© RICH
an object-
‘a three year
tion
versity of
nt for rapid
research is noi
roperties
/ are
tual
1odelled. Th:
lels made to
1 Automatic
ation and
york. The
abase
t (SNE) data
ite and
imagery.
lution mult
tellite sensors
ion content
action from
odel as the
ar features and
use of
h form the
, to offer à
ext of its
ategy for
Gary Priestnall
» THE NATURE OF THE PROBLEM
Tue problem CREED foams ox raction has proved complex and may benefit from the incorporation of contextual
clues similar to those used by human interpreters of imagery. Often feature recognition algorithms work at local levels,
in a bottom-up fashion and lack the higher level control that would allow a more global understanding of parts of the
image. The overall aim of the project is to extract vector linear map objects from imagery with a minimum of user
intervention. Traditionally, the approach has been to adopt one or more algorithms, many of these employing local
operators such as edge detectors alone to attempt to extract the linear components of the imagery. One aim of the
project is to explore the use of more contextual information to aid the extraction and classification of linear features
from a range of imagery.
Linear feature recognition has often been primarily concerned with pixel-based or region-based attribute-value
representations, such as in Brown and Martin (1995), and Baraldi and Parmiggiani (1994). Linear objects are often
represented in imagery by many pixels in width and as such their shape becomes dominant over the resolution of the
raster image. However, not only is the radiometric character of pixels important, but the shape, dimensions and context
become vital.
The extraction of objects from imagery is a problem common in other domains and is essentially a machine vision
problem. Machine vision has been successful in non-remote sensing applications such as engineering but less common
in remote sensing and GIS where, until recently, it has been limited to land cover classifications. Within the domain of
engineering drawing recognition, object extraction can be very successful due in part to the high contrast and low noise
of drawings on paper, but also because they are created by following an accepted code of practice, namely the drawing
standards. Many of the same rules can be adopted by computer vision strategies. In the field of industrial object
inspection the number of different object types under study is often limited and their shape and dimensions are
predictable. Geographical objects do not conform to the rigorous standards of engineering drawings nor can the same
assumptions be made regarding shape, size and orientation as with industrial inspection. The shape, size, orientation,
complexity and reflective properties of geographical objects vary considerably, however some of the techniques and
ideas applied successfully within other domains can be applied to geographical information. One idea crucial in other
domains is the use of contextual clues such as the relationship between one object and its neighbours, or the
containment of an object within a region. In the case of engineering drawings the containment of a linear feature within
the linework already recognised as representing a machined part will, for example, limit the possibilities for labelling
this linear feature. Also key features such as junctions with arrowheads are vital contextual clues which suggest that
certain line types may be found connecting to these features. The role of similar contextual knowledge in aiding an
object recognition strategy for geographical information, primarily linear features such as roads, railways and rivers, is
the focus of this paper.
3 THE ROLE OF CONTEXT
3.1 Defining context
The ability of humans to recognise objects or parts of objects can be illustrated by viewing the images in Figure 1 from
left to right. At the pixel level in the first image, it is difficult to identify the object, but as the view is widened then
more ‘supporting evidence’ from the immediate context of the object is incorporated and so the object can be identified
more confidently.
Figure 1 ; The introduction of context (© NRSC, 1996)
The word tontext'can be used in several different ways. One use of the word is to describe the existence of
knowledge not only about the object of interest but also about other relevant facts and their relations to the object of
Interest. With respect to image analysis for example, every piece of information about the image (resolution, sensor),
Image conditions (viewing, lighting), object (geometry, radiometric properties, functionality), and relations between
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 741