11-02-2010, 11:47 PM
I saw a few threads about this on other forums. Part of my profession is signal processing (I'm a geophysicist) and I think I should comment here.
When a new camera comes out, people react in two different ways. One group says: "18 megapixels on a crop camera, that's crazy! Useless! You'll run into all sorts of trouble with diffraction and such". Another group says: "Great! Now I must have a set of lenses that squeezes 18 megapixels worth of information out of that sensor on every possible shot. How do I yank out the AA filter?"
The more reasonable way to go about it is like this. I like to make prints a certain side, and 12 megapixels works well for this. So, your best bet is to buy a 18 megapixel camera. It's oversampled but that's okay. This way the blur due to the AA filter, Bayer de-mosaicing, or diffraction become non-issues.
Suppose that, as a geophysicist, I want to record signals up to 80 hertz. I'll sample at 4 milliseconds which gives me a Nyquist frequency of 125 Herz. I'm way oversampled but that's okay. Disk space is cheap. I know that my AA filter has ramps, and while it completely suppresses 125 Hz and up, it still attenuates stuff in the 80-125 Hz range. Being oversampled I can use simpler sample interpolation schemes (think of this being analogous to demosaicing). There are other advantages. Never would I want to design my experiment so that I'm close to Nyquist.
Finally, once you're aliased you've got a big problem. Removing Moire is not a trivial task. Mathematically it's a problem involving non-uniqueness. Like phase unwrapping, there are schemes going back fopr decades in the literature but all are cumbersome, all produce their own artifacts, and none are satisfactory. If you can avoid aliasing in the first place then it's worth it.
When a new camera comes out, people react in two different ways. One group says: "18 megapixels on a crop camera, that's crazy! Useless! You'll run into all sorts of trouble with diffraction and such". Another group says: "Great! Now I must have a set of lenses that squeezes 18 megapixels worth of information out of that sensor on every possible shot. How do I yank out the AA filter?"
The more reasonable way to go about it is like this. I like to make prints a certain side, and 12 megapixels works well for this. So, your best bet is to buy a 18 megapixel camera. It's oversampled but that's okay. This way the blur due to the AA filter, Bayer de-mosaicing, or diffraction become non-issues.
Suppose that, as a geophysicist, I want to record signals up to 80 hertz. I'll sample at 4 milliseconds which gives me a Nyquist frequency of 125 Herz. I'm way oversampled but that's okay. Disk space is cheap. I know that my AA filter has ramps, and while it completely suppresses 125 Hz and up, it still attenuates stuff in the 80-125 Hz range. Being oversampled I can use simpler sample interpolation schemes (think of this being analogous to demosaicing). There are other advantages. Never would I want to design my experiment so that I'm close to Nyquist.
Finally, once you're aliased you've got a big problem. Removing Moire is not a trivial task. Mathematically it's a problem involving non-uniqueness. Like phase unwrapping, there are schemes going back fopr decades in the literature but all are cumbersome, all produce their own artifacts, and none are satisfactory. If you can avoid aliasing in the first place then it's worth it.