help for deblock (qed or other)

[byme]

I can not remove blocks so effective
the DeBlock creates jagged, the DeBlock_QED seems to be ineffective

source
http://d.imagehost.org/t/0307/source_1.jpg (http://d.imagehost.org/view/0307/source_1)

deblock_qed
http://b.imagehost.org/t/0820/deblock_qed_1.jpg (http://b.imagehost.org/view/0820/deblock_qed_1)


source
http://b.imagehost.org/t/0195/source_2.jpg (http://b.imagehost.org/view/0195/source_2)

deblock_qed
http://b.imagehost.org/t/0122/deblock_qed_2.jpg (http://b.imagehost.org/view/0122/deblock_qed_2)

you have some advice?

[Didée]

It seems the level of degradation is too bad that deblock_qed would be appropriate. However, seeing how very little change there is, I suppose you didn't use it the right way. You should not filter before qed, and you must not resize before qed, and you must not crop before qed.

You can try a plain Deblock(50), that will be more hard on the blocking.

You could also try Bicubicresize(104,56).bicubicresize(1024,576,1,0). This will most definetly remove all of the blocking. :p

[byme]

It seems the level of degradation is too bad that deblock_qed would be appropriate. However, seeing how very little change there is, I suppose you didn't use it the right way. You should not filter before qed, and you must not resize before qed, and you must not crop before qed.

You can try a plain Deblock(50), that will be more hard on the blocking.

You could also try Bicubicresize(104,56).bicubicresize(1024,576,1,0). This will most definetly remove all of the blocking. :p

excellent ... as always :eek:

thanks for other answers...

[Didée]

There's lots of other options to try.

- fft3dfilter
- dfttest
- tnlmeans
- frfun
- dctfun
- ...

There is no one filter that will turn those TV-caps into gold. Removing the artifacts will also destroy even more of the little detail that actually is there, and/or will make everything more blurry. And sharpening isn't an easy task either, since the source already comes along with halos.

[byme]

thanks, I'll try them all

the frfun I've already tried, but raising the value flattens too... and I took

[Bi11]

However, seeing how very little change there is, I suppose you didn't use it the right way.
That's probably my fault.
I lowered the defaults in an attempt to reduce blurring/smoothing of film sources, but it seems those defaults may have been too low.

I have now updated the defaults in the Deblock_QED script (http://forum.doom9.org/showthread.php?p=1402777#post1402777) to be more effective, but still within sensible limits.
I've also replaced mt_LutSpa()'s legacy parameter 'relative' with the new 'mode' parameter introduced in MaskTools 2.0a45.

[Mounir]

That's probably my fault.
I lowered the defaults in an attempt to reduce blurring/smoothing of film sources, but it seems those defaults may have been too low.

I have now updated the defaults in the Deblock_QED script (http://forum.doom9.org/showthread.php?p=1402777#post1402777) to be more effective, but still within sensible limits.


Show us the result

[Didée]

Don't hold your breath, it's not for such bad sources like shown in the OP, anyway. QED is made to mainly remove blocking at block borders, it does hardly touch the interiour of blocks, in order to keep detail. (Back when deblock_qed was born, my point was that most other deblockers smooth too much detail when you try to remove blocking of otherwise good quality sources.)

With such low quality sources, the problem is not only block borders, but also lots of DCT-crap *inside* of the blocks. And again, QED doesn't even try to clean the block interiours....
Hence, for such sources a plain deblock() is more suited, or mpeg2source()'s deblocking, or fft3d/dfttest/tnlmeans. And yes, all those will make such sources more blurry when they remove the blocking.

That's how it is. Cleaning bad sources comes at a price.

You can't make an omelet without breaking eggs. Period.

[byme]

the dfttest , before Sharpen, creates "rings" on color, characteristic of YV12

http://file2.net/thb/101019/8970ug.png (http://file2.net/101019img8970ug.png.html)

is there a way to remove them?

the GradFun2DBmod in fact eliminates them pretty good, but creates a grain that I do not like

sets so: GradFun2DBmod(str=0.0, strC=0.0, temp=-1, adapt=-1, thr=2.4, mode=2)

[Mounir]

My opinion: You should try msu smart deblocking in RGB, that's the best deblocker i know and then clean after with whatever denoiser you like.

[cretindesalpes]

the dfttest , before Sharpen, creates "rings" on color, characteristic of YV12
is there a way to remove them?
Use some kind of unsharp mask in the first place. A decent sharpener or sharpening combo shouldn't amplify +/-1 pixel value differences.

[Usedocne]

byme
Maybe try deblocking the luma at a lower strength than the chroma.

It might help by:
- Retaining more perceived detail.
- Less blurring.
- Less sharpening needed. (so as not to enhance haloes)

Example:

dfttest(Y=true, U=false, V=false, sigma=16) #Luma deblock
dfttest(Y=false, U=true, V=true, sigma=32) #Chroma deblock

or

FFT3DFilter(sigma=4, plane=0) #Luma deblock
FFT3DFilter(sigma=8, plane=3) #Chroma deblock

note: sigma strengths aren't tuned.