Is it possible to reconstruct sound from high-speed video images?
Part of this video was sponsored by LastPass: bit.ly/2SmRQkk
Special thanks to Dr. Abe Davis for revisiting his research with me: abedavis.com
This video was based on research by Dr. Abe Davis and colleagues. I found out about this work years ago and was fascinated by the way he was able to capture vibration information in image-only video. I always imagined the motions of objects would be visible as when recording a tuning fork in slow motion - so deriving sound from high speed images seemed a feasible task. But the reality is much more difficult.
Sound vibrations only cause objects to wiggle by about a micrometer. This is much smaller than a pixel, so the algorithm must understand the characteristics of the image. A move in one direction should cause some pixels to lighten slightly, while others darken - and this behavior is correlated along the edges of the image. So noise can be reduced because it's random over the image and there are enough places to sample that you can get it to cancel out.
Something I'm wondering now is - would it be possible to capture sound in a single image? I'm thinking it would have to be an image of a large object or space because the wavelengths of typical sounds are quite long. Maybe a high frequency sound could be imaged in a suitable medium...
Animations by Alan Chamberlain
Music from epidemicsound.com "Seaweed"