A Single Pixel Camera

BuzzSkyline writes "Scientists at Rice University have developed a one pixel camera. Instead of recording an image point by point, it records the brightness of the light reflected from an array of movable micromirrors. Each configuration of the mirrors encodes some information about the scene, which the pixel collects as a single number. The camera produces a picture by psuedorandomly switching the mirrors and measuring the result several thousand times. Unlike megapixel cameras that record millions of pieces of data and then compress the information to keep file sizes down, the single pixel camera compresses the data first and records only the compact information. The experimental version is slow and the image quality is rough, but the technique may lead to single-pixel cameras that use detectors that can collect images outside the visible range, multi-pixel cameras that get by with much smaller imaging arrays, or possibly even megapixel cameras that provide gigapixel resolution. The researchers described their research on October 11 at the Optical Society of America's Frontiers in Optics meeting in Rochester, NY."

Comments

Voyager worked (still works?) like that

[flyingfsck]

Early space cameras were single pixel and scanned their surroundings by their rotation.

Early fax machines worked the same way, but spun the paper around while the single photocell moved linearly left to right.

Hmmfff - Guess I'm giving my age away...

Other wavelengths

[vespazzari]

I have often thought that it would be really neat if you could get a visual image of radio waves like around for example 2.4ghz and be able to see exactly how your surroundings block/absorb/reflect those wave - in addition to seeing sources of the waves. They mention that might be possible by throwing a different sort of detector instead of a ccd in there? anyone know - would that be possible? do 2.4ghz waves bounce off anything else like light does mirrors, without getting scattered?

Re:Nothing for nothing

[The Panther!]

I think you may be missing the point (har har).

What they are recording is not solely a pixel, I would suspect, but the configuration of mirrors that achieved that point. So, there is a significant amount of information that they can extrapolate from just a random number seed and the final color. The plenoptic function that describes the transfer of light from the environment to the plane of the sensor is 4D. By capturing from many different non-parallel input rays onto a sensor, you can extrapolate a lot about the environment that isn't present in a purely parallel data set.

What I suspect they're goal is, is ultimately getting an array of mirrors onto a consumer-grade camera, and having it take three or four shots in rapid succession, then merge the information gained from each so that the result is more like having a High Dynamic Range image (well beyond the capabilities of any consumer-grade sensor) and use a tone-mapping algorithm to bring it back into a typical 8-bit range per component. It's complicated, but not impossible. Similar such things that are only a year or two old in the graphics community (flash + non-flash images being merged to give good color in low-light situations, multiple exposure images merged for HDR, etc) should come out in a couple of years as automatic modes for color correction, probably even on low-end cameras.

Of course, I still have a 6 year old point and shoot, so what do I know? :-)

Practical uses. Why the stupid comments?

[mattr]

Pretty surprised at all the dumb comments on this story. The scientists involved are not demeaned by consumers being used to cheap megapixel cameras, nor by a secret lab having done something that sounds similar, nor by some patent existing. Slashdot really sucks!

If you are interested you can find out a lot about the really fascinating and cutting edge science of computationally assisted optics, or whatever is the correct term. It is the same field as the people who have been experimenting with giant arrays of cheap cameras, capturing entire light fields that can be sliced in time and space and reprojected later on, etc. It is computers plus physics and a big dose of creativity, which is why it is related to SIGGRAPH too.

Anyway this is interesting and is based on different principles from current megapixel cameras, which is why they think it might improve current cameras too. Just like the way the spaghetti physicists were laughed at by Harvard's igNobel, even though they finally solved something Feynman couldn't crack and have discovered a new method for focusing energy.

Just off-hand, the one pixel camera and compressive imaging theory they have looks very interesting:

• A one-chip computer with transmitter, battery and 1 pixel camera could be worn on your cuffs or collar and capture/assemble from random angles through which it is jangled your entire surroundings.
  
• Could be used if mounted on a wire tip and wire oscillated giving many views of an object for cheap 3d scanning
  
• Camera could include one pixel per range of spectrum, recording a full electromangetic spectrum
  
• They are doing only some simple compression right now. If your current camera could do wavelet compression within the ccd you could certainly get much better pictures and reduce the storage needed.
  
• If current cameras can do all the work needed in 1/500 of a second that means they could be doing a lot more if only compression, transmission and storage are solved, that is what they are working on.
  
• The one pixel camera uses random projections to achieve a certain density of information that seems to be constant throughout the light field they are capturing. This means if they store orientation and time accurately, their data can be sliced at constant quality in any direction, so it is homogenous data which is good. Imagine slicing diagonally through Kraft cheese block or through swiss cheese.
  
• Compressive imaging might help video camera manufacturers wrap their heads around recording at far higher frame rates, including side radio bands for orientation, or combining multiple image sources. Compression in the imaging chip means less data to handle elsewhere.
  
• If you read some of the bibliography (the Architecture one) you will see use of Haar wavelets to reconstruct an image from a 3-dimensional (200,000 voxel) data structure which performs much better than a 2-d one due to the sparseness of data. This paper also talks about the use of bands for which CCD use is impossible.