Filter white specks with Avisynth

[AlanHK]

I've got a video that is not bad except for white dust specks.
Colour, 464x352.

I'd like to remove all the 2x2 or smaller white spots that last just one frame.
E.g., as image, top right.


I tried the RemoveDirt() AVSI, but while it gets rid of the spots it destroys too much detail. Even extreme sharpening after can't restore the original detail.
It's very good if you have a crappy source, but is overkill for this.

Is there a different filter that is more specific or does less collateral damage?
Or a tweak to RemoveDirt -- the "repmode" perhaps? This doesn't seem to be documented though.

[manono]

RemoveSpots, or RemoveSpotsMC?

[jagabo]

And you could use a mask based on the luma so that only the brightest areas are treated.

AviSource("spot0.avi") 
ConvertToYV12()
src=last

lmask = mt_binarize(150).mt_expand().mt_expand().Blur(1.0) # build a mask based on luma
Overlay(last, RemoveSpotsMC(), mask=lmask) # only apply RemoveSpotsMC() to bright areas
StackHorizontal(src, last) # show before and after, side by side
AssumeFPS(3) #slow so we can clearly see it

[manono]

That's pretty cool. Can you explain how only bright spots are treated? Is that what the 'mt_binarize(150)' does? It calls RemoveSpotsMC to be used only in areas where the luma is greater than 150?

To sidetrack this a bit, I'm much more likely to have dark spots needing removing. How would you adapt the script to use RemoveSpotsMC only in areas below ... say ... 50?

[poisondeathray]

That's pretty cool. Can you explain how only bright spots are treated? Is that what the 'mt_binarize(150)' does? It calls ReplaceFramesMC to be used only in areas where the luma is greater than 150?

To sidetrack this a bit, I'm much more likely to have dark spots needing removing. How would you adapt the script to use ReplaceFramesMC only in areas below ... say ... 50?

Do you mean RemoveSpotsMC ? Because ReplaceFramesMC() wouldn't work so well with a luma mask in general because the timing of the replacement overlay wouldn't match the baseclip

There is a lumamask() function that you can set black and white points with b=, w=

But mt_binarize creates a binary mask ("binary" as in on/off or white/black) . If you use mt_binarize(50) , the cutoff will be 50, but 0-50 will be black, 51-255 will be white. Use upper=true to "invert" the mask (so 0-50 is white, 51-255 is black)

Often you don't want such a "harsh" or "abrupt" cutoff for the transitions for the mask overlay , ie . you want to "feather" the mask, so the overlay "blends" in . That's what jagabo is doing by expanding the mask and blurring it; as usual there are many approaches . You can "view" the mask by returning "mymask" in the example below. Of course replace somefilter() with your intended filters

eg

orig=whateversource()

orig
mt_binarize(50, upper=true)
mymask=last

orig
somefilter()
filtered=last

overlay(orig, filtered, mask=mymask)
#mt_merge(orig, filtered, mymask, luma=true)

I was discussing with gavino , and mt_merge() is preferred to overlay() in the strict YV12 case, because it's faster and uses less memory (overlay has more uses when using other colorspaces, or using various layer blend mode operations)

[manono]

Damn, I read and reread it before posting but still... Yes, I meant to say RemoveSpotsMC. I'll fix it above.

Ah, Upper=true does it. No matter how good the RemoveSpotsMC, it can't help but soften the picture a bit and this sort of thing can lessen the 'damage' outside of the areas being filtered. Thanks a lot to you both.

[Betelman]

AlanHK makes sure to take extra care of his chocolate