The basic image (pictorial) data comprise distinct features and 
other non-distinct (i.e., more homogeneous) regions. The former have 
high information content and provide a datum at image matching. They 
can support the neighbouring non-distinct regions (having lower 
information content). 
5. Problems originating in images 
Quality of output data depends on that of the input data and on 
quality of processing. Hence, input data and algorithms should be 
mutually balanced. 
Problems originating in images are partly geometric and partly pic- 
torial (intensity). Geometric sources of disturbances are terrain 
discontinuities, occlusions, roughness, ambigous levels, thin ob- 
jects and periodic features. 
Pictorial sources are poor image quality, different reflectance (of 
the same objects), and lack of signal variation. 
Systems can be provided with self-adaptive capabilities to cope with 
some of these problems, though not with all of them. 
6. On-line versus off-line systems 
On-line operation is time-comstrained; hence, time delay between 
input and output is assumed to be insignificant. Real-time operation 
is a requirement for handling dynamic scenes, such as in some indus- 
trial (robots) and medical fields. Real-time systems are also appli- 
cable to mapping (and other static scenes), though real-time opera- 
tion is not required. 
The properties of the real-time systems include dedicated hardware, 
great data flow, instantenous output, limited flexibility, and rela- 
tively low storage requirement. 
Off-line operation is time-relaxed, and thus suitable for static 
scenes. Off-line systems have the following properties: general 
purpose hardware can be applied, great flexibility, phased and pa- 
rallel operation, but considerable storage capacity is required. 
The individual main stages of the process, and for each stage the 
corresponding operations are discussed below in an ordred sequence. 
III. PRE-PROCESSING 
The aim is to condition raw image data for determining parallaxes 
(i.e., image disparities) and subsequent calculation of DTM data. 
Actual choice of the pre-processing operations should be in accor- 
dance with the strategy for the process as a whole. The list of pre- 
processing operations is open-ended. In the context of a specific 
strategy, however, a selected subset of operations is implemented. 
A tentative list of operations is: merging (or linking) different 
data sets, excluding regions, correcting data, resampling, segmen- 
ting and structuring data, filtering and thresholding, synthesising 
data, image analyses, feature (or signature) extraction, integration 
of external (a priori) information, and predicting approximate con- 
jugacy. 
334