A free high quality resampling tool for ImageMagick users

[Bart_van_der_Wolf]

Here is what is hopefully a sufficient fix of the "high frequency detail lightening": use gamma 1.25 instead of gamma 3.
This version does not suppress dark halos as much as earlier ones, but I think that, when one does not pixel peep, it may suppress them enough.

Hi Nicolas,

While subject to taste, I do not like the deeper undershoot halos that result from the reduced gamma.

No free hot lunch: If you want less halos, push the gamma up from 1.25 (and fix its reciprocal). If you want more "lightness preservation" in high frequency areas, push the gamma down toward 1 (and fix its reciprocal).

Yes, it's up to the user to pick his/her own poison, although for a balanced amount of under- / over-shoot halo with boosted sharpening after returning to the original gamma, it's hard to beat the original script Version 1.2.2 settings (given the restraints of 16-b/ch processing). The reduced gamma settings will clip dark side halos pretty quickly when sharpening is increased. Only floating point processing might allow to repair some of that before saving the output.

This runs considerably more slowly than Bart's original, because it uses a rank filter, and rank filters are slow.

With the original gamma settings, there is virtually no difference between the outputs. So I'm a bit doubtful if it really helps to use this slower method, although I do like the simplicity of clamping with an upper and lower bounds version through Median averaging.

Cheers,
Bart

[Bart_van_der_Wolf]

Jan, Cem,

Good thinking! Might be very usable for many more people who have a Lightroom oriented workflow.

Of course the script, as it is, is intended to be very flexible. When things stabilize, it would be possible to simplify things for certain types of output (e.g. down-sampling or upsampling centric, with preferred settings for sharpening being hard-coded).

Cheers,
Bart

[JRSmit]

Hi Jan,

This is a great idea. It can be further improved by skipping the re-import/synchronize step as follows:

1) Search the text "%~dpn1_%mSize%%~x1" in the bat file (occurs multiple times) and replace all of them with "%~dpn1%~x1".

2) Afterwards, when processing a file from within LR, just select to "edit a copy" (instead of the original in case it is a tif or jpg file).

3) This will create a new file using your chosen rename scheme and the output file from the bat process will just override the copy file. LR will then display the end result without having to re-import or synchronize.

Hope this help.

Cem, thanks for the tip. Removes one step from this activity.

A first test was an image well captured, low noise, modest capture sharpening applied. I uprezzed a little over 2 times, sharpening 50 .  Cropped both the source tiff and uprezzed tiff , a same area (using auto sync in LR). Then printed both with the same print size. Source tiff at 170 ppi and uprezzed at360 ppi. Printed on hahnemuhle photo silk baryta with an epson4900 set at 360ppi. Prints are indistinguishable visually. So no b enefits . Perhaps the 360 ppi is the limiting factor.
Next test will l be at 720ppi with fine details on. With enough pixels this setting usually shows better resolution.

[NicolasRobidoux]

Looking at the histogram, there is little difference, also the softproofing shows practically the same oog areas. Also visually on screen there is no directly visible change in lightness etc.
...

The "lightening of high frequency areas" affects some of the schemes when downsampling. I don't expect it to be an issue when upsampling.

[NicolasRobidoux]

...
With the original gamma settings, there is virtually no difference between the outputs. So I'm a bit doubtful if it really helps to use this slower method, although I do like the simplicity of clamping with an upper and lower bounds version through Median averaging.

We're beating around one bush. At this point, I don't expect a "Wow!" improvement to what we have already. Just pushing the envelope a little.
This being said, I really don't like the lightening of high freq. content when downsampling. Staying away from pixel peeping, it is my (changeable) opinion that this is more of an issue than dark halos.

[Bart_van_der_Wolf]

We're beating around one bush. At this point, I don't expect a "Wow!" improvement to what we have already. Just pushing the envelope a little.
This being said, I really don't like the lightening of high freq. content when downsampling. Staying away from pixel peeping, it is my (changeable) opinion that this is more of an issue than dark halos.

Nicolas, I'm not quite sure what lightening you are seeing. Do you have a specific example? Maybe we are talking about something different, and are thus not focusing on the same issue that should be addressed?

When you try and down-sample e.g. one of Norman Koren's targets, in particular this one, and down-sample it to 25%, it should show how a sinusoidal (and for the masochists a bi-tonal) progressive grating of increasingly higher spatial frequencies will perform all the way past the Nyquist frequency. That should give a repeatable image for testing different scenarios.

Alternatively, a zoneplate 'rings' target does a similar thing in all angles, even at modest down-sampling percentages (because the original has 0.5 cycles/pixel diagonal resolution in the corners). Anything below 70% size will challenge the aliasing suppression in all directions. I could make a lower amplitude version if under- / over-shoots need to be quantified at lower spatial frequencies.

Cheers,
Bart

[Bart_van_der_Wolf]

Cem, thanks for the tip. Removes one step from this activity.

A first test was an image well captured, low noise, modest capture sharpening applied. I uprezzed a little over 2 times, sharpening 50 .  Cropped both the source tiff and uprezzed tiff , a same area (using auto sync in LR). Then printed both with the same print size. Source tiff at 170 ppi and uprezzed at360 ppi. Printed on hahnemuhle photo silk baryta with an epson4900 set at 360ppi. Prints are indistinguishable visually. So no b enefits . Perhaps the 360 ppi is the limiting factor.

Hi Jan,

Thanks for the feedback. The upsampling algorithm attempts to avoid creating artifacts, while maintaining as much of the original resolution as possible. It does not create additional resolution, so a source of 170 PPI will remain just that.

You can also experiment with initial '0' sharpening, which then allows to add your own deconvolution radius and amount, instead of using the built-in modification of the resampling window to 'sharpen'the image.

Next test will l be at 720ppi with fine details on. With enough pixels this setting usually shows better resolution.

That should allow to add a higher amount of a relatively modest radius deconvolution.

Cheers,
Bart

[NicolasRobidoux]

Nicolas, I'm not quite sure what lightening you are seeing. Do you have a specific example? Maybe we are talking about something different, and are thus not focusing on the same issue that should be addressed?
...

I'm talking about this:
First, the results with sharpening 50 and 100 with "Bart's downsampling scheme" (EWA RobidouxSoft through linear RGB followed by sharpening with one tuned DoG for gamma 1 and one for gamma 3, auto-level luminance blended), and "Nicolas' tweak" (EWA RobidouxSoft through linear RGB followed by sharpening in gamma 1 and gamma 3 using the DoG tuned for gamma 1, followed by per channel median; I'm using gamma 3, which I now think is overdoing it, so as to compare apples to apples) applied to the fly http://upload.wikimedia.org/wikipedia/commons/8/85/Calliphora_sp_Portrait.jpg.

convert \( input.jpg -set colorspace sRGB -colorspace RGB \
  -define filter:c=0.1601886205085204 -filter Cubic -distort Resize 25\% \) \
  \( -clone 0 -gamma 3 -define convolve:scale=50^,100% \
  -morphology Convolve DoG:3,0,0.4981063336734057 -gamma 0.3333333333333333333 \) \
  \( -clone 0 -define convolve:scale=50^,100% \
  -morphology Convolve DoG:3,0,0.4806768770037563 \) \
  -delete 0 \
  \( -clone 1 -colorspace gray -auto-level \) \
  -compose over -composite \
  -set colorspace RGB -colorspace sRGB LWGBautolevelGamma3s50.png

convert \( input.jpg -set colorspace sRGB -colorspace RGB \
  -define filter:c=0.1601886205085204 -filter Cubic -distort Resize 25\% \) \
  \( -clone 0 -gamma 3 -define convolve:scale=100^,100% \
  -morphology Convolve DoG:3,0,0.4981063336734057 -gamma 0.3333333333333333333 \) \
  \( -clone 0 -define convolve:scale=100^,100% \
  -morphology Convolve DoG:3,0,0.4806768770037563 \) \
  -delete 0 \
  \( -clone 1 -colorspace gray -auto-level \) \
  -compose over -composite \
  -set colorspace RGB -colorspace sRGB LWGBautolevelGamma3s100.png

convert \
  \( input.jpg -set colorspace sRGB -colorspace RGB \
     -define filter:c=0.1601886205085204 -filter Cubic -distort Resize 25\% \) \
    \( -clone 0 -define convolve:scale=50^,100% \
     -morphology Convolve DoG:3,0,0.4806768770037563 \) \
  \( -clone 0 -gamma 3 -define convolve:scale=50^,100% \
     -morphology Convolve DoG:3,0,0.4806768770037563 -gamma 0.33333333333333333 \) \
  -evaluate-sequence median \
  -set colorspace RGB -colorspace sRGB MGBgamma3s50.png

convert \
  \( input.jpg -set colorspace sRGB -colorspace RGB \
     -define filter:c=0.1601886205085204 -filter Cubic -distort Resize 25\% \) \
    \( -clone 0 -define convolve:scale=100^,100% \
     -morphology Convolve DoG:3,0,0.4806768770037563 \) \
  \( -clone 0 -gamma 3 -define convolve:scale=100^,100% \
     -morphology Convolve DoG:3,0,0.4806768770037563 -gamma 0.33333333333333333 \) \
  -evaluate-sequence median \
  -set colorspace RGB -colorspace sRGB MGBgamma3s100.png

[NicolasRobidoux]

Now, the results of downsampling with schemes that should preserve the lightness in high frequency areas better: Bilinear done correctly, EWA Quadratic, EWA Robidoux and EWA RobidouxSharp:

convert \
    input.jpg -set colorspace sRGB -colorspace RGB \
    -filter Triangle -resize 25\% \
    -set colorspace RGB -colorspace sRGB bilinear.png

convert \
    input.jpg -set colorspace sRGB -colorspace RGB \
    -filter Quadratic -distort Resize 25\% \
    -set colorspace RGB -colorspace sRGB EWAquadratic.png

convert \
    input.jpg -set colorspace sRGB -colorspace RGB \
    -distort Resize 25\% \
    -set colorspace RGB -colorspace sRGB EWArobidoux.png

convert \
    input.jpg -set colorspace sRGB -colorspace RGB \
    -filter RobidouxSharp -distort Resize 25\% \
    -set colorspace RGB -colorspace sRGB EWArobidouxSharp.png

Look at the eyes or measure the average luminance in a patch. Some of this is probably caused by the sharpening "bottoming out" less than it "tops up" but my guess is that at least some of it comes from mixing gamma 3 in.
It looks to me like the halo suppression comes with a subtle form of moire.
P.S. Hopefully it's not a side effect of using a Q16 version of ImageMagick without care or worry. If so, I'll feel like a fool.

[NicolasRobidoux]

...
When you try and down-sample e.g. one of Norman Koren's targets, in particular this one, and down-sample it to 25%, it should show how a sinusoidal (and for the masochists a bi-tonal) progressive grating of increasingly higher spatial frequencies will perform all the way past the Nyquist frequency. That should give a repeatable image for testing different scenarios.
...

First, with the gamma 3 versions of Bart's (LWGB...) and Nicolas' (MGB...).
(Hopefully I'm not violating copyright law by posting downsampled versions. Let me know if I'm wrong and I'll take the results down.)

[NicolasRobidoux]

... now with more neutral schemes.

[Bart_van_der_Wolf]

First, with the gamma 3 versions of Bart's (LWGB...) and Nicolas' (MGB...).

And if we make a profile plot through the sinusoidal grating we get something like this (see attachments).

As can be seen, due to blending of unresolvable pixels in linear light, the resulting average amplitude in '2.2 gamma space' (actually 1/2.2, or gamma 2.2 pre-compensated) is higher than the plain average between minimum and maximum.

What also can be seen is that deconvolution with an amount of 50 allows to maintain higher resolutions (higher spatial frequencies) with the restored maximum amplitude of the original input. There comes a point where resolution will diminish as we come close to the Nyquist frequency where amplitude goes to zero, and at even higher spatial frequencies we see some aliasing artifacts (which will fold back (mirror) to lower spatial frequencies as aliases).

We can also see that at a deconvolution amount of 50, the highest non-halo adding level, there is little difference between the luminance blending and the median blending approaches. At a higher deconvolution amount, which will add some haloing (predictable by filter design), we also see that Median blending fails to use most of the sharpened versions, and takes the unsharpened middle ground, and thus is less effective in sharpness restoration. The average amplitude level remains at the same elevated level as the others, because that is caused by linear light blending of unresolvable pixels.

Because of the less effective sharpening, the maximum amplitudes are lower (which could be interpreted as less lightening, but is actually just a lower maximum amplitude, also at the minimum values which become lighter).

So all effects can be attributed to less effective restoration of the original input signal amplitudes near high spatial frequencies.

Cheers,
Bart

[Bart_van_der_Wolf]

For those who are not familiar with such a profile plot, this (see attachment) is how the original/input file looks, given that not enough pixels were plotted on the horizontal axis. Important is to note that the amplitudes of the input signal is basically maximum for all spatial frequencies. Ideally, although not achievable without creating ugly artifacts, we aim to keep as much of the original input amplitude in our down-sampled images as possible. It is inevitable that as we approach the limiting spatial frequencies near the Nyquist frequency, the amplitude starts to fall to zero amplitude, a flat average.

Cheers,
Bart

[Bart_van_der_Wolf]

... now with more neutral schemes.

Bilinear has overall lower amplitudes, produces duller images with reduced (micro-)contrast, but with the same average amplitude as the more complex methods.
EWAQuadratic has slightly lower amplitudes than Bilinear (IM Triangle filter), but also lower amplitude of aliasing.
EWARobidoux has significantly higher amplitudes at higher spatial frequencies (and aliasing), and we know that the different Keys Cubic filter produces a different trade-off between resolution and artifacts.
EWArobidouxSharp with different Keys Cubic filter settings has better preserved amplitudes (is much sharper) than the others, but also with more artifacts.

The trouble with the aliasing artifacts is that they get mixed in with the actual signal levels, and can afterwards not be separated to reduce them. That's why I have chosen a filter/blending method that produces low levels of artifacts, and allows significant restoration of detail (with lower levels of aliasing in the mix).

Cheers,
Bart

[JRSmit]

Hi Jan,

Thanks for the feedback. The upsampling algorithm attempts to avoid creating artifacts, while maintaining as much of the original resolution as possible. It does not create additional resolution, so a source of 170 PPI will remain just that.

You can also experiment with initial '0' sharpening, which then allows to add your own deconvolution radius and amount, instead of using the built-in modification of the resampling window to 'sharpen'the image.

That should allow to add a higher amount of a relatively modest radius deconvolution.

Cheers,
Bart

What would be a good tool for deconv sharpening?

[Bart_van_der_Wolf]

What would be a good tool for deconv sharpening?

The script file will offer a simple one-pass deconvolution when sharpening is initially set to '0' or less (negative sharpening amounts will blur).

For superior deconvolution quality you can use a PS plugin like FocusMagic. But you can also use TopazLabs 'Detail', for detail enhancement when there is not enough native resolution to challenge the printer.

Remember, deconvolution is basically a restoration tool, so it can restore some of the seemingly lost resolution, but it does not invent new detail (unless pushed too far where it will start producing halo artifacts). Topaz Detail does not produce halo artifacts, but it selectively boosts detail amplitude to mimic resolution, which perceptually is very effective due to the absence of halos.

Then there are tools like 'Photozoom Professional' and 'Perfect Resize' which add detail and effectively increases resolution of edges to higher levels than available in the source image. Care must be taken to not exaggerate the edge resolution because that will look artificial when there is no structural/surface detail to match the edges.

Cheers,
Bart

[Bart_van_der_Wolf]

I'm talking about this:
First, the results with sharpening 50 and 100 with "Bart's downsampling scheme" (EWA RobidouxSoft through linear RGB followed by sharpening with one tuned DoG for gamma 1 and one for gamma 3, auto-level luminance blended), and "Nicolas' tweak" (EWA RobidouxSoft through linear RGB followed by sharpening in gamma 1 and gamma 3 using the DoG tuned for gamma 1, followed by per channel median; I'm using gamma 3, which I now think is overdoing it, so as to compare apples to apples) applied to the fly http://upload.wikimedia.org/wikipedia/commons/8/85/Calliphora_sp_Portrait.jpg.

Attached is a crop animation of the resampled and the original version. The down-sampled version was then upsampled again (by using Nearestneighbor) to approx. align with the original version as a layer. This upsampling allows to avoid effects caused by our human visual system, and reduces (but not eliminates) the risk of gamma effects from viewing the image at non-native (100% zoom) size. There is also a slightly lower risk of the gamma effect of viewing LCD displays from the 'wrong' angle.

Do note, that it should be viewed at a 100% zoom level (use a dedicated image viewer/editor, because a browser may still resize even at 100% setting!), and  from the correct angle to avoid gamma shift by some LCD display types. While there is a bit of luminance shift, it is both in the highlights and in the shadows of high frequency detail in the faceted eyes.

There is no significant lightening other than specific features that align with the pixel grid, and gamma effects as described above.

Cheers,
Bart

[NicolasRobidoux]

Bart:
Trying to sort things out at the perceptual level, it appears to me that EWA RobidouxSharp through linear RGB is generally both sharper and (this is the important bit) has less noticeable haloing than LWGBautolevelGamma3s100. And yet it introduces less moire.
EWA Mitchell through linear RGB, for example, looks really good in the sharpness/halo ratio department. Nice balanced scheme.
Am I missing something? (Away from my large calibrated monitor, but I don't think it matters.)
P.S. OK. Maybe LWGBautolevelGamma3s100 is a nudge sharper. It's less sharp than EWA CatRom through linear light, though, and CatRom does not seem to add an amount of haloing that matches the additional sharpness.
What I don't see is obvious advantages of the new downsampling methods over simply pushing the Keys alpha up from 0 until the sharpness is sufficient.
Again: Am I missing something? (I hope I am.)
(I reserve my opinion w.r.t. upsampling. Different game.)

[Jack Hogan]

Bilinear has overall lower amplitudes, produces duller images with reduced (micro-)contrast, but with the same average amplitude as the more complex methods.
EWAQuadratic has slightly lower amplitudes than Bilinear (IM Triangle filter), but also lower amplitude of aliasing.
EWARobidoux has significantly higher amplitudes at higher spatial frequencies (and aliasing), and we know that the different Keys Cubic filter produces a different trade-off between resolution and artifacts.
EWArobidouxSharp with different Keys Cubic filter settings has better preserved amplitudes (is much sharper) than the others, but also with more artifacts.

The trouble with the aliasing artifacts is that they get mixed in with the actual signal levels, and can afterwards not be separated to reduce them. That's why I have chosen a filter/blending method that produces low levels of artifacts, and allows significant restoration of detail (with lower levels of aliasing in the mix).

Cheers,
Bart

Bart, interesting chart.  What does bicubic (no sharp) look like?

[Bart_van_der_Wolf]

Bart:
Trying to sort things out at the perceptual level, it appears to me that EWA RobidouxSharp through linear RGB is generally both sharper and (this is the important bit) has less noticeable haloing than LWGBautolevelGamma3s100. And yet it introduces less moire.
EWA Mitchell through linear RGB, for example, looks really good in the sharpness/halo ratio department. Nice balanced scheme.
Am I missing something? (Away from my large calibrated monitor, but I don't think it matters.)

Hi Nicolas,

After my initial investment in time, I tried to avoid changing too many things too fast. I preferred waiting for some user feedback, likes, dislikes, the dust to settle as it were. It's just a case of not trying to aim at a moving target until it slows down enough, to take a shot, and then aim at the next target.

As I said in an earlier post, I'd like to test more Keys Cubic settings if time allows. However, until the deconvolution (if any) is automated, it takes a lot of time to do right. I therefore wanted to optimize the setting that you found to be useful (for 'minimizing the worst case deviation from the original value'), before benchmarking against (less robust?) alternatives, and also use the insights that were gained in the course of experimentation and code-optimization.

P.S. OK. Maybe LWGBautolevelGamma3s100 is a nudge sharper. It's less sharp than EWA CatRom through linear light, though, and CatRom does not seem to add an amount of haloing that matches the additional sharpness.
What I don't see is obvious advantages of the new downsampling methods over simply pushing the Keys alpha up from 0 until the sharpness is sufficient.
Again: Am I missing something? (I hope I am.)

We'll have to see how well they perform when scrutinized for aliasing(!)/ringing/blocking behavior (which may also affect rendering of down-sampled noise), and the risk of clipping due to halo under- / over-shoots. I would think that a user controllable Keys alpha input (on a scale from soft to sharp/crisp, or 0% to 100%) to adjust the behavior of the resampling filter should be a parameter for a modern resizing tool. Of course, the negative side effects (from softness to artifacts) should also be minimized in the background.

(I reserve my opinion w.r.t. upsampling. Different game.)

Absolutely, different game, different rules.

Cheers,
Bart

P.S. Different values for Keys alpha will not only require different deconvolution filter values, but also different gammas than 3.0 for a balanced halo under- / over-shoot. High alpha values will require significantly higher gammas, which may be risky in a non-floating point processing environment, but more experimentation will be required to make sure.