Forum

Forum



SearchSearch   Users   Registration   Entrance
Today: 30.11.2025 - 12:41:04
Pages:  1  

Rev398 explanation

MessageAuthor

Changeset [398] by pengvado dct-domain noise reduction (ported from lavc) This means that I can remove denoiser filter from the avs script?




ryan_george

user




Statistics:
Messages: 32
Registration: 11.23.2003
20.02.21 - 11:40:24
Message # 1
RE: Rev398 explanation

Unfortunately, you have to be a member (or Google) to read it, and I'm not. I'd really like to check it out, but I suppose a summary from Pengvado would work.




Reed

user




Statistics:
Messages: 444
Registration: 05.01.2002
20.02.21 - 11:49:08
Message # 2
RE: Rev398 explanation

So this is an internal progression, not a seperate tag that can be enabled/disabled, right? According to the synopsis you provided, foxyshadis, it seems like this will be of great benefit to AVC encoding, but it didn't do anything for me at q35 (but by then there might be no need for filtering anyway--all significant, noisey detail is gone ;)).


-----------------------
5000k HID Startup Clip-Minnesota Sunset

///Samson

user




Statistics:
Messages: 229
Registration: 05.29.2002
20.02.21 - 11:55:36
Message # 3
RE: Rev398 explanation

Summary for those without an ieee account (actually, I didn't read that paper until just now, but it's similar enough): 1) Estimate the noise-to-signal ratio of each frequency band: We assume a constant amount of noise (specified by --nr), and thus the ratio is just inversely proportional to the amount of signal (= average coefficient magnitude). 2) Filter: Before quantization, bias each coef of the residual towards 0, with strength determined above. Try --nr values on the order of 100-1000. It may or may not match the quality of a good avisynth-based denoiser, but is much faster.


-----------------------
Previous 10 cars owned 1993 325is ITS/JP race car02 S20001995 M3 track/street car90 Civic Si beater1986 Porsche 944 ES SCCA Solo II car2001 Boxster S2000 Boxster AS Solo II car1993 Honda Civic EX beater1993 Porsche 968 AS SCCA Solo II car1996 318ti s

singletrack1

user




Statistics:
Messages: 580
Registration: 04.02.2002
20.02.21 - 12:01:59
Message # 4
RE: Rev398 explanation

akupenguin, so where did you find the algorithm or info about that dct?


-----------------------
91 Integra LS - For Now -LS/VTEC, I/H/E, 93 GS-R trans, Tokico Blues, Spoon SW388's until someone stole them 2000 Cosmos M Coupe - Someday

entityM

user




Statistics:
Messages: 146
Registration: 08.29.2002
20.02.21 - 12:12:53
Message # 5
RE: Rev398 explanation

@Episodio1: in libavcodec




Cru Jones

user




Statistics:
Messages: 248
Registration: 01.16.2003
20.02.21 - 12:20:11
Message # 6
RE: Rev398 explanation

Noise filter is superb at removing noise from analog tv caps 




dcllee

user




Statistics:
Messages: 273
Registration: 05.19.2003
20.02.21 - 12:25:19
Message # 7
RE: Rev398 explanation

A common problem with some of my old analog tv caps used to be problem of thick and thin lines running across the width of the frame.Problem was only with couple of channels.If I complained to my cable guy, the problem was shifted to some other channel. In the worst cases, could never find filters which would clean without noticeable blurring. I have been trying some of those bad cases with noise filter set at 50 (100 caused some blurring to my eye,though noise removal was almost total) along with "removedirt().turnleft().removedirt().turnright()" (which was removing thin lines without blurring). The videos are now eminently suitable for watching. The quality of the filter almost matches the neatvideo filter for virtualdub but is very much faster.:thanks:


-----------------------
93 325i Sedan - S50B32 powered

3MTA3

user




Statistics:
Messages: 78
Registration: 09.13.2002
20.02.21 - 12:35:04
Message # 8
RE: Rev398 explanation
Pages:  1  

Message
Name
Message

Click on the dot next to the name to address the participant

           
   
The administrator has prohibited guests from replying to messages! To register, follow the link: register


Participants

inside: IP.Board © 2025 5.1.1