Simulating small and large-pixel Bayer CFA cameras

Started Jul 1, 2014 | Discussions thread
Flat view
JimKasson Forum Pro • Posts: 17,363
Simulating small and large-pixel Bayer CFA cameras

This is a continuation of the sub-thread in this now-closed (due to number of posts) thread:

concerning simulation studies of digital cameras of varying pixel pitches. If you're new to this topic, it might be a good idea to go poke around the previous thread now.

Back already? Good. Before I get started here, let me say that I really appreciate all the suggestions for things I can do to make the sim better. If I don't follow up on yours, it's because I don't have enough time. What I'd really like is one or more people with Matlab on their computer and time on their hands to help extend the model. Please?

OK, enough sniveling. there were several open questions from the last discussion.

1) Am I unduly handicapping the large-pixel sensors by using a crude demosaicing algorithm like bilinear interpolation. Thanks to crames for this pointer:

There's a lot there, and it's going to take me a while to separate the wheat from the chaff. I'm starting with AHD, since I have experience with it in DCRAW, and it appears to be a modern, if not state-of-the-art demosaicer. By the way, I wouldn't know a state-of-the-art demosaicer if I ran into one on the street. I am a demosaicing neophyte, and wish to remain so as much as possible, consistent with making this simulation project worthwhile. A bar that any demosaicer that I use will have to jump over is that there is easy-to-use Matlab code available. Fortunately, as Matlab appears to be the lingua franca of the demosacing community, just as it is of the color science community (which is one of the reasons I'm using it).

Suggestions for demosaicing algorithms to use or not to use are appreciated.

2) What's the best way to upsample images to common, large-printer-oriented sizes? I'm currently using Lanczos-3 in sRGB (gamma = 2.2). The right algorithm will deal will with preserving detail in the presence of noise.

3) What's the best way to downsample images to common, small-printer-oriented and large-screen-oriented sizes? I'm currently using Lanczos-3 in sRGB (gamma = 2.2) with and without a pre-scaling Gaussian blur. The right algorithm will deal will with preserving detail in the presence of noise. I have read somewhere that gamma = 1 may be better for downsizing. Almost as much as in the case of demosacing, I am inexpert on resampling, although I do have more desire to learn thatn in the case of demosacing.

By the way, the model currently runs in the space domain, using convolution for filtering. I have successfully run similar models in the frequency domain, but I don't like to switch back and forth much in the same sim, so let's stay in space for the time being.

4) After some chiding from Eric Fossum, I'm dropping the use of the word "sensel". I'll be using "sensor pixel" or just "pixel" if it's clear in context that I'm talking about the sensor. I just picked "sensel" up a few months ago 'cause I thought it made me sound like I knew what I was talking about. Who knew that it did the opposite, at least to Eric?

5) I'm going to be moving to standard pixel pitches at the smaller sizes: 1.1 and 1.4 um. I'll probably do powers of two above that, but I could use the pitches of particularly interesting cameras if there aren't too many of them. Any suggestions?

6) It's been pretty conclusively demonstrated that, with the image development that I'm using and the simulate Otus, that there's not much improvement in resolution at 1.25 um vs 2.5 um. What if we threw lens blur deconvolution into the mix?

That's about all I can think of for now.

Have fun, everybody!


-- hide signature --
 JimKasson's gear list:JimKasson's gear list
Nikon D5 Fujifilm GFX 50S Sony a9 Nikon D850 Sony Alpha a7R III +4 more
Flat view
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow