Oscar Firschein, Visiting Scholar (previously Manager, ARPA Image
Understanding Program)
Seminar abstract
Specialized applications involving image and text databases can sometimes use
techniques not available to general image-text database systems. An unusual
indexing mechanism, the "site model," for an intelligence application was
described. A site model is a stored 3-D model of a region of intelligence
interest, such as a port or a military base. If images and reports are keyed
to parts of a site model, it is possible to use the site model as an indexing
mechanism for both images and text. For example, one might ask for recent
intelligence reports related to a feature of the site or for images that show
a particular roof or a parking area. The talk described the ARPA-sponsored
RADIUS project that uses the site model approach, and indicated the database
aspects of the system. RADIUS uses semi-automated and automated image
understanding techniques to aid in image exploitation. The ultimate goal is
to provide the IA with a seamless system capable of examining images and
answering queries in a database. Several novel technical approaches have been
used in RADIUS:
Model-supported exploitation provides the key to change detection, and
also serves as an indexing mechanism to both image and text databases.
The context-based approach to image understanding explicitly represents
the assumptions made by each algorithm, and uses the context of the current IA
task to select the most appropriate algorithms for solving that task.
Introduction
Image exploitation is carried out by an intelligence analyst (IA) who reviews
images taken from satellites or aircraft using a variety of sensors. The IA
determines significant changes at a geographic site of intelligence interest,
follows and reports on events such as movement of military forces, follows
trends, and monitors and counts objects such as vehicles in a parking area or
fuel drums at a loading dock. The IA often uses other information, such as
maps, previous IA or intelligence reports, and previous imagery, and other
collateral information, to make these determinations.
While previously image exploitation was carried out by analyzing film products
on a light table, there is increasing use of "softcopy," i.e., images in
digital form that are analyzed on softcopy workstations using computer-based
techniques. Because of increasingly effective sensor platforms, the growing
volume of intelligence imagery is overloading the IA, leading to unreviewed
images or lack of timely analysis. RADIUS is designed to alleviate this
problem.
A Future Scenario
Although the current RADIUS system does not have the capability to carry out
the following scenario, the scenario serves as a good indication of what the
next generation of RADIUS might provide:
The IA logs onto the system in the morning. Overnight, the system has reviewed
the new imagery and provides the IA with the following information ordered in
perceived importance:
Event. Site 325 armor movement. Location:Radistan
Count. Site 5948 amount of petrol exceeded Location: Lallia
Change: Site 978 excavation Location: Trufis
New: Site 1342 unidentified object Location: Brazistof
The first item indicates that an important event, armor movement, has
occurred. The second item indicates that a count of petroleum drums that has
been carried out over the past 6 months has now exceeded the threshold alarm
value. The next item is a change to a site that has been detected, namely an
excavation. Finally, the last item indicates an unidentified object at a
site. The IA can now take action on these indicators, reviewing the imagery,
calling up further collateral, and making new tasking assignments of the
system.
Automatic Selection of Image Exploitation Algorithms
The analyst must be able to perform the usual image manipulation tasks such
image rotation, scaling, and enhancement as well as the registration of two or
more images. In addition to these image manipulation capabilities, the analyst
must have easy access to both image and textual databases, and the ability to
use image understanding algorithms to aid in the exploitation.
A basic problem in providing computer aid to the IA for image exploitation is
the proper selection of an IU algorithm and parameters to carry out a
particular task. Without the proper settings, the algorithm may work poorly
or not at all. Because the IA cannot be expected to understand the
characteristics of each IU algorithm, an interface must be provided that
enables the IA to easily specify what he wants the system to do. This
specification is then used to select the algorithm and the appropriate
parameter settings. The basic approach used is to represent explicitly the
assumptions made by each algorithm, and to use the context of the present task
to select the most appropriate algorithms for solving that task. The
mechanism for selection of algorithms and parameters can be a table of context
items to be satisfied versus appropriate algorithms for each row of the table.
A more efficient mechanization is to express the table in PROLOG rules and
have the PROLOG system solve for the appropriate algorithms and settings.
This approach to algorithm and parameter selection is of great general
interest and can be used whenever an unskilled user must utilize sophisticated
algorithms. The basic requirement is that each algorithm be labeled with the
context required by the algorithm and the paarameter settings for that
context.
Database Aspects of RADIUS
The site model can be used as the basic indexing mechanism for both text and
imagery. The user can indicate a location in a site model to select an object
of interest and can then indicate whether text or imagery related to that
object is desired. In addition, the IA can set a trigger on an object in the
site model that will automatically alert the analyst when a new image arrives
that satisfies that trigger. The triggers are stored in the site model
database for each site.
For example, the IA might set a trigger on a parking lot and task the image
understanding system to count vehicles in that parking lot. When a new image
arrives for that site, the parking lot is analyzed and vehicles counted. When
a threshold number of vehicles is exceeded, the system can notify the IA.
Similarly, triggers can be set for new construction, new structures on a roof,
etc. Thus, the following sequences of actions occur when a new image arrives:
The system first registers the image to the appropriate site model so that
each object in the image is identified.
Then the site model database triggers assign IU algorithms to analyze the
appropriate portion of the new image.
The image understanding algorithm carries out the analysis, and detections
or counts are reported to the IA.
Because an important question of interest is "when did a change at a site
first take place?" it is important that the database time-stamp all versions
of a site.
Summary
The RADIUS database ties together both text and imagery so as to provide the
IA a seamless system that can manipulate, register, and exploit imagery, and
can retrieve text information relevant to the site objects. The site model
concept offers a powerful approach for any system that must deal with images
and text related to relatively fixed three-dimensional objects.
Reference
Papers on RADIUS can be found in the Proceedings of ARPA Image Understanding
Workshop, 1995, 1995, and 1996, published by Morgan Kaufmann Publishers,
340 Pine St., San Francisco, CA 94104-3205.