finally, the conclusions will be drawn.
994 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
