adding noise to images
Orsola De Marco wrote on Jun 15, 2006
Hello -
I have obtained arc spectra using 300 sec exposures and I want to know how much I can reduce this rather long exposure time and still get the same wavelength calibration precision.
I presume I can add noise with mknoise (although I have not yet been fully successful in doing that either). Assuming that I can add noise with mknoise and that I find what amount of noise I can add and still get a good calibration, how do I determine what exposure time corresponds to that increased-noise image?
Thanks!
Orsola
I have obtained arc spectra using 300 sec exposures and I want to know how much I can reduce this rather long exposure time and still get the same wavelength calibration precision.
I presume I can add noise with mknoise (although I have not yet been fully successful in doing that either). Assuming that I can add noise with mknoise and that I find what amount of noise I can add and still get a good calibration, how do I determine what exposure time corresponds to that increased-noise image?
Thanks!
Orsola
Francisco Valdes wrote on Jun 15, 2006
Hi Orsala,
Post back if you are still having trouble with mknoise.
Your question depends on understanding your detector. Most detectors are treated as having a Poisson noise model plus a fixed readout noise. Your first question is whether the readout noise is important. Usually you want to expose long enough that it is not. Then the noise increases with the square root of the exposure time.
Your approach of simulating is the most reliable one when complex reductions are involved. You should start with a sample raw exposure that has high signal to noise. Then, as you have done, you would add Poisson noise equivalent to some lower exposure time. Mknoise can do this by exposure time. Then reduce the frame. The "bootstrap" method, which is computationally intense, says to do this many times with the same original exposure and the same test lower exposure time but with different random number seeds. The scatter in your final answer is then interpreted to get the estimate you want. In your case you would want something like 70% of the trials to give you the calibration to be within the accuracy you want.
I hope this helps.
Yours,
Frank
Post back if you are still having trouble with mknoise.
Your question depends on understanding your detector. Most detectors are treated as having a Poisson noise model plus a fixed readout noise. Your first question is whether the readout noise is important. Usually you want to expose long enough that it is not. Then the noise increases with the square root of the exposure time.
Your approach of simulating is the most reliable one when complex reductions are involved. You should start with a sample raw exposure that has high signal to noise. Then, as you have done, you would add Poisson noise equivalent to some lower exposure time. Mknoise can do this by exposure time. Then reduce the frame. The "bootstrap" method, which is computationally intense, says to do this many times with the same original exposure and the same test lower exposure time but with different random number seeds. The scatter in your final answer is then interpreted to get the estimate you want. In your case you would want something like 70% of the trials to give you the calibration to be within the accuracy you want.
I hope this helps.
Yours,
Frank
Last post on Jun 15, 2006