Demosaicing and bit dimensions

Started 8 months ago | Discussions thread
DSPographer
Senior MemberPosts: 2,236
Like?
Re: Demosaicing and bit dimensions
In reply to John Sheehy, 8 months ago

John Sheehy wrote:

Roland Karlsson wrote:

Mark Scott Abeln wrote:

Ah, I defined oversampling as collecting more pixels at the sensor than the lens can sharply deliver.

Oh, I think you need some more to make it interesting. You really want to be able to apply some digital filtering to improve the quality of the image. Then you need at least 4x more than the lens can deliver I assume. Or maybe it is enough with 2 or 3?

Several years back I did some simulations with blurred B&W edges, box filtering them

A box filter is a horrible choice. If you choose a decent filter like Lanczos then you don't need much oversampling.

at various sizes, and upsampling them back to original size and also pixelating them, and it really wasn't until it took 4 pixels inclusive (or 3 exclusive) that luck of alignment had no significant distortive effect on the shape of the edges. That's my standard; virtual analog. Anything less is under-sampling, IMO.

I don't believe the common "wisdom" of Nyquist sampling, and reconstruction. It doesn't work.

As I have shown you before, it most certainly does work. Remember this thread?

http://www.dpreview.com/forums/post/31401611

It doesn't even work perfectly well for audio, but we get away with it with audio because audio is experienced after the fact and we can not do the auditory equivalent of staring at and studying the waveform, and real-world sounds rarely have perfectly stable pitches, and have some frequency modulation, so artifacts are very short in occurrence. For imaging, even the slightest distortion is visible, especially with video, as it shimmers with slight camera/subject registration changes.

Not if you do the processing properly.

You do need to be a bit above Nyquist for reconstruction, ex. 2.5 samples per period. There is also a consideration of the effect on the noise of a pixel's area integration versus extreme oversampling followed by optimal filtering: but the difference becomes negligible above 4 samples per period (2x oversampled versus Nyquist). So the extreme oversampling that you recommend is just not warranted.

Reply   Reply with quote   Complain
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark post MMy threads
Color scheme? Blue / Yellow