Interlacing and channel bandwidth

[ppppenguin]

AFAIK there's nothing very scientific about the Kell factor. As oldtvnut says, it's determined subjectively without a great deal of theoretical backup. My feeling FWIW, is that it depends significantly on the vertical scanning aperture. In tube cameras this is gaussian which gives a falling vertical spatial frequency response. Usually corrected, at least partially by vertical aperture correction. Like wise in CRT receivers but without VAK. LCD displays and CCD cameras have a very different vertical aperture, much squarer.

The Kell factor has been used to justify decisions about choosing H bandwidth wrt number of lines. Has any good experimental work been done with modern cameras and displays? In any case all HD systems we use square pixels so if there is still a Kell effect there is a shortfall of vertical resolution.

[old_tv_nut]

Originally Posted by ppppenguin

...In any case all HD systems we use square pixels so if there is still a Kell effect there is a shortfall of vertical resolution.

Not sure what you mean here. "Square pixels" is a lousy term, as it does not refer to the display elements having a certain shape. This phrase is used to mean that the centers of the pixels have the same spacing vertically and horizontally. Pixels are point values of color and do not have shape or dimensions.

If you mean the display elements have square shapes, then, yes, this implies a certain vertical and horizontal spatial frequency response, different from that with a Gaussian CRT spot.

[ppppenguin]

Originally Posted by old_tv_nut

Not sure what you mean here. "Square pixels" is a lousy term.....

This is a commonplace term to indicate that the pixel spacing is equal on H and V axes. By contrast to SD digital systems where the spacing is not equal.

As oldtvnut correctly points out, in all sampling theory the idealised sample is infinitesimal in length. (Dirac delta function if anyone is that interested). In TV this is generalised to 2 dimensions rather than one. Practical pixels have a finite size and shape. For LCD displays and CCD cameras this ideally approaches a square having the same dimensions as the pixel spacing. This gives a zero order hold function and hence a loss of HF response on both axes which follows a sin(x)/x curve.

The point I am trying to make is that the assumptions which underpin Kell Factor stem from the days when H scanning was a continuous function while vertical scan was sampled. These assumptions may well not apply when the picture is inherently sampled at the sensor on both axes. As a thought experiment consider a sensor and/or display where each pixel can be individually addressed. They can then be read or written in an arbitrary sequence*. I can conceive that this might affect motion protrayal (motion above a very slow rate is aliased in TV systems) but I cannot see how it might affect our perception of H and V resolution. Hence the Kell factor of a progressively scanned system using modern techniques should be unity.

I may have overlooked something here. For example unless there is some kind of optical filter before the sensor there can be H and V aliasing. Or there may be performance problems of the sensor that affect the axes differently.

*In doing this thought experiment I was influenced by BBC Research Report 1991/4 "Image Scanning using a Fractal Curve" by John Drewery. http://www.bbc.co.uk/rd/publications..._1991_04.shtml John Drewery had a superb understanding of scanning, sampling and spectra. Back in about 1975 I remember him demonstrating the 3 dimensional spectrum of TV signals (PAL in this case) using some wonderful models that he had the BBC Research Dept workshop make from pieces of coloured PTFE. Nowadays this would have been done by computer graphics.

[old_tv_nut]

Originally Posted by ppppenguin

...Hence the Kell factor of a progressively scanned system using modern techniques should be unity.

I may have overlooked something here. For example unless there is some kind of optical filter before the sensor there can be H and V aliasing. ...


*In doing this thought experiment I was influenced by BBC Research Report 1991/4 "Image Scanning using a Fractal Curve" by John Drewery. http://www.bbc.co.uk/rd/publications..._1991_04.shtml John Drewery had a superb understanding of scanning, sampling and spectra. Back in about 1975 I remember him demonstrating the 3 dimensional spectrum of TV signals (PAL in this case) using some wonderful models that he had the BBC Research Dept workshop make from pieces of coloured PTFE. Nowadays this would have been done by computer graphics.

The Kell factor relates to how close to unity you can come, even with progressive scan. You can't get unity because the phase of sampling is restricted to the positions of the sampling points. That is, if the details lie exactly on the sampling points, you get full amplitude, but if they lie halfway between sampling pints, you get zero amplitude. The sampling becomes equivalent to a synchronous detector. So, the Kell factor says how close you can come to unity and still perceive a repetitive pattern correctly given that the filtering in the system consists of the sensor spot or element shape, the display spot or element shape, and the human eye optical function. Interlace makes it worse, but it's not unity for progressive.

For a full explanation of 3-dimensional spectra resulting from scanning, I also recommend an out-of-print book by Pearson:
http://www.amazon.com/Transmission-D...n+transmission

[ppppenguin]

Originally Posted by old_tv_nut

You can't get unity because the phase of sampling is restricted to the positions of the sampling points. You can't get unity because the phase of sampling is restricted to the positions of the sampling points. That is, if the details lie exactly on the sampling points, you get full amplitude, but if they lie halfway between sampling pints, you get zero amplitude.

This would apply equally to both X and Y axes on a sampled system, so resolution is degraded euqally on both axes. It is also a misunderstanding of sampling theory. This works equally well in space as well as time. If you sample a signal using infinitesimal size samples at more than the Nyquist limit the original signal can be reconstructed exactly. This is also applicable to 2 dimensional sampling, as noted by Mertz and Gray in their famous 1934 paper. If the sampling aperture is finite you get a falling frequency response, the exact response depending on the shape of the sample aperture. For example a square aperture competely filling the sample pitch would give a sin(x)/x response with a little over 2dB drop as you approach the Nyquist limit. In the real world it is difficult to put a sharp cutoff optical filter ahead of the sensor so there will be aliasing. This suggests using a sensor with more pixels than needed for your TV system and filtering the output. This is equivalent to using oversampling ADCs. It is also one reason why some of the best SD pictures are obtained by downsampling an HD input.

[NewVista]

Originally Posted by ppppenguin

. This suggests using a sensor with more pixels than needed for your TV system and filtering the output..

This should be a no-brainer yet only movie people use 4k cameras!
Always wondered why HDTV cameras 2k? Naive?

[ppppenguin]

Originally Posted by NewVista

This should be a no-brainer yet only movie people use 4k cameras!
Always wondered why HDTV cameras 2k? Naive?

Depends on the tradeoffs when you're making the sensors. In the early days of CCDs they were struggling to make them work at all, you got as many pixels as you could reasonably make work at all. There are also 2 tradeoffs that may well be fundamental. For a given image size on the chip, if you have more pixels each one is smaller and hence collects less light. Also the fill factor, the fraction of the chip's surface that's sensitive to light, goes down. While a theoretically ideal sampler has infinitesimally small pixels it would also have infinitesimal sensitivity. Hence the sensor designer strives to fill as much of the space as possible with pixels and leave minimum space between them.

Image size is important. For cine camera replacement you want to be able to use your existing stock of 35mm prime lenses. Hence the sensor size needs to replicate 35mm film area. For TV the sensors are smaller.

I haven't looked at what size sensors Super Hi-Vision uses but the fundamental resolution is about 8k x 4k. I saw a demonstration a few days ago at BBC Broadcasting House, some recordings from the Olympics. NHK and BBC have worked together to televise parts of the olympics on this new system. Only 3 cameras so a refreshing return to old fashioned production values, lots of lingering wide shots, minimal pans or zooms. You don't need closeups when you have that much resolution available. From my seat, about 30 feet from a 25 foot screen the pictures were perfectly detailed and flawless, even under difficult lighting conditions such as fireworks.

The pictures were also being relayed to Bradford, Glasgow, Washington DC, Tokyo and Fukushima so some of you may have had a chance to see them.

[Penthode]

Some good points here. I recall John Watkinson wrote a paper in 1998 on Video Oversampling. He stated that because of the optical filtering ahead of the sensor, it is not necessary to use so many lines to deliver HD. If on the otherhand, the number of pixels on the sensor is substantially higher followed by the optical low pass filter, rescaling to fewer lines will not result in loss of spatial resolution. The only caveat is that oversampling only really works with non-interlaced video.

I believe we now underestimate the resolution of Image Orthicon video since resolution was limited by the structure of the target element and not by a digital imager's pixel array. Hence higher number of pixel imagers, rescaling and progressive scan is the future.

Nevertheless, I would have liked to have seen what a 4" IO tube could yield in terms of spatial resolution.