New approach to fine detail sharpening without edge halos. Very fast, HD also.

[poisondeathray]

Originally Posted by Hotte

Unfortunately in chroma there is no significant improvement. For some strange reason flat blue here and there becomes flat magenta. But all in all it is not a big problem. You have to have a very, very close look to see it.

You must be looking at the full video or testing on another example, or you didn't quite use the same script - because he replaced it with a mask. If you are getting areas that still have it, then the mask isn't perfect (this can easily happen because the mask was for a still image, on other frames you're bound to get holes in the mask.) You can use amplified subtract or amplied differences (or other variations) to verify this, or view the mask directly to see the relationship. There is zero pixel difference in that region on the still, but maybe a few stray pixels near the borders of the edges

eg.
subtract(filtered, original).Levels(127, 1, 129, 0, 255)

But how useful a mask like that would be on actual footage instead of a still is debatable... or at minimum you would have to tweak different masks for different shots



If you still feel something in masktools is contributing to it, and it's not from the pre-exisiting artifacts (which exist in all 3 channels, you can visualize the chroma channels by UtoY() and VtoY() , then test on a clean clip with similar blue gradient. Even a CG gradient will demonstrate an issue if there is a math calculation bug. I just think it's not merely a coincidence that you get "blocks" of magenta here and there that just happen to coincide with the h264 compression artifacts

The other sharpening algorithm in Edius probably isn't using this approach of blurring and subtracting. That probably has something to do with it. And even if it is, "subtract" is defined differently in avisynth than other programs like photoshop, after effects , probably Edius etc... .Not only is the formula different, those other programs are working in RGB for their channel operations

[feisty2]

thats not a "still" mask
mt_edge creates mask based on the edge of "every" frame
the mask changes by every frame unless the clip itself is still
so, the mask is a "video" too, not a still picture

[poisondeathray]

Originally Posted by feisty2

thats not a "still" mask
mt_edge creates mask based on the edge of "every" frame
the mask changes by every frame unless the clip itself is still
so, the mask is a "video" too, not a still picture

Obviously

What I meant is you refined the mask for a still, not the video. Unless he uploaded the video somewhere? He mentioned something that suggests holes in the mask. Well there were hardly any holes when I looked, there were no magenta shifts like the unmasked version. It was a very good mask for that still in terms of using masktools, almost as good as if you rotoed it. So I am assuming he is checking and referring to the actual video.

What I am saying is there are other scenes, other content, other characteristics to the video, other types of edges. And that perhaps you need to adjust the settings in other shots, threshold values - Or did you think this would work perfectly for everything with those settings without looking at them ?

[feisty2]

holes?
I used expand.inpand to remove holes
maybe do it some more times?

[poisondeathray]

Originally Posted by feisty2

holes?
I used expand.inpand to remove holes
maybe do it some more times?

We don't know what is going on for sure, because he didn't post a video sample, only an image

But I don't see any big holes with your script on the still image. But what he says suggests either he's messed it up, or actually checking the video (not the stilll), or another video

"For some strange reason flat blue here and there becomes flat magenta" - I don't see that with the mask

Since we don't know what is going on for sure it's hard to suggest something, so I would wait for clarification




I think the artifacts have everything to do with this. I would treat the underlying problem first

[foxyshadis]

I wouldn't be surprised if it was a case where moving the U and V by 1 value each changed the color visibly, if it's that flat and you flipped back and forth enough.

[Hotte]

Hi everybody - thank you for your precious contribution.

I am most happy do upload a short video sample. Would 1 or max. 2 seconds be ok ?, because I have very very low upload speed and canopus avi are huge.

Would it be best to just upload the out of camera original ?

Could you recommend a storage site where I could upload without any registersoftwaredownloadwhatever hassel ?

I can do it only on the weekend cause I got that stuff at home.

Thanks.

[feisty2]

yeah, paid online hosting sites

[creaothceann]

Originally Posted by Hotte

Could you recommend a storage site where I could upload

MEGA, MediaFire

[feisty2]

found FQSharp (deconvolution filter)
now everything is done under avisynth
cool

[feisty2]

if someone really kind would like to port nnedi3 eedi3 fqsharp mvtools2 to 16bpc precision
that would be even better

[Hotte]

Hi,

I am back. After some research I realized, that I made a terrible mistake. Though I have to admit, it was tricky: The magenta artefacts did not derive from the sharpening nor from the quality of the original MOV, it happened in virtual dub! Without being clear how important they were I used default values for color depth which were:

Decompression format: Autoselect
Output format to compressor/display: 24 bit RGB (888)

The latter caused a significant drop of fine color resolution hence magenta artefacts and fince color blocks. It occured with a simple avisource() command already. After having changed output format to "Same as decompression format" the artefacts and tonal influences were gone!

I am so sorry, guys, for having driven you mad with this. Anyway I appreciated a lot that you were trying to help me.

I am sure now that all we discussed will make a very good fine sharpening solution - miles better than standard or in-camera sharpening.

I'll be back soon with a summary. Thank you all!

[feisty2]

comparisons (mcdeconvsharp (yeah, I decided to name dat bad ass slow filter this) screenshot will be added when the process is done since its just wicked slow, changed some things, the result would be different from (slightly better than) the screenshot showed earlier, and it runs even slower now, 0.000fps according to x264)

[Hotte]

Hi,

thank you for this comparison. mcdeconvsharp clearly makes the race!

However watching at this samples I come to the conclusion that our finesharpening approach is not the right solution for low-res material like this clip. It enhances small blocks way to much. Bottom line I find LSFmod better.

But things are completely different when it comes to 1920x1080 material. Here #1 (Hotte) and #25 (feisty2) of this thread are doing a very good job to pull out fine detail.

1. Original. Intentionally sort of soft:
https://mega.co.nz/#!AlBHAZQC!vUTmGb...97f_hLW5wtnUJw

2. Edius internal sharpener with moderate strength of 22. No bad detail but too much edge halo, shimmery stones, too much detail contrast. Not so pleasing:
https://mega.co.nz/#!hspGCRBQ!nKwjH0...t5u7qzAVuxgEsA

3. #25 from this thread. Very good detail enhancement. Well controlled edge halos/detail contrast:
https://mega.co.nz/#!x14zmYJT!pDgzym...KP0iSif-IRliXg

4. My #1 from this thread with s=70, w=1 plus dehalo_alpha to reduce edge halos and supersampling=1.5. Very fast. sharpening technique seems similar to 3. dehalo-alpha improves edge halos.
https://mega.co.nz/#!ZlYV1D5C!28oXZN...lpcfC9AfZ-vzbg

5. LSFmod (defaults="slow"). For me seems much closer to 2. than to 3. or 4.
https://mega.co.nz/#!UxZXAByY!nO4vGn...vb5lX-bKidU3jo

Fine detail can be seen best in the trees of the forest on the left side under the description text. Edge halos show nicely at the edge of the castle wall.

[feisty2]

yeah, the kernel of mcdeconvsharp is deconvolution, sure it brings hell lot of noises, but noise is not the concern of this (sharpen/deconvolution) step, all sharpen filters enhance noise more or less, just deconvolution enhances noise more wildly
thats why it aint done yet, i'm writing a denoise suite to clean the craps (including deconvolution noise)
i dont see noise as an artifacts at the step of "sharpening" cuz we are gonna remove it right afterward

[feisty2]

personally I think the most important feature that a sharpening filter must have is keeping the image looking natural not oil painted after the process
thats the reason i decided to write a new sharpening filter at the first place

[Hotte]

I think oilpainted effects are in most cases a result of edge halos in mid detail regions where outlines get doubled into one another. With the castle clip this makes stones "shimmer". The used approach does not really provoke the effect with FullHD material and Dehalo_alpha helps to put this down additionally. I feel this is very controlled here. I also tried to reduce grain with Mdegrain2 and Mdegrain3. The effect is very subtle. Is there something more effective with motion compensation ?

[feisty2]

are you asking is there something more effective than mc?
I honestly dunno

[Hotte]

feisty2, I am trying to understand how degrain works here. (How) can the very subtle degraining effect in this code being made stronger ? Changing thSAD only makes a very minor difference. Thanks.

source = (source clip)
sharp0 = (sharpened clip)
sharpD = mt_makediff(source,sharp0)
zeroD = sharpD.mt_lut("x",Y=-128)

sup1   = source.MSuper(pel=2,sharp=2)
sup2   = sharpD.MSuper(pel=2,sharp=2,levels=1)
bv1    = sup1.MAnalyse(isb=true, delta=1,overlap=4)
fv1    = sup1.MAnalyse(isb=false,delta=1,overlap=4)
bv2    = sup1.MAnalyse(isb=true, delta=2,overlap=4)
fv2    = sup1.MAnalyse(isb=false,delta=2,overlap=4)
bv3    = sup1.MAnalyse(isb=true, delta=3,overlap=4)
fv3    = sup1.MAnalyse(isb=false,delta=3,overlap=4)

sharpD.MDegrain3(sup2,bv1,fv1,bv2,fv2,bv3,fv3,thSAD=400)
source.mt_makediff(last,U=2,V=2)

[feisty2]

larger time radius
use "mdegrainN"
or use mdegrain3 multiple times