denoise after scaling or before?
Hello, I just have an general question: As I know denoiser are specialized to specific noise pattern. Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling. In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way. Have a nice day! Jonathan
Am 23.04.2019 um 11:43 schrieb Jon bae:
Hello, I just have an general question:
As I know denoiser are specialized to specific noise pattern.
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way.
You could try both versions simultaneously and compare them with hstack: ffmpeg -i in.mov -i in.mov -filter_complex "[0]scale=640:400,atadenoise[a];[1]atadenoise,scale=640:400[b];[a][b]hstack" out.mov There are at least six different denoisers available: atadenoise, bm3d, dctdnoiz, fftdenoiz, owdenoise and vaguedenoiser Michael
On 4/23/19, Michael Koch <astroelectronic@t-online.de> wrote:
Am 23.04.2019 um 11:43 schrieb Jon bae:
Hello, I just have an general question:
As I know denoiser are specialized to specific noise pattern.
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way.
You could try both versions simultaneously and compare them with hstack:
ffmpeg -i in.mov -i in.mov -filter_complex "[0]scale=640:400,atadenoise[a];[1]atadenoise,scale=640:400[b];[a][b]hstack" out.mov
There are at least six different denoisers available: atadenoise, bm3d, dctdnoiz, fftdenoiz, owdenoise and vaguedenoiser
You missed nlmeans.
Am Di., 23. Apr. 2019 um 12:17 Uhr schrieb Paul B Mahol <onemda@gmail.com>:
On 4/23/19, Michael Koch <astroelectronic@t-online.de> wrote:
Am 23.04.2019 um 11:43 schrieb Jon bae:
Hello, I just have an general question:
As I know denoiser are specialized to specific noise pattern.
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way.
You could try both versions simultaneously and compare them with hstack:
ffmpeg -i in.mov -i in.mov -filter_complex
"[0]scale=640:400,atadenoise[a];[1]atadenoise,scale=640:400[b];[a][b]hstack"
out.mov
There are at least six different denoisers available: atadenoise, bm3d, dctdnoiz, fftdenoiz, owdenoise and vaguedenoiser
You missed nlmeans.
Ok, thank you, comparing is a good idea too! In general I thought there must be a more technically accurate way, but I guess this depends on the selected filter.
On 2019-04-23 12:17, Paul B Mahol wrote:
On 4/23/19, Michael Koch <astroelectronic@t-online.de> wrote:
There are at least six different denoisers available: atadenoise, bm3d, dctdnoiz, fftdenoiz, owdenoise and vaguedenoiser
You missed nlmeans.
Removegrain too. Maybe a bit basic but it is still there. On the original point: it used to be considered best practice to do your filtering before scaling.
Am 23.04.19 um 11:43 schrieb Jon bae:
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way. If I see this correctly, most denoisers define a spacial area to work on. So denoising before downscaling in the end will affect a smaller spacial area then in opposite order.
So it's your decision, how aggressive you want the denoiser to work. -Ulf
Am 24.04.2019 um 10:43 schrieb Ulf Zibis:
Am 23.04.19 um 11:43 schrieb Jon bae:
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way. If I see this correctly, most denoisers define a spacial area to work on. So denoising before downscaling in the end will affect a smaller spacial area then in opposite order.
What are the pros and cons of the 8 available denoisers? Which of them is the best choice for which purpose? Are there any recommendations? I did try a few of them (but not all) and found that atadenoise is good for high-ISO videos of the night sky. This testing is very time consuming because most denoisers have several parameters that must also be optimized. Michael
On 4/24/19, Michael Koch <astroelectronic@t-online.de> wrote:
Am 24.04.2019 um 10:43 schrieb Ulf Zibis:
Am 23.04.19 um 11:43 schrieb Jon bae:
Makes it a different when I denoise after scaling a video, or is the quality better when I denoise before scaling.
In terms of speed it would be better for me, denoising after scaling, but when the quality is better, then I would go that way. If I see this correctly, most denoisers define a spacial area to work on. So denoising before downscaling in the end will affect a smaller spacial area then in opposite order.
What are the pros and cons of the 8 available denoisers? Which of them is the best choice for which purpose? Are there any recommendations? I did try a few of them (but not all) and found that atadenoise is good for high-ISO videos of the night sky. This testing is very time consuming because most denoisers have several parameters that must also be optimized.
atadenoise: very fast, temporal only with no motion compensation; LGPL hqdn3d: fast, both spatial and temporal, does basically lowpass by destroying high frequencies, blurs with extreme settings; GPL nlmeans: very slow, currently implemented as spatial only, algorithm considered as one of the state of art denoisers; LGPL bm3d: very very slow, currently implemented as spatial only, algorithm considered as one of the state of art denoisers; LGPL vaguedenoiser: slow, spatial only, pretty good, wavelet; LGPL dctdnoiz: very very slow: spatial only, blurs too much; LGPL fftdnoiz: slow, spatial and limited temporal, using Fast Fourier Transform, may have introduce ringing with bad settings; LGPL owdenoise: very very very slow, spatial only, wavelet; GPL removegrain: fast, spatial only, limited usecase
participants (5)
-
James Darnley -
Jon bae -
Michael Koch -
Paul B Mahol -
Ulf Zibis