On Wed, Aug 5, 2009 at 6:01 AM, Larry Colen<[email protected]> wrote: > What I'd love is a program that I could give a JPEG, or a PEF and get > back a measurement of the SNR, though I suspect with images you are > interested in both the amplitude of the noise, and the percent of > pixels that have that much noise.
I don't know if anyone's written something like that (that you can get your hands on), or how easy it is to get at the data through dcraw or libraw or something. But in principle, here's how I would start: You definitely want to work with raw, not JPEG, and preferably non-debayered raw data. I would take a raw frame, and convert it to 3 or 4 single-color images (a "red" image consisting of only the red pixels, and "blue" image consisting of only the blue pixels, and one or both "green" images consisting of green pixels). These will be half the size, in each dimension, of the original image. Take two identical exposures of the scene, and convert them to single-channel images as described above. Subtract them (e.g. red image 1 minus red image 2). You now have a "noise" image. Because of the subtraction, the noise in this image is sqrt(2) higher than in the original image. (I am assuming Gaussian noise throughout.) If you divide the noise image by sqrt(2), and then calculate the standard deviation of pixel values in some region that's uniform in the original image (like your gray square or black square, but not both together), you'll get a measure of the noise in that channel, in that region of the images. For signal-to-noise, you basically want to divide that value by the mean of the original (red, green, or blue, whichever you're working with) in the same region, but I think there are complications here. The "signal" image has to be at zero for "no signal" (no photons). There may be some bias that makes it read more than zero for no signal. In a research CCD, there's an "overscan" region that reads out bins (pixels) that don't actually exist (literally, you have 2048 pixels and read out 2096 or so... the last 48 recorded no photons, since they don't exist). You can subtract off that overscan to debias the frame. I don't know exactly how you would do that with DSLR images... I've never worked with them at this level. Note that with debiased data, the pixel values in dark parts of the scene may be below zero (due to the effect of noise on "true" values near zero). Be sure to use a signed data type. -- PDML Pentax-Discuss Mail List [email protected] http://pdml.net/mailman/listinfo/pdml_pdml.net to UNSUBSCRIBE from the PDML, please visit the link directly above and follow the directions.
