7A-2-2 
: f! 
(a) Edges detected from 
(b) Edges detected from 
(c) Edges detected from 
Figure 1(c) 
(d) Edges detected from 
Figure 1(d) 
Figure 2. Results of edge detection from images in Figure 1 
at the initial scale 
Figures 3(a) to (d) are edge detection results at refined scale on 
images in Figure 1. 
(c) (d) 
Figure 3. Results of edge detection from images in Figure 1 at the 
refined scale 
(c) (d) 
Figure 4. Segmentation results from images in Figure 1 
Because of noise and dispersive refraction, the image 
segmentation results does not delineate region boundary 
accurately. Gradients carry most of the region boundary 
information in an image and are relatively unaffected by changes 
in image contrast and radiometry. We improve the above image 
segmentation result by treating the regions as topographic 
surfaces and calculating their image gradients. The detected 
watershed lines are then accepted as region boundaries. 
We use grayscale morphological transformation to form gradient 
imagery (Serra, 1982). Suppose original image is represented as 
F , B is a spherical structure element, the gradient image of 
F can be calculated as 
G(F)={F®B)-{FQB) (5) 
where © refers to grayscale dilatjon and 0 grayscale erosion. 
Figures 5(a) to (d) are gradient images of images in Figure 1 
computed using Equation (5). 
3. REGION AND MORPHOLOGY BASED IMAGE 
SEGMENTATION THROUGH MULTI-LEVEL 
THRESHOLDS 
The method introduced above is capable of detecting edges 
representing objects such as roads and other linear features. On 
the other hand, region information (e.g. reasonable uniform 
brightness) can be used to detect features by image segmentation 
that usually divides the image into homogeneous regions. 
The basic method for image segmentation is to set thresholds to 
classify image pixels into different groups. A threshold value 
may be the minimum point of the histogram, e.g. between its 
bimodal peaks (Pratt 1991). However, due to the complexity of 
mapping imagery, segmentation may not be finished using one 
threshold. Instead, a multi-level threshold method was 
developed, in which the thresholding result of a previous step is 
adopted if and only if its histogram forms bimodal peaks. A 
parabola is used at both of the histogram peaks to model a curve 
segment, and a threshold is selected at the intersection of the two 
parabolas. The steps of the multi-level threshold procedure are 
described in (Li et al. 1998). Figures 4(a) to (d) are the multi 
level threshold segmentation result from Figure 1. 
(c) (d) 
Figure 5. Gradient images derived from Figure 1 
Watershed transforfnation of an image is analogous to flooding 
process within a topographic surface (Beucher and Mayer 1993), 
so the algorithm of watershed transformation is easy to 
implement in an iterative procedure. The algorithm searches for 
catchments in the gradient images in Figure 5 by analysing 
minima in regions with various gradient levels. The catchment 
boundaries provide a set of improved lines that can be used for 
feature extraction (Figure 6).