Algorithm Reveals Objects Hidden Behind Other Things In Camera Phone Images

KentuckyFC writes "Imaging is undergoing a quiet revolution at the moment thanks to various new techniques for extracting data from images. Now physicists have worked out how to create an image of an object hidden behind a translucent material using little more than an ordinary smartphone and some clever data processing. The team placed objects behind materials that scatter light such as onion skin, frosted glass and chicken breast tissue. They photographed them using a Nokia Lumina 1020 smartphone, with a 41 megapixel sensor. To the naked eye, the resulting images look like random speckle. But by treating the data from each pixel separately and looking for correlations between pixels, the team was able to produce images of the hidden objects. They even photographed light scattered off a white wall and recovered an image of the reflected scene--a technique that effectively looks round corners. The new technique has applications in areas such as surveillance and medical imaging."

Fascinating, but with limits

[BillX]

I don't claim to be an imaging expert, but a few odd details about the experimental method jumped out at me. It's been known for some time now that diffusive and other scene-perturbing objects (e.g. grossly distorting 'lenses' such as a Coke bottle) can be nullified using a structured light technique to characterize and effectively 'undo' the perturber. A simple structured light example is to replace the light source with a DLP projector and take multiple images with only one pixel illuminated at a time. More clever implementations can replace the single pixels with speckle patterns, zebra stripes, etc., and replace the 2D imager with a single-pixel photocell. Other neat tricks can then be performed such as reconstructing the image from the POV of the light source [stanford.edu] rather than the imaging device.

The experimentals shown in this paper all seem to have two things in common: 1) the "object" in each case is a backlit, 2D binary pattern on a transparency film or similar, with a relatively small illuminated area, and 2) an extremely narrowband (laser, actually) light source is used. The paper does mention several times that the light source is non-coherent, but it is a laser under the hood. This explains the numerous references to "speckle" in the images, which may leave most readers scratching their heads since things don't normally speckle when looked at through a slice of onion under ordinary light. Speckling is a laser (de)coherence phenomenon [repairfaq.org] where the rays are put slightly out of coherence so as to interfere constructively and destructively.

These things suggest to me that while the paper is definitely interesting, there is no need to worry about the neighbors snapping passable nudes through your shower door or Feds cataloging your grow farm via pictures of a blank wall through your window. This sounds more like a modest extension to what's already been done stirring coherent and structured-light in a pot with convolution and autocorrellation methods.

Since the coherence length of cheap semiconductor lasers (e.g. laser pointers) can be on the order of 1mm or less, it's possible to call even a straight-up laserbeam "non-coherent narrowband light" with a somewhat straight face. Likewise, the quasi-point-sources created using a sparse geometric 2D aperture in transparency film, backlight by the aforementioned source, is pretty close to structured light for practical purposes. The takeaway message is these are very special lighting and "scene" conditions that are not representative of everyday photographic circumstances. So not to worry just yet :-)