Edge Based Region Growing 
M.J.P.M. Lemmens & R.J. Wicherson 
Lab. of Photogrammetry & Remote Sensing, Faculty of Geodetic Engineering, 
Delft University of Technology, Thijsseweg 11, 2629 JA Delft, The Netherlands. 
Commission III 
Abstract We develop a two-stage edge based region growing technique. The first stage consists of 
a half plane edge detector, that acts as a predictor to group pixels into regions. In a subsequent stage 
adjacent similar regions are merged. Since the method is edge based, edges are located accurately, 
which is an obvious improvement compared with the common region growing technique such as split- 
and-merge that tend to dislocate boundaries. Common edge detection techniques find fragmented 
boundaries; our method has the apparent advantage to trace closed boundaries. The method is 
extensively described and many experimental results are presented. One of the immediate application 
area's we see, is the highly desirable improvement of multispectral classification. 
Keywords: Edge Detection, Region Growing, Multispectral Classification, Image Segmentation. 
1 Introduction 
Our ultimate aim is to arrive at an automation of the up- 
dating of topographical databases. This problem can be 
readily approached as an image understanding problem us- 
ing GIS knowledge as a priori information source (Lem- 
mens, 1990). One of the main problems to be tackled is 
the segmentation problem. That means, the partioning of 
an image into meaningful segments that are relevant with 
respect to object space and task domain. It is generally 
recognized that segmentation of natural images is a severe 
problem that is far from being solved, although a plethora 
of segmentation scheme’s are developed last decades. 
We present a new segmentation method based on (1) 
a prediction stage where a half plane edge operator, exam- 
ines each pixel on the presence of an edge and (2) a merging 
stage. The method is entirely based on statistical reason- 
ing, assuming (limited) a priori knowledge about the image 
noise. If the predicted value and the actual value are suf- 
ficiently close together, the pixel is assigned to the region 
under examination. If the pixel doesn't fit into any of the 
surrounding regions, a new region starts. The result of the 
prediction stage is a set of homogeneous regions. However, 
they are not maximal homogeneous. For that a subsequent 
stage is required, where adjacent regions that show suffi- 
cient similarity are merged. 
One of the immediate application area's we see is the 
highly desirable improvement of multispectral classification 
of satellite images. The Bayesian MSC classifiers presently 
commonly used by the remote sensing community classify 
an image only on a pixel-by-pixel base without incorporat- 
ing neighbourhood information. Consequently, these meth- 
ods are severely prone to error. Aggregation of these pixels 
into regions highly improve the classification accuracy (cf. 
Lemmens and Verheij, 1988; Janssen et al., 1990). 
The paper is organized as follows. First we consider re- 
gion growing. Next we treat some smoothing filters. Than 
we present the theoretical background of our edge based 
region growing method. In section 5 we discuss the com- 
puter implementation, illustrated by an extensive example. 
Section 6 gives experimental results. 
793 
2 Region Growing 
Conceptually, region growing concerns: 
1. The splitting of a region Ry, that doesn’t fulfil a pre- 
scribed homogeneity measure, into two or more re- 
gions; 
. The grouping or merging of neighbouring regions R; 
and R,, that fulfil a predefined similarity measure, 
into larger regions. 
A common measure for the homogeneity of R, is the 
adjusted grey value variance 67, and for the similarity the 
absolute difference of the mean grey values of R, and R,. 
The thresholds of the decision rules are usually determined 
by trial-and-error. 
One of the most well-Known region growing scheme’s 
is the split-and-merge technique of Horowitz and Pavlidis 
(1974), (see also, e.g. Ballard and Brown, 1982; Pavlidis, 
1977). The scheme is based on a quad tree representation 
of the image function. À square image segment is split into 
four new square segments if its elements violate the ho- 
mogeneity condition. If for any four appropriate adjacent 
regions the homegeneity condition is fulfilled, then they are 
merged into a single region. This first stage requires a sec- 
ond stage, in which adjacent ’blocky’ segments are merged 
if they fulfil the homogeneity condition. One of the conse- 
quences of the above approach is that a few grey values in 
the square region that deviate from the trend of the other 
grey values, are smeared out. Also, deviating pixels lo- 
cated at the borders of the quad, and which may be due 
to the presence of another region, will not be traced. This 
is the reason for the dislocation of boundaries, which is an 
inherent disadvantage of the split-and-merge scheme. Our 
method doesn’t show this drawback. 
3 Noise Reduction 
Noise and textures may impede severely the performance 
of segmentation. So, we need methods to effectively reduce 
noise and textures without affecting relevant segments.