A single-pixel digital camera, scientists at Rice Universitybelieve, will reduce power consumption and storage space withoutsacrificing spatial resolution. The new approach aims to confrontone of the basic dilemmas of digital imaging, namely the huge wastefactor.
Considerthat a megapixel camera will, when you take the picture, capture andmomentarily store a million numbers (the light levels from thepixels). No camera can store that much information for hundreds ofpictures, so an immediate data compression takes place right thereinside the camera. A tiny microprocessor performs a Fouriertransform; that is, it converts the digital image into a weightedsum of many sinusoid waves. Instead of a million numbers, therepresentation of the image can now been compressed into somethinglike 10,000 numbers, corresponding to the most importantcoefficients from the mathematical transformation. These are thenumbers actuallyretained for later processing into pictures.
The Rice camera saves space and energy by eliminating the firststep. It gets rid of the million pixels. Instead, it goes right to atransformed version (about 10,000 numbers rather than a million) byviewing the scene prismatically with a single pixel. No, the lightfrom the object doesn't go through a prism, but it is viewed about10,000 different ways. The light, in a quick succession of glances,bounces off the myriad individually driven facets of a digitalmicromirror device, or DMD(see Wikipedia entry). Themirrors of a DMD (only a micron or so in size) do not image anobject or record data but merely steer light; they can beindividually angled in such a way that the light strikes a photodetector or not, depending on whether the light is representing adigital 1 or 0 at that moment.
The main idea is that the DMD is acting as a sort of analog opticalcomputer. Each time the pixel views the object, a different set oforientations is imposed on the array of micromirrors. And, in aninteresting twist, the Rice camera uses random orientations.Looking like the haphazard splotch of black and white squares of acrossword puzzle, the DMD's surface is reflective here and darkthere; some of the mirrors will faithfully reflect light from theobject to the pixel while others will, in effect, appear black.Then the object is viewed again with a different micromirroractivation pattern; again the pixel will record an overall lightlevel. This process recurs about 10,000 times.
Later, offline on acomputer, the single pixel light levels,along with the micromirror patterns are processed using newalgorithms to reconstruct a sharp image. This isn't quite the oldtype of imaging process, the kind used in X-ray crystallography orCAT scans (which also convert pinpoints of data into images), but anew kind of imaging called compressive sensing that is only abouttwo years old.
To summarize, the acquisition of imaging data is reduced many-fold(saving on data storage), only a single pixel is needed (freeing upvaluable space in the primary detector), and the bulk of theprocessing can be offloaded to a remote computer rather than a chipinside the camera, thusgreatly reducing power needs and extending the usefulness ofbatteries.
Rice researchers Richard Baraniuk (richb@rice.edu) andKevin Kelly (kkelly@rice.edu) say that an additional virtue of thecamera is that with only a single pixel, the detector (a photodiode)can be as fancy as you want. It can even accommodate wavelengthscurrently unavailable to digital photography, such as X-rays,terrahertz waves, even radar.
A working camera prototype has beenbuilt. One of the main tasks is to reduce the time it takes torecord an image; the price for compressing space, pixels, and poweris to spread everything out in time since the cyclops-like pixelmust blink ten thousand or more times to capture the image. AsBaraniuk says, the Rice form of photography is multiplexed in time.The Rice results were reported last week at theFrontiers in Optics Meeting of the Optical Society of America (OSA) held in Rochester, N.Y.
Pictures of the setup and imaging results on theRice Web siteand in theresearch paper (PDF)
Contact Richard Baraniuk (richb@rice.edu) or
Kevin Kelly (kkelly@rice.edu)
Rice University
