"In the 75 year transition from standard definition to 8K, the pixel capacity of video has increased by a factor of 100. Over the same time period, the computer was invented and the processing, communications and storage capacities of digital systems improved by 6-8 orders of magnitude. The failure of video resolution to develop a rate comparable to other information technologies may be attributed to the physical challenge of creating lenses and sensors capable of capturing more than 10 megapixels. Recently, however, parallel and multiscale optical and electronic designs have enabled video capture with resolution in the range of 0.1-10 gigapixels per frame. At 10 to 100 gigapixels, video capacity will have increased by a factor comparable to improvements in other information technologies."
I am not convinced by their argument that they can save power by using compressive sampling. Sure, they can take less samples, but say if you compress the image by 1000 (so you take only 100M samples instead of 100G), you still need to perform non-linear reconstruction of 100M samples. You can of course do this via deep-learning as they suggest, but I don't see how the overall power consumption will decrease (if all of this has to happen on-chip of course, for off-chip reconstruction this argument might hold).
ReplyDeleteI disagree with your use of the word reconstruction. Its not reconstruktion, but purely guesswork. Some guesswork are better then other guesswork but its still guesswork. Reconstruction is more like constructing again what was there before.
DeleteIt is reconstruction if the error epsilon is infinitesimal small, which for obvious practical reasons it can't be. But indeed, due to the nature of this "reconstruction" (which is more an optimization problem really) you can never really recreate the original image you would get with sampling all pixels.
Delete