How to remove huge but transparent watermark?

[Athari]

Original video: http://www.youtube.com/watch?v=RVQev9V-Ybw
Screenshot:


There's a huge watermark over the whole screen, but it's transparent (only around 15% opacity) and it's quite easy to reproduce roughly.

I've tried using DeLogo to analyse the whole video with this DeBlend mask, but results were unsatisfactory. Some frames looked quite good, some looked terrible. The plugin came to wrong conclusion that some areas need to be repaired etc.

I've also tried applying reverse alpha blend using my own application and this reproduction of original watermark (it's white with alpha channel, so you may see only white) and got this result:

While not perfect, all I need now is to use DeBlend to repair small edges. (I believe it can be done, but I haven't tried it yet.)

The problem is, my app is for still images, but I need a filter. Is there a filter for VirtualDub which can apply alpha blending in reverse order, that is, (Val1–Val2*Alpha)/(1–Alpha) instead of normal Val1*(1–Alpha)+Val2*Alpha?

By the way, if you know a plugin for “reversed” applying of layers for PhotoShop, please tell me. I don't really like reinventing wheels. This method works perfectly in case it's possible to exactly reproduce original watermark.

P.S. Sorry for possible mistakes in my English, my native language is C++. Er... I mean, Russian.

[jagabo]

Do you have a sample frame in the video with the watermark over a black background? Like from a fade to black sequence. With that and the proper settings Delogo should be able to get rid of the logo almost completely.

Can you provide a short video with the watermark and that black shot?

[Athari]

Do you have a sample frame in the video with the watermark over a black background? Like from a fade to black sequence. Can you provide a short video with the watermark and that black shot?

Unfortunately, I don't have it. There's no fade to black. And Jamba (whoever they are) seem to use different watermarks every time, sometimes they don't use it.

Well, you can get this “frame” by applying the watermark PNG drawn by me (see above) to the black background. However, it's not pixel-perfect.

[jagabo]

Well, you can get this “frame” by applying the watermark PNG drawn by me (see above) to the black background. However, it's not pixel-perfect.

If it's not pixel perfect it won't work very well. And the link didn't work for me.

<edit>All I saw was a white page -- but I see now it was a white image on a white page. So I saw nothing but white. I'll see if I can do anything with it.
</edit>

[budwzr]

Mask the watermark and composite it on top using brightness and contrast to match.

[Athari]

Mask the watermark and composite it on top using brightness and contrast to match.

Hmm... While it's possible to get (Val1–Val2*Alpha)/(1–Alpha) from subtraction (Val1*(1–Alpha)–Val2*Alpha) and brightness/contrast (something like Contrast*Value+Brightness I think), I need a few more things. First, I need exact formula for brightness/contrast to precisely calculate (1–Alpha) coefficient. Second, I need a way to apply a masked brightness/contrast filter to the image.

Okay, I got the result, similar to the one from my app, by playing around with a masked brightness/contrast adjustment layer in PhotoShop (values I used: legacy mode, -40 brightness, +30 contrast; couldn't get it without legacy mode, somehow it's contrast adjustment is too limited).

Still, I don't know how to reproduce it in VirtualDub. Values in its brightness/contrast filter have different ranges: black-norma-white for brightness, 0%-200% for contrast (exact numbers aren't even displayed!). And I don't see any way to apply the effect only to a portion of the image (to the watermark area that is).

Anyway, I'd greatly appreciate a filter which can simply use the exact formula to reverse alpha composition, instead of playing around with multiple filters and masks trying out a lot of different combinations of numbers.

[budwzr]

Hmmm...that's interesting. Let me try too.....

[budwzr]

Hmmm...that's interesting. Let me try too.....

No luck, so I turned it into a generic movie trailer. http://youtu.be/HuD3wlm1U5U

[poisondeathray]

The problem is, my app is for still images, but I need a filter.

That's not a problem, video is just a string of images. You could convert to a sequence of .png for example, apply your app (I'm assuming it's some CLI app), then reassemble the images either in vdub , avidemux, or avisynth etc...

[Athari]

That's not a problem, video is just a string of images. You could convert to a sequence of .png for example, apply your app (I'm assuming it's some CLI app), then reassemble the images either in vdub , avidemux, or avisynth etc...

I can just write a filter plugin for VirtualDub. But I can hardly believe I'm the first person ever who wants to remove a watermark from a video or a picture by reversing the formula of alpha composition.

[lordsmurf]

reversing the formula

That doesn't even make sense.
You may as well just reverse the polarity of the neutron flow, while you're at it.

[Athari]

That doesn't even make sense.
You may as well just reverse the polarity of the neutron flow, while you're at it.

It would be nice of you to read the thread first and try to understand it.

[jagabo]

reversing the formula

That doesn't even make sense.

It makes perfect sense. Simplified (a 50:50 mix):

B = blended image:
O = original image
L = logo image

B = (O + L) / 2

Reversed:

O = B * 2 - L

I've been playing around with this a bit with AviSynth's Overlay() but don't quite have the weighting right yet:

[intracube]

Here's my attempt:


BTW, does anyone know how to upload images, but only have them appear as links on my post? I've got several large images that I don't want to clutter the thead with.

[Athari]

Here's my attempt:

How did you get it?

[intracube]

Here's my attempt:

How did you get it?

GIMP, Blender, mencoder.

I'll post a detailed explaination and the image masks I used later - I've got some enchiladas in the oven that need my attention

There's also a couple of issues I want to fix with the processing.

[poisondeathray]

That's not a problem, video is just a string of images. You could convert to a sequence of .png for example, apply your app (I'm assuming it's some CLI app), then reassemble the images either in vdub , avidemux, or avisynth etc...

I can just write a filter plugin for VirtualDub. But I can hardly believe I'm the first person ever who wants to remove a watermark from a video or a picture by reversing the formula of alpha composition.

Maybe that's the one of the formulas used "under the hood" ?

Have you checked any of the plugins in avisynth or vdub ? There must be a dozen of them

I haven't really used any of them, but I know some use inpainting and might not apply a simple equation




@intracube - very nice repair

[budwzr]

Yes, that was a nice job.

However, I'm assuming you painstakingly "hand-crafted" the mask, yes?

I think the OP, as well as myself, is looking for an algorithm that can be automated, using standard compositing operators. That would be a very useful routine indeed.

My suggestion is anyone that can come up with that, should patent it before releasing it here.

[budwzr]

reversing the formula

That doesn't even make sense.
You may as well just reverse the polarity of the neutron flow, while you're at it.

Hahaha, that's funny.

But Scotty reversed flow of the dylithium crystals a lot, and it always worked.

[intracube]

@poisondeathray - thanks

Yes, that was a nice job.

However, I'm assuming you painstakingly "hand-crafted" the mask, yes?

Yup. Although I wouldn't say it was painstaking to recreate. I've traced fonts before, so this was relatively quick; 6 unique characters, and all but 2 comprised of straight lines. I had to make adjustments to my traced copy of the text to get it to line up properly which took a bit of time.

I think the OP, as well as myself, is looking for an algorithm that can be automated, using standard compositing operators. That would be a very useful routine indeed.

It would have to be a mighty sophisticated algorithm to do that

Part of the delay this afternoon was dealing with the odd resolution YT video - mediainfo reports the clip as 854x468. I did most of this last night, and for reasons unknown I chose a project resolution of 854x476 which complicated things...

Blenders node compositor did most of the processing. Here's the mask I used:

The image above influenced how much effect this CRGB curve had on the image:

which removed most of the text.

Then I overlayed a blurred (3x3 radius) copy of the video using this image to control the amount of overlay:

to try and hide any remaining visible edges. This trick only worked because the video was soft - if it had been pin sharp, you'd see a blurred silhouette of the text.

The heavy compression of the YT video causes the edges of the overlayed text to vary/move slightly which makes it almost impossible to completely remove the text - even with a perfect mask. Applying the procedure to an uncompressed or lossless compressed stream would give better results.

I'll upload the processed video shortly.

[poisondeathray]

Another reason for the edge problems (along with compression issues) is the original overlay was probably done in RGB, and the source was originally RGB (because it's CGI material). The subsampled chroma when encoding to 4:2:0 YUV means you will never get a "perfect" overlay on the edges

intracube - did you just "eyeball" the amount for the curves until it looked ok ? or apply some formula

[intracube]

intracube - did you just "eyeball" the amount for the curves until it looked ok ? or apply some formula

I made the adjustments by eye. I temporarily bumped up the contrast on the output which helped show up any slight mismatch in levels/alignment.

megaupload is saying the file I uploaded is 'temporarily unavaiable'. I'll give it an hour, in case they've got problems at their end.

[lordsmurf]

It would have to be a mighty sophisticated algorithm to do that

Yep, that's what I was getting at.
There's no magic "remove me" brush. You're pretty much forced to paint frames manually.

[intracube]

It would have to be a mighty sophisticated algorithm to do that

Yep, that's what I was getting at.
There's no magic "remove me" brush. You're pretty much forced to paint frames manually.

The 2 masks only have to be created once. The transformation is then applied to the video without any manual adjustments needed for individual frames.

Processed video:
http://www.megaupload.com/?d=43RZ01AL

[budwzr]

It would have to be a mighty sophisticated algorithm to do that

Yep, that's what I was getting at.
There's no magic "remove me" brush. You're pretty much forced to paint frames manually.

What I meant by "algorithm" is a formula. Like a way to create a difference, outline, or luminance mask as opposed to a hand drawn one. What if the "watermark" is George Washington's bust? Kapeesh?

[budwzr]

I downloaded the final, and that's some great work there intracube.

[poisondeathray]

^ yes , nice job on the video

[Cauptain]

Unbelievable result...It´s MAGIC?

Congratulations intracube


Claudio

[lordsmurf]

The pop-up ads from Megaupload can come with malware payloads. Not good.
So I've attached your video here instead.

There's still a number of frames that need manual repaint. But it is a really decent job for removing a translucent simple color-less alpha layer. I would be surprised if this method still worked for colored text, or slightly less translucent objects.

Too many people want to treat logos as if it were something that could be fixed by hitting a magic "undo" button.

Still interesting to see what people attempt. Better than your average noise blob.


..

[AlanHK]

I can just write a filter plugin for VirtualDub. But I can hardly believe I'm the first person ever who wants to remove a watermark from a video or a picture by reversing the formula of alpha composition.

You're not.
I use NoLogo in Avisynth.
Made in 2003 still works.
http://forum.doom9.org/showthread.php?t=64031

Since the original site is gone, I attached the files.

I use the simplest mode.
First get a clean shot of the logo on a black background, save as png.

Then :

AVISource("3_~_Play_Me.avi")
loadplugin("P:\AviSynth 2.5\Old\LoadPluginEx.dll")#to load an old (pre2.5) plugin
loadplugin("P:\AviSynth 2.5\Old\logotools.dll") 
ConverttoYUY2() #need to change the colour space
logo=ImageSource("logo302.png").ConverttoYUY2() #create the logo clip from the image
NoLogoAuto(logo,1) #apply the filter
ConvertToYV12() # and change it back

The only hard part can be finding a clean image of the logo.
Often I have to PhotoShop a screengrab, desaturate, use Levels to adjust the background around the logo so it's true black and fill the rest of the screen with black. But with luck it's over the credits on mostly plain black.