18-04-2012, 02:35 PM
Evaluation of Image Quality using Neural Networks
v48-46.pdf (Size: 581.12 KB / Downloads: 44)
INTRODUCTION
TRAINING simulators, such as those employed in
vehicle and aircraft simulation, have been used for an
ever-widening range of applications in recent years.
Visual displays form a major part of virtually all of these
systems, most of which are used to simulate normal
optical views, but many (particularly in military training
simulators) are used to present infra-red, sonar or radar
information. The increasing diversity of applications can
often make the design of new systems difficult, because
there is a lack of previous experience on which to base
decisions.
In graphics-based simulators, the way that
images are presented is critical to the effectiveness of the
system. Therefore, to achieve maximum effectiveness,
the design of the graphics system which generates the
displays should be determined by the requirements of the
application, rather than just by what is already available
commercially. Unfortunately, these requirements are
initially defined in terms of the utility of the images for
performing a given task and this does not translate
directly into a graphics system specification.
RESOLUTION AND SAMPLING
Many training simulators, such as periscope
simulators, anti-aircraft missile simulators, and air traffic
control simulators, are performance-limited by the
resolution of their displays. To match the resolution of
the human eye, these displays would need to be replaced
with ones having over 5,000 scan lines. Unfortunately,
World Academy of Science, Engineering and Technology 48 2008
this would not only increase the requirements of frame
buffer memory and processing during rasterisation, but
would also increase beyond reach the scanout bandwidth
needed to refresh the display; more advanced and greatly
more expensive scanout technology would have to be
developed. Even with such high resolution, aliasing
phenomena would still be detectable if the images were
not properly sampled. Since it is not possible to match
the eye’s resolving power economically, training
programmes must be designed around the capabilities of
simulators as they are. Typically, display resolution and
sampling accuracy affect training tasks such as the
detection, recognition and attitude determination of
vehicles at long range.
A. Resolution Effectiveness
This study concerned the recognition and
attitude determination of three types of fighter aircraft
when drawn at low resolution. The aircraft that were
used are models of the Harrier, Mig 27 and F-16 that
were actually employed in periscope simulators
developed by Ferranti Simulation and Training during the
1980s. They are shown in Figure 1. The objectives of
the study were to determine: (i) what happens to image
quality when the resolution of the image is near its limit;
(ii) how ‘basic’ (i.e. non-anti-aliased) images compare
with anti-aliased images; and (iii) what the differences
are between human and neural network perception. In
achieving these objectives, image quality was first
evaluated using neural networks. This was then
compared with the evaluation of humans, as detailed in
the following sections.
B. Neural Network Evaluation
The neural network evaluation involved training
the image analysis system to categorise images of the
three aircraft according to type and orientation over a
series of low resolutions. The quality of the images over
the range of resolutions chosen is illustrated by Figure 2.
For each resolution, the image analysis system was tested
to determine how well it had learnt (i.e. how well the
images could be categorised from the information they
contained). By plotting network performance against
image resolution, the effect of resolution on image quality
could then be determined. To see if anti-aliasing
provided any significant advantage, the evaluation was
carried out first with basic images and then with antialiased
images.
SHADING ALGORITHMS
In the real world, the way surfaces are shaded
depends on the position, orientation and characteristics of
the surfaces and the light sources illuminating them.
Shading provides essential depth cues to the 3D structure
of objects as well as indicating the materials from which
the objects are made. During the evolution of 3D
computer graphics, to allow objects to be represented
realistically in computer-generated images, a number of
illumination, reflection and shading techniques have been
developed. So far, however, only the simpler methods
have been applied to real-time systems. Space does not
permit the detailed description of these algorithms in the
present paper. The reader is referred to any good
textbook on computer graphics for details, for example
[Foley et al., 1990].
CONCLUSIONS
The studies presented here show that neural
network analysis in general gives the result that might
have been expected intuitively. This is reassuring,
although the superficial conclusion might be that such
analysis can be left to experience and educated
guesswork. There are, however, one or two slight
surprises. Although, on reflection, these can be
understood, they might easily be missed had less
thorough methods been relied on.
The full potential of the method can only be
realised when it is applied to real life requirements
analysis. The facility to define the limits of
performance in such a case, and to flush out any
subtle problems, should prove invaluable.