"Correct" default for Post-Resize Sharpening


I have a couple of questions: (I am using 5.3)

  1. What is the correct default for Post-Resize sharpening, USM or RL deconvolution? The default choice is USM on the Sharpening tab?

  2. If it is not a bug and its supposed to be RL deconvolution for post resize, what’s the rational for having the USM default on initial sharpening and RL deconvolution, for post resizing?

Was it because no one could decide so we did one of each ? :smile:

(Morgan Hardwood) #2

We found those defaults to work best in the majority of cases.


Interesting I am finding the opposite, but my captures suffer from blurring caused by my image not being all that sharp (it was scanned), and the USM at the end after resizing works well.

Is there away to force the post sharp resizing without actually resizing.

i.e. input is 1000x1000 --> RL deconv --> no resize or resize to 1000x1000 --> usm mask

(Ingo Weyrich) #4

I always use RL deconvolution with its default values for Post-Resize sharpening.

I wouldn’t exclude that :wink:


More follow up questions as the RL deconvolution documentation is a bit sparse…

(Perhaps If get some good answers I could update the documentation?)

  1. RL deconvolution is driven by a Point Spread function (PSF) , and in the case of RT it uses a Gaussian PSF as opposed to say a specific one. Probably because a Guassian PSF is the best if you don’t know what the “correct” PSF to actually use. Is this a reasonable statement?

  2. Radius, makes some sense if the my first point was correct, though some other examples on the web use some kind of matrix e.g. 5x5

  3. Amount, is that a percentage, as in 100 attempts to remove 100% of the possible Gaussian blur for each iteration?

  4. Damping, I am this guessing this acts as a sort of threshold?

  5. Iteration, this is easy to understand, from my reading on the topic if the deconvolution works, then 20 or 30 iterations is enough. Though some of the examples on the web use hundreds of iterations.

Any clues appreciated :slight_smile:

(Ingo Weyrich) #6
  1. RL deconvolution in RT uses Gaussian PSF

  2. Radius is Sigma of gaussian

  3. Amount is a simple blend factor between before(unsharpened) and after(sharpened)

  4. Damping reduces the effect of the deconvolution at each iteration (Contrary to amount, which is just a blend between before first and after last iteration)

  5. Iterations is what you wrote

(Morgan Hardwood) #7

RawPedia updated:


Thanks, both Moran and Ingo!

I would like to add the following that may assist others.

I am finding the RL deconvolution a better algorithm to use initially with film scans. In this case my digital camera has difficulty achieving the necessary resolution (flatbed). I use the USM mask for post size resharpening.

  1. The effect of changing the radius is very noticeable and can be fine tuned to remove visible blur. The blur around dust on a scan works well to set the radius for a given magnification or scan resolution. It works noticeably different to the USM mask in this respect.

  2. I find using 100 is working well for the amount.

  3. I set the damping to at least 33, otherwise the deconvolution tends to make film grain of dye clouds more noticeable.

  4. I leave the iterations on the default

(Ingo Weyrich) #9

I detected, that usm with settings [0.4;0.45] for radius and amount set to 1000 also works good.


I was using something similar .61 for radius and 350 to 500% for amount. This actually gives a visually more striking result when viewed at 100%, but unfortunately also intensifies the film grain in an unrealistic way.

The camera I am using (a flat bed) uses pixel shift, and the RL deconvolution seems to be able to actually remove some of the blur better, without intensifying film grain in the same way. i.e. achieving a result a little closer to a traditional optical print.

This is my observation for the scanner/resolution I am using.

(Ingo Weyrich) #11

The values I mentioned for sharpening actually are based on pixel shift images. Just try them. Then we can discuss. I don’t say they are the best settings. but imho they are better (at least for pixel shift) than the current settings in pixel shift profiles (which I set)


Those setting are good too. Thanks for that. But the difference that I seem to be noticing is that RL is better at not intensifying film grain (with damping set appropriately). With ‘your’ settings it is absolutely essential to only sharpen edges. With the scanner the pixel size is in the order or 5 -15 μm where as the silver particles are in the order of .5μm and the silver clumps are little bit bigger, but still considerably smaller than a pixel.

There is a good proprietary algorithm called GEM now owned by Kodak (from the same group that made ICE), to reduce some of this “aliasing” but it not available in a standard alone package.

I will post some crops later today, if any one wants to have a crack at de blurring them.

(Ingo Weyrich) #13



OK here is a selection, I made sure each of these has small dust fragment to deblur as well. You can see the smearing around the dust. These are 16bit tiffs.

examples.zip (7.2 MB)

These are all taken with 4800dpi flatbed which is almost the start of the art that can be easily purchased new… These don’t really compare with something dedicated but they are still useful, and could certainly benefit from some enhancement so perhaps a good challenge… This scanner uses pixel shift; two rows of pixels at 2400dpi in the X dimension and the stepper motor that steps in increments to match 4800dpi. It has a wide angle lens with hopefully enough depth of field to cover the film. When you choose the lower resolution you can’t be sure exactly what the scanner does to achieve this lower resolution.

These are some of the settings that seem to give good results:

for the 4800dpi scan: RL radius 2, amount 100 damping 50, iterations 30
for the 2400dpi scan: RL radius 1, amount 100 damping 50, iterations 30

then resized and USM (using the RT defaults)