The downsizing challenge - Part 1: The Problem

[Andrea B.]

From the various downsizes I made, I've selected a strip. Looking at the strips side-by-side might give some insight as to which downsize is "best". It will take me a minute to paste together the strips and post them. I just realized this forum software does not permit display of photos side-by-side.

It is very obvious this way which downsizes are not good.

[Jan Anne]

I was one of the "complainers" stating that my 12MP a7S files looked better on the web than the ones from the 42MP a7RII.

As there is more detail in the "in focus" area with 42MP files I was expecting an improvement in separation from the OOF background, instead my a7RII files look less sharp overal when exported from Capture One. The general idea was that there's too much downsizing needed with the high res files they start to lose details.

Looking at my own situation this is probably caused by my unfamiliarity with Capture One and the apparent lack of additional sharpening options when exporting a 1200pix JPG, something which I used to do with CNX2.

That said I don't add any sharpening besides the default C1 sharpening settings, when working with these files there seems no need to as a little nudge on the Clarity, Structure and Contrast sliders usually does the trick while looking at the RAW files at roughly the same resolution as the exported JPG (1440 vs 1200).

[Andrea B.]

The general idea was that there's too much downsizing needed with the high res files they start to lose details.

Yes!! But I'm not sure that resharpening after downsizing helps give the illusion of detail. Hence all these experiments.

It should be that oversampling helps produce a better smaller file. I'm just not seeing it yet.

[Andrea B.]

Here are some two more excerpts from my experiments. This time look at the bark on the tree trunk to see if it is still detailed. The pink flowers are hopeless. Too much pink hides the details.

It is easy to see the really bad resizes. It is easy to see the soft-ish resizes. It is not so easy to pick out the best of the remaining resizes.

[Andrea B.]

I tried to narrow it down by selecting the best 2 from each panel.
Turns out that I chose the following downsizes.

Shrub & grass composite, panel 1.

• Left:  Photo Ninja conversion with no sharpening or detail slider.
  Resize in PSE using Bicubic Sharper followed by PSE Adjust Sharpen 25-1.
• Right: Photo Ninja conversion with no sharpening or detail slider.
  Resize in PSE using Bicubic Sharper followed by High Pass Overlay 1-20.

Shrub & grass composite, panel 2.

• Left:  Photo Ninja conversion with Sharp=75, Detail=15.
  Resize in PSE using Bicubic Sharper followed by PSE Adjust Sharpen 25-1.
• Right:  Photo Ninja conversion with Sharp=75, Detail=15.
  Resize in PSE using Bicubic Sharper with no further sharpening.

Trunk & flower composite, panel 1.

• Top: Nx-D conversion with no sharpening.
  Resize in Nx-D followed by Nx-D sharpening USM 10-5-5.
• Bottom:  Photo Ninja conversion with Sharp=75, Detail=15.
  Resize in Photo Mechanic with no further sharpening.

Trunk & flower composite, panel 2.

• Top:  Photo Ninja conversion with no sharpening or detail slider.
  Resize in PSE using Bicubic Sharper followed by High Pass Overlay 1-20.
• Bottom: Photo Ninja conversion with Sharp=75, Detail=15.
  Resize in PSE using Bicubic Sharper followed by High Pass Overlay 1-20.
  [ADDED:  I don't like this one anymore.]

Do we have a winner here yet??

• 7 out of 8 votes for Photo Ninja conversion.
• 6 out of 8 votes for resizing in PSE using the Bicubic Sharper algorithm.
• 6 out of 8 votes for resharpening after resizing.B
• And it is an even split for sharpening during conversion or not.
• Of the 6 resharpened versions, 3 used HiPass Overlay.
• 4 out of 8 votes for the No-sharpening-during-conversion followed by resharpening after resizing.
• But also 4 votes for Sharpening-during-conversion.

As I am the only "voter", this may not be a fair election. 

[Andrea B.]

I actually made progress in this quest today!!!

Test EIGHT: Pixel peep for differences in resizes as described below.

I used 4 methods of resizing. It turns out that there is little difference between certain pairs. I'll try to show you this in the next 3 posts.

Test EIGHT:  PART 1
The first part investigates some Resizes which have no addition sharpening after the resize.
The subject is a Peritoma arborea shrub in full bloom found in the Mohave Desert.

Note:  The photos were resized to 1200 pixels width. Please expand your browser or click up the photo to best see it.

Photo 1) Convert photo in Photo Ninja. Add initial Sharpening = 75 and Detail = 15.
Resize in Capture Nx2 with NO resharpening.


Photo 2) Convert photo in Photo Ninja with no initial Sharpening nor Detail.
Resize in Capture Nx2 with NO resharpening.

Methods 1 and 2 produce resized photos which are very nearly identical when you pixel peep. It must be that the resizing drowns out any initial sharpening that was done. When you layer photo #1 as a difference layer over photo #2, nothing shows up even when blown up to 3200%. The actual pixels (next) show how minor the differences are.

Detail 3) 3200% blowup of an area from Method 1 resize which had some initial sharpening of the raw photo. This blowup has a very very very small amount more contrast (apparently from the initial sharpening?) than does the Method 2 blowup shown next.

Detail 4) 3200% blowup of an area from Method 2 resize which had no initial resharpening.

[Andrea B.]

TEST 8:  PART 2
Part 2 investigates some Resizes to which resharpening was applied after the resize in the form of USM 15-4-4. Later we can figure out the best kind of resharpening and the necessary amount. That was not my goal today.

Note:  The photos were resized to 1200 pixels width. Please expand your browser or click up the photo to best see it.

Photo 5) Convert photo in Photo Ninja. Add initial Sharpening = 75 and Detail = 15.
Resize in Capture Nx2 followed by application of USM-15-4-4.

Photo 6) Convert photo in Photo Ninja with no initial Sharpening nor Detail.
Resize in Capture Nx2 followed by application of USM-15-4-4.

Just as before, methods 3 and 4 produce resized photos which are very nearly identical when you pixel peep. Again I suppose the resizing drowns out any initial sharpening that was done. When you layer photo #3 as a difference layer over photo #4, nothing shows up even when blown up to 3200%. The actual pixels (next) show how minor the differences are.

Detail 7) 3200% blowup of an area from Method 1 resize which had some initial sharpening of the raw photo followed by USM resharpening after the resize. This blowup has a very very very small amount more contrast from the initial sharpening than does the companion area next.

Detail 8 ) 3200% blowup of an area from Method 2 resize which had no initial resharpening but did have USM resharpening after the resize.

[Andrea B.]

TEST EIGHT:  PART 3
If you've manage to sort out the preceding, I hope that you have seen that it is beginning to look like that the only real factor we need to look at is whether the resizing of an image is followed by additional sharpening - or not - because any initial sharpening or detail enhancement seems to be almost all lost during the resize.

So I'm now going to compare one photo from each of the preceding two posts.

Photo 9) PN converted photo with no inital sharpening and with no sharpening after the Nx2 resize.
Shown above as Photo #2.

Photo 10) PN converted photo with no initial sharpening but followed by USM15-4-4 after the Nx2 resize.
Shown above as Photo #6.

There are - finally - some differences between these two photos. Not as much as you would think, but certainly enough to be able to distinguish between the two resizes.

Detail 11) This is an excerpt from the difference layer between the two photos used simply to show where the differences are between the sharpened and the unsharpened photos. It is interesting there are few differences due to the sharpening, but it is noticeable when you look at the actual photos. The human eye is very sensitive to contrast. [Note:  1000 pixel width]

Composite 12) I layered the photos and cut windows in the top layer. Inside the windows is the unsharpened version. It is (to me anyway) clear which is the sharpened version.
Composite 13) Flip-flop. It is sharp inside the windows in this version. These are both a little silly but I had fun making them.

[Andrea B.]

Some actual pixels blown up to 3200% from the preceding test.

Detail 14) No sharpening during conversion nor after resize.
Detail 15) No sharpening during conversion but USM 15-4-4 applied after resize. No surprise that these pixels are brighter and more contrasty.

[simsurace]

I propose the following method to evaluate the resizing/re-sharpening objectively:

You can download this image of a Siemens star here:  https://www.dropbox.com/s/hijmrelhpaesijt/Siemens.png?dl=0

I prepared the file to have the same resolution as a D800 file.
The red circles show equal increments of spacial frequencies.
The innermost circle is at Nyquist, the next circle at 1/2 Nyquist, the next at 1/3 Nyquist etc.
The stuff inside the innermost circle can be ignored. It looks nice, but is just false detail. The ideal pattern that is being represented would have infinitely thin rays as you approach the center, but because the file has a finite number of pixels, the rays cannot be resolved inside the circle.
There is a little bit of false details even outside the inner circle (but inside the second), which is due to the fact that the pixel grid is vertical/horizontal, but the rays are slanted. This is an issue that is very relevant since we rarely align all fine detail horizontally or vertically.
It would therefore be wise to go a bit below Nyquist with regards to our expectations of the finest representable detail.

You can test your algorithms on this file. You can test what happens to the detail during resizing. Let's say your resize is to 920px width (a factor of 8 ). The Nyquist frequency is now at circle no. 8. Look at the detail outside circle no.8. Depending on the resizing method, this detail will be more or less contrasty. After re-sharpening, the contrast should be higher (as high as further out) but not as high to give additional false patterns.

This should give you a way to fine-tune your sharpening strength/radius independently of image content.
I haven't given this much thought, but I would guess that a sharpening radius of 0.5 will be close to ideal, since it will operate at the Nyquist frequency of the resized image.

[beryllium10]

Based on everything in this thread so far, my conclusions are (i) downsize to the exact pixel dimensions, if known, (ii) sharpen to the desired degree after downsizing, (iii) different people seem to be happy with different levels of sharpness in the downsized images.  Not too surprising, I think, and echo comments on other NG threads about sharpening - defer sharpening until the final size and intent are known.  Andrea - I also discovered that I personally have a hard time judging the appropriate level of sharpening unless I can see a complete image.  I can tell that some of the test strips you produced are sharper than others, but I can't tell how much detail and local contrast is going to be too much in the final image without seeing it as a whole.  For me this is really true for the "pixel peeping" crops.  I really can't discern anything from these except pixel-by-pixel contrast, hence can't tell whether the 100% image will look under- or over-sharpened (do "pixel peepers" really base their arguments about lens sharpness on staring at 3200% enlargements?).  Amusingly, the 9 x 7 px crops in the final post would correspond to a 100% image 104 m wide x 81 m high (assuming a 24 MP sensor).  That would be impressive! 

Cheers,  John

[Andrea B.]

Well, the purpose of the 3200% enlargements was certainly not to judge sharpness. It was merely to illustrate that there were only the most minor of differences between two different conversion/downsizing/sharpening methods. And that there is definitely some difference between resized images one of which was subsequently sharpened and one not.
Nothing else can be deduced from such ridiculously enlarged samples. 

The complete images from which the strips are taken are shown in earlier posts.

The whole exercise was undertaken to explore the commonly accepted wisdom about resizing by performing the experiments myself because I have been unhappy with my D810 downsized photos.

Some of the commonly accepted wisdom about resizing seems to be invalid.
Example:  It appears not to matter whether you do or do not sharpen the raw before downsizing it. Both downsizes were very nearly identical - and will still need sharpening after the downsize.. But only tested that on landscapes so far. May not hold for other types of shots.

I'm not sure what it means to "downsize to the exact pixel dimensions, if known". Please explain, thanks.

[beryllium10]

>I'm not sure what it means to "downsize to the exact pixel dimensions, if known". Please explain, thanks.

Some of the earlier discussion on the thread was about re-sizing photos for posting on web sites such as facebook.  In such cases you might know in advance the pixel dimensions at which your image will be displayed.  If so, it seems to be good advice to downsize to those exact pixel dimensions, and sharpen to your taste.  Otherwise someone else's code will re-size your image, and may apply sharpening or other rendering adjustments beyond your control.

Cheers again,  John

[Andrea B.]

Oh yes, I understand. And certainly agree! It's best to do our own resizing to meet requirements rather than be at the mercy of sometimes crude forum software resizing.

**********

Simone, thank you for the Star link. I'm going to explore that next. It will be interesting to play with the Nyquist frequency thing we had discussed in Scotland. And it is time to resize something other than a landscape. 

[simsurace]

Some of the commonly accepted wisdom about resizing seems to be invalid.
Example:  It appears not to matter whether you do or do not sharpen the raw before downsizing it. Both downsizes were very nearly identical - and will still need sharpening after the downsize.. But only tested that on landscapes so far. May not hold for other types of shots.

I don't know about how commonly accepted it is, but the sharpening before downsizing is unsurprisingly not going to matter a lot in massive resizes as discussed here.
With the exception of regular patterns such as architectural or textile textures where interference patterns could show up more if the full-size image has very high contrast at spacial frequencies that are going to end up way above Nyquist in the final image, I do not expect a lot of visible differences, and your examples confirm that.

What could be interesting to test out, particularly with textiles or such, is blurring before downsizing. An efffective blurring is implicit in most resizing algorithms, but in cases where excessive interference patterns are obtained, additional blurring to suppress high-frequency detail and make the final resize look smoother might be in order. The final re-sharpen will thus only increase contrast below Nyquist and not accentuate any false patterns that have emerged due to the fine pattern of the cloth interferring with the pixel grid of the resized image.