IMAGE RE-SEGMENTATION APPLIED TO URBAN IMAGERY 
Thales S. Korting, Leila M. Fonseca, Luciano V. Dutra and Felipe C. Silva 
Image Processing Division 
National Institute for Space Research 
Av. dos Astronautas, 1758 
So Jos dos Campos, SP, Brazil, 12227-010 
{tkorting,leila,dutra,felipe}@dpi.inpe.br 
http://www.dpi.inpe.br/ 
Commission III/4 
KEY WORDS: Image interpretation, Segmentation, Computer Vision, Feature detection, Machine vision 
ABSTRACT: 
This article presents a new approach for image segmentation applied to urban imagery. Re-segmentation is so called because it uses 
a previous over-segmented image as input to generate a new set of objects more adequate to the application of interest. Concerning 
urban objects such as roofs, building and roads, the algorithm tries to generate rectangular objects by merging operations in a weighted 
Region Adjacency Graph. Objects whose union generate larger regular objects are merged, otherwise remain unmerged for further 
analysis. To verify the potential of the method, two experimental results using Quickbird images are presented. 
1 INTRODUCTION 
Segmentation of urban remote sensing imagery still is an open 
problem in the image processing community. Such process is 
also an important taks in various image processing and computer 
vision applications, since it represents the first step of low-level 
processing of an image. Many approaches have been proposed in 
the literature (Lucchese and Mitra, 2001). Due to environmental 
conditions, objects are segmented in more than one region, like 
roofs, or streets. This fact compromises further analysis, such as 
classification, or urban planning, since the shape of regions is an 
important feature. However, finding in images regions that repre 
sent objects noticed by humans is a challenging task. The aim of 
this work is to perform re-segmentation, considering by input one 
or more images and a set of segments, from a previous segmen 
tation. Re-segmentation will be shape-based, finding rectangular 
shapes for roofs, and contiguous regions, considering other urban 
objects, such as trees or roads. 
(Chen et al., 2006) define segmentation as partitioning of an im 
age into a subset of fairly homogeneous closed cells. Here we re 
fer to “closed cells” as regions or objects. Each region must have 
its own characteristics such as spectral variability, shape, texture, 
and context, which can be distinguished from its adjacent neigh 
bors. Several algorithms use mainly the region spectral proper 
ties to segment an image. More elaborated approaches also deal 
with contextual and multiscale segmentation (Baatz and Schape, 
2000). 
The details in a high resolution image holds its spectral variabil 
ity and may decrease the segmentation accuracy when traditional 
segmentation methods are used. In urban scenes, one can observe 
that regular shapes such as rectangles can efficiently represent the 
structure of a street, or a roof, for instance. 
Therefore, this paper aims to present a novel approach for high 
resolution image segmentation. The proposed methodology takes 
into account shape attributes besides the spectral ones to produce 
more accurate segmentation. 
This paper is organized as follows. Section 2 presents a brief seg 
mentation review focusing to graph-based approaches and some 
aspects related to urban imagery. Section 3 presents the proposed 
method nemed re-segmentation. We also describe how to build a 
Region Adjacency Graph and discuss the procedure to find regu 
lar shapes on it. Finally, some results and conclusion are shown 
in Section 4 and Section 5, respectively. 
2 GRAPH-BASED RE-SEGMENTATION 
The proposed segmentation method is called re-segmentation be 
cause its input is a previously over-segmented image and a merg 
ing strategy is applied to generate a new regions set. Methods 
such as watershed (Duarte et al., 2006, Felzenszwalb and Hutten- 
locher, 2004, Tremeau and Colantoni, 2000) and region growing 
(Bins et al., 1996) can be used to produce the input segmenta 
tion. Spectral properties of the regions are also input data and 
each region can be connected to its neighbors when succeeding 
topological operation “touch” (Egenhofer and Franzosa, 1991) is 
applied. Such connections are stored in an undirected graph and 
the distance between the nodes, also called weights, is defined by 
the difference of their attributes. 
Subsequently, a graph processing stage is performed. Connected 
regions are merged when their attribute values are similar. The 
graph is built in a structure called Region Adjacency Graph 
(RAG) (Schettini, 1993). The strategy used to join the nodes is 
the principal characteristic of our re-segmentation approach, dis 
cussed with more detail in Section 3. 
2.1 Region Adjacency Graph 
A Region Adjacency Graph is a data structure which provides 
spatial view of an image. One way to understand the RAG struc 
ture is to associate a vertex at each region and an edge at each 
pair of adjacent regions (Tremeau and Colantoni, 2000). Figure 
1 depicts a simple RAG of a synthetic image. 
The RAG can be covered, merged, and partitioned in differ 
ent manners in accord with the expected results. For example, 
(Tremeau and Colantoni, 2000) cover the graph and join a re 
gions set (or vertices) if its spectral distance is enough small.