This article discusses video comb filters, what they are, how they work and why not all comb filters are created equally. It was first published in EDN magazine in July 2012.
Revisiting the Analogue Video Decoder
Brushing Up on Your Comb Filters
The basic function of the video decoder is to accept analogue video from a wide variety of sources such as broadcast, DVD players, cameras and video cassette recorders, in either NTSC or PAL format and still occasionally SECAM, separate it into its component parts, luminance and chrominance, and output it in some digital video standard, usually BT656, a multiplexed Y/Cb/Cr video format with embedded timing signals running at a clock rate of 27MHz.
Most major semiconductor companies, for example Texas Instruments, NXP (formerly Philips Semiconductors) and Analog Devices, and a large number of smaller companies manufacture an analogue video decoder. In addition many companies have integrated this function into SoC (System on Chip) integrated circuits.
With such a large number of video decoders in the market it might seem an unnecessary indulgence to spend time looking again at the design of this fundamental but apparently obsolete building block and certainly new designs are not appearing on the market and haven’t done so for 3-4 years.
However the block is still a requirement of most video designs, from televisions to video recorders because large swathes of the planet are yet to convert to digital only broadcasting and because of the legacy requirements. The block is however being marginalized with most of the design effort spent on newer features such as higher definition and 3D.
The analogue television broadcast system was designed as a complete system, from the psycho-visual compromises made at the encoding end through to the persistence of the phosphor on the cathode ray tube (CRT). The advent of large flat screen displays, whilst in many ways being the element that allowed the widespread adoption of high definition television broadcasting, placed additional demands on the analogue video decoder where conventional sources were viewed, and for a large number of viewers that was still the case; that yellow RCA plug was still in widespread use.
The most obvious result of viewing analogue video sources on a large display is that any artifacts are, of course, larger and visually more apparent. For larger displays the analogue video decoder actually has a more stringent requirement. This problem is compounded because the flat screen displays require additional processing of the analogue source before it can be properly displayed, namely de-interlacing and scaling. The de-interlacer in particular can amplify any artifacts left from the video decoder. This is because the de-interlacer is sensitive to motion in the image and residual artifacts and noise left from the analogue decoder cannot be discriminated from real motion in the image. The result is the de-interlacer may make the wrong mode decision resulting in additional artifacts.
The majority of the video decoders use a line comb decoder to separate the luminance from the chrominance component because there is no need of a large frame size memory for such an architecture. The separation relies of the repetitive line to line phase relationship of the colour subcarrier used to encode the chrominance component. For example, for NTSC, each line in the field has a 180° phase shift so adding or subtracting the video signal results in cancellation or reinforcement of the chrominance. That statement is only true, of course, if the colours are of exactly the same hue and even if they are we are not acting upon spatially aligned pixels because not only are we looking at pixels a line apart, but two lines apart because there is the interlaced field line that we do not have access to. Effectively we are looking at pixels 2 lines apart and the situation is worse for PAL because of the additional 90° line based subcarrier phase shift: without specialized and complex processing it is necessary to have a 4 line spacing.
It is possible to detect most of the conditions under which a line comb filter based video decoder will fail, and under these circumstances a simple notch filter is usually reverted to, but Figure 1 shows the result of not detecting this condition.
Figure 1 'Dot Crawl' : The running line of dots highlighted in the left image is caused by poor comb adapatation, failure to detect the line comb failure condition. The right image shows a decoder without these artifacts.
Although the magenta/cyan boundary seems obvious to the eye, a lot of comb failure detection uses luminance values only to determine the comb mode, and in this case whilst the hues are very different the luminance values are similar so the comb failure is not detected.
As mentioned, one of the issues with these artifacts is that they are not static and cannot be removed easily with further post processing such as noise reduction; because the artifacts are moving they interfere with the motion adaptation of the television de-interlacer.
A similar issue arises if the output of the video decoder is to be compressed as all MPEG compression methods effectively send only the motion of an image. Unable to discriminate between artifacts, video source noise and ‘real’ image motion it can be shown that up 20% of satellite and cable digital broadcast bandwidth is utilized to send unnecessary information. This is extremely useful bandwidth that is especially useful given the high compression ratios used by today’s broadcasters and is the difference between the viewer seeing the highly visible MPEG artifacts such as blocking, or not.
One large improvement to the video decoder that has been made by some manufacturers is to add a 3D comb filter. The filter utilizes the same phase relationship in the subcarrier but a pixel accurate frame delay ensures the pixels are exactly spatially aligned. Even on the most complex images near perfect, artifact free decoding is the result: (I say near perfect because frame combs are very sensitive to clock jitter and clock jitter as little as 1ns over the frame delay period can result is residual subcarrier. PAL also has an addition subcarrier offset which means perfect cancellation cannot occur even with a frame comb).
Figure 2 shows the artifact free image achievable with a well designed frame comb.
Figure 2 The left image shows a zone plate with a line comb filter. The colours in the image are caused by 'failures' in the line comb. The right image shows the 'clean' image from a frame comb filter.
The zone plate image in Figure 2 may seem rather esoteric, but you may be more familiar with the shimmering colours created by a newsreader’s check shirt, something they seem to have a peculiar penchant for wearing.
But of course, the frame comb does not answer all our problems. We are now looking across comb filter taps of one frame (for NTSC) and 2 frames for PAL, a delay of 80ms in the latter case. Whereas any difference spatially across the 2 or 4 lines caused the line comb filter to fail, now any difference temporally across 33ms or 80ms will cause the frame comb to fail, with similar artifacts being created. The only recourse open to most video decoders is to choose the line comb, or if that is also failing to go to the notch filter mode, a clean but low bandwidth fall back.
For PAL in particular it can be shown that for a large amount of video material the frame comb cannot be chosen because of image motion; it becomes an expensive luxury good only for those static demonstration zone plate images. However we have yet one more, under-utilised comb option and that if the field comb. The field comb has an aperture of 262 lines for NTSC and 312 lines for PAL and is therefore much closer spatially than the line comb and closer temporally than the frame comb (especially for PAL).
Figure 3 shows the image of a woman wearing one of those fine check blouses. The parts of the image coloured green are where the frame comb can operate, the parts of the image coloured magenta are where the field comb is operating and the parts of the image shown red are where the line comb or notch filter is operating. As expected where there is some motion the frame comb fails and with decoders without a field comb, only the line comb or notch filter is the available option with the resulting loss of resolution. The field comb offers a relatively artifact free, yet full bandwidth image even in those areas that have small amounts of motion.
Figure 3 Comb Modes: Green = frame comb, Magenta = field comb, Red = line comb/notch filter.
To avoid a complex memory architecture it is also possible to forgo the usual symmetrical architecture employed in 3D comb filters where the frame before and after is available to the filter. An asymmetric comb filter where only the past frame (or field) tap is available can yield similar results, simplify the memory interface and allow the design of a video decoder with minimal delay. Another advantage of this architecture is that no corresponding audio delay need be employed to prevent lip-sync issues.
Another market segment that is putting additional demands on the traditional analogue video decoder is the mobile reception market.
The mobile market will utilize a silicon tuner and digital video demodulator. For the video decoder this means that it has to accept digital composite video at a fixed sample rate and conceivably with non-zero IF offset. Most analogue decoders employ a line locked architecture where the clock is multiple of the horizontal line frequency. Conventionally the analogue to digital converter (ADC) samples the analogue video signal at this frequency but this is not possible if the input is already digitized at a fixed clock frequency, (unless we convert that signal back to analogue and re-sample it). However it is possible to re-architect the video decoder to adapt to this new input, and approach that also yields other benefits.
The chrominance demodulation can be modified to run at the new fixed sample rate. The output of the demodulator will then be low-pass filtered chroma and also a notched luma signal (the composite signal has the remodulated chroma subtracted from it to yield a notched and subcarrier free output). This fixed sample rate signal can then be sample rate converted to a line lock clock domain for the comb filter and to comply with the line locked clock requirement of the BT656 standard.
Whilst this approach solves the problem of fixed sample rate inputs it also yields other benefits. With conventional architectures, noise or errors in the horizontal frequency (such as from mechanically scanned sources such as VCRs) give rise to errors in the line locked clock which, in turn, give rise to errors or jitter in the chroma. Using a fixed sample rate for the chroma demodulation avoids this issue.
A fixed sample rate for the chroma demodulation also makes the design of the SECAM demodulator, if required, simpler.
Such an architecture also yields advantages in the comb filter. The output of the comb filter (when not failing) is now clean (combed) chroma and high frequency luminance, effectively what was removed in the simple notch filter. Instead of relying on the comb filter to remove the chroma this architecture uses the comb filter to replace the high frequency luma. The difference is subtle. In the latter case, should be fail to detect the comb failure, we will not remove the chroma from the composite signal leading to very visible artifacts. In the new architecture we never have a composite signal, we always have a ‘clean’ luma signal and the comb only adds back high frequency luma. The effect is that, should the comb failure circuit fail to detect an error signal the output artifacts are much reduced.
The fixed sample rate chroma demodulation also allows us to detect non-zero Ifs from the video demodulator, using the subcarrier frequency as the reference, the one absolute in all analogue video sources, (typically most chroma phase looked loops with only have a lock range of 100-200ppm).
The architectural changes described above allow a much improved image quality suitable for MPEG compression and for large screen displays, and also interface to front ends of mobile reception devices, yet impact very little on silicon area or cost. Rumours of the death of analogue video sources were premature.