Image Processing Reference
In-Depth Information
Notice how costly the localization phase is in terms of computation time, since it
conducts a local analysis on the entire image. This gives us an understanding of how
useful the focusing phase is, since it will quickly provide the system with more limited
areas of analysis, thus reducing the computation time. This is because localization is
only conducted once for the entire image, even in the case of a sequence of images,
and only for the first concept we are searching for. The result of the segmentation is
mediocre. While vehicles are detected despite how little information is available, the
rate of false positives is significant (44%). The next logical step therefore is to better
control the analysis by using one or several focusing strategies.
10.6.2.
Indirect analysis: two focusing strategies
In order to limit the domain we have to search for in order to find relevant informa-
tion, we have introduced models that express the contextual relations between the var-
ious concepts found in the scene. These relations, attached to a concept, are expressed
in the form of focusing strategies. They are specified in the knowledge base. As we
have said before, we have defined three focusing strategies. It is possible to focus:
“on” a support, “next to” a support, or also “elsewhere” in the image. In this section,
we present the experimental results obtained with two focusing strategies.
The first idea is to search for large structures in the image where the system is likely
to find vehicles. Therefore, in the previous example (section 10.6.1), we chose roads
as our initial objective, knowing that the associated focusing strategies will consist of
searching for vehicles on and next to roads. Although the knowledge base remains
virtually unchanged compared to the previous example, it does, however, include two
additional focusing rules that specify whether to search for vehicles (concept 1) on
(“O”) or next to (“N”) roads (concept 2). The file associated with this knowledge base
is described in Figure 10.14.
{DOMAIN} ROOT:IMAGES/SOURCE/Vol/vol.%02d.gif
EXTENSION:gif: BEGINNING:0: NB:1:
INCREMENT:1: --------------------------
{GOAL}
CONCEPT:2:
--------------------------
{VALIDATION} VAL:0.10
--------------------------
{CONCEPT 1} NAME:VEHICLE: SIZE:1:
L:1:GREY:1:114:255: L:1:HOMOG:3:1:10:
L:1:AREA:0:10:100: D:1:DETECT:1:0:0
{CONCEPT 2} NAME:ROAD: SIZE:1:
L:1:HOUGH:1:0.1:0.5: L:1:GREY:4:80:255:
L:1:AREA:1:250:4000: D:1:DETECT:2:0:0
F:1:FOCUS:S:3 F:1:FOCUS:A:3
Figure 10.14.
Knowledge base for an indirect analysis
of the vehicle concept
Search WWH ::
Custom Search