InpaintDelogo - advanced logo removal & hardcoded subtitles extraction

[kedautinh12]

Originally Posted by VoodooFX

If you understand then I don't understand what you are asking.

You said: "join parts with no logo, and delogo and join the animated parts". And i ask my scripts above correct with "delogo part" as you said

[VoodooFX]

Originally Posted by kedautinh12

You said: "join parts with no logo, and delogo and join the animated parts". And i ask my scripts above correct with "delogo part" as you said

Sorry, I don't understand what you write.

[kedautinh12]

Originally Posted by VoodooFX

Sorry, I don't understand what you write.

Lol, just forget about it

[pcroland]

If anyone is interested I created a small script that can extract forced cues from a full subtitle based on the folder that SubsMask2Img created:
https://github.com/pcroland/substoforced

❯ ./stf.py -h
substoforced 1.0.0

USAGE: stf.py [-h] [-v] [-s SUB] [-f FOLDER]

FLAGS:
  -h, --help                  show this help message.
  -v, --version               show version.
  -s, --sub SUB               specifies srt input
  -f, --folder FOLDER         specifies a folder where SubsMask2Img generated timecodes (optional)
                              you should remove the junk from there manually

[VoodooFX]

Originally Posted by pcroland

If anyone is interested I created a small script that can extract forced cues from a full subtitle based on the folder that SubsMask2Img created:
https://github.com/pcroland/substoforced

Could you describe more specifically what it does?

[pcroland]

It keeps the cues from the full subtitle that overlaps with any of the timecodes generated by SubsMask2Img. Lets say you have an English movie with French parts that are hardsubbed and you have a dub in another language where French parts aren't translated and you only have a full subtitle for it. You can generate the forced automatically for it.
Another use case would be to have a video with hardsubs that are also present in the full subtitle. You could create a subtitle based on your full subtitle without OCRing image (so it has less errors).

[VoodooFX]

Now it's clear. I was thinking what you are extracting there when you have srt already.

[Emulgator]

First off, many thanks to you, VoodooFX for building and maintaining that huge and useful script.

A suggestion: While having to handle plain white text logo with black borders (subtitle style) I came across the need to give expand=3.
expand=2 would still leave inpainting artifacts, which postprocessing could mend only a bit.
Going into script version 2.17 line 1178 it was well possible to change the assert line to acommodate expand up to 5 without breaking the underlying function balloonPX.

buginflate = (inflate < 0 || inflate > 2) ? true : false

changed to

buginflate = (inflate < 0 || inflate > 5) ? true : false

Maybe worth a suggestion to have that in the coming version.

[VoodooFX]

@Emulgator

Thing is that I never saw a practical case where I would need it more than 2...
You would need to share your logo case so I can see it with my own dead eyes to change my mind on that option.

Btw, you are not the first one asking this, and all of them were doing something wrong and wanted to "expand" stuff more to cover mistakes.

[Emulgator]

It was a subtitle style logo. White body, plus black borders, plus a bit halo.
Photosobbing -> Curves -> Isolating the white body left my mask without the black borders, let alone the faint halo, so too thin
To get that inpainted properly it would either take photosobbing again to pad up the lost borders and halo, or take the borders into account first.

Well, more work...or simply hand that padding job over to Inpaint Delogo, which did work nicely.

I guess the old fat bordered Sony Handycam style time code rendering would benefit from that too, still got to capture a case to publish.
Badly needed for stabilising such footage, because stabilising alone leaves you with shaky timecode, so timecode got to be removed first before stabilising.

[VoodooFX]

Originally Posted by Emulgator

Photosobbing -> Curves -> Isolating the white body left my mask without the black borders

So it's problem of you not making a proper basemask with Photozhop. If logo has black borders then select a frame with a bright background for edits. Use Magic Wand [there is Contiguous option] and Paint Bucket.
Or just use Automask.

Originally Posted by Emulgator

I guess the old fat bordered Sony Handycam style time code rendering would benefit from that too

Inflate is only for static logos, for a dynamic mask there is DynInflate, it's limited to "4".

[Emulgator]

Good, I will try that.
Automask got it wrong, no pure logo on black to be had, so I would have liked to tune thresholds and such, but at the moment I am only a Delogo beginner...

[VoodooFX]

Did you tried to use aMix option with Automask? Are you sure logo is not transparent?
Without video I can't tell you much, share it by PM if you want.

[Guilbert]

Hi, @VoodooFX.

Can you please take a look to the video I send you?
I can't get a good result with InpaintDelogo to OCR these hardcoded subtitles.

I tried with these parameters, with no luck:

InpaintDelogo(Loc="280,560,-280,-40", Show=4, DynMask=4, DynTune=210, DynMask4H=60, DynMaskUp=2)

Thank you for your help!

[VoodooFX]

Originally Posted by Guilbert

I tried with these parameters, with no luck:

InpaintDelogo(Loc="280,560,-280,-40", Show=4, DynMask=4, DynTune=210, DynMask4H=60, DynMaskUp=2)

Did you read the manual how to set those parameters? In this video a text brightness is way over "210" and a halo brightness as "60" is way too low.

DynTune aka text brightness - set high as possible [till text doesn't disappear].
DynMask4H aka halo brightness - set low as possible [till letters doesn't disappear].
KillBlobs - set low [ignore "-"] as possible [till letters doesn't disappear, good indicator are capital "M" and "W" letters].

v=LWLibAvVideoSource("D:\video.1080p.mkv")
x=v.crop(360,820,-360,-80)

o1=v.InpaintDelogo(Loc="360,820,-360,-80", Show=4, DynMask=4, DynTune=210, DynMask4H=60)
o2=v.InpaintDelogo(Loc="360,820,-360,-80", Show=4, DynMask=6, DynTune=240)
o3=v.InpaintDelogo(Loc="360,820,-360,-80", Show=4, DynMask=4, DynTune=240, DynMask4H=140)
o4=v.InpaintDelogo(Loc="360,820,-360,-80", Show=4, DynMask=4, DynTune=240, DynMask4H=140, KillBlobs=-4, KillNoise=-5)

Stackvertical(
\ x,
\ o1.ConvertToYV12().AddBorders(0,2,0,0,$F0F080).Subtitle("1: your settings"),
\ o2.ConvertToYV12().AddBorders(0,2,0,0,$F0F080).Subtitle("2: only DynTune"),
\ o3.ConvertToYV12().AddBorders(0,2,0,0,$F0F080).Subtitle("3: DynMask=4"),
\ o4.ConvertToYV12().AddBorders(0,2,0,0,$F0F080).Subtitle("4: DynMask=4 + KillBlobs & KillNoise"))

Frame 12270:




Here is the extracted result ["00_08_09_640__00_08_12_360.png"]:

InpaintDelogo(Loc="360,820,-360,-80", Show=4, DynMask=4, DynTune=240, DynMask4H=140, KillBlobs=-4, KillNoise=-5, DynMaskUp=2, Extract=1, ImgDir="D:\Images", ImgInflate=1)

ImgInflate=1 because DynTune=240 is a bit too high for some sub lines [there are some small "holes" in text], so ImgInflate=1 fills them in.

[Guilbert]

No, I didn't read the manual yet, but I will, thank you for the advice.
Thank you as well for the explanations and those parameters, it looks amazing!

Do you know if Google Cloud Vision is able to detect italics, by any chance, please?
(The API was added recently in Subtitle Edit. )

[VoodooFX]

Originally Posted by Guilbert

Do you know if Google Cloud Vision is able to detect italics, by any chance, please?
(The API was added recently in Subtitle Edit. )

I dunno. You can ask at Subtitle Edit's thread.

[Zetti]

I have tested the subtitle extraction function with your settings from post 435 like this:

LoadPlugin("C:\MeGUI-x64\tools\ffms\ffms2.dll")
FFVideoSource("D:\Test\Test.mp4", threads=1)
InpaintDelogo(Loc="164, 362, -150, -38", Show=4, DynMask=4, DynTune=240, DynMask4H=140, KillBlobs=-4, KillNoise=-5, DynMaskUp=2, Extract=1, ImgDir="D:\Test", ImgInflate=1)

So i can see what it is doing with my video test, but i got this result out of it:



What can i have done wrong?

[VoodooFX]

Originally Posted by Zetti

I have tested the subtitle extraction function with your settings from post 435 like this:
...
What can i have done wrong?

Most likely - subtitles in your video are not the same type as subtitles in Guilbert's video, you need to set parameters fitting your subtitles.
Cut [no re-encode] and post a short few mins sample of video if you want advice [or a whole video by PM].

[Zetti]

Here is small sample: https://www.file.io/9gdT/download/Xm0V7j2kmDMc