[opendtv] Re: Math of oversampling - a simple comparison

At 7:36 AM -0400 4/28/05, Tom Barry wrote:
>Your filter really is quite impressive.  But I tend to visualize
>the sudden loss of detail just by visualizing a possible MTF
>curve.  If there was only mild detail in the original above 544
>then filtering that amount would not be noticeable.  But possibly
>when you filter at 480 then the cut off point starts climbing the
>steeper shoulder of the MTF curve, suppressing frequencies that
>really exist in the image to some degree.  And the actual
>perceived sharpness is the square of this ratio, so it is more
>like (480/544)^2 = .778.

Tom is on the right track here. The difference between 480P and 544P 
is indeed visible to the average viewer. This is my take on why, and 
why this is very important.

As I suggested (hopefully correctly), the ability to represent high 
contrast details in an image is related to the cut-off frequency of 
the required filters to meet the requirements of the sampling 
theorem. As we increase the density of the sampling grid, we can 
represent higher frequency details with increased contrast, up to the 
point where the samples accurately reproduce the intensity of the 
information "seen" by the camera (or created via a synthetic test 
signal).

What I believe we are dealing with here is the relationship between 
the visual acuity of the viewer and the ability of any sampling grid 
to equal or surpass that level of acuity. There are many metrics at 
work here. The number of vertical lines in an image determines the 
level of vertical detail that can be passed (the use of interlace and 
the required filtering also influences this). Likewise the number of 
samples per line determines the level of horizontal detail that can 
be represented. When the world agreed to standardize on a common 
sampling rate for SDTV - i.e. ITU-R.BT601 - we made the horizontal 
resolution the same for everyone, but we left the rather large 
difference between 480 and 576 lines in place. 576 line systems 
traded off temporal resolution for increased vertical spatial 
resolution - to me these images have ALWAYS appeared to be sharper.

Based on many tests that I have performed over the years, it appears 
that both PAL and NTSC are not capable of saturating the acuity of 
the average viewer at the designed viewing distance. If we eliminate 
interlace and compare high resolution images that are downsampled 
respectively to 480P and 576P there is a very noticeable difference 
in vertical detail. This example is a bit more extreme than the 
difference between 480P and 544P that Jeroen noted, but it relates to 
the same phenomenon.

Somewhere in the continuum from 480P to 544/576P to 720P we begin to 
saturate the human visual system, at least in terms of the highest 
level of detail that the average viewer can perceive. Clearly things 
keep improving along this continuum, but once we get passed the stuff 
that everyone can see easily, the addition of more detail become less 
critical or noticeable. I think Tom is correct about "the square of 
the ratio." As we raise the area under the MTF curve, we reach the 
point where everyone perceives a sharp image - as we keep going each 
incremental step becomes less obvious.

This is one of the reasons that I have long believed that we should 
raise the floor on image quality to 544/576P as a baseline. 480P cuts 
off well within the average acuity of viewers. With 576P we see very 
significant improvements. But we also need to increase the level of 
horizontal detail to match. With 601 the limit of 720 samples per 
line becomes a significant factor. There is a huge difference in 
perceived detail when we increase from 720 x 576 to 1024 x 576. &20 
samples per line was adequate for an interlaced system since the 
vertical detail is so "low," but when we remove the interlace 
footprint we need to bring the horizontal resolution up to match.

If Tom and Jeroen have time to play, I would like to see a down/up 
sample pass through 1024 x 576P. My educated guess is that it will be 
nearly impossible to see the difference from the original.

Regards
Craig

 
 
----------------------------------------------------------------------
You can UNSUBSCRIBE from the OpenDTV list in two ways:

- Using the UNSUBSCRIBE command in your user configuration settings at 
FreeLists.org 

- By sending a message to: opendtv-request@xxxxxxxxxxxxx with the word 
unsubscribe in the subject line.

Other related posts: