|
|
| valja |
| Posted: Jun 2 2003, 07:08 PM |
 |
|
Advanced Member
Group: Members
Posts: 179
Member No.: 66
Joined: 1-August 02
|
I'm trying to improve a terrible rip. It has quite high spatial resolution 696x408, but looks like there was used temporal "blending" of two consecutive frames (probably to decrease noise). Scene changes show this clearly - there is always one frame that looks like 50:50 mix of previous and next scene. On static scenes the picture is quite clear, but all moving parts look terrible.
Ahybody have heard about some filter that can "restore" unblended frames? Theoretically it should be possible to "remove" such blending (as example by substacting previous frame from the next) , but practically it looks like "uncorrect" task - precision of luminance and chrominance of all pixels is limited - hard to say what will happen after even first hundreds of such substractions.
|
|
| Morsa |
| Posted: Jun 3 2003, 04:51 AM |
|
|
Moderator of the Vdub support board
Group: Moderators
Posts: 640
Member No.: 246
Joined: 9-September 02
|
Sorry, but I don't think it is posible up today.Think about deconvolution, which in fact is a lot easier than what you need to do.I guess it will require a lot of advanced mathematics and also guess nobody wants to do such a programming.
An Idea: what will happen if you work with a range higher than 256 and perform what you said?
At least after correcting the results the blended frames between scenes should be restored. |
|
| phaeron |
| Posted: Jun 3 2003, 05:47 AM |
|
|
Virtualdub Developer
Group: Administrator
Posts: 7773
Member No.: 61
Joined: 30-July 02
|
It is approximately doable, but you won't like the solution.
A two-tap box filter (average frame pairs) annihilates the highest frequency (alternation), so perfect reconstruction is impossible. Second, a box filter is a pure blurring function. That means any inverse is going to have to be a pure amplification function (minus the unrecoverable frequencies, of course), and thus going to correspondingly amp noise at the same levels as well.
|
|
| valja |
| Posted: Jun 7 2003, 06:52 PM |
|
|
Advanced Member
Group: Members
Posts: 179
Member No.: 66
Joined: 1-August 02
|
| QUOTE (Morsa @ Jun 3 2003, 07:51 AM) | An Idea: what will happen if you work with a range higher than 256 and perform what you said?
At least after correcting the results the blended frames between scenes should be restored. |
Don't think it would help a much. Even if I temporarily convert luma/chroma to say 512 or 1024 levels and convert back after calculations, the luma/chroma is still quantised. And the worst thing, in half of cases the blended pixel is already rounded to the nearest integer - when we blend say 13 and 14, after blending we will have 13 or 14 (depending of rules), but not 13.5. So if we have in the "blended" frame some level 13, we don't know, was it result of blending 13 and 13 or 13 and 14, so we can restore "unblended" pixel only with precision plus-minus one level. And such errors will probably accumulate during long chain of calculations.
@phaeron
Thanx for explanation. That was clear from beginning, that pefect reconstruction is impossible, I had hope only about some approximation. And increasing of the noise isn't fatal problem - it can be decreased using "temporal smoother" I often use. Algorithm it uses is much better than simple blending of frames.
|
|
