Pixel density revisited

Started Oct 22, 2008 | Discussions
Iliah Borg Forum Pro • Posts: 25,841
Re: Pixel density revisited

The point is, all the comparative tests you give are based on 100%
crops,

I always test the noise and sharpness on actual prints of the same size. If I need to present a test I scan the prints.

Is it one of the changes we need?

-- hide signature --
alanr0 Senior Member • Posts: 2,171
Lanczos resampling?

DSPographer wrote:

I never tested Photoshop's resize. It's disappointing that it doesn't
properly filter first. I wonder what the bicubic smoother in the
newer versions does? Your technique of first applying a Gaussian blur
is much better than no filter but we should be able to do still
better than that. Since the Fourier transform of a Gaussian is also a
Gaussian the result has droop in the pass band while still retaining
some energy that will alias. I think the filter window should look
more like a sinc function in the vertical and horizontal directions.

Something like Lanczos resampling perhaps?
http://en.wikipedia.org/wiki/Lanczos_resampling

There is a list of software known to use this lower down the Wiki page (XnView, Irfanview, Gimp, Faststone etc.).

Cheers.
--
Alan Robinson

rhlpetrus Forum Pro • Posts: 25,860
not sure if pixel density but A900

shows noise at base ISO. I had seen it in the image used for RAW comparisons here at DPr, now confirmed by IR tests as well. The stutio test shot shows lots of noise at base ISO:

http://www.imaging-resource.com/PRODS/AA900/FULLRES/AA900hSLI0200.HTM

Check darker areas, lots of noise, bot h chroma and luma. Too bad, the resolution and DR seem pretty good. Hope Nikon doesn't use this sensor.
--
Regards, Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/

You cannot depend on your eyes when your imagination is out of focus
(Mark Twain)

 rhlpetrus's gear list:rhlpetrus's gear list
Leica D-Lux (Typ 109) Nikon D7000 Nikon 1 V1 Nikon AF-S DX Nikkor 12-24mm f/4G ED-IF Nikon AF Nikkor 35mm f/2D +4 more
DSPographer Senior Member • Posts: 2,466
Re: Lanczos resampling?

The Lanczos kernel can work well if it is organized for rescaling and not just resampling. That means that the window is applied to the data before resampling but the units of x in the function are in samples at the final spacing. So, if we want to change the image size by a factor of 1/2 with a Lanczos window with a=5 we would apply the following kernel centered at every other point in each of the vertical and horizontal directions:

0.0077 0 -0.0335 0 0.0811 0 -0.1822 0 0.6262 1 0.6262 0 -0.1822 0 0.0811 0 -0.0335 0 0.0077

If a program only uses the Lanczos method to resample the point locations without scaling the kernel to adjust the filter pass band size then it wouldn't avoid aliasing. Too low a value of a would also result in a less effective filter. These effects could easily be tested by making an image with a frequency sweep to Nyquest before re-sampling.

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
Jonathan Demarais Veteran Member • Posts: 3,607
Re: Pixel density revisited

Phil Askey wrote:

Did you read the rest of the review? The original point was that the
pixel density figure provided in the specs database should not be
used as a predictor of image quality alone.

So lets draw a line. What pixel density, given current sensor quality, what is the highest pixel density where you can stiil get reasonably good images?

-- hide signature --

'I cried because I had no E-3. Then I met a man with no E-510'

Olympus E-410, E-330, Nikon D100 (IR) & Pentax K20D.
57 lenses of various types from most brands.

bobn2
bobn2 Forum Pro • Posts: 61,441
Re: Pixel density revisited

Jonathan Demarais wrote:

Phil Askey wrote:

Did you read the rest of the review? The original point was that the
pixel density figure provided in the specs database should not be
used as a predictor of image quality alone.

So lets draw a line. What pixel density, given current sensor
quality, what is the highest pixel density where you can stiil get
reasonably good images?

I'm not sure that the position of this limit has been definitively defined yet. If you look at this thread from John Sheehy:
http://forums.dpreview.com/forums/readflat.asp?forum=1018&thread=28607494&page=1

You will see a demonstration that the 1.97um pixels from a Panasonic FZ40 outperform on an area for area basis the pixels from a Canon 40D. That's a pixel density of 26 MP/cm². If you made a FF sensor form those pixels it would have 222MP, and you could extract a pretty decent image from it.
--
Bob

DSPographer Senior Member • Posts: 2,466
Gimp Lanczos has aliasing

I downloaded Gimp and tried its Lanczos rescaling and although the aliasing is less than with cubic there is still far too much aliasing compared to a properly scaled Lanczos with either a=3 or a=2.

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
bobn2
bobn2 Forum Pro • Posts: 61,441
Re: Pixel density revisited

bobn2 wrote:

Jonathan Demarais wrote:

Phil Askey wrote:

Did you read the rest of the review? The original point was that the
pixel density figure provided in the specs database should not be
used as a predictor of image quality alone.

So lets draw a line. What pixel density, given current sensor
quality, what is the highest pixel density where you can stiil get
reasonably good images?

I'm not sure that the position of this limit has been definitively
defined yet. If you look at this thread from John Sheehy:
http://forums.dpreview.com/forums/readflat.asp?forum=1018&thread=28607494&page=1
You will see a demonstration that the 1.97um pixels from a Panasonic
FZ40 outperform on an area for area basis the pixels from a Canon
40D. That's a pixel density of 26 MP/cm². If you made a FF sensor
form those pixels it would have 222MP, and you could extract a pretty
decent image from it.
--
Bob

Apologies, failed to spot the rhetorical question. Phil?
--
Bob

Iliah Borg Forum Pro • Posts: 25,841
Re: not sure if pixel density but A900

rhlpetrus wrote:

shows noise at base ISO. I had seen it in the image used for RAW
comparisons here at DPr, now confirmed by IR tests as well. The
stutio test shot shows lots of noise at base ISO:

http://www.imaging-resource.com/PRODS/AA900/FULLRES/AA900hSLI0200.HTM

What raw converter you used?

Check darker areas, lots of noise, bot h chroma and luma. Too bad,
the resolution and DR seem pretty good.

DR is limited by the noise. One can't have good DR and bad noise

Hope Nikon doesn't use this sensor.

Wow.

-- hide signature --
DSPographer Senior Member • Posts: 2,466
Resampling test examples

Here is the results of resampling an image of a frequency sweep. If the resampled images had no aliasing then the right half of the downsampled images would be flat grey.
Here is the original image:

Here is it downsampled to 1/2 width with Gimp cubic interpolation:

Here is is using Gimp's Lanczos3 interpolation:

Here it is with a Lanczos (a=3) properly scaled to avoid aliasing:

And finally here it is with properly scaled Lanczos with a=2:

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
alanr0 Senior Member • Posts: 2,171
Re: Resampling test examples

DSPographer wrote:

Here is the results of resampling an image of a frequency sweep. If
the resampled images had no aliasing then the right half of the
downsampled images would be flat grey.
Here is the original image:
http://farm4.static.flickr.com/3271/2967277913_cef297caf4_o.png
Here is it downsampled to 1/2 width with Gimp cubic interpolation:

Here is is using Gimp's Lanczos3 interpolation:

Thanks for posting. Looks like one or two short cuts too many.

Good to know that Gimp's Lanczos seems to work better than cubic, even if both need a spot of Gaussian blur to avoid aliasing.

Cheers.
--
Alan Robinson

DSPographer Senior Member • Posts: 2,466
Re: Resampling PS Elements 3.0

Here is the result of downsampling with Photoshop Elements 3.0 bicubic:

It looks to me to have less aliasing than Gimp's Lanczos3.

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
DSPographer Senior Member • Posts: 2,466
Re: There is a resized comparison

As you can see by my test image Photoshop Elements 3.0 does a proper resampling. I just tested it on random data and with a 50% horizontal and 50% vertical resize the standard deviation is reduced to 43% of the original value. The extra reduction below the expected 50% is due to the response rolling off a bit below the new Nyquest rate. I expect that Photoshop CS-X resample with Bicubic smoother should have similar results.

bobn2 wrote:

Missed that. Thanks. Unfortunately, its usefulness is compromised by
use of Photoshop resize, which I'm told (by John Sheehy amongst
others) decimates rather than resamples. The result is you don't lose
the noise but you do lose the detail.
--
Bob

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
rhlpetrus Forum Pro • Posts: 25,860
Re: not sure if pixel density but A900

Iliah Borg wrote:

rhlpetrus wrote:

shows noise at base ISO. I had seen it in the image used for RAW
comparisons here at DPr, now confirmed by IR tests as well. The
stutio test shot shows lots of noise at base ISO:

http://www.imaging-resource.com/PRODS/AA900/FULLRES/AA900hSLI0200.HTM

What raw converter you used?

These are in-camera jpegs they posted. I compared it to the in-camera jpegs posted there from 1DsIII. These are completely clean, as are the in-camera jpegs from D3.

Check darker areas, lots of noise, bot h chroma and luma. Too bad,
the resolution and DR seem pretty good.

DR is limited by the noise. One can't have good DR and bad noise

The DR which is claimed to be good is also the jpeg test from Dpr, a little above 9stops, but any camera these days provide 9 stops with good conversion. This shows that Sony engineers have designed a nice tone curve for in-camera jpegs.

OTOH, again from DPr's test, D700's default jpeg curve seems awful, even though the sensor is known to be about state-of-art re SNR.

Hope Nikon doesn't use this sensor.

Wow.

Too harsh a judgement? Maybe a bit too early to tell, but it doesn't look good. If you shoot landscapes and are not into multiple shots HDR merging, you need latitude to keep HLs from clipping and then good lowlights recovery in PP'ing. If noise creeps up already at low ISOs, that will mean recovery won't go well. This is the may strength opf the D3/D700. It's RAW file has very good performance at 3200 ISO or even at 6400 ISO, meaning if you shoot at ISO200 and undexpose to preserve HLs, in PP'ing you'll be able to bring details from dark areas very well.

Check this test comparing A900 with Hassy. At ISO100 (wrong ISO, reduces DR, but also reduces noise a bit), the A900 shows noise in darker areas, check water. These were converted with ACR, NR off, from post.

http://forums.dpreview.com/forums/read.asp?forum=1037&message=29778684

http://picasaweb.google.com/DouglasPBoyd/A900VsH1P30#5260166084550839010

-- hide signature --

Regards, Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/

You cannot depend on your eyes when your imagination is out of focus
(Mark Twain)

 rhlpetrus's gear list:rhlpetrus's gear list
Leica D-Lux (Typ 109) Nikon D7000 Nikon 1 V1 Nikon AF-S DX Nikkor 12-24mm f/4G ED-IF Nikon AF Nikkor 35mm f/2D +4 more
Iliah Borg Forum Pro • Posts: 25,841
Re: not sure if pixel density but A900

rhlpetrus wrote:

Iliah Borg wrote:

rhlpetrus wrote:

shows noise at base ISO. I had seen it in the image used for RAW
comparisons here at DPr, now confirmed by IR tests as well. The
stutio test shot shows lots of noise at base ISO:

http://www.imaging-resource.com/PRODS/AA900/FULLRES/AA900hSLI0200.HTM

What raw converter you used?

These are in-camera jpegs they posted. I compared it to the in-camera
jpegs posted there from 1DsIII. These are completely clean, as are
the in-camera jpegs from D3.

I do not think it is the right thing to do - to compare OOC jpegs and make a verdict on sensor. In-camera jpeg engine can be improved with firmware updates, by the way. But if you are comparing OOC jpegs in regards to noise, compare resolution too, and not on a black-and-white target.

The DR which is claimed to be good is also the jpeg test from Dpr,

And do you know what are the results of DPR and Seconic tests of OOC jpegs on 5D original? You may be amazed by the difference.

I have A900, and I'm completely satisfied with this camera for my intended use. I would be very happy to have that sensor in D-one-digit body because the lens choice with Sony is quite limited and with the wider choice I would use that camera much more.

-- hide signature --
Jonathan Demarais Veteran Member • Posts: 3,607
Re: Pixel density revisited

bobn2 wrote:

Jonathan Demarais wrote:

Phil Askey wrote:

Did you read the rest of the review? The original point was that the
pixel density figure provided in the specs database should not be
used as a predictor of image quality alone.

So lets draw a line. What pixel density, given current sensor
quality, what is the highest pixel density where you can stiil get
reasonably good images?

I'm not sure that the position of this limit has been definitively
defined yet. If you look at this thread from John Sheehy:
http://forums.dpreview.com/forums/readflat.asp?forum=1018&thread=28607494&page=1
You will see a demonstration that the 1.97um pixels from a Panasonic
FZ40 outperform on an area for area basis the pixels from a Canon
40D. That's a pixel density of 26 MP/cm². If you made a FF sensor
form those pixels it would have 222MP, and you could extract a pretty
decent image from it.

Complete with a beltpack 260G hard drive to hold the images!
--

'I cried because I had no E-3. Then I met a man with no E-510'

Olympus E-410, E-330, Nikon D100 (IR) & Pentax K20D.
57 lenses of various types from most brands.

J A K Forum Pro • Posts: 15,833
Re: Pixel density revisited

Iliah Borg wrote:

I always test the noise and sharpness on actual prints of the same
size. If I need to present a test I scan the prints.

AFAIC that's the "only" thing that counts when comparing camera-A versus camera-B.

Although it's a real PITA I always download the little crops that appear in the DPR reviews and then resize them to an equal size before I look at them on my monitor.

When I do my tests at home for purposes of A/B comparisons, I do resize the images for equality before looking on my monitor, but eventually I make prints. It takes very little effort on my part to print at equal sizes and then peep; this tells the whole story PERIOD.

Regards,

Joe Kurkjian

Galleries: http://www.pbase.com/jkurkjia

SEARCHING FOR A BETTER SELF PORTRAIT

attomole Regular Member • Posts: 290
Re: There is a resized comparison

It does seem to me that this sort of comparison, although how it is done is open to debate, is more real world than zooming in to an equivalent number of pixels, that is decisions about how big to print are not dictated by the pixel count of the sensor, rather the size of the picture or paper have , Therefore for most this would be scaled to say A3.

My contribution to the sensor size / pixel density debate, its about a trade off, the bigger the pixel the greater the information in the light to dark domain, the higher the pixel density the greater the information in the resolution domain, the two are antagonistic, to improve both. You would have to increase the size of the sensor or the efficacy of the photosensitive elements of the device.

Atto

attomole Regular Member • Posts: 290
Yes

We view a picture most commonly as a print, how a picture looks under normal viewing is the most important to me, not what it looks like at the pixel level, so yes I woulds like to see that bottle label magnified to the same physical size not number of pixels

bobn2
bobn2 Forum Pro • Posts: 61,441
Re: Pixel density revisited

Jonathan Demarais wrote:

bobn2 wrote:

Jonathan Demarais wrote:

Phil Askey wrote:

Did you read the rest of the review? The original point was that the
pixel density figure provided in the specs database should not be
used as a predictor of image quality alone.

So lets draw a line. What pixel density, given current sensor
quality, what is the highest pixel density where you can stiil get
reasonably good images?

I'm not sure that the position of this limit has been definitively
defined yet. If you look at this thread from John Sheehy:
http://forums.dpreview.com/forums/readflat.asp?forum=1018&thread=28607494&page=1
You will see a demonstration that the 1.97um pixels from a Panasonic
FZ40 outperform on an area for area basis the pixels from a Canon
40D. That's a pixel density of 26 MP/cm². If you made a FF sensor
form those pixels it would have 222MP, and you could extract a pretty
decent image from it.

Complete with a beltpack 260G hard drive to hold the images!

Couple of years time you'll get 260G on an SD!

More seriously, capturing the image with that resolution doesn't mean you are forced to store it at that resolution, and remember, the size of each sample would be smaller. I haven't done the sums, but a camera such as that might operate perfectly satisfactorily with an 8 bit ADC or smaller.

-- hide signature --

Bob

Keyboard shortcuts:
FForum MMy threads