"Denoise" Patches - for GH2 Hack - Experimental

apefos

This is experimental.

Maybe increasing the DC number and keeping same or similar numbers in the matrix can deliver a denoised image more video look, which would be useful to save denoise time in post production and to be good for web upload (web upload likes clean images due to low datarate).

first matrix idea:

024 024 024 024

any number repeat can be tested: 12, 16, 18, 20, 24, 28, 32 or 36 in all positions for Luma ( I am thinking about to try multiples of 4 or 6 which are good dc numbers for the matrix I found in "The Patch" topic.

I think deblocking tables can be low, because quantization in main matrix is already high in first digits (near dc number). I am thinking to try 1, 2, 3, 4 or 6 in all positions for intra and inter.

I think there is no need for fallback, also due to high quantization near dc number. Or fallback can be implemented same way the patch did, to allow iso increase using the fallback for it.

first patch: matrix = 24, no fallback, deblocking = 1, based on "the patch" ini file, experimental

frullaccia

@apefos I am curious... By the way, the e patch for PAL hbr users is interesting, but quite close to the I version.

apefos

@frullaccia I did a reply in the correct topic about e vs. i, please keep on topic, thanks.

apefos

In the need of testers for these patches, please read:

I do not have enough time to do this careful tests now, so please if there is someone willing to help, please read and test:

denoise patches - experimental - 10 versions

an attempt to get denoise images from camera and less sharpness

if it works, it can be good for web upload videos, more clean, for low datarate web videos, dvd...

denoise inside camera saves post production time and avoids image trembling from denoise softwares

also image will be ready to apply your film grain in post

the matrix desing uses a high number in first position (the dc number) and repeat it in all matrix

there is no fallback due to high values in first digits of matrix seems to be enough for quantization

there is no deblocking tables (all positions set to 1) because the scaling matrix desing are already in a deblocking desing and with enough quantization

all digits are multiples of 4 or 6 or both.

all recording modes have the same matrix in each version, this is better for tests

versions:

seta = denoise 12

setb = denoise 16

setc = denoise 18

setd = denoise 20

sete = denoise 24

setf = denoise 28

setg = denoise 30

seth = denoise 32

seti = denoise 36

setj = denoise 40

how to test:

put camera on tripod, interior shoot, low light, artificial light, no changes in light.

put camera in manual mode, fix shutter speed to 1/60 or 1/50

select a recording mode

start from iso 160 and shoot 10 seconds in each iso until iso 12800

keep changing the diafragm for each iso to keep same exposure

do this for these three recording modes: 24H, HBR and SH

see images in timeline with 400% zoom and take notes

the goal is to find the best version for each recording mode

and also which iso range the versions works for the recording modes

this way it will be possible to build one version with fallback for iso increasae and different matrix for each recording modes

report here...

producer

@apefos: Sorry to post in this thread since it's about denoising, but is it possible to modify the matrix for higher sharpness too? Not like Moon T4 for example. What I mean is to increase the micro sharpness instead of total sharpness.

apefos

@producer you are on topic, as you are talking about sharpness, related to these patches. I think the answer for your question is on the "the patch" topic. Please see there the new dcn4 patches and compare them to previous "the patch" versions. And then you can talk there about them.

frullaccia

@apefos Rainy afternoon, teleconverter on, full zoom, no tripod. Trees at 200-250 meters. Version 24 from this denoise patch vs version I from "The Patch". MP4 files; same compression. Link: https://bft.usu.edu/rf6mj https://bft.usu.edu/d8q4z

apefos

@frullaccia Interesting test!

Even with the MP4 compression, it seems that the "denoise 24" is showing less noise than the "the patch i" even in good light parts of the video. Also I perceived less trembling in trees when they are under good light near the road in the "denoise 24", the handheld helped to see this.

What did you perceive when you saw the original mts videos on the screen? Are this things I said correct?

Was iDynamic disabled in both shoots?

Did you use same iso for both shoots? As you said you used iso 160 for the "i", so it seems that for denoise the denoise24 is better.

Would be good idea to redo the test in iso 1600.

It seems that denoise patches have a future...

Thanks a lot.

apefos

This thing can have an interesting solution for denoise, if my theory is correct.

I did some tests with neatvideo right now. What happens is:

In a iso 2000 video with luma and chroma noise, if you apply just spatial denoise, there is no tremble in denoised video. If you apply temporal denoise, some areas of image trembles. The more the amount of temporal denoise, the more the tremble. NeatVideo levels for temporal denoise are 0, 1, 2, 3, 4 and 5. 0 is disabled, 1 is one frame before, one frame after, and so on... the temporal denoise when comparing the previous and after pixels changes the position of the rendered image in some areas.

The "the patch" are tweaked to show the noise from the sensor, not to hide it, in an attempt to deliver a cinematic, filmic image, good for the big screen.

The "denoise 24" is tweaked to compress more the image and mud the noise and texture into a flat look, better for low datarate web videos.

The videos uploaded by @frullaccia have a MP4 recompression from the original MTS videos. The MP4 recompression does some kind of temporal compression using information from one frame to another from the original I, P and B frames, to generate new more compressed I, P and B frames. This way the high compression of MP4 acts as a denoise, because high compression muds the noise and textures. This allows we to can explain why the images from the "the patch I" trembles after recompression and from the "denoise 24" does not tremble: it is because there is less noise in the denoise 24 video, then the MP4 recompression does not need to mud any significant noise, so when using information from before and after frames to generate the new I, P and B frames the parts of image keeps in the same place. So the "denoise 24" patch is proving to be better for recompressed videos for web upload.

One thing I do not know yet: would be better to disable the deblocking tables or would it helps even more to avoid noise and macroblocking? Here it is the "denoise 24" now with a deblocking table enabled and also a better tweaked chroma matrix. Probably the "denoise 24" is the sweet spot because the quantization in luma matrix is 1.5 and in the chroma matrix is 2.94 resulting in a final quantization near 2. This can handle all the isos. Also 24 is 4x6 which is interesting maths. 24 is also the last digit of the great 24p matrix in the "the patch" which delivers the best texture in the dcn6 series. Now deblocking is following these maths in the "denoise 24".

Based on these tests I also did another version of the "denoise 36" following same maths ideas. The higher quantization on it can be better for in camera denoised videos from high iso. It would be interesting to see a comparison of the new "denoise 24" vs. the new "denoise 36" in iso 2000 which is a good high iso limit for the GH2. Both compressed to MP4 using same compression level.

@frullaccia the handheld streess tests you do are great to develop the "denoise" patches. you show same areas and movements in each shoot, great for comparisons. the recompression to mp4 also helps to see which patch will look better after editing and recompression. Thanks!

One question: which picture profile are you using and which noise reduction level are you using in the picture profile? Thanks again.

these versions are not working ok, use upload from previous post, without fallback and deblocking tables

apefos

this is a version using 24 as main matrix and 36 as fallback. I do not know if it improves because the quantization from the matrix 24 is already high, so I do not know if the camera will use the fallback.

also I used two deblocking tables, 24-4 as flag table and 36-6 as high table.

all recording modes uses the same matrix configuration, I also do not know yet if there will be benefits with different matrices for different recording modes.

if you want to do careful tests to perceive difference from denoise 24 to denoise 36, it will be better to compare the two versions in previous post.

this version is not working ok, use version uploaded in post above.

frullaccia

@apefos All these versions are interesting. The texture is always very good. I cannot see big differences between 12 and 40. But I should test again in low lights at high iso.

Probably Denoise 24 has a little less noise than I.

IDynamic was disabled. OIS and autofocus single were on. The profile of the camera was in Standard mode (-2 -2 +1 -2). ISO was 160 (I also made a try at 800 ISO).

HBR is also very well performing.

Talking about mp4 compression, I think you're right. I noticed a strange denoise compression from the 24version file, the first time I made these copies. Once they were compressed, the difference between 24 and "I" was not real. The 24 version being much more neat than the original mts file. I made this second copy, which is more realistic (I didn't change any compression set. I just made a second copy from the original file, to see what happened).

apefos

@frullaccia Thanks for answers. So, after the mp4 recompression, you like more the "denoise 24"?

frullaccia

Yes. After the first compression, denoise 24 was much more neat. I made a second copy from the original, which came out with a different result. I don't know why. It is the copy that you have seen.

apefos

The copy I saw I liked the "denoise24" more.

frullaccia

So...

I can tell, all these versions show a very fine texture.

apefos

So... it is the end of developing.

I did a fast test, the denoise24 shows the less chroma noise I ever saw. texture is ok.

avoid the last versions with fallback and deblocking tables, they do not work good, use the versions uploaded together from denoise 12 to 40

apefos

the difference in chroma noise between denoise and the patch is not so significant, but I used some ideas from denoise in the the patch fallback to improve chroma noise reduction, so I will keep using the patch.

Howdy, Stranger!

Categories

Tags in Topic

Top Posters