Thoughts about a new IVTC algorithm

[Zarxrax]

This idea just happened to pop into my head a while ago, so I thought maybe I'd share it. I'm not a coder, so I don't know where this idea is any good or not. Maybe it might would be usefull for detecting those interlaced mouths in anime or something...
Anyways, couldn't it be possible to detect interlaced frames by first blending the fields, then comparing this to the original frame? The non-interlaced parts of the image would stay mostly the same, while the interlaced areas would change a good bit. Depending on the ammount of change it detects, perhaps it could find interlaced areas?

[RadicalEd]

sorry for being OT but hey zarxrax, I didnt know you were a member here :D

[neuron2]

Your idea is not clear because if you blend two fields, you still have a field. How can you then compare that to a frame?

These are the ways I know of to detect combed areas:

1. Interframe comparison. Compare pixel to corresponding pixel in the previous frame. If the difference exceeds threshold, it's "moving" and should be deinterlaced.

2. Interfield comparison 1. Compare the pixel to the "corresponding" pixel in the other field. If the difference exceeds threshold, it's "moving" and should be deinterlaced.

3. Interfield comparison 2. Actually look for combing by making a detector that looks for spatial frequency of 2 pixels, e.g., Gunnar Thalin's approach that he calls area-based, and which I adapted from him and called field differencing.

Please explain your idea more fully and clearly. Thank you.

[Zarxrax]

@RadicalEd:
Yea, been here for a while, just dont post often :)

@neuron2:
Hmmm, its probly a dumb idea. I guess it couldn't really do anything more than your "interfield comparison 2".
Anyways, heres original frame and field blended frame. I was thinking that by just making a comparison between the 2 it would be able to tell which areas contained the interlacing.
Interlaced (http://www4.ncsu.edu/~baclontz/interlaced.jpg)
blended (http://www4.ncsu.edu/~baclontz/blended.jpg)

[neuron2]

@Zarxrax

You didn't tell us how you made the blended frame, but I have to guess you just used the VirtualDub internal deinterlace set to blend (or an equivalent thereof). But then differencing that against the original frame is going to produce output everywhere, because the blended frame is basically just the original with a vertical blur. This is going to be useless for detecting combing, as I confirmed by actually differencing the two sample frames using Avisynth Subtract().

[Zarxrax]

Ok, I guess it was just a pointless idea then :)