How can image sensors - the most complicated and
expensive part of a digital camera - be made cheaper and less
complex? Easy: take the lid off a memory chip and use that
instead.
As simple as it sounds, that sums up a device being developed by
a team led by Edoardo Charbon, of the Technical University of
Delft, in the Netherlands. In a paper presented at an imaging
conference in Kyoto, Japan, this week, the team say that their
so-called "gigavision" sensor will pave the way for mobile phones
and other inexpensive gadgets that take richer, more pleasing
pictures than today's devices. Crucially, Charbon says the device
performs better in both very bright light and dim light, conditions
which regular digital cameras struggle to cope with.
While Charbon's idea is new and has a patent pending, the
principle behind it is not. It has long been known that memory
chips are extremely sensitive to light: remove their black plastic
packages to let in light, and the onrush of photons energises
electrons, creating a current in each memory cell that overwhelms
the tiny stored charge that might have represented digital
information. "Light simply destroys the information," says Martin
Vetterli, a member of the EPFL team.
A similar effect occurs aboard spacecraft: when energetic cosmic
rays hit a cell in an unprotected memory chip they can "flip" the
state of the cell, corrupting the data stored in the chip.
What Charbon and his team have found is that when they carefully
focus light arriving on an exposed memory chip, the charge stored
in every cell corresponds to whether that cell is in a light or
dark area. The chip is in effect storing a digital image.
All very clever, you might say, but why would anyone want to do
that? The answer is that the two types of sensor chips used in
today's digital cameras store the brightness of each pixel as an
analogue signal. To translate this into a form that can be stored
digitally, they need complex, bulky, noise-inducing circuitry.
The charge-coupled device (CCD) sensors used on early cameras
and camcorders, and the cheaper and more modern complementary metal
oxide semiconductor (CMOS) type both operate on a similar
principle. On each, the area that forms an individual pixel can be
thought of as a small charge-containing "bucket". The size of the
charge contained in one of these buckets depends only on the amount
of light falling on it.
In a CCD, the contents of each bucket of charge are "poured"
into the bucket next door, and then the next until the signal
reaches the edge of the chip. There, an analogue-to-digital
converter (ADC) typically assigns it an 8-bit greyscale value,
ranging from 0 to 255. In a CMOS sensor, the charge is converted to
a voltage local to each pixel before being shunted off to an ADC at
the edge of the chip - where it too is assigned a greyscale value
between 0 and 255.
A memory chip needs none of this conversion circuitry, as it
creates digital data directly. As a result, says Vetterli, the
memory cell will always be 100 times smaller than CMOS sensor
cells; it is bound to be that way because of the sheer number of
signal-conditioning transistors the CMOS sensor needs around each
pixel. "Our technology will always be two orders of magnitude
smaller," he says.
So for every pixel on one of today's sensors, the memory-based
sensor could have 100 pixels. A chip the size of a 10-megapixel
camera sensor will have 100 times as many sensing cells if
implemented in memory technology - hence the choice of the
gigavision name.
But don't expect a gigapixel camera any time soon. Unlike the
pixels in a conventional sensor, which record a greyscale, the
cells in Charbon's memory-chip sensor are simple on-off devices:
they can only store a digital 0 or 1, for which read either light
or dark. To build a sensor that can record shades of grey, EPFL
engineer Feng Yang, who presented the Kyoto paper, is developing a
software algorithm that looks across an array of 100 pixels to
estimate their overall greyscale value.
It's a technique called spatial oversampling - and while it's
early days, he's getting somewhere. "It's turning out to be a lot
more accurate than the greyscale values you get from regular CMOS
sensors," says Vetterli. "Analogue to digital conversion gives only
poor estimates of the actual analogue light value."
The EPFL team have found that the more binary pixels they have,
the better their chips perform in rendering deep shadow and bright
highlights. "Gigavision cameras do not saturate anywhere near so
easily, so we'll be able to use it for high dynamic range
applications like medical imaging," says Yang.
"This is not pure academic interest," says Vetterli. "We're
hoping to have a big version of a gigavision memory chip fabricated
late this year and working early next."
They'll have their work cut out, observers say. A major problem
they will have to overcome is that of the poor sensitivity of their
pint-sized pixels. Their size means the number of photons that can
be scooped up by each of them will be small - and that can make for
a very noisy signal.
The prospect of producing image sensors as cheaply and easily as
memory chips is bound to attract attention, says Alexis Gerard,
chief executive of the consultancy Future Image in San Mateo,
California, which organises a conference on imaging technology
called 6Sight. "It will be pretty interesting if they can make
these sensors using regular memory-chip-making technology."