杏吧原创

WHERE are the security locks that open only for a known face? Or the smart
bombs that target their prey using video input from an on-board camera? Or the
cash machines that deliver in response to a winning smile? The cameras are
there鈥攐n street corners, behind cash machines and even built into personal
computers. No one doubts that computers see. But they find it almost impossible
to do something even the dumbest creature can manage: recognise what that image
represents.

So how can a computer be made to recognise a face in a crowd, match a
fingerprint with thousands in its library, or pick out a camouflaged tank hidden
in dense scrub. The conventional approach has been to break down images picked
up by electronic sensors into picture elements or 鈥減ixels鈥�, and convert the
signals from these pixels into digital data. By comparing the data generated by
one image with the data from another, a computer should be able to recognise
whether they are the same thing. That was the theory, but experts in optical
image processing now say this approach is doomed, not least because of the huge
processing power it demands.

So they are turning their backs on digital processing, and instead are using
an analogue technique that processes an entire optical image in one go. The
technique is called optical correlation and, in principle, it can outperform the
combined processing power of dozens of supercomputers. Researchers are
developing ways of carrying out optical processing in a package the size of a
deck of cards. At present such devices cost of tens of thousands dollars, but if
they could be put into mass production, prices should tumble.

To see how optical correlation works, think of a task like sorting through a
series of images to discover if any of them contains a pattern of dots like the
five on the face of a die. One way to do this would be to start with a picture
of the classic five-dot pattern, superimpose each of the images on the original,
and look for areas where they overlap. Obviously, if the images overlap exactly,
they must match.

But things seldom work out that easily. Even if the candidate and the target
are the same, they might be translated or rotated slightly, or viewed on a
slant. If any or all of these distracting things have happened, the images won鈥檛
superimpose, even if both of them contain the five-dot pattern. Of course,
computers can resize, rotate and squeeze the images until a best match appears.
But this brute-force approach is time consuming. And if it鈥檚 this complex for
simple arrangements of dots, just imagine how difficult it becomes for a human
face in dim light.

Magical essence

It looks hopeless. But what if there were some extraordinary underlying
property of an image that remains the same regardless of its orientation and
positioning鈥攁n 鈥渆ssence鈥� that could somehow be extracted from each image
in turn, and then compared? With this magical essence to work with, there would
be no need to mess around precisely aligning the images. Pattern matching would
suddenly become easy: if the underlying essence of the two images is the same,
they must match. And the good news is that scientists have known for years that
just such a fundamental property鈥攖he 鈥渆ssence鈥濃�攅xists.

To understand what this essence is, first imagine a simple sine wave and then
add to it successive harmonics鈥攕ine waves with frequencies that are
multiples of the original (see Diagram).
The result is a square wave that
has a period equal to that of the original sine wave. Now imagine reversing this
process: start with a square wave, and extract from it the family of sine waves
from which it is built up. This technique is called Fourier analysis, and it is
well known and well understood.

Fourier analyses are concerned only with the shape of the waveform. A shape
can vary either with time or in space and the difference is crucial. Imagine an
image formed by a long narrow slit that is lit from behind. In the same way that
the composite wave described in the previous paragraph is a square wave in time,
the image of the slit is a square wave in space. Beyond the edges of the slit,
the intensity of light is zero. At the edge itself it rises to a maximum, and it
remains at this level until it drops sharply to zero at the opposite edge.

Just as a wave that varies in time can be subjected to Fourier analysis, so
can a wave that varies in space. The harmonic sine wavefronts that created it
are called spatial frequencies and together they make up what is called a
鈥淔ourier transform鈥� of the original wave. And the Fourier transform is the
sought-after essence of the original image. Importantly, the information
contained in a Fourier transform鈥攖he spatial frequencies of the component
waves and their proportions鈥攊s not affected by the position of the
original images. This makes it much easier to compare Fourier transforms than to
compare the images. All a computer has to do is to check whether the Fourier
transforms it is comparing contain the same spatial frequencies in the same
proportions.

A computer charged with the task of looking for a triangle in an image
containing many different shapes can create a Fourier transform and then hunt
for the spatial frequencies that form the characteristic 鈥渇ingerprint鈥� of a
triangle. 鈥淚t reduces a complex problem to a much simpler one,鈥� explains David
Carrott, a program manager specialising in optical correlation at Litton Data
Systems near Los Angeles.

Practical problems

This is all very well in theory, but it still leaves the problem of how
exactly you generate the Fourier transform in the first place. One way is to go
back to good old computer methods and do the job digitally. But for this, the
computer must compare every pixel in an image with every other pixel. For small
images of maybe a few hundred pixels this is reasonably straightforward. But as
the pixel count rises, the number of calculations rapidly gets out of hand. In
some images the pixel count is in the millions.

This problem arises from a fundamental incompatibility between optical
sensors and digital computers, Carrott observes. Sensors produce data from every
pixel simultaneously, while computers have to process it one bit at a time. 鈥淭he
bottom line problem has been that the digital world is always trying to catch up
with the sensor world,鈥� he says.

But there is actually no need for digital processing. This is because in the
right circumstances, an ordinary lens can perform a Fourier transform directly
on an optical image. If the image is imposed on a laser beam鈥攃ollimated
light of a single frequency with all the waves completely in step鈥攖hen the
laws of optics take over. These dictate that where the intensity of light
changes rapidly, the light diffracts just as it would around a physical edge.
This diffracted light interferes with light diffracted from other parts of the
image to create a giant diffraction pattern. In effect, the lens is comparing
every part of the image with every other part: it is an analogue mechanism for
generating the Fourier transform. And instead of requiring hours of calculation
as a digital computer might, it comes up with the answer instantly.

The interference pattern is a kind of map of the spatial frequencies
distilled from the original image. And these maps have some special
characteristics that make it easy for a computer to handle. They are all centred
around the optical axis of the lens, making it easy to accurately superimpose
one map on another.

Another feature is that different parts of the map don鈥檛 represent particular
areas of the original image鈥攖hey represent a particular characteristic.
For example, for light to have reached the outer edges of the map it must have
been diffracted by a large angle. For this to have happened, it must have come
from an area of the original image that behaved like a well-defined edge.
Similarly, light near the middle of the map has only been slightly diffracted,
so it must have come from a less well defined part of the image. Thus the
distance from the optical centre is a measure of spatial frequencies in the
image. Similarly, the angular position of features in the Fourier transform is a
measure of the orientation of the edges in the image that produced them.

Maps and overlaps

Comparing these maps involves superimposing them and looking for areas that
overlap. If they were produced from images that contain the same fundamental
features, then the maps will overlap in a very clear way.

杏吧原创s have known for decades about optical Fourier transforms and their
amazing applications in pattern recognition. The challenge of the past twenty
years has been to find a viable way of exploiting the idea. It has been no
simple task.

Cheap lasers have been a crucial factor in the last few years. Lasers are
important because they generate flat coherent wavefronts. But somehow the image
has to be imprinted onto this wavefront before the Fourier transform can go
ahead.

The obvious way is to make a physical template or mask of the image, and
shine the laser through that. This will work, but it is hardly elegant. Creating
these templates is so fiddly and time consuming that it negates the speed
advantages of optical correlation. What scientists need is a programmable device
that imposes the images onto the laser wavefronts as they pass through it, and
which can then be reprogrammed in a fraction of a second. One promising option
is a very high-resolution liquid crystal display.

Another hurdle is building devices that can record the Fourier transform. The
latest, very high-resolution light-sensitive arrays are up to the job. 鈥淭he
technology is now at the point where things can be built at the card level:
that鈥檚 the main breakthrough,鈥� says Carrott.

Engineers at Litton have been working on optical correlators for military
applications for years. Among the specifications they have had to meet is for
the system to be robust, and for it to use the minimum of power. In 1996, they
came up with a device Litton calls the Miniature Ruggedized Optical Correlator.
Costing about a quarter of a million dollars, it is the size of a shoe box and
can carry out pattern recognition in real time. Its optical pathways are carved
from a solid block of glass, and the electronic screens, lenses and laser are
bonded to the block to prevent vibration. Once calibrated, says Carrot, the
device stays aligned. Litton hopes it will be used inside missiles to hunt for
targets in real time.

Lockheed Martin Astronautics, based in Denver, Colorado, is another company
that cut its teeth on expensive optical correlators for the military. In the
hope that prices are about to fall, it is now turning to industry and the
medical world. According to Scott Lindell, an optical engineer with the company,
the beauty of optical correlators is that they offer tremendous processing power
while needing very little electrical power. The instruments his company is
building now 鈥渁re equivalent to about 40 Cray computers of processing power, and
they run on less power than some light bulbs that are in your house鈥�, he
explains.

Working with a healthcare firm, Lockheed Martin is developing optical
correlation as a way of recognising cancerous breast lesions. The design uses a
digital neural network to analyse the Fourier transforms from the correlator.
This setup already performs up to 30 per cent better than the average
radiologist, and works ten times faster than existing electronic systems. Best
of all, it is only half the price. Another project aims to build a device that
can look for poorly soldered or badly placed components on electronics assembly
lines. It is designed to check 1000 parts and 8000 solder joints in only 20
seconds.

The secretive world of security is another area that has huge promise for the
makers of optical correlators. Hamamatsu Photonics based in Hamakita City,
Japan, has built fingerprint recognition systems for banks and other
security-conscious companies to help identify employees and clientele. Other
manufacturers are close on Hamamatsu鈥檚 heels.

As optical correlation becomes cheaper, the availability of accurate
pattern-recognition systems will have far-reaching implications. At home, people
will hunt through digital video archives for treasured family pictures.
Intelligent entry systems for apartment blocks will allow access only to those
authorised to get in. Security cameras might hunt out a known terrorist鈥檚 face
in a crowded airport. And perhaps the next generation of search engines on the
Web will be able to track down pictures, not just text.

Carrott is convinced that such wonders are on the way. 鈥淭he applications we
see now are just the beginning,鈥� he says. 鈥淭here is enormous potential to
肠辞尘别.鈥�