Log in

View Full Version : a psycho idea


trbarry
11th March 2002, 17:23
As a psycho-visual enhancement it seems you could give slightly more resolution towards the center of the screen where the most interesting things are happening. Do any codecs or filters do this?

It seems it would be easy to write an Avisynth (or whatever) filter that could save some bits by filtering around the edges.

- Tom

Guest
13th March 2002, 04:35
Interesting idea, Tom!

movmasty
13th March 2002, 07:53
from a my old post (http://forum.doom9.org/showthread.php?s=&threadid=15777&highlight=center)
would be nice also to apply a filter only to the borders of screen,not just at the center like unsharp(kinda of reverse region option?)
as would be good to sharph the center of a movie, leave all around it untouched, and smooth just the borders.
or apply a filter only over,or below,any luminance threshold,also using a pic like mask,like Ulead media studio can do
actually you could smooth all the screen, and then sharp just the center with unsharph mask.

in general to apply a filter only in some region is good.

b0b0b0b
14th March 2002, 07:14
A temporal filter whose threshold varied according the the x,y would be one way of doing something like this, if temporal filters are good at saving bits.


If lower resolution means saving bits, then before compressing you could apply a warp-resize to the image that shrinks the overall dimensions but retains more information in the visual hotspot (like a fisheye but not necessarily spherical could be cylindrical), and then after decompression, an inverse-warp-resize is applied in the zoomplayer filter chain.

trbarry
14th March 2002, 13:12
I thought of using the WarpedResize warp formula, but just using it as a parm to a softening filter. But I don't know which softening filter would have the greatest effect on the eventual bit rate compared to the effect on the eye.

- Tom

movmasty
15th March 2002, 05:50
the two aspects of noise removal and size reducing are not really the same thing, and have to be treated separately,
eg, the SpatialSoften of avisynth is impareggiable in removing noise w/o blurring, but doesnt reduces size at all, cause it smooths and sharps at same time.

the internal smoother of vdub causes heavy arifacts,and could be used only at low threshold,say 3/4,thus limiting its effects.

the smarth smoother of D.Graft "cartoonize" the video at high settings,
so you have to use at light setts again, and it does some sharpening too.

the "smarth smoother hq" by klaus post is quite good but effective only at high radius,where it is VERY slow, slower than applying 2/3 different smooth filters.

then you have heavy blurring filters,3x3average really removes noise,bring size down,and is fast,it does better than blur imo,
but also blurs the screen noticeably.
maybe would be good for borders, or you could apply a sharp filter after it
the blur of avisynth is like that of vdub but settable,on my system its incompatible with the use of others filter.

to smooth and reduce size i use often the p bilinear resize,2 times,
magnify per 2 and then reduce per 2,
good to denoise and reduce size,its faster than smarth smoother hq,
but its a resize filter, and cant be applyed to only a region.

trbarry
15th March 2002, 13:55
Just out of curiousity, are there any filters that do an 8x8 block DCT, lower the high frequency components, and then IDCT again?

It seems that might be good for compression. If one of these existed maybe it could be modified to reduce the high frequency components more around the edges.

- Tom

b0b0b0b
15th March 2002, 18:14
That sounds like a filter that could, in a controlled fashion, reduce the number of bits it would take to encode a picture. The threshold for clipping the high frequency components might have a well-defined relationship to the savings in size. This would be a lot better than the handwaving that currently goes on about applying lots of softeners or doing black magic with resizing and re-resizing.

I guess the flip side is a filter that would make the motion vectors of a picture simpler to encode. Existing temporal filters don't do this well.