New Sharpening Function

[Archimedes]

Some Experiments.

Original (upscaled from 512x384 to 1024x768 with NNEDI3) >>

SSSharp >>
ReCon >>

[Didée]

Originally Posted by Archimedes

It's much more the "difference.downsize" which I dislike in combination with normal sharpening.

That implies you never downscale a [too-] large image to a smaller size? Because downscaling does necessarily look bad?

Perhaps the problem comes from using overly aggressive settings, which make the result look bad already at the supersampled stage, so much that the final downscale can't rescue the damage anymore ...

In case it's because of the inherent blurring that happens during downscaling - that effect can be combat'ed, too. Just highpass the downscaled difference before applying it.

[*.mp4 guy]

There shouldn't be anything in ReCon that makes it produce more aliasing than an equivalent strength unsharpmask. Everything that inherently needs supersampling effectively already has it. The butterfly has aliasing because the sharpening settings used on it are of the same strength as unsharp(vary=4, str=6), which also produces aliasing.

The vast majority of the time ReCon is just sharpening aliasing that was already there. I can't prevent that without also preventing it from sharping details that occupy the same space.

If you look at the example image, you can see that the lines were not completely alias free before being sharpened. sssharp barely touches the lines, and doesn't reveal much aliasing. ReCon does sharpen the lines, and reveals aliasing. The two white dots indicate the places where ReCon actually screws up.

Originally Posted by Didée

In case it's because of the inherent blurring that happens during downscaling - that effect can be combat'ed, too. Just highpass the downscaled difference before applying it.

The inherent blurring would probably be the only thing causing giving aa in this case. Anyway, the correct thing to do would be to rework the sharpening kernel to avoid sharpening high frequencies, which would do the same thing. Only problem is that I want to sharpen the high frequencies.

[foxyshadis]

The only thing that currently turns me off to ReCon relative to SSSharp is the way straight lines become dirty (looks like a cat chewed on them). Otherwise it's overall better. I looked at the problem but could think of no way to correct it without "correcting" most of the sharpening.

[*.mp4 guy]

Originally Posted by foxyshadis

The only thing that currently turns me off to ReCon relative to SSSharp is the way straight lines become dirty (looks like a cat chewed on them). Otherwise it's overall better. I looked at the problem but could think of no way to correct it without "correcting" most of the sharpening.

Could you post a really bad example of this?

[Archimedes]

Having problems with bright areas after sharpening (see branches in the middle of the pictures). To show the effect, I started with an aggressive setting (example 1). But the effect is also visible with a less aggressive setting (example 2).

Original >>
Example 1 >>
Example 2 >>

[rfmmars]

Originally Posted by Archimedes

Having problems with bright areas after sharpening (see branches in the middle of the pictures). To show the effect, I started with an aggressive setting (example 1). But the effect is also visible with a less aggressive setting (example 2).

Original >>
Example 1 >>
Example 2 >>

Uploaded with ImageShack.us



Uploaded with ImageShack.us



Uploaded with ImageShack.us
My try at them.

[Undead Sega]

Originally Posted by Archimedes

Some Experiments.

Original (upscaled from 512x384 to 1024x768 with NNEDI3) >>

SSSharp >>
ReCon >>

SSSharpen never fails to amuse me, I wish I can see more examples or do them myself but I just dont know how to do so.

[SilaSurfer]

Originally Posted by Didée

In case it's because of the inherent blurring that happens during downscaling - that effect can be combat'ed, too. Just highpass the downscaled difference before applying it.

Didée, could you elaborate please? I'm interested in combating blurring while downscaling from a 1080p Bluray sources. I know this is for Upsize->Filter->Downsamle, but I'm only Downsampling.

[Jenyok]

Very interesting result (for me) is achieved using following code in following sequence (not another).
.

MCSharpening()
ReCon(str=4, rad=4, lmode=4)   # various parameters

.
function MCSharpening(clip source)
http://forum.doom9.org/showthread.php?t=153170&page=3
.

function MCSharpening(clip source)
{
str = 1   # strength of predictionfilter, don't change

#source = last

#comment out tblurnl if you don't have residual low frequency blotchiness which fft3d loves to leave on everything
pred_   = Source.fft3dfilter(bw=6, bh=6, ow=3, oh=3, bt=1, sigma=str, plane=4). \
                 fft3dfilter(bw=216, bh=216, ow=108, oh=108, bt=1, sigma=str/8, sigma2=str/4, sigma3=str/2, sigma4=str, plane=4).TblurNL(rad=2, thresh=4)
pred    = pred_.ttempsmooth(maxr=7).gradfun2db(1.01) #.hqdn3d(0.1, 0.1, 1.5, 1.5).gradfun2db(1.01) is faster and nearly as good, except on scene changes

super   = pred.MSuper(pel=2, hpad=16, vpad=16, sharp=2)

backward_vec2 = MAnalyse(super, isb = true,  delta = 2, overlap=4, truemotion=true, dct=5)
backward_vec1 = MAnalyse(super, isb = true,  delta = 1, overlap=4, truemotion=true, dct=5)
forward_vec1  = MAnalyse(super, isb = false, delta = 1, overlap=4, truemotion=true, dct=5)
forward_vec2  = MAnalyse(super, isb = false, delta = 2, overlap=4, truemotion=true, dct=5)

maskp1 = Source.mmask(forward_vec1 ,kind=1, ysc=255).UtoY()
maskp2 = Source.mmask(forward_vec2 ,kind=1, ysc=255).UtoY()
maskp3 = Source.mmask(backward_vec1,kind=1, ysc=255).UtoY()
maskp4 = Source.mmask(backward_vec2,kind=1, ysc=255).UtoY()
maskf  = average(maskp1, 0.25, maskp2, 0.25, maskp3, 0.25, maskp4, 0.25).spline36resize(source.width, source.height)

smooth  = pred_.gradfun2db(1.01)
source2 = MT_Merge(source, smooth, maskf)

source3 = source2.MDegrain2(super, backward_vec1,forward_vec1,backward_vec2,forward_vec2,thSAD=400)
source3

# remove temporally blurred high frequency crud, can cause blocking upon reencode
VagueDenoiser(method=4, nsteps=10, wavelet=2, Wiener=true, auxclip=pred, percent=95, chromaT=1.0, wratio=0.75, threshold=0.5)

return gradfun2db(1.01)
}

.
ReCon(str=4, rad=4, lmode=4) # various parameters
http://forum.doom9.org/showthread.php?t=155030