Fine Tuning FieldDeinterlace

[Piper]

The last few days, I've made an honest attempt to get to know FieldDeinterlace a little better to help improve the quality of my analogue tv caps (352x480). I'm using the latest decomb with 2.5.1.

Not all frames are interlaced so I chose full=false.

I prefer interpolation rather than blended frames (looks sharper)`so I chose blend=false.

Initially, I found FieldDeinterlace's defaults didn't catch all interlaced frames. Viewing sections of my clip frame by frame using show=true & map=true, I gradually reduced the threshold down to 6.

The same went for dthreshold. I slowly backed it down until I found better field blending. The value I ended up with is 4.

My filter now looks like:
FieldDeinterlace(full=false,threshold=6,dthreshold=4,blend=false)

When viewing the clip back, I saw a distracting "jumpyness" I hadn't noticed before. When theres a mixture of progressive and interlaced frames it really becomes noticable, like the filter is attempting to deinterlace the whole image, not just interlaced areas. Visually, it looks like a sharpening filter is applied only on the interlaced frames so when it's played back in sequence it gets quite distracting.

I then started to tweak my settings more, first by advancing the threshold values back up. I also experimented with chroma=true which also helped. My new filter string becomes:

FieldDeinterlace(full=false,threshold=19,dthreshold=4,blend=false,chroma=true)

Now things look a lot more stable, and I could live with the increased threshold if it weren't for the following. On frames where there's horizontal titling, logos etc flickering "ants" appear which move horizontally along the tops & bottoms of the titling. This effect also appears throughout on the bottom couple of lines of the frame. This effect also seems to affect white titling the most. The ants appear grey & white.

So, if I back off threshold values, the ants go away, but the picture becomes less stable. Playing with dthreshold didn't appear to help either, so I've left it at 4. So far I haven't found a combination where both are acceptable. Perhaps I'm missing something else? Is achieving this balance simply not attainable?

Can anyone help?

[Guest]

It's an unfortunate artifact of the field differencing. It can be reduced somewhat by adding edge avoidance code as in Gunnar Thalin's Area-Based Deinterlacer. But it would reduce performance quite a bit. As time allows I will experiment with adding it to FieldDeinterlace. In the meantime, you might want to try TomsMoComp().

[Piper]

@neuron2 - thanks for the reply.

Unfortunate indeed. I really like how fast FieldDeinterlace is and how much detail it retains. This is why I spent the time to get to know it a little better.

I have been using GreedyHMA with relatively heavy msharpen to achieve similar results to FieldDeinterlace. I have avoided TomsMoComp() (perhaps unfairly) after reading other comments regarding its tendancy to remove too much detail.

[Piper]

Ok, now I feel right foolish. :)

I just gave TomsMoComp() a try. I tried 3 different settings and I'm done. TomsMoComp(1,15,0) looks brilliant. Tom, where do you get off writing filters like this? Amazing stuff, thank you.

[Guest]

Yes, TomsMoComp() is great stuff, which is why I recommended it. :)

Keep things in perspective, though, you're going to run into cases where it hoses up, and only FieldDeinterlace() works right (i've posted clips demonstrating that in the past). You just can't use a hammer for everything!

BTW, Tom usually says anything beyond (1,5,0) doesn't buy you much.

[trbarry]

I think detail gradually increases by a small amount between 5 and 15 but so does the frequency of random 1 pixels artifacts from false motion search matches. That's why I really don't recommend anything over 5 anymore.

- Tom

[Piper]

I must've picked up 15 from an old post. Thanks for pointing that out.