Over-resolution fuzzy pixels / AA filter maybe is good afterall?

[torger]

At some point resolution of the sensor becomes so high that the limitations in the lens resolving power start to show, and also diffraction. Modern APS-C sensors is a good example of this 18 megapixel on the Canon 7D and 24 megapixels on the Sony SLT-A77, scaled to full-frame 135 it would be 45 - 56 megapixels.

The common view on this is that it is a waste and that you instead want high contrast sharp images down to the pixel level, that is lower resolution (unnecessary to outresolve lenses / or sample diffraction) and no AA filter.

However, from recent prints and experiments with upscaling software I've made it seems to me that you'd not want sharp images at the pixel level, because they won't scale up well. Jaggies in sharp diagonal contrasty lines will show as artifacts / false detail after upscaling. The most visually pleasing upscaling result I get is from simple bilinear (or perhaps rather bicubic) interpolation when the original image is fuzzy enough at the pixel level so that no jaggies exist in the original image, and then no false detail is introduced after upscaling.

Thus, to me it seems better for image quality to print an image at say 300 ppi where pixels are a bit fuzzy than print at 200 ppi with sharp pixels even if actual detail is the same. And if you need to scale up the image, having sharp pixels is not necessarily an advantage. For a print, the image usually need to be rescaled some to match the printer resolution, there is rarely an exact match with the photo pixel resolution. (The c-printer I use seems to use bilinear scaling to match the input with printer resolution.)

Another aspect is if you do lens distortion correction in post (correct barrell distortion etc) or rotate the image etc, this will show less artifacts if the image is not too sharp at pixel level.

Sure you can blur down the pixelsharp image before upscaling etc, but it seems to me that you will then lose more resolution than if you from the start had a suitable over-resolution which do not require blurring. Where the limits are I'm not sure though. On full-frame 135 I'm quite sure you'd rather have 40 megapixels with AA filter than 20 megapixels without if you want to make large prints.

What do you think?

[NikoJorj]

Yes, that's just oversampling and I'd think it's a good practice common to many kinds of A/D conversions...

[hjulenissen]

Yes, that's just oversampling and I'd think it's a good practice common to many kinds of A/D conversions...

It seems to me that what the OP describes is spatial pre-filtering, a prerequisite to do proper Nyquistian sampling of general bandwidth signals.

"Oversampling" is the use of a multiplum of that sampling rate in order to move pre-filtering complexity from the optical/analog domain and into the digital domain (the end result is filtered first by an analog filter, then sampled, then filtered by a digital filter, then digitally filtered and decimated). In order for cameras to use oversampling, they would have to:
a)Blur the image before spatial discretization (via lense diffraction, olpf,...)
b)Downsample the resulting image

The decimated formats (mRaw, sRaw or whatever) could perhaps in principle be called oversampled. The regular raw formats could in my opinion not.
http://en.wikipedia.org/wiki/Oversampling

-h

[NikoJorj]

Thanks for the precisions! I thought the blur & finer discretization was enough to use the term.

[torger]

Here's an example, a simulation. Both viewed at 200% pixel size.

To the right 100% crop from a 70mm Canon 7D 18 megapixel APS-C corresponding to 45 megapixel fullframe.
To the left a simulation of half the megapixels without AA by zooming in to 200mm and then downrezing without filter so you get very sharp image at the pixel level.

Both has been resized to print in native resolution 400 ppi, but assuming the high res version only could do 236 ppi and thus the low res sharp could only do 167 ppi. Bicubic and unsharp mask. Local contrast was hard to get the same on both, so it is a bit higher on the low res version.

The problem I have with the low res sharp pixel version is that you start to see jaggies in the upsized version. You could apply a spline-type upsizing but then it looks like a painting with very sharp edges, and you can get strange spline-tracing in places.

[Fine_Art]

I don't understand your logic. You are arguing against higher res. cameras because the details don't upscale well? Why do you want to rely on up-scaling? File size?

[Fine_Art]

On re-filtering your OP you prefer fuzzy high res to sharp low res due to the ability to do upscaling smoothly. Is that right?

Don't you want the proper amount of detail so no upres is required?

[torger]

I don't understand your logic. You are arguing against higher res. cameras because the details don't upscale well? Why do you want to rely on up-scaling? File size?

No, I'm (potentially) arguing against keeping down megapixels and not having AA filter to get a pixel-peep sharp file from the camera instead of putting in many megapixels on the sensor (more than the lens can resolve with high contrast) and an AA filter so you get a smooth file at pixel level which is much more flexible when stretched and bent.

Upscaling happens whether one wants it or not when you print large to printers that do 300 - 400 ppi natively. Reshaping the image happens when distortion correction is applied. Or when you stitch a mosaic. So it nice to get files from the camera that responds well to such treatment. That is either have an AA filter or very high resolution or both.

[torger]

On re-filtering your OP you prefer fuzzy high res to sharp low res due to the ability to do upscaling smoothly. Is that right?

Don't you want the proper amount of detail so no upres is required?

Yes I prefer fuzzy high res, for the reason regardless of upsizing amount and regardless of how close you watch the print it will not look pixelated or strange, just smooth.

Ideally you'd even have more resolution than the printer of course, but I rarely see that as feasible. Not all of us are working professionals that can afford an IQ180 back, and even 80 megapixels can fall short. Say you want a 30x30 inch square format, with IQ180 that would be 260 ppi, which may be upscaled to 400 ppi native resolution of a durst lambda c-print. With a 5Dmk2 that would be 125 ppi. I think 135 DSLRs have a potential to have a much higher pixel density than they have today.

[ErikKaffehr]

Hi,

I share Torger's opinion.

Large pixels without AA-filtering is a sure recipe for aliasing. An image with more but smaller pixels with correct AA-filtering will scale better, sharpen better and have less aliasing then larger pixels on a sensor of the same size.

What is lost with smaller pixels is SNR (Signal/Noise ratio) in the darkest areas.

Best regards
Erik

At some point resolution of the sensor becomes so high that the limitations in the lens resolving power start to show, and also diffraction. Modern APS-C sensors is a good example of this 18 megapixel on the Canon 7D and 24 megapixels on the Sony SLT-A77, scaled to full-frame 135 it would be 45 - 56 megapixels.

The common view on this is that it is a waste and that you instead want high contrast sharp images down to the pixel level, that is lower resolution (unnecessary to outresolve lenses / or sample diffraction) and no AA filter.

However, from recent prints and experiments with upscaling software I've made it seems to me that you'd not want sharp images at the pixel level, because they won't scale up well. Jaggies in sharp diagonal contrasty lines will show as artifacts / false detail after upscaling. The most visually pleasing upscaling result I get is from simple bilinear (or perhaps rather bicubic) interpolation when the original image is fuzzy enough at the pixel level so that no jaggies exist in the original image, and then no false detail is introduced after upscaling.

Thus, to me it seems better for image quality to print an image at say 300 ppi where pixels are a bit fuzzy than print at 200 ppi with sharp pixels even if actual detail is the same. And if you need to scale up the image, having sharp pixels is not necessarily an advantage. For a print, the image usually need to be rescaled some to match the printer resolution, there is rarely an exact match with the photo pixel resolution. (The c-printer I use seems to use bilinear scaling to match the input with printer resolution.)

Another aspect is if you do lens distortion correction in post (correct barrell distortion etc) or rotate the image etc, this will show less artifacts if the image is not too sharp at pixel level.

Sure you can blur down the pixelsharp image before upscaling etc, but it seems to me that you will then lose more resolution than if you from the start had a suitable over-resolution which do not require blurring. Where the limits are I'm not sure though. On full-frame 135 I'm quite sure you'd rather have 40 megapixels with AA filter than 20 megapixels without if you want to make large prints.

What do you think?

[hjulenissen]

What is lost with smaller pixels is SNR (Signal/Noise ratio) in the darkest areas.

At least on a per-pixel basis. There is some controversy over whether it is the case on a per-image basis:
http://forums.dpreview.com/forums/read.asp?forum=1000&message=38717540

Nevertheless, to the human eye, at the same technology generation, there is a "sweet spot" where the sense of IQ peaks, which is usually at the highest pixel density (ha ha ha) that a sensor maker will choose to make, not surprisingy.

-h

[Fine_Art]

Yes I prefer fuzzy high res, for the reason regardless of upsizing amount and regardless of how close you watch the print it will not look pixelated or strange, just smooth.

Ideally you'd even have more resolution than the printer of course, but I rarely see that as feasible. Not all of us are working professionals that can afford an IQ180 back, and even 80 megapixels can fall short. Say you want a 30x30 inch square format, with IQ180 that would be 260 ppi, which may be upscaled to 400 ppi native resolution of a durst lambda c-print. With a 5Dmk2 that would be 125 ppi. I think 135 DSLRs have a potential to have a much higher pixel density than they have today.

You can stitch multi-shot. 3x3 3x5 or even 5x5 should give you plenty of resolution with your money still in the bank.

[torger]

You can stitch multi-shot. 3x3 3x5 or even 5x5 should give you plenty of resolution with your money still in the bank.

Yes, I actually do quite a lot of that. A good panorama head is nice to have :-).

[ErikKaffehr]

Hi,

Regarding shot noise it is essentially irrelevant if the photons are collected over many small pixels or fewer larger pixels. DR is more tricky. There is a technical definition of DR defined as Full Well Capacity/(SNR=1) where SNR=1 in general corresponds to read noise. If we double pixel size, read noise will be constant but full well capacity will double. Binning to pixels in software would still double photo count, but read noise would add in quadrature, so read noise would actually increase about 41%. So software binning is very effective for shot noise but less so for readout noise.

The creenshots for DxO-mark illustrates the issue, on Phase One IQ180 that does binning in software:

DR in "print mode" shows some advantage of binning
Tonal range "in print mode" shows no advantage of binning
Tonal range "in screen mode" shows big advantage of binning

The interpretation is that hardware binning has some advantage over software binning in DR but not on tonal range.

Best regards
Erik

At least on a per-pixel basis. There is some controversy over whether it is the case on a per-image basis:
http://forums.dpreview.com/forums/read.asp?forum=1000&message=38717540

-h