The principle problem with "classical" digital imaging

Started 1 month ago | Discussions thread
Flat view
FDecker Senior Member • Posts: 1,826
The principle problem with "classical" digital imaging

The threads about image quality a pixel level, aliasing, AA filters, Foveon vs. Bayer, show the principle problem of digital photography, the representation of an analog picture in a discrete way, approximated by a regular grid of separated samples.

It doesn't matter how much the MP count is raised. At pixel level, the quality is not great and is getting even more disappointing the higher the pixel count (with a given sensor format).  The Foveon, and this is why I really like that approach, is clearly better in this respect because the acuity at pixel level is higher for the reasons we all know here. But that's just a mitigation, not a solution.

The best solution to overcome the underlying problem is clearly oversampling a signal (image) which is limited in spatial resolution by an optical filter (or by other means as shaking the sensor while acquiring the image) so that no aliasing happens and using digital filters and spatial decimation (downsampling) afterwards to a spatial frequency which is considerably below the optical filter bandwidth.

But and this is a big but, especially for Bayer patterns, this means that the decimation factor in each direction is in the range of 4 for Bayer. For a Foveon, a factor of 2 might be sufficient. That means, the "MP resolution" is reduced by a factor between 4 (Foveon) and 16 (Bayer).

And this would be a huge step back in the stupid MP race which is still going on. Instead of 50-60MP on a Bayer, one would be limited to e.g. 6-8MP.

One could, of course, use the very small pixel pitch as seen on mobile phones. This would allow for 200-400 MP Bayer sensors (FF) which would then be filtered down to 12-24 MP output. There is no penalty in S/N or DR because the digital filter serves as an averaging function which brings the S/N and DR of the "combined" small pixels to the level of the hypothetical large pixel they represent in the downsampled final image.

The people buying the camera would not even have to know that this technique is used. We would no longer talk about the MP count of the sensor but the MP count of the output. This would also set an end to the discussion about Bayer vs. X-Trans vs. Foveon vs. whatever. Technically, it would turn out that a Foveon needs less MP on the sensor to create the same high quality output (in terms of resolution and acuity, not S/N) of a given MP count than a Bayer sensor. But in the end, both would create the same output MP, so who cares?

Unfortunately, in the "classical" photography market, everything is centred around the raw data from the sensors. This is something to overcome because it leads into the wrong direction. Here, the mobile phones are nowadays much closer to a reasonable approach. Some call it computational imaging. But the direction is the right one. The camera delivers a high quality image. The way this is done in the digital processing of the data is not really relevant. Of course, with the tiny sensors of mobile phones, it is not possible to get the quality one expects from a FF or MF camera. But without these techniques, the mobile phones would not deliver the quality they do.

Now, one could say that all I was writing about can be done in post-processing. Unfortunately, this is not true because some things must be done when acquiring the raw data. Especially the AA filter must be present because aliasing happens at the earliest stage and is not completely removable from an image afterwards. And some people won't do it properly and a lot of people won't do it at all because they like so much to pixel peep into the details (look ! by zooming in, I can even read the number plate of that car...yes, it is crappy, unsharp and with aliasing, but I can read it...).

In my point of view, the people interested in photography hurt themselves by promoting the pixel race and the market's focus on the raw sensor data instead of focusing on the final image.

As such, it was a bold move of Sigma to use and promote the Foveon sensor, focusing more on the high quality output than on the maximal pixel count. But we all know that they felt forced to play the MP game because the market was not ready to fully understand the implications of that concept compared to Bayer.

Note: take the oversampling factors I mentioned as a ball park, not as technically proven lower limits.

2nd note: this must be one of the longest threads I ever wrote here. Hope it was worth it.

3rd note: I am sure, these things have already been mentioned by some forum members in other discussions. After reading one of the latest threads about aliasing, I thought it might be good to summarize this topic and post it here.



Flat view
Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow