FreeLists / opendtv / [opendtv] Re: Sampling Frequency of Image Signals

[Jeroen Stessen <jeroen.stessen@xxxxxxxxxxx>]

Hello, 

Dan Grimes wrote:
> "It's also possible to use non-linear quantization to advantage."
> This refers to floating point, no?  Or are there new forms of
> non-linear quantization being used?

Like I once said to Prin: "at the end of the day any non-linear 
transfer function, like a gamma function, is just a quantisation 
table". We were talking about Digital Cinema, where they convert 
the linear-light signal to the gamma domain with a pure power law 
with an exponent of 1/2.6, and then encode to 12-bits (X',Y',Z'). 
This function makes the quantization steps approximately equally 
visible (or rather: invisible). But under which conditions of eye 
adaptation ? It also allocates a lot of codes to intensities that 
may never occur in the particular scene, or to parts of the image 
that the eye can not adapt to (like a dark window in a bright wall). 
I think that you would get better results with an adaptive scale 
of lower precision, i.e. a compression-type encoding. 


Bert Manfredi: 
> Jeroen said that with linear coding, our eyes want at least 14 bits of
> resolution. If the bits are weighted differently, we can do fine with 8
> bits.

Actually, my 14-bits example was the linear-light equivalent of a 
10-bits gamma-domain signal. In order to get away with only 8-bits 
we would need to have a sufficiently high level of analog noise... 
It makes no sense to quantize a noisy signal any more accurate than 
the level of its noise. As the signals get cleaner, and our eyes 
get more chance to adapt temporally and spatially to the scene 
(larger viewing angles !), then we will need more bits. 


About the original subject, sampling frequency: anti-aliasing is 
needed in order to be able to represent a signal at arbitrary 
positions (i.e. in any possible phase). That alone already gives 
you the infinite variety of positioning of edges. If not, then 
you'll see jitter and moving jaggies on (slow) moving edges, and 
that is precisely aliasing. I see a lot of that every day ! 

A decent anti-aliasing filter (or anti-imaging, for that matter) 
with a mild roll-off needs some space for its transition band. 
Typically this lies between 1/3 and 1/2 of the sample frequency, 
therefore the usable bandwidth is only 1/3 and not 1/2. Nyquist 
was an optimist. In audio steeper filters can be used, because 
we do not perceive ringing at 20 kHz very well and also the energy 
that would kick the filters into such ringing is mostly absent. 
So for audio it is not too hard to approach 45% of the sample rate 
(20 kHz / 44 kHz). Not so in video: any edge will be surrounded by 
ringing. And low-ringing filters are necessarily not steep. 

I find it ridiculous that some people put on-off-on-off patterns 
(i.e. the Nyquist frequency) on test disks, and then other people 
expect that this frequency should pass unattenuated through the 
entire system. This frequency can be represented only in two 
phases (i.e. 0 and 180 degrees), therefore it can not move... 
It is much safer to suppress it entirely than to pass it ! 

Best, 
-- Jeroen

+------------------------------------------+------------------------------------+
| From:    Jeroen H. Stessen               | Phone:  ++31.40.27.40246  |
| Deptmt.: Philips Applied Technologies    | Mobex:  ++31.40.27.99650  |
|          Digital Systems & Technologies  | Mobile: ++31.6.4468.0021  |
| Address: High Tech Campus 5 - room 5.025 | Skype:  Jeroen.Stessen      |
|          5656 AE Eindhoven  - Nederland  | VoIP: 
Jeroen.Stessen.at.Philips  |
| Website: http://www.apptech.philips.com/ | E-mail: 
Jeroen.Stessen@xxxxxxxxxxx |
+------------------------------------------+------------------------------------+