| Version 2 (modified by , 2 years ago) ( diff ) |
|---|
Denoise
Denoise is removing the signal insignificant or unsensible [1] in intended view. Which shall improve compressibility due to reduced entropy (effective amount of information); and subjectively improve quality.
[ [1] "insignificant" is that cannot be seen. "unsensible" is that does not make sense. ]
Typical denoise is value clamp based: Transform the data into frequency domain (via Fourier Transform alike), or Wavelet: [ https://scikit-image.org/docs/stable/auto_examples/filters/plot_denoise_wavelet.html ] And allow passing only certain range.
FFT (Fast Fourier Transform) is fast implementation of DFT (Discrete Fourier Transform): which asserts 2^n input length (to speed-up). [ ^ https://www.ni.com/docs/en-US/bundle/diadem/page/genmaths/genmaths/calc_fouriertransform.htm#d20e91 ]
DCT (Discrete Cosine Transform) is cut-down variant of DFT that uses only cosine transform; thus of reduced complexity but also minimum resolution: DFT: 2 * π / N DCT: 2 * π / (2 * N) ; somewhat comparable to "full_chroma_int": (not complexity-wise) https://github.com/MasterInQuestion/attach/commit/e18e5cd26b9d27bacefa685ba70d0d41bc7d7cbc#diff-75a0130f1814ba26c23e3a6fc6ce5046620163061d392086bc55e2c13c145bf9L264
Discrete Cosine Transform, Fourier Transform:
- https://users.dimi.uniud.it/~antonio.dangelo/MMS/2013/lessons/L11lecture.pdf
- https://users.dimi.uniud.it/~antonio.dangelo/MMS/2013/lessons/L03lecture.pdf
- https://www.youtube.com/watch?v=k3byqIaULb8
- https://www.google.com/search?hl=en&gl=ca&num=100&filter=0&q=site:users.dimi.uniud.it/~antonio.dangelo/MMS+inurl:lessons
Fourier vs. Wavelet Transform:
This article currently mainly covers video denoise. Info on audio denoise shall be added later.
Video
FFmpeg supports 8 dedicated [1] video denoise filters:
- ͏"bm3d" (BM3D: Block-Matching 3D)
- ͏"nlmeans" (NLM: Non-Local Means)
- ͏"fftdnoiz" (FFT then frequency clamp)
- ͏"dctdnoiz" (DCT)
- ͏"hqdn3d"
- ͏"vaguedenoiser"
- ͏"owdenoise" (Overcomplete Wavelet)
- ͏"atadenoise" (Adaptive Temporal Averaging)
[1] Not counting ͏"pp" (much defunct "libpostproc"), general manipulation ͏"fftfilt" alike.
. More info:
- https://stackoverflow.com/questions/74831282#75348463
- https://forum.shotcut.org/t/8631/12
- https://old.reddit.com/r/ffmpeg/comments/a5ejgi?sort=old#ebyev3t
"bm3d" and "nlmeans" would be of most interest: for best quality.
More details:
- https://stanford.edu/class/ee367/Winter2020/report/zhou_m_report.pdf
- https://webpages.tuni.fi/foi/3D-DFT/BM3DDEN_article.pdf
- https://github.com/gfacciol/bm3d/blob/master/bm3d.cpp
- https://www.ipol.im/pub/art/2011/bcm_nlm/revisions/2011-09-13/bcm_nlm.htm#index1h2
- https://www.dirk-farin.net/projects/nlmeans/
- https://forum.handbrake.fr/viewtopic.php?t=30135#p139077
- https://github.com/HandBrake/HandBrake-docs/blob/master/source/docs/en/latest/technical/filters-summary.markdown#denoise]
"bm3d"
BM3D pre-filters via Block-Matching for similar blocks: Then applies tuned traditional denoise for each. . Thereby workaround the overdone of applying typical denoise vastly.
Using denoised input for Block-Matching may somewhat aid similarity branching. But using complex algorithm (BM3D, NLM alike) for this phase mostly helps not, besides pointless computation. ]
"nlmeans"
[ Quote <Non-local means denoising for preserving textures>:
https://scikit-image.org/docs/stable/auto_examples/filters/plot_nonlocal_means.html
Non-Local Means replaces pixels' value by an average of selected pixels': Small patches centered around pixels are compared to the one of interest: with the average performed only when the patches share enough similarity. As result, the algorithm may restore well textures that would be otherwise blurred. ]
[ Quote <Comparison of CNN based and self-similarity based denoising methods> # Introduction:
https://stanford.edu/class/ee367/Winter2020/report/zhou_m_report.pdf
NLM averages pixels with a similar neighborhood. Like other traditional methods, NLM can produce a high denoising quality. But has the drawback of low computational efficiency; and requires manual choice of some parameters: Such as the patch size, search window size, and cut-off distance. ]
<^> This is primarily for major NLM implementations seem to be false NLM:
Not really "Non-Local" thus involve the patch size, search window etc. non-essential. [ Quote <Non-Local Means Denoising> # Introduction:
https://www.ipol.im/pub/art/2011/bcm_nlm/article.pdf
Since the search for similar pixels will be made in a larger neighborhood, but still locally, the name "non-local" is somewhat misleading. ] ]
"favg"
Hypothetical new filter derived from NLM: The NLM related papers didn't well explain why the patch similarity procedure was even necessary. Wouldn't just mere averaging all pixels within specific value threshold: repeat for each unprocessed pixel group; adequately work? (Full-Image Averaging) [ #11019 ] ]
Caveat the factual validity of some linked.
Audio
FFmpeg supports ? dedicated audio denoise filters: ...
See also
https://github.com/FFmpeg/FFmpeg/blob/master/libavfilter/[av]f_${Name}.[ch] E.g. https://github.com/FFmpeg/FFmpeg/blob/master/libavfilter/vf_nlmeans.c
