Building a Standards Converter

[ppppenguin]

The difference between 2 and 3 line interp isn't big. One day I must do some coefficient sets for 3 and 4 lines and see the difference. My own converter has a 4 line interp but I've only ever used it at 2 lines. According to BBC work, there was very little difference between 3 and4 lines but significant difference between 2 and 3. While you can definitely see that on test card you'd be hard pressed to see it on pictures.any set that can resolve the scan lines will show the interp quality. You don't need good horizontal resolution.

8 bit coefficients and 5 bit sub line position are amply good enough.

[`FRANK.C]

any set that can resolve the scan lines will show the interp quality. You don't need good horizontal resolution.

That's interesting and good to know. I had assumed the horizontal resolution would make a big difference.

Frank

[ppppenguin]

Look at the circle on the test card. Shows interp errors nicely. Also diagonal lines especially at shallow angles.

[`FRANK.C]

I have a DVD of This advert and was using it on repeat to view how well the interpolater was doing.  As the TV's moved around the screen it was easy to see the ripples created by poor interpolation, move along the horizontal(ish) edges of the TV's.  As the interpolater got better the ripples got less and less, but then I noticed that the ripples wasn't leaving some TV's. I couldn't understand why this was happening, that was until I watched it on 625.
The third TV in is particularly bad.  The DVD is of better quality but the ripples are as prominent.

Frank

[ppppenguin]

Yes, you certainly get ripples on 625

[`FRANK.C]

Its easy to understand how a 2 line interpolater works and in a 3 line interpolater it is easy to imagine how the output line can straddle 3 input lines as the output line is wider than a input line. From a 4 line interpolater up it gets difficult to understand how the extra input lines have a positive effect on the output line.  I thought it would be interesting to compare interpolaters  with different numbers of input lines.

In a BBC Research  Department document titled "DIGITAL LINE STORE CONVERSION: determination of the interpolation aperture function"  there is a graph showing what was considered the optimum aperture for 2,3,4,6 and 8 lines. I took measurements from the graph and made coefficients for the interpolater as accurately as I could.  

I have rewrote the interpolater so can now be switched between 2,3,4,6 and 8 lines.

What I have observed so far is
On Test card C I cannot not see any difference between 2 and 3 lines.

There is what looks like an improvement between 2 and 4 lines, looking at the diagonal lines on TCC they get sharper with the whites getting whiter, similar to the contrast being turned up.

There is little difference between 4,6 and 8 lines.
6 and 8 lines display slight ghosting above and below certain areas. a caption is pictured below displaying it. I don't think it would be very visible on a moving picture.

It is easier to compare the 4,6 and 8 lines to 2 lines as the majority of the information is coming from the same two input lines with the additional lines just adding to that.
With the 3 line the majority of the information is coming from different lines to the 2 line. This has an effect that is very noticeable when switching between 2 and 3 lines if viewing captions, as the letters change in height, they either get taller or shorter depending on what lines they are on.
A photo below shows the same section of a caption  viewed on 2 and 3 lines. The thickness of the horizontal bar of the "T" varies quite a bit between the two.

Frank
 

[ppppenguin]

The optimal aperture shape for more than 2 lines is a matter of debate. It's always a compromise, if only because there may be vertical aliasing on the input picture. A sharp transition from white to black going down the screen is aliased. Interlace makes things worse as you can't readily interpolate between spatially adjacent lines. If you do, then moving pictures are truly horrible. The Pineapple converter had an option to do this (easy with a framestore) and the results were great on still images, otherwise useless.

The interpolation aperture is theoretically infinitely wide and truncated to 2, 3 or more lines in practice. Truncating an aperture gives spurious effects which is why windowing is used. Hamming, Hanning and other windows are common. Look it up and prepare to be confused. However this doesn't help much here, the visual impression is what matters so there is no theoretically optimal aperture. Make the aperture too sharp, with excessive negative coefficients, and the effect will be unpleasantly edgy and nasty. Make it too soft, perhaps with all positive coefficients, and vertical resolution will suffer.

The BBC research reports (cited in one my articles, see my website, and all downloadable from the BBC R&D website) show their conclusions that led to the CO6/509 converter with its 4 line aperture. They reckoned that 3 lines was a perceptible improvement on 2 while 4 was just about visible over 3 on the most critical material. I think some later work suggested going up to 6, 9 or even more lines but I don't think this gave any improvement for ordiary standards conversion.

Don't forget that as you add more lines to the aperture the picture will move down the screen. In theory the lines before the centre of the aperture are in negative time and the centre stays still. In practice the centre of the aperture moves downwards. Just a matter of tweaking the output vertical position slightly.

[`FRANK.C]

Hi Jeffrey

The picture has moved down quite a bit. I haven't adjusted the output yet.
Regardless of how many lines are been interpolated all 8 lines are in circuit with '0' value coefficients being given to the unused lines. This way so long as there is an even number of lines interpolated the picture remains in the same vertical position.

you can't readily interpolate between spatially adjacent lines. If you do, then moving pictures are truly horrible. The Pineapple converter had an option to do this (easy with a framestore) and the results were great on still images, otherwise useless.

That gives rise to an interesting question,  what if, a combination of field and line interpolation was used?
The decision to use field or line being made on a pixel by pixel basis.
To decide if a pixel was to be formed from field or line interpolation all the pixels in the current frame needed to do the interpolation for that pixel, would be compared to the same pixels in the the previous frame.
If they were all the same do field interpolation but if any had changed do line interpolation.
It would require quit a bit more memory, if using FIFO type memory it would probably need 3 of them for the field stores.


Frank

[ppppenguin]

Yes of course you can use a mixture of lines from different fields. This is exactly what you do when changing field frequency. The IBA's DICE used a 2 field 5 line aperture. The BBC's ACE used 4 field and 4 lines. I used something very close to the ACE aperture on a design I did for a client some years ago. With a fixed aperture there is always a compromise between different artefacts. Some of the ACE research was done at BBC Research Labs while I was in my gap year there in 1974/5. I knew what they were doing was clever but I didn't understand it. Probably in a research report or 2 somewhere.

Later converters used adaptive apertures which also have their limitations as you can often see the aperture being switched as the speed of motion changes. Later still there was motion vector compensation, a technique now widely used in compressed video bitsteams. Even this isn't 100% perfect as in an interlaced system there is always some ambiguity between temporal and vertical information.

Unless you are really determined, I would strongly recommend you not to experiment with adaptive apertures. I've tried it in a small way with horrible results. The cure was worse than the disease. Don't even think of trying motion vectors yourself. I wouldn't touch them unless there's a some freely available intellectual property that shows you how. Really need to do the work with computer based simulation with digital files for input and output.

[`FRANK.C]

Hi Jeffrey
Thanks for the explanation,  you have done your bit for the environment today and stopped the pile of scrap in the corner of my workshop getting bigger
It makes sense now that changing the aperture during a field would produce unwanted effects on the picture.

Frank