signal processors (DSPs) are more widely used. In fact, DSPs are evolving into real commodit
y multimedia processor workhorses.
Specialized DSPs
DSPs typically execute most instructions in one clock cycle and often execute multiple instructions per clock cycle. For example, today's multimedia DSP can execute a multiply/accumulate (MAC) instruction -- the fundamental operation for all audio and video processing -- in a single clock cycle. An algorithm implementing a similar instruction on a Pentium processor requires 11 clock cycles. Hence, a $30 66-MHz DSP may have the multimedia performance of a 133-MHz Pentium processor or better. However, with the advent of multimedia extensions (MMX) technology, general-purpose processors are catching up.
To deal more efficiently with real-time digitized streams of video, innovative multimedia processors such as the MicroUnity Mediaprocessor, the Philips Trimedia, the Chromatic Mpact Media Engine, and Texas Instruments' TMS320C6x are adopting techniques used by DSPs (see "Chip Fashion," November 1995 BYTE). Indeed,
they often resemble hybrid CPU/DSPs that blend CISC, RISC, and DSP architectures in fascinating ways.
Just as your PC takes on different personalities depending on whether you've loaded a word processor, spreadsheet, or database, a programmable multimedia engine takes on different personalities by loading DSP programs for filtering, interpolation, edge detection, warping, or image compression. These multimedia chips can sustain 1 to 2 billion operations per second (BOPS) for most multimedia functions and have peak rates for some pixel operations of up to 20 BOPS.
The trend to fast, cheap multipurpose hardware in image processing is especially important to industrial applications, because the ever-increasing demand for cost-effective, high-quality goods can be fulfilled only by very sensitive, automated quality monitoring. For example, many production lines in the steel and paper industries run at speeds of up to 1000 meters per minute. To automatically detect any surface defects on goods, more tha
n 40 digital cameras may be involved. The inspection of surface defects in such an environment requires data transfer rates of up to 600 MBps and a total on-line pixel-processing power of about 90 BOPS.
Flexibility Is Key
On the other hand, in today's volatile and competitive business environments, industrial applications need the flexibility to adapt to fast-changing production methods. With multipurpose hardware and mainstream OSes, this is much easier. "Everybody is trying to get fast graphics without special hardware," notes Peter Mowforth, marketing manager and image-processing connoisseur at The Turing Institute. "There are lots of very good reasons for avoiding special-purpose boards."
Although many people in the industry believe Unix continues to be the best platform for image processing, we are starting to see a move to Windows NT. As Pieter P. Jonker, associate professor in the pattern recognition group at the Delft University of Technology in the Netherlands, says, "T
he best OS for IP is Unix. However, NT is often a cheaper solution than Unix." Another reason for the move toward NT is the wide range of inexpensive statistical-analysis, drawing, display, and standard image-manipulation applications.
Another major challenge facing the image-processing sector is the need for fast prototyping. Thus, we need more flexible and comprehensive image-processing libraries that include modules for cognitive recognition.
One example of such a library is the TimWin toolbox, from the Delft University pattern recognition group. It lets you extract and analyze the information hidden in images. TimWin includes a C-like programming environment and provides functions for the automatic counting of objects, measuring of sizes, tracking of moving objects, and analysis of gray-scale values.
According to a study published last year by the ministry of economics and technology of the German state North Rhine-Westphalia, image-processing tools fall into one of the following categor
ies: tools using industry-standard programming languages such as C or Pascal, programs based on interpreters and proprietary macro languages, and systems that allow for the graphical programming of analysis processes.
Examples of products in the first category are Minos 3.0, from Stemmer PC Systems, and WinTas, from Technische Interaktionen. Programs such as NeuroCheck 1.1, from Data Translation, and Visual Workshop 1.1, from Leutron Vision, fall into the second category. Products in the third category include AdOculos 2.0, from DBS, and Stemmer's WIT.
Process Monitoring
All these tools run in a Windows environment and are geared toward either quality-control process monitoring or surface inspection in industrial applications. They support digital cameras or line scanners and a variety of off-the-shelf frame-grabber boards. In heavy-duty real-time applications, however, a single system can't handle the data streams of several hundred MBps.
Systems integrators and hardware
developers have addressed the challenges of constantly accelerating process speeds, increasing optical resolutions, and more complex classification criteria in the industrial area with proprietary hardware (i.e., ASICs), specialized frame grabbers and DSPs, or arrays of RISC workstations working in parallel.
Parsytec, a German manufacturer of massively parallel computers, is pushing the imaging envelope with its
cognitive automation
initiative. Applications include check reading, mail-order processing, surface inspection, quality control, and machine-vision systems used for sorting waste.
The hardware side of Parsytec's image-processing CC/ipp systems uses standard Pentium Pro or PowerPC motherboards and PCI-based multimedia processors to accelerate pixel operations. In the current CC/ipp version, Parsytec deploys Adaptive Solutions' CNAPS processor. Parsytec is also evaluating the use of such standard multimedia processors as the Trimedia and the Mpact Media Engine. "With the move to ind
ustry-standard media processors, we expect to leverage the pixel performance and at the same time bring the cost of systems down," says Fritz Luecking, test equipment product manager at Parsytec.
When a single-node system does not meet power requirements, the CC/ipp can become a multiprocessor system. In such a case, each node is a fully functional PC, running Windows NT and having its own drives and multimedia processors. (CC/ipp versions that are based on PowerPC 604e motherboards run AIX.)
The high-speed links between the processor nodes sit on a standard PCI board and use an ATM-like (asynchronous transfer mode) technology standardized as IEEE-1355, which was designed to transfer data between the nodes through a router-based network. The result is a hybrid multiple instruction/multiple data/single instruction/multiple data (MIMD/SIMD) architecture that couples Parsytec's massively parallel systems with the immense pixel-operation capabilities of new multimedia processors.
The CC/ipp syst
em operates a multitude of routine image-preprocessing functions on these multipurpose multimedia processors to recognize and classify patterns such as defects in pure software on very fast parallel computers. "The combination of SIMD and MIMD brings the price/performance ratio to a level that so far only dedicated, nonprogrammable hardware could achieve," says Parsytec's Luecking, "while at the same time gaining software flexibility and time-to-market benefits."
Develop Applications
To develop imaging applications for CC/ipp machines, Parsytec provides an application framework called ParsyFrame, which includes general data flow modeling as well as direct access to multimedia-processor functions. "Because the application logic is implemented in software, it's freely programmable and therefore much more flexible," a Parsytec spokesperson notes.
"The use of Windows NT on a parallel, real-time platform is a real challenge," says Luecking. "Memory requirements are far beyond the dem
ands of real-time kernels." To circumvent these problems, Parsytec put much effort into making ParsyFrame available for NT. Says Luecking, "Now any communications backbone, even standard Ethernet, can be used to form a parallel image-processing system based on NT."
The use of image processing in industrial applications is growing. This growth will accelerate as more sophisticated and flexible image applications move out of the research labs and into factories.
The separation of application logic and routine image processing on multimedia processors will help make these solutions affordable. The appearance of smarter chips that, for example, comprise a complete digital camera on one die or perform pattern recognition (see the sidebar "A Chip That Sees") will enable new kinds of intelligent applications.
Where to Find
Data Translation
Bietigheim-Bissingen, Germany
Phone: +49 7142 95 31 0
Fax: +49 7
142 95 31 13
DBS
Bremen, Germany
Phone: +49 421 33591 0
Fax: +49 421 33591 80
E-mail:
sales@dbs.de
Internet:
http://www.dbs.de
Get the Right Picture
