GenericFilters [Now part of VS core]

[Chikuzen]

GenericFilters-0.7.0.zip

https://github.com/chikuzen/GenericFilters

GenericFilters is a set of common image-processing filters.

I think that the necessity of using masktools2 of avisynth will decrease considerably if this plugin and std.functions are used together.

Cross reference list of masktools2 and GenericFilters

mt_edge - generic.Convolution/generic.Sobel/generic.Prewitt/generic.TEdge
mt_invert - generic.Invert
mt_binarize - generic.Binarize
mt_expand - generic.Maximum
mt_inpand - generic.Minimum
mt_inflate - generic.Inflate
mt_deflate - generic.Deflate
mt_clamp - generic.Limiter
mt_convolution - generic.Convolution/generic.ConvolutionHV
mt_hysteresis - generic.Hysteresis
mt_lut - std.Lut/std.Expr
mt_lutxy -std.Lut2/std.Expr
mt_lutxyz - std.Expr
mt_merge - std.MaskedMerge

EDIT:
20130812 update to 0.7.0

[Myrsloik]

It is my duty to shamelessly bump this post since it renders masktools almost completely obsolete. We're now one big step closer to getting completely native vapoursynth scripts for useful things.

[Daemon404]

Originally Posted by Myrsloik

It is my duty to shamelessly bump this post since it renders masktools almost completely obsolete. We're now one big step closer to getting completely native vapoursynth scripts for useful things.

I dream of a native OS X / Linux QTGMC.

[sl1pkn07]

https://github.com/chikuzen/GenericFilters/issues/1

[06_taro]

Originally Posted by HolyWu

Another thing is, I use the following script to compare the output of mt_edge and generic.Convolution.

clip = something
c1 = core.avs.mt_edge(clip, '1 2 1 0 0 0 -1 -2 -1 1', 0, 255)
c2 = core.generic.Convolution(clip, [1, 2, 1, 0, 0, 0, -1, -2, -1], planes=0)
clip = core.std.Interleave([c1, c2])
clip = core.std.ShufflePlanes(clip, planes=0, format=vs.GRAY)

I originally think the result of the two will be identical, but it's not. However I am no expert at matrix convolution so I may misunderstand its functionality and use it the wrong way.

I believe you intended to compare between avs.mt_convolution(clip, '1 2 1', '1 0 -1') and generic.Convolution(clip, [1, 2, 1, 0, 0, 0, -1, -2, -1], planes=0), which looks quite identical to me....

[Chikuzen]

hmm...

generic.Convolution always saturate the output on 8bit format.
mt_edge("1 2 1 0 0 0 -1 -2 -1 1", 0, 255) is same as mt_convolution("1 2 1", "1 0 -1", saturate=false)

currently, I'm working on imprementing Sobel and Prewitt proper function by the request from Myrsloik.
Since I'm not familiar with SIMD, I may require time for beginning it.

[Chikuzen]

Updated to 0.2.0

* Add new function 'Sobel'.
* Add new function 'Prewitt'.
* Convolution/HV: add new option 'saturate'.
* Minimum/Maximum: add new option 'coordinates'
* various bug fix.

[Revgen]

The coordinates option and saturate option fix the issues HolyWu described. Thank You.

[Chikuzen]

Originally Posted by HolyWu

But the output of generic.Sobel/generic.Prewitt is quite different from the equivalent mode of mt_edge.

from masktools2 docs

"sobel" uses the kernel "0 -1 0 -1 0 1 0 1 0"

I've never heard calling this "Sobel" except masktools.

"hprewitt" is equivalent to : mt_logic(mt_edge("1 2 1 0 0 0 -1 -2 -1 1"), mt_edge("1 0 -1 2 0 -2 1 0 -1 1"), mode = "max"),

hprewitt? I think this is Sobel
and, Prewitt that I wrote is 'Prewitt compass edge detection' which requires 8 kernels.

[Myrsloik]

Originally Posted by Chikuzen

from masktools2 docs

I've never heard calling this "Sobel" except masktools.


hprewitt? I think this is Sobel
and, Prewitt that I wrote is 'Prewitt compass edge detection' which requires 8 kernels.

Now you made me go check the reference literature...

Sobel is indeed defined as 1 2 1, 0 0 0, - 1 -2 - 1 and then the horizontal and vertical filters are usually summed up as sqrt(sh^2+sv^2) or abs(sh) + abs(sv) for speed. There are also variants where the filter is rotated when looking for diagonal edges but I've never seen it used.

Now on to prewitt. Apparently the prewitt guy was busy and suggested several similar things. But the one you want is described here: http://en.wikipedia.org/wiki/Prewitt_operator
and should be just about identical to sobel except that it's 1 1 1.

I really can't find any mention of masktool's so called sobel filter. I can however tell you this, it really is implemented like a single 0 1 0, 1 0 -1, 0 -1 0 convolution and I think it looks like a horrible idea. It's easy to see that edges get a VEEEEEERY different weight depending on directionality. Just imagine a diagonal gradient in the \ and / directions and see the difference in result.

[Chikuzen]

I found some typo and wrong understanding on Sobel and Prewitt.

updated to 0.2.1

* Sobel/Prewitt: fix invalid logic
* Sobel/Prewitt: Change default values of 'min' and 'max'

[aegisofrime]

So is there an updated version of QTGMC implementing these generic filters yet?

[Revgen]

^Not yet.

The current Vapoursynth QTGMC script is here. http://forum.doom9.org/showthread.php?t=166582

[Chikuzen]

updated to 0.2.2

* Sobel, Prewitt: add new option 'rshift'
* Sobel, Prewitt: fix 9/10/16bit format processing

[Are_]

Link to 0.2.2 does not work.

[Chikuzen]

Originally Posted by Are_

Link to 0.2.2 does not work.

fixed

[Chikuzen]

updated to 0.3.0

* add new function 'Blur'

[kolak]

[kolak]

Looks like there is a bug in blur when format is 16bit- footage has some green artefacts (10bit and 8bit are ok).

[Chikuzen]

(Probably,) fixed.
genericfilters_0.4.1.dll