DATATRAN.NOV–Data Translation–pm
PC-based image processing
Faster, more flexible boards capture images with increasing accuracy
BY MICHAEL TRAVIS Data Translation Marlboro, MA
Recently introduced image-processing boards show a definite focus on
high-quality image acquisition, speed, and flexibility. These new boards
fall into two groups. The first is basic frame grabbers that rely on the
host CPU for image-processing operations. The second is more-flexible,
higher-performance boards that use special circuitry to accelerate
arithmetic operations.
Capturing the image One of the biggest limitations to using a frame
grabber as part of an image-processing system has been the inaccuracy of
the digitized image. This inaccuracy is caused both by time-base errors
introduced while the image is being digitized and by noise. Although the
A/D conversion process has been improved and A/D converters have shrunk to
make room for more components on a board, the real strength of these
frame-grabber boards is the flexibility of their front ends. Several
boards now have circuitry that capture an image that is actually cleaner
and more accurate than the incoming video signal. Features to watch for in
this area are adjustability of the input signal range, special sync
clock-generator circuitry, adjustable input filters, and real-time frame
averaging on input. The main objective here is to get more signal and less
noise in the captured image.
Adjusting the input signal range
A weak video signal can be adjusted in two ways. The first is to adjust
the offset and reference points of the A/D converter to match the input
voltage range of the incoming signal. Ideally, the A/D's offset should
match the lightest point of the image, and its reference should match the
darkest point of the image. Although popular, this method has a distinct
disadvantage: it cannot be used to adjust for an effective gain of greater
than 2. Gain adjustments of greater than 2 produce an increase in noise
that is directly proportional to any further increases in signal. This
method, therefore, is not recommended for digitizing low-light or
low-contrast images. The second method, which provides a better S/N ratio
than the first, uses a programmable gain amplifier to boost the incoming
video signal. Some newer boards use this type of range adjustment because
it provides an excellent means of adjusting for extremely low light
conditions. These boards use a gain amplifier to boost the signal level in
increments of 2, 4, or 8, or to reduce the signal by half before the A/D
conversion on the board (see Fig. 1). The offset and reference are then
used to fine-tune the adjustment. This method can provide up to 16,000
ranges on input.
Locking on to unstable inputs To get an accurate measurement of a video
signal, both the amplitude and the position in time of each portion of the
waveform being sampled must be precisely measured. Timing, therefore, is
everything. One of the biggest problems plaguing image-processing users is
that of grabbing images from unstable or intermittent video sources like
paused VCRs, asynchronous cameras, and rapidly multiplexed signals. The
standard video timing circuit on image-processing boards has been a
phase-locked loop that can't lock on to unstable signals. The
phase-locked loop typically uses a phase comparator to match the incoming
signal to a reference. This method results in a minimum horizontal timing
error of 20 ns in even the most stable video signal. In worst cases–such
as VCRs in pause mode-the timing error with a phase-locked loop will be
far greater. At NTSC frequencies, a single pixel width on a 640-pixel line
has a duration of 80 ns. Here, using a phase-locked loop as a timing
reference can result in a 25% value error for each pixel sampled. For
higher-resolution frame grabbing, the percentage of error increases
proportionally with the increase in frame-grabber resolution. This error
in the sample value is called pixel jitter. Certain newer boards, like
Data Translation's DT3851 (see Fig. 2), use a special programmable timing
circuit, or digital clock, to provide extremely precise timing for
grabbing images. The digital clock on these boards works as follows. When
the horizontal sync pulse is read, the digital clock circuit instantly
matches the clock phase to the horizontal sync to within 10 ns (1/8 of a
pixel). This is twice as accurate as the best-case timing achievable with
a standard phase-locked loop under ideal conditions. This digital clock
uses a sync generator running at a constant frequency to provide a
rock-solid time base for image grabbing and can lock to any incoming sync
pulse in less than 10 ns. The generator resets the clock to the leading
edge of each horizontal sync pulse to adjust the phase of each scan line
independently, but it maintains a constant preset frequency. This achieves
at least 50% less noise and pixel jitter than is possible under the most
ideal conditions with a phase-locked loop. For intermittent video sources
like asynchronous cameras, this digital clock circuit allows the board to
lock instantly to the first incoming sync pulse. Phase-locked loops can
require as much as a full field to lock to the signal. They don't provide
the degree of stability required for use with asynchronous or unstable
video sources.
Input resolution Some newer boards support up to 1-K x 1-Kpixel
resolutions. They may also accept variable scan rates. Boards with a
programmable timing circuit also allow the user to input from line-scan
cameras, SEMs, and asynchronous cameras that use nonstandard scan rates
and do not follow the NTSC or PAL standards for resolution. Programmable
timing allows adjusting the board to the exact aspect ratio of the camera,
which may vary by as much as 5%. This adjustment is crucial for
applications where measurements are being taken from the captured image
because it allows the user to acquire an image with perfectly square
pixels. By capturing square pixels, the user ensures that the image being
processed has the same aspect ratio as the original object.
Analog vs. digital cameras
Several image-processing boards input digital as well as analog video
signals. Digital cameras are most useful where more than 10 bits of
monochromatic data are required. An 8-bit monochrome board with a highly
adjustable front end that includes a gain amplifier can provide resolution
effectively equivalent to 10 bits of data. The 8-bit board saves the user
the considerable expense of all-digital equipment. However, for
situations requiring more than 10 bits, digital input is the best
solution. Also, where there is a significant distance between the video
source and the host computer, a digital input eliminates the noise and
shielding problems associated with the transmission of an analog video
signal.
Processing the image Image grabbing and image processing have
traditionally been two separate functions. An image had to be digitized
before processing operations could be performed on it. A new breed of
boards has special on-board processors to accelerate arithmetic functions
and perform certain filter operations in real time or near real time (see
“???,” by Spectrum Signal Processing). This filtering enables the user, in
some cases, to see the effect of an operation on an image before it is
captured. Products from Data Translation, Imaging Technology, Inc., and
Matrox Electronic Systems (see table) use on-board arithmetic logic units
(ALUs) to perform real-time frame averaging to reduce noise in the image.
This technique can improve the signal-to-noise ratio to where it is
actually better than the original video signal (see Fig. 3). Some products
can also generate image histogram data from the incoming video signal in
real time. Another recent innovation is an image-processing board that
either acts as a display board for the host computer or combines the host
computer's VGA display signal with image data from a frame buffer on the
board. The board allows users to view the application interface and image
data on a single screen. This method, preferred by many users, can be
implemented in two different ways. Key to the method are variable display
resolutions of both the screen interface and the buffered image, overlay
buffers that let the user superimpose graphics onto the image without
destroying image data, and an on-board TMS34020 digital signal processor
to enable real-time scaling of captured images to the display. The two
ways of seeing the application interface and image data on a single screen
are VGA passthrough and on-board graphics generation. VGA passthrough uses
the VGA card in the host computer to generate the graphical interface (see
Fig. 4). In this scheme, the system VGA card is connected directly to the
frame-grabber board. Screen graphics generated by the VGA card are then
passed to the image-processing board where they are combined with image
data from the image buffer. The biggest disadvantage to this method is
that throughput limitations of the interconnect between the VGA card and
the frame-grabber card limit the VGA display resolution to 800 x 600
pixels. On-board graphics generation uses either an on-board VGA chipset or
VGA emulation with a custom display driver to drive an on-board TMS34020
chip. One advantage to using this method is that it frees up an extra slot
on the computer, although many machines now ship with a VGA controller
built onto the motherboard. The key advantage to this is that it allows
both higher display resolutions of up to 1,024 x 768 and nondestructive
image scaling when combined with a double-buffering scheme.
Top-of-the-line offerings from the three largest contenders–Data
Translation, Matrox, and ITI–all have a Texas Instruments TMS34020 chip.
Most major board suppliers also have some type of direct interconnect
between the base image-processing board and a coprocessor board. This
direct connection allows users to optimize the functionality of the board
to suit specific requirements. The Data Translation boards support the
DT-Connect open bus interface, which connects to boards from third parties
as well as their own boards. ITI boards support the VISIONbus, which
connects to their own boards and those of third parties. Matrox provides
IMAGE-Bus, a direct interconnect that allows connection of their own
processing boards in a daughterboard configuration.
Software support These boards all ship with basic drivers and libraries
that allow OEMs and systems integrators to package them into systems.
Manufacturers are also beginning to develop software oriented toward the
end-user rather than programmers and systems developers. One such product
is Global Lab Image for Microsoft Windows. Produced by Data Translation,
Global Lab Image directly supports Data Translation boards, but it doesn't
require any special hardware beyond a super VGA display.
CAPTIONS:
Fig. 1. Using a gain amplifier before adjusting the range on the A/D
converter allows precise adjustment of a board's input to the range of the
A/D converter.
Fig. 2. The Data Translation DT3851 frame-grabber board uses a special
programmable timing circuit, or digital clock, to provide extremely
precise timing for grabbing images.
Fig. 3. Noise in each successively averaged frame is reduced by half over
the previous frame.
Fig. 4. Single-monitor display options include VGA passthrough and two
types of on-board graphics generation. Double buffering allows the
real-time image to be scaled to the display.
Products from the following companies were mentioned in this article.
For more information, call the contact or circle the reader service
number.
Data Translation Marlboro, MA Michael Travis 508-481-3700
Dipix Technologies, Inc. Ottawa, Ontario, Canada Vijay Dube
613-596-4942
Epix, Inc., Northbrook, IL Charles Dijak 708-498-4002
Imaging Technology, Inc. Bedford, MA Tom Hospod 617-275-2700
Matrox Electronic Systems, Ltd. Dorval, Quebec, Canada Sales Dept.
800-361-4903
Univision Technologies, Inc. Burlington, MA Bonnie Pietragallo
617-221-6700
Advertisement
