The promise of optical computing
Thirty years ago, when researchers were using DEC VAX 11/750 minicomputers to develop code, image processing was a CPU-intensive task. Von Neumann architectures, not specifically suited to high-speed multiply-accumulate operations...
by Andy Wilson
The idea that electrical signals first need to be converted to optical form may be flawed.
Thirty years ago, when researchers were using DEC VAX 11/750 minicomputers to develop code, image processing was a CPU-intensive task. Von Neumann architectures, not specifically suited to high-speed multiply-accumulate operations, would take minutes, sometimes even hours, to process 512 x 512 x 8-bit images depending on the algorithm used.
Wth the advent of early bit-slice architectures and low-cost memory, the speed of performing image-processing algorithms was increased by using discrete multiply-accumulators and look-up tables on Multibus and VME boards. Today, with the use of high-speed CPUs and interface buses, many frame grabbers and image-processing boards have abandoned on-board processing capability, relegating the function to the host computer.
Despite this setup, many image-processing applications still require functions, such as Fast Fourier and multiscale wavelet transforms, to be performed at high speed. To perform such functions digitally, developers can opt for CPUs or faster digital-signal processors (DSPs). Alternatively, they can turn to application-specific processors. In using this hardware, developers must partition their designs to maintain the throughput of the fastest part, a process that may result in effective, but costly, implementations.
Electrically, such designs may still not be fast enough for applications such as target tracking or medical image analysis. To perform these operations, off-the-shelf cameras and frame grabbers are often used with spatial light modulators in hybrid electronic/optical systems. Although this approach is effective, the conversion of electrical images to optical form and back to electrical signals for display restricts system throughput.
What is required to perform true optical image processing, say many researchers, is a means to convert high-speed digital data to optical form. After conversion, a programmable optical processor capable of performing a number of image-processing functions could be used to process these images at light speed. Once processed, optical images would again be converted back to electrical signals and at high speed.
However, the idea that electrical signals first need to be converted to optical form may be flawed. In the early days of machine vision, many systems integrators used cameras with broadcast-standard output. In the design of such cameras, images captured from CCDs are encoded into RS-170 output. At the frame grabber, these RS-170 signals are then re-converted back to digital format for processing—an encoding and decoding process that can add errors to the captured image.
PROGRAMMABLE OPTICAL PROCESSORS
Today, systems integrators are more fortunate. They can purchase cameras with digital outputs that interface to frame grabbers using standard interfaces such as RS-644 or Camera Link. No encoding or decoding is required, allowing images with higher fidelity to be captured, processed, and displayed. Similarly, in the optical domain, placing an optical image processor in the optical path before images are digitized would alleviate the need for any electrical-to-optical or optical-to-electrical conversion, reducing errors and system cost.
Unfortunately, such programmable optical processors still remain the domain of university and laboratory research. If they do emerge as commercial products, they will certainly leverage some of the technical developments currently occurring in telecommunications. But like their spatial-light-modulator predecessors, initial applications will probably be expensive military or medical equipment. Until then, industrial designers will still be faced with overcoming the problems of converting between electrical signals and optical formats when building ultrahigh-speed image processors.