PDA

View Full Version : Noise Reduction


twolfe18
30th May 2007, 22:26
i have been using MeGUI for a while now, and i use the automated avisynth script creator most of the time. i have a question about the order that i should do my noise reduction though.

for reference, the encoding i will be doing will be from hd mpeg2 (captured OTA) to very high quality h264 (1024x576, 2200kb/s).

the automatic script creator always puts the noise reduction after the resize, but i was wondering if the other way around might be a better idea. it seems to me that the noise reducing filter (i forget which one i use), would be more effective (ie, be less destructive to the original video) on the full size hd than the resized video. it also seems like the resize filter might work better on a denoised image as well.

is any of my speculation accurate? should i change the order of my denoising? also, could using the denoising filter twice (once before resize, one after) be a good idea?

:thanks:

Blue_MiSfit
31st May 2007, 04:26
That's a decent idea. Don't use it twice without a very good reason :)

There's lots of great noise reduction filters, a personal favorite of mine is fft3dgpu because if you have a decent video card, the speed hit is essentially non-existant, even with two instances (one for luma and one for chroma), and it can provide both a very powerful and configurable de-noise operation, and even some basic sharpening.

For HD, it's another story. Any filter on HD will be very slow, much more so than on the SD material I typically work with. But yes, definitely de-noise and sharpen if you can / need to before the resize. Scaling should usually be the last step. I would reccomend a neutral scaler like Spline36 as well, Lanczos might introduce some aliasing for big downsizing like 1080i to NTSC.

~MiSfit

twolfe18
1st June 2007, 03:15
well, if i do change the order, i assume it will take significantly longer, but i was wondering if the settings used with the sized down denoising filter will apply the same for the un-resized video.

like, i am not sure exactly how the filters work, but i figure at least one of the parameters deals with the size of the noise (which will be different if not resized). any recommendations for how i should change the parameters (bigger, smaller).

i also have not tested the speed of the changed order, but i plan to do that tonight, i will report back.

ronnylov
1st June 2007, 15:44
Denoising before downsizing will be less strong than doing it after the downsizing so you will probably need stronger settings to get the same amount of denoising when doing it before the downsizing. The downsizing itself will also reduce some noise. The actual encoding may also reduce noise.

In my experience too strong denoising hurts more than no denoising at all so you may want to try without denoising. Do some short samples and see what you like most.

twolfe18
1st June 2007, 22:29
well, i encoded a sample twice, once with the denoising first, once with it second. in the one where i denoised first, you could see a small improvement. in terms of speed, the results were as follows:

FPS
__________Denoise First______Denoise Second___% Slower
Pass1_____5.98_____________7.94_____________24.7
Pass2_____1.66_____________1.60_____________-3.61
Pass3_____1.65_____________1.73_____________4.62
Overall______________________________________3.40

don't ask me why i did a 3 pass, but those are the results. i used the FluxSmoothST(7,7) filter in both encodeds. the 3.40% slower overall is for total time to encode (not a numerical average of the FPS's). also, if i had done a 2 pass (and i averaged the second and third passes together to count as the second pass for the theoretical 2 pass), change in order would have caused a 6.52% overall slow down. i would have to test it again though to tell for certain. to me, this seems like a pretty small and worthwhile time trade off, and i am going to continue this.

foxyshadis
2nd June 2007, 06:22
Since first pass plays fast and loose, you can usually get away with using the faster script, as long as it doesn't make huge changes to the nature of the video - depending on your video, no denoising or removegrain may be sufficient. Of course using the same radius before and after will look very different, such is the nature of 3x3x3 filters, but only you can figure out what you want. =D