Vision chip shines
Technology Research News
The video cameras and complicated image
processing software that are used to give machines the ability to see
are relatively bulky and expensive. Many research teams are working toward
a better solution -- eyes-on-a-chip.
Researchers from the State University of New York at Buffalo and
Stanford University have built a silicon retina that uses a timing signal
to mimic a form of data compression performed by biological eyes, and
transmits high-speed optical rather than electrical output.
The silicon retina could be used to give small robots a better
understanding of their visual environment, according to Albert Titus,
an assistant professor of electrical engineering at the State University
of New York at Buffalo. The electronic retina could also be used in smart
sensors and remote monitoring cameras, where its ability to sort out important
information would allow reduced amounts of data to be analyzed, transmitted
Like its biological forerunners, the electronic retina processes
the larger amount of data that makes up an image in order to transmit
a smaller amount of key information. The silicon retina provides information
about the edges of images rather than a whole picture. Edge information
is usually sufficient for detecting and tracking objects.
The device's pixels are an array of light detectors made from
metal oxide semiconductor. The array takes in an image, processes the
information, and passes the compressed output to a liquid crystal spatial
light modulator on the chip. Spatial light modulators pattern light, in
this case allowing it through in positions corresponding to pixels that
generate an electrical "on" signal.
The spatial light modulator enables output from each receptor,
or pixel, to be transmitted optically, which allows the process to take
place nearly in real-time. "Optical output... allows for maximum parallelism
of the data output, and requires no wires to send the data," said Titus.
Otherwise, with electrical output, an array of 4,096 pixels would
require 4,096 output wires from the chip, or it would have to slow down
the process by sending more than one output per wire, said Titus. "These
are both unrealistic approaches," he said.
Biological retinas use two types of light receptor cells -- rods
and cones -- to convert optical energy into electrochemical responses
that can be processed by nerve cells. Cones are sensitive to color, and
work best in bright light. Rods allow for vision in dim light.
Three other types of cells -- amacrine, bipolar and horizontal
-- work together to share signals between receptors, and to transmit the
signals to nerve cells. The result is the light pattern the retina picks
up gets transformed, or filtered, into a more concise set of information
for the ganglion cells that make up the optic nerve, said Titus. "There
are anywhere from 10 to 1,000 [times] fewer ganglion cells than receptors,
so there is a significant amount of data compression that occurs between
the light input and what is transmitted to the brain," he said.
One form of retinal data compression is "a response that corresponds
to the edges of objects," said Titus. "If you break an input image into
objects represented by just their edges -- changes in intensity -- then
you remove a lot of information from the image, but you still have quite
a bit of information about the scene," he said.
The researchers' design models the function of the receptors and
the bipolar, amacrine and horizontal cells, said Titus. "Our silicon retina
produces information about the edges and performs edge enhancement based
on motion," he said.
Edge detection is a common capability of artificial retinas. The
researchers' design is unique because it uses a clock signal to synchronize
the pixels, which allows the chip to work efficiently, according to Titus.
A pixel in the artificial retina is about 10 times faster than a photoreceptor
in a biological retina, so it can perform several operations for every
photoreceptor operation, said Titus. This helps the artificial retina
perform edge detection using a relatively small number of pixels, he said.
The chip also draws very little power. Each cell requires less
than one ten thousandth of a Watt to turn on and off at speeds of a few
kilohertz, or thousand times a second.
The researchers have built a prototype that contains 256 pixels,
and are working to make a more complete silicon-based system that can
be used in autonomous robots and smart sensors, said Titus. They're also
aiming to use the silicon retina in cameras for remote monitoring for
safety, identification and biometrics purposes, he said.
The researchers are also working on artificial retinas that do
more than just edge detection, said Titus.
The silicon retina could be used in practical applications within
one to five years, according to Titus. Applications using optical output
will have to wait 10 years or so until optical interconnects are available
for interchip communications, he said.
Titus's research colleague was Timothy J. Drabik. The work appeared
in the August, 2003 issue of Optical Engineering. The research
was funded by Displaytech, Inc.
Timeline: 1-5 years, 10 years
TRN Categories: Computer Vision and Image Processing; Optical
Computing, Optoelectronics and Photonics
Story Type: News
Related Elements: Technical paper, "Optical Output Silicon
Retina Chip," Optical Engineering, August, 2003.
September 10/17, 2003
Display brighter than film
Sponges grow sturdy
Electron teams make
Vision chip shines
VR system grabs 3D
Gold speck highlights
Neural net tracks
Light drives electron
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog
Buy an ad link