imfuse: Finding good parameters: fly on yellow background

[mviereck]

Here I show the progress of finding a good parameter set for a stack provided by @holgerd001. Much thanks for the stack!

quick result with enfuse

First try: Generating an enfuse result

imfuse --background=enfuse

I've marked and numerated some points of interest:

• (1) A hair behind the eye appears in front of it
• (2) Dark areas of the eyes appear a bit too bright
• (3) The leg in background looks well
• (4) Between leg and eye the bokeh is somewhat wrong: yellow instead of leg
• (5) Yet another blurry bokeh area

Try a mask: --dog, --log or --wavelet

Good mask options for macro stacks in general are --log, --dog and --wavelet. Now I use --dog.

imfuse --dog

Visible artefacts:

• (1) hair in front of eye
• (2) eyes darker than with enfuse, but with some noise
• (3) noise in the leg / bad bokeh

fix hair in front of eye: --less --close

Fix the hair in the eye (1) with --less --close

imfuse --dog --less --close

• (1) hair is behind eye, well done
• (2) less noise in darke area of eyes
• (3) bad noisy bokeh at the leg

low contrast area detection: --threshold=50

Now its up to look for the bokeh.
Low contrast areas that might cause artefacts are below 50% contrast.
To get an idea of the location of low contrast areas I use --threshold=50

imfuse --dog --threshold=50 --background=red

blur bokeh with --finalblur

Option --finalblur, currently with default --finalblur=w50,s2, blurs the bokeh.

imfuse --dog --less --close --finalblur

Overall a good result, but the leg (3) looks somewhat incomplete

A high blur sigma improves the leg (3) / bokeh appearance, but causes light artefacts in the eyes (2):

imfuse --dog --less --close --finalblur=w50,s5

Reducing the percent value of finalblur helps, we have luck that the black area in the eyes (2) has a bit more contrast than the leg (3):

imfuse --dog --less --close --finalblur=w47,s5

I'd say this is a good result.

light balance with --fakehdr

As a final tweak some light balance with --fakehdr:

imfuse --dog --less --close --finalblur=w47,s5 --fakehdr=40

[holgerd001]

With --finalblur=w50,s5 there seams to be a yellow glue between the hairs.

Now I have used imfuse --dog --less --close --finalblur=w47,s5 --fakehdr=40 with the non resized images. It looks very different to the 20% images.

[mviereck]

With --finalblur=w50,s5 there seams to be a yellow glue between the hairs.

Indeed. I didn't noticed it before.
Maybe it is better to use --soft instead of increasing the sigma of --finalblur

imfuse --dog --less --close --finalblur=w47,s2 --soft

The bokeh of the leg is still not great. Increasing the softness in lowcontrast with:

imfuse --dog --less --close --finalblur=w47,s2 --soft=-75x100

The bokeh of the leg is still not really great.

[holgerd001]

The bokeh of the leg is still not really great.

Yes, just like in the full size image.

Perhaps you should use the full size images. I can send you. But it takes much more time 14-39 minutes by my experience.

And what I don't understand are the differences between 20% and full size input.

[mviereck]

Perhaps you should use the full size images. I can send you. But it takes much more time 14-39 minutes by my experience.

Yes, that would be appreciated. I'll test them, too.

And what I don't understand are the differences between 20% and full size input.

I am a bit surprised, too.
The --close option might make the difference. It takes an absolute pixel value and a multiplication factor.
Default is --close=2x2, respective a 4 pixel radius (multiplied 2x2=4).
You could try a greater value, maybe --close=10x10; that would outweight the 20%. But be warned: --close with a high radius takes a lot of time. --close=2x50 or --close=4x25 will be faster.

[mviereck]

Here a rude attempt to fix the bokeh:
Running with --threshold=50 and using a background.

imfuse --dog --less --close --finalblur=w47,s2 --soft --threshold=50 --bg=transparent

imfuse --dog --less --close --finalblur=w47,s2 --soft --threshold=50 --bg=min

I've used --background=min instead of --background=enfuse because the eyes in the enfuse result have a white flare that would not fit here.

[mviereck]

Looking closer at your full-sized image I could image that --close=5x5 might be enough to fill the hairs.
The other ugly areas, especially in the eyes, might be fixed with --background=min, with or without --threshold.
I assume it is the enfuse background that just does not fit for this case.
Edit: Maybe omit --soft, too, if using --threshold.

[holgerd001]

Yes, that would be appreciated. I'll test them, too.

The same way as last time. Your email address is still alive?

[mviereck]

The same way as last time. Your email address is still alive?

Yes, it is a permanent address.

[holgerd001]

Looking closer at your full-sized image I could image that --close=5x5 might be enough to fill the hairs.

I tried imfuse --verbose --dog --less --close --finalblur=w47,s2 --soft --threshold=50 --bg=enfuse aligned_C_c60_m_order_00??.tif and got this:

---

imfuse note: Setting --output=/mnt/Daten/holger/Fotos/stacks/20210207_Fliege_Kopf_gelb_Mikroskopobj/imfuse.dog=s0.3-s0.48-close2x2_soft0x100_less60_finalblur=w47-s2_threshold50_bg=enfuse.a6ec4f.tif

imfuse: Mask options:
NUM=1,METHOD=dog,s=0.3,S=0.48,CLOSE1=2,CLOSE2=2,LEVEL,TYPE=mask

imfuse note:
Image number: 64
Image width: 3897 px
Image height: 2595 px
Image memory size: 202253 kb
Image list memory size: 12944192 kb
Free Memory: 12899835 kb

imfuse: Generating transparent image:
nice magick -quiet -size 3897x2595 canvas:transparent -type TrueColorAlpha /mnt/Daten/holger/Fotos/stacks/20210207_Fliege_Kopf_gelb_Mikroskopobj/cache.imfuse/imfuse.dog=s0.3-s0.48-close2x2_soft0x100_less60_finalblur=w47-s2_threshold50_bg=enfuse.a6ec4f.tif
imfuse: Mask generator kernels:
Kernel "DoG" of size 5x5+2+2 with values from -0.2219 to 2
Forming a output range from -1 to 2 (Normalized)
0: -6.14023e-08 -4.12689e-05 -0.000361493 -4.12689e-05 -6.14023e-08
1: -4.12689e-05 -0.0276556 -0.2219 -0.0276556 -4.12689e-05
2: -0.000361493 -0.2219 2 -0.2219 -0.000361493
3: -4.12689e-05 -0.0276556 -0.2219 -0.0276556 -4.12689e-05
4: -6.14023e-08 -4.12689e-05 -0.000361493 -4.12689e-05 -6.14023e-08
magick: image sequence is required `-composite' @ error/operation.c/CLIListOperatorImages/4021.

imfuse note: Generating enfuse base image

---

It's still running.
Version="0.6.6.5-beta"

[mviereck]

magick: image sequence is required `-composite' @ error/operation.c/CLIListOperatorImages/4021.

The error message is harmless at this point, only part of the verbose kernel output. I'll fix that.

Are you running with the default --close [=2x2]? For the full size image rather use --close=5x5, otherwise the hairs will be transparent.

[holgerd001]

Are you running with the default --close [=2x2]? For the full size image rather use --close=5x5, otherwise the hairs will be transparent.

I use --close=5x5 for the full size image.

[mviereck]

The download for full-sized stack is running, thank you!

I found a setup that gives a reasonable good result combining --dog --log --wavelet:

imfuse --dog --log --wavelet --less --close --alpha --finalblur --background=min

None of --dog --log --wavelet was satisfying on its own, but together they give a nice bokeh.

Works also well without --alpha --background=min, but some light flare appears in the eye (2) and the leg bokeh is a bit less well.

Note that I use --alpha instead of formerly --soft according to #9 (comment)

Remaining artefacts:

• The yellow area between eye and leg (4)
• The hairs left of the left eye have a gap between hair and eye.

[holgerd001]

imfuse --dog --log --wavelet --less --close --alpha --finalblur --background=min

That's what I did too.

I found noisy areas

• at the leg (3)
• in the right eye (at the view of the fly)
• in the both legs in the foreground
  It looks like paint splatters.

But in normal magnification only the noise on the leg (3) is visible.

It's a real high art to find good parameters.
And a question of patience. The image above took nearly 23 minutes.

[mviereck]

Adding --threshold=50 might fix some of the issues, especially in the eye.
--finalblur might take a higher sigma for big sized images, e.g. --finalblur=s5