TL;DR. Most uncooled camera chips give you maybe 10 or 11 bits of dynamic range, and light is subject to Poisson noise, meaning the brighter a pixel, the noiser it is in absolute (not relative) terms. If you have to solve a big giant matrix inversion to do the job of a collimating lens, you're composing each pixel as a sum of many others instead of just itself, some of them being way brighter than the reconstructed image, meaning your reconstructed pixel is always noisier. Cool idea, and certainly has its a
If you have to solve a big giant matrix inversion to do the job of a collimating lens, you're composing each pixel as a sum of many others instead of just itself, some of them being way brighter than the reconstructed image, meaning your reconstructed pixel is always noisier.
Not really.
When you average a large number of samples the noise tends to partially cancel out while the signal keeps adding up. Though the noise goes up with more samples, the signal goes up more, improving the signal to noise ratio.
No, the original poster was more correct. They're not averaging together a bunch of pixels, but applying an inverse matrix , which will weigh pixels differently, and quite frequently can involve very high weights assigned to noisier signals. This can result in an emphasis that amplifies noise. There is a lot of work done on different ways of clipping or modifying such matrix equations to make it slightly less accurate in an ideal world, but much less noisy in the real world.
Also, no averaging of noise will happen if you try to produce images with similar pixel count to the number of detectors.
I agree completely there. (I'd say there is averaging but you're averaging in as much extra noise from other pixels as you're averaging out from multiple samples of the target pixel - and even if the noise were merely proportional the pixel brightness, rather than disproportionate as they get brighter, the bright ones would noise up the dim ones.)
They're not averaging together a bunch of pixels, but applying an inverse matrix , which will weigh pixels differently, and quite frequently can involve very high weights assigned to noisier signals. This can result in an emphasis that amplifies noise.
And that sounds reasonable so I'll defer to your expertise.
Dynamic range? (Score:5, Insightful)
Re: (Score:5, Informative)
If you have to solve a big giant matrix inversion to do the job of a collimating lens, you're composing each pixel as a sum of many others instead of just itself, some of them being way brighter than the reconstructed image, meaning your reconstructed pixel is always noisier.
Not really.
When you average a large number of samples the noise tends to partially cancel out while the signal keeps adding up. Though the noise goes up with more samples, the signal goes up more, improving the signal to noise ratio.
Re: (Score:3, Informative)
No, the original poster was more correct. They're not averaging together a bunch of pixels, but applying an inverse matrix , which will weigh pixels differently, and quite frequently can involve very high weights assigned to noisier signals. This can result in an emphasis that amplifies noise. There is a lot of work done on different ways of clipping or modifying such matrix equations to make it slightly less accurate in an ideal world, but much less noisy in the real world.
Also, no averaging of noise w
Re:Dynamic range? (Score:2)
Also, no averaging of noise will happen if you try to produce images with similar pixel count to the number of detectors.
I agree completely there. (I'd say there is averaging but you're averaging in as much extra noise from other pixels as you're averaging out from multiple samples of the target pixel - and even if the noise were merely proportional the pixel brightness, rather than disproportionate as they get brighter, the bright ones would noise up the dim ones.)
They're not averaging together a bunch of pixels, but applying an inverse matrix , which will weigh pixels differently, and quite frequently can involve very high weights assigned to noisier signals. This can result in an emphasis that amplifies noise.
And that sounds reasonable so I'll defer to your expertise.