ng 
copic mode. 
ents, as well 
thm and ob- 
using wave- 
present the 
es template 
ly identified 
rced to pro- 
ject outline 
ng provides 
introduction 
1981]: 
lity in prop- 
out particu- 
termination, 
1 at reliable 
cision in de- 
ite approxi- 
able. 
arators from 
mplex com- 
[Suetens et 
'al methods 
xt extraction 
netric semi- 
TING 
lentification 
task of a type | operator is performed manually on a 
single image, while a special automated digital mod- 
ule performs the tracking task of a type Il operator. 
More specifically, a human operator is used to iden- 
tify an object from an on-screen display of a digital im- 
age, select the particular class to which the current 
object belongs (e.g. road, house etc.) and provide a 
rough approximation of the object outline. Typically, 
this approximation consists of loosely identifying on- 
screen nodes (e.g. corners for houses, breakpoints 
for curvilinear objects etc.) of the outline. Subse- 
quently, these pieces of information are used as the 
necessary approximations for the automatic, precise 
edge positioning task. By repeating this process, any 
objects within an image can be identified and pre- 
cisely positioned. The degree of automation varies 
according to the extent of the required human opera- 
tor contribution (e.g. how many nodes have to be pro- 
vided for successful object extraction and how close 
to the actual outline breakpoints). 
Judging from experience in both analytical photo- 
grammetric data collection and digital image feature 
extraction, such use of a human operator within the 
broader object extraction strategy is considered opti- 
mal. Humans perform the identification task flawles- 
sly and almost effortlessly, and thus, their intervention 
optimizes achieved accuracies without imposing time 
burden. At the same time, the task of precise object 
outline positioning and tracking, which experience 
Shows to be the most time-consuming and error- 
prone part of photogrammetric data collection, is per- 
formed automatically in a fast and objective manner. 
In the next sections we will present the mathematical 
foundation and implementational issues for the fol- 
lowing semi-automatic object exraction methods: 
road extraction from wavelet transformed SPOT 
imagery, 
OU the use of active contour models (snakes) for 
outline extraction, 
least squares template matching, and 
Q globally enforced least squares template match- 
ing. 
3. ROAD EXTRACTION USING WAVELET 
TRANSFORMED SPOT IMAGES 
The semi-automatic strategy for the extraction of 
road networks from SPOT images combines a wave- 
let decomposition for road sharpening with a linear 
feature extraction algorithm based on dynamic pro- 
gramming (Fig. 1). 
Wallis filter preprocessing is used to enhance the 
available image and facilitate subsequent object 
  
  
Image Preprocessing by Wallis Filter i 
Road Sharpening by Wavelet Transformation i 
Manual Selection of Seed Points a 
Road Detection & Tracking by Dynamic Programming 
  
  
Fig. 1: Semi-automatic road extraction strategy 
extraction processes by locally forcing the gray value 
mean and contrast (dynamic range) to fit certain tar- 
get values. Filtering is performed in image blocks of 
kx I pixels by modifying the original gray value g(i.j) 
of a pixel to a new gray value f(j,/), as 
[g(,j) - mg cs, 
(i,j) = 
+ bm + (1-b) m, (1) 
where m.,, m, are the old (original) and new (target) 
block mean gray values respectively and s S, are 
the old and new block gray value standard devia- 
tions. The histogram enhancement parameters c and 
b, which respectively affect contrast and brightness, 
are selected such that image noise is limited. A gray 
value transformation for any pixel is performed by bi- 
linearly interpolating the filtering parameters of its 
four neighboring image blocks [Baltsavias, 1991]. 
After preprocessing, wavelet transformation is ap- 
plied on the image to emphasize features of interest, 
while suppressing other details [Grossmann & Mor- 
let, 1984], [Mallat, 1989]. A particular wavelet ¥() is 
uniquely defined in the frequency domain by specify- 
ing its associated filter function H(w) as 
© © 
V(o) = G(3) (>) (2) 
where 
®(®) = [I H(27Po) (3) 
p=1 
G(o) = 6 “H(o+n) (4) 
147