MPEG2 compression for broadcasting ala Zapex..........
Custom compression culls video streams
August 26, 1997
Electronic Engineering Times via Individual Inc. : The opportunity for broadcasting
digital video content has never been greater. Recent improvements in the broadcast
infrastructure, including satellite TV, DVD and ever-increasing bandwidth on corporate
networks, are making it practical for content providers to easily deliver digital video at
quality levels never previously possible. Broad acceptance of MPEG-2 has created a
standard that has allowed the creative energies of the industry to solve many of the
problems historically associated with digital video transmission.
While early attempts at creating general-purpose MPEG-2 compression technology
simplified the building of compression engines, the task of creating production- quality
engines has remained problematic. General-purpose compression techniques present the
compression engine designer with a number of unresolvable limitations, most notably
lack of optimization for the MPEG-2 standard, long design cycles associated with
working at the chip level and a lack of support for custom compression features.
There are a number of IC vendors who have created general-purpose video compression
processing chip sets that can be modified to do virtually any compression standard in
microcode-MPEG-1, MPEG-2, motion JPEG, etc. The problem with general-purpose
technology is that it is not optimized for any single compression standard. For example,
no chip set can provide the pure processing power required to execute a number of
MPEG-2 compression techniques in real-time. As a result, these chip set compression
systems cannot reach the level of quality and efficiency that technology developed by
Zapex is able to deliver.
In an effort to reduce the cost of MPEG-2 compression, other suppliers have created
single-chip designs that offer limited performance. The real estate of a single chip simply
does not have the horsepower or sophistication to implement the features necessary to
produce optimal MPEG-2 quality and efficiency. This approach is problematic, in that
MPEG-2 compression is for entertainment, broadcast and DVD, where quality is an
overriding issue for content providers while efficiency (bandwidth requirements) is an
overriding issue for broadcasters.
System platform is another important consideration when evaluating different
compression alternatives. In general, there are two basic approaches: a plug-in card that
uses standard workstations such as a PC, Macintosh or Unix system and a dedicated
black-box hardware engine.
In a typical add-in card scenario, an MPEG-2 compression board set is added to a PC in
order to create a complete compression system. Although there are a number of
advantages to the plug-in card approach, such as platform cost, there are significant
drawbacks. Most importantly, the reliability of the entire compression system is
dependent on the reliability of the PC system. When the PC fails, an event that happens
all too frequently, the compression system fails along with it. Given increasing reliance
on MPEG-2 systems for business- critical, revenue-generating applications, this
approach is unacceptable for many environments. Consider the impact of system
downtime on a satellite television company that is broadcasting digital video
programming from analog film content in real-time. Beyond simple customer
dissatisfaction, this case of system downtime would result in tangible lost revenues.
An alternative design approach for MPEG compression technology is based on custom
compression algorithms optimized for the MPEG-2 standard. These software algorithms
can be executed either on a general-purpose digital-signal- processing platform or
through custom hardware. In addition to providing a higher degree of compression
efficiency (which results in both improved video quality and reduced
transmission-bandwidth requirements) the custom algorithm approach also enables the
designer to more easily implement future product generations and incremental product
improvements.
The outcome of this approach is a new type of compression device-the plug-in
compression engine. Unlike board-level compression technology, which typically relies
on a PC platform, or a chip-level compression technology, which must be integrated into
an overall compression system, the compression engine is highly efficient, reliable and
can be easily integrated with other digital video products and transmission components.
Perhaps most importantly, the compression engine is optimized to provide the most
efficient MPEG-2 techniques possible. Two of the most important techniques are unique
to the compression engine approach: macroblock-based and conventional adaptive
field/frame (AFF) encoding.
MPEG-2 compression is based on motion-estimation algorithms, where video with
different motion characteristics requires different encoding techniques. For example,
scenes that have little motion (a television news program, talking heads, etc.) have
compression requirements different from scenes with rapid motion (such as sports
events, camera pans, etc.). In response, the MPEG-2 standard provides the field/frame
encoding technique, where different compression algorithms, field techniques for rapid
motion and frame techniques for stills are used for the different types of scenes.
Conventional adaptive field frame selects one coding type for each picture. This
approach is severely limited because real-world pictures are always a mix of motion and
still regions. Picture a football game where there is rapid motion on the field, pockets of
motion in the stands and little motion elsewhere. The question is, which is the best
encoding technique for this picture?
The answer is neither. It is with this situation in mind that Zapex developed
macroblock-based AFF encoding. With this approach, the compression system makes
the adaptive decision on which compression algorithm to use on a fine 16-x-16- bit
macroblock basis, as opposed to a rough 720-x-576-picture basis. As a result, the Zapex
technology is able to find "still" sections in otherwise "dynamic" pictures and vice
versa.
In the football game example, the players on the field and certain blocks of fans are
encoded using field-compression techniques, while the empty sections of the field and
most sections of the stands are encoded using frame compression. As a result, the
encoder is able to compress the video more efficiently and better utilize available bits.
Another technique used by MPEG-2 compression engines is motion estimation, where
the engine tracks motion by matching macroblocks in neighboring frames. The motion
vector describing the macroblock displacement is used by a decoder to generate a
motion-compensated prediction for the current macroblock.
The prediction error is the difference between the current macroblock and this
prediction. The better the estimate, the smaller the prediction error and the better the
video quality.
Traditionally, motion estimation is done by block matching a 16-x-16-pixel macroblock
against previous and/or future reference frames. The larger the search window and the
more exhaustive the block matching, the better the compression efficiency. One problem
that results from this situation is the processing horsepower necessary to execute this
matching in real-time. As a result, most conventional motion-estimation techniques trade
off the accuracy of motion estimation in order to maintain a relatively large search
window. The motion-estimation pattern is typically done in a hierarchical pattern, where
comparisons are made on a rough grid. While this enables the compression engine to
meet the single-pass, real-time compression requirement, it only considers 1/16 of the
possible motion-estimation matches.
Full search
An alternative approach is full-search pixel-by-pixel motion estimation. A full-search
technique, also known as exhaustive motion estimation, displaces each macroblock a
pixel at a time, over the entire 128-x-64 search window in the reference frame and
conducts a pixel-by-pixel comparison between the reference and displaced macroblocks
over the entire 720-x-576-pixel screen. The best match is then refined to half-pixel
accuracy. With this approach, the full- search compression engine is able to check for
every possible match before selecting the best. The result is better prediction and higher
video quality.
One of the primary design challenges of real-time full-search motion estimation is the
sheer compute power that it requires. A full-search algorithm implemented with a
128-x-64-pixel search window on a 720-x-576 frame at 25 Hz requires an amazing 85 billion
operations per second. Systems based on general- purpose compression chips simply
cannot offer this level of checking in real- time.
A typical off-the-shelf 14-chip compression chip set offers a maximum of 35 billion
operation per second. Although the digital video industry is still in its infancy, market
acceptance is already growing at a significant pace.
In the near future, the Internet and corporate networks will offer a vast new opportunity
for video publishers. Interestingly, as these trends emerge, the importance of video
compression efficiency will grow, rather than diminish.
Copyright 1997 CMP Media Inc.
<> |
