88
REMOTE SENSING IMAGE SEGMENTATION
BASED SELF-ORGANIZING MAP AT MULTI-SCALE
Zhao Xi-an, Zhang Xue-wen Wei Shi-yan a
Dep. of Surveying and Mapping and Urban Spatial Information, Beijing Institute of Civil Engineering and
Architecture, Beijing China, zhaoxian@bucea.edu.cn
Commission VI, WG VI/4
KEY WORDS: Directional Wavelets, Multi-dimension Feature Vectors, SOM Image Segmentation
ABSTRACT:
There exist some problems in multi-scale wavelet transforms based on Mallat fast transform of discrete wavelets» in which image
features may be only acquired in horizontal, vertical, and diagonal directions, and in k-nearest neighbor (k-NN ) segmentation the
relation between image features are not considered. We propose a novel approach in which remote images may be segmented in two
steps based on the directional wavelet transform at multi-scale. Firstly, the images are transformed by the directional wavelet at
coarser scale, multi-dimension feature vectors are constructed, and the images are coarsely segmented by k-NN algorithm. Secondly,
the result of coarse segmentation is used for prior messages , at fine scale the image segmentation is performed again by self
organizing map algorithm (SOM) . Finally the paper approach has been compared with the histogram threshold segmentation
and K-means segmentation. It is showed in the experiments that the novel approach is very effective for extracting the block objects
in remote sensing images.
1. INTRODUCTION
With a large number of high-resolution remote sensing images
being acquired, based on remote sensing images automatically
updated GIS data, as well as the important status in military and
civilian and so to, there is urgent request in automatic target
recognition of remote sensing images. So far, the bottleneck in
the digital photogrammtry, remote sensing technology and
application and machine vision areas is still image extraction
and automatic target recognition. Image extraction and
recognition is the object extraction from the image data by
manual or automatic / semi-automatic way extracting the
identifying target, including artificial and natural features
objectives. From the original images through a variety of
mathematical algorithms extracting, segmenting and classifying
image features, such as the structural features, edge features,
texture features, shadow features is the basis of object
recognition. Interpretation of remote sensing images is
interpretation of images and classifying the target by artificial or
computer methods. Interpretation of remote sensing images can
use all kinds of statistic methods based on gray of the image,
use the classification methods based on image characteristics,
and reason from the image understanding and various
knowledge. These methods can be combined to form mutual
various intelligence methods. Multi-scale wavelet remote
sensing image segmentation (1996; Li, 2000; Chio et al., 2001;
Fosgate et al., 1997; Rezaee et al., 2000; Song, 2004) is based
on Mallat fast transform of discrete wavelets, extracting in
various scales the high-frequency characteristics of images in
the horizontal , vertical and diagonal direction, with the low-
frequency information of images wavelet transform in the same
scale composing wavelet feature tree, in different scales window
segmenting, gradually acquired segmentation results at fine
scale. Image features, in particular texture features, may be in
any direction. Image analysis methods based on Mallat
transform of wavelet can not describe the other direction
features. In the process of remote sensing image object
extraction, sometimes we are interested in features of the
certain direction (for example: along the road heading, building
parallel direction, etc.), here the limitation of multi-scale
wavelet remote sensing image analysis is on the surface. In the
classification and segmentation of remote sensing image, there
exists certain conjunction and influence between images
features and this effect is likely to be nonlinear. Usually
neighboring segmentation algorithms do not take into account
this character of the image features. Bayesian classifier
(Pemkopf, 2005; Romberg et al., 2001; Bouman et al., 1996)
regards image function as realization of Random field X, based
on the maximum similar the features with the same probability
density function been marked or clustering. Bayesian classifier
takes into account the interrelated relations between the images
features, but it needs to know joint probability density between
imaging features. Authors propose a novel approach in which
remote images may be segmented in two steps based on the
directional wavelet transform at multi-scale. Firstly, the images
are transformed by the directional wavelet at coarser scale,
multi-dimension feature vectors are constructed, and the images
are coarsely segmented by k-NN algorithm. Secondly, the result
of coarse segmentation is used for prior messages, at fine scale
the image segmentation is performed again by SOM (Doucette
et al., 2001; Haykin, 2001; Kohonen et al., 1996; Kohonen,
1997). SOM is a two-tier network architecture, including input
models layer and competition output layer. Feature vectors of
input image in the process of network disposing the winner will
be priority marked, classified and until competition level output.
The character of SOM side inhibition not only take into account
the non-linear relationship between the images features, and
also make sure the winners priority marked. In the paper
experiment uses histogram threshold segmentation, K-means
segmentation and multi-scale self-organizing image
segmentation for comparing. It is showed in the experiments
