NHW Project - a free state-of-the-art image/video compression codec

[nhw_pulsar]

Originally Posted by benwaggoner

A more relevant comparison would be something like x264's --tune stillimage.

I think I found something like this provided good web quality at a small file size:

x265 --preset placebo --range full --keyint 1 --fps 1 --crf 18 --aq-mode 4 --aq-strength 2 --psy-rd 3 --psy-rdoq 5 --rd-refine

--no-sao and a lower crf can improve detail if a bigger size isn't an issue.

Hi,

I forgot to thank you for your 2 still frame encoding commands with x264 and x265.

Sadly, I won't have time to test them my side, very sorry, but if some people could test them and let us know how they would (visually) compare for example to AVIF -s 0, then I would be very interested.

Cheers,
Raphael

[nhw_pulsar]

Hello,

Just a quick update to remind you that NHW codec is only developed for -q9 to -q23 settings and it is absolutely not ready for extreme compression.

But maybe it should be my new focus now, to develop -q0 to -q8 compression settings?

Because some people seem to be testing at extreme compression, for example I had last year a quick conversation with a MPEG video compression expert, and apparently video compression experts seem to test at extreme compression, because he tested NHW at something like -q4 setting if I remember correctly, and he answered me that he was not convinced of NHW compared to VVC, but I answered him that NHW was absolutely not developed for -q0 to -q8 settings (!), and then our discussion stopped.

I did not really test but it seems that VVC is extremely interesting at extreme compression, and when we learn that ECM is 40% better than VVC, yes it's a real revolution that could be announced: to have good quality images at very small/tiny sizes.That could be game changing, but on the other hand ECM is very slow to decode which is a major drawback.

So would some of you really need that I develop -q0 to -q8 settings? Really don't know how far I could go there, I can first start to revise the psychovisual model parameters for these settings, they are too severe and so wash too many things out... I also think that a post-processing filter will really become important at extreme compression.

Cheers,
Raphael

[nhw_pulsar]

Hello,

Just a quick update, yes I have taken a lot of delay in releasing the version 0.3.0 of NHW... I however found some time to work on NHW this month of July but I did not manage to improve it... I start to think now that the next essential step that will have a real visual impact, is to design a post-processing filter that will remove aliasing from decoded images and that would also improve reconstructed quality.But designing such a beast is above my forces currently, so I think there's nothing much to expect for NHW for the rest of this year...

But I can be wrong, as some people seem interested in NHW, if some of you managed to improve NHW, it would be great if you could share it!

BTW to finish, in 2015 an engineer told me that the vast majority of low-power IP cameras were using MotionJPEG, and that they could be interested in NHW as it was lower power than JPEG, but this project never materialized.I wanted to know, in 2023 does MJPEG is still the leading codec for the low-power IP cameras market and does the niche industry could be still interested in a lower power codec for this market? Also any niche application suggestion where the low-power NHW codec could be interesting, is very welcome!

Cheers,
Raphael

[Selur]

I would recommend checking first that there are still low-power IP cameras out there that use 512x512,.... (I doubt it)

[nhw_pulsar]

Hi,

Actually, I think that the engineer that contacted me was interested in the NHW technology for its low power aspect.And if his company had decided to develop NHW, then I also think it would have been possible that they delegated ressources to adapt NHW to any image size (maybe with my help).But sadly, this project never materialized...

Cheers,
Raphael

[benwaggoner]

Originally Posted by Selur

I would recommend checking first that there are still low-power IP cameras out there that use 512x512,.... (I doubt it)

Yeah, 512x512 is really a thumbnail image size these days. Smaller thumbnails have value in many applications, but they're not the "main thing" people are optimizing for.

[nhw_pulsar]

Originally Posted by benwaggoner

Yeah, 512x512 is really a thumbnail image size these days. Smaller thumbnails have value in many applications, but they're not the "main thing" people are optimizing for.

Hi Ben,

I would like to end this debate.I am not trying to find applications for a 512x512 codec, but I just would have liked to find a company/organization that could be interested in NHW technology and that would like to develop it.

Cheers,
Raphael

[nhw_pulsar]

Hello,

Ok, that's right that if since 2008, none company/organization was interested in the NHW codec, it very means that there is an extremely big problem with my approach.Yes that's right that compression bodies told me they can not evaluate my codec because of its restricted 512x512 size, it can not encode their test/evaluation images.

But again so unfortunately, adapting NHW to any image size is _for me_ an extremely big task, and I absolutely won't be able to do that massive effort.So NHW will stay an experimental playground for image compression enthusiasts. I think I will let "sleep" NHW for the rest of this year, as I have other priorities now.

Cheers,
Raphael

[jmartinr]

If you ever pick it up again, would a fixed size of 2048 x 2048 or 4096 * 4096 be of some value? Or even 8192 x 8192?

Big sizes fast seems more interesting than small sizes fast, and you can always just mask it and use a smaller part of the canvas.

[nhw_pulsar]

Originally Posted by jmartinr

If you ever pick it up again, would a fixed size of 2048 x 2048 or 4096 * 4096 be of some value? Or even 8192 x 8192?

Big sizes fast seems more interesting than small sizes fast, and you can always just mask it and use a smaller part of the canvas.

Hi,

Thank you for your suggestion.

Yes that would be better, but actually going beyond 512x512 will require quite some code modifications.Actually for a 512x512 image, you have a 256x256 first order wavelet DC image, and I was very lazy but I coded all the processings on this wavelet DC 256x256 image, like notably the residual coding and the loop-feedback correction, with char variables (8bit). So going larger will require some important code adaptation/modification, and I am very sorry but I don't have the forces to do it...

I am aware that this is a very experimental version, but contrarily to some people who say that NHW will really degrade at larger/big size, I think that NHW should perform the same for larger sizes just as 512x512.

Cheers,
Raphael

[nhw_pulsar]

Hello,

I could find some time these last weeks to work on NHW... I tried to improve the global tuning, I coded some new processings, but none of that worked.My aim is not to increase neatness (as it tends to oversharpen, for me) but to keep the current neatness and improve more and more precision.... But I didn't manage to improve that aspect.Now it seems very hard to achieve a real visual improvement in that regards (for me), but actually I am happy with the visual results of NHW v0.3.0-rc1 last version.I know I really shouldn't say this and you'll consider me as an unserious troll, but as my eyes seem very sensitive to neatness, then I visually prefer the results of NHW v0.3.0-rc1 than the latest version of AVIF and VVC VTM 12.3... -so maybe I'll also freeze this version and bump the number to the official v0.3.0 release...-

As I said (yes many times), I think now that an extremely important and beneficial processing to develop would be a post-processing filter that will remove aliasing from decoded images and that would globally improve reconstructed quality. But I am a total newbie with it... I have read that there are anti-aliasing filters, but would they work on NHW decoded images? I also could think of machine learning for this design, and for example train a model on aliasing patterns (which are often the same)? I would also like that this post-processing enhancement filter stays compatible with my very fast decoding/high speed goals if possible.As I will have less time now to work on NHW, it would be great if I could find a collaboration for this so important post-processing design.If you could help, do not hesitate to show up!

Cheers,
Raphael

[nhw_pulsar]

Hello,

Quite many people urged me to adapt NHW to any image size, I know that it is essential but I am very sorry again, but it is a very too much colossal task for me (and my spare time) currently.

Now, I can find little time to work on micro core improvements for subjective visual quality. I am actually puzzled because I tried some new ideas that on the paper should work, but when I visually reviewed them, images were not better, and so surprisingly, it's hard for me to visually improve the current NHW v0.3.0-rc1 version...

Yes some people also told me that NHW is uninteresting and quite bad because it has very bad PSNR and SSIM results which are the reference criteria of evaluation for the industry, but all I can answer to that, yes it's very controversial, but VVC VTM 12.3 (slowest preset) has extraordinary PSNR and NHW has extremely bad PSNR, but when you actually look at the images then I visually prefer the NHW images than VVC ones, mainly because they have more neatness.VVC has more precision, but to my taste, neatness is visually more pleasant than precision... -Not to say that NHW has no post-filters for now that would remove aliasing...-

Cheers,
Raphael

[nhw_pulsar]

Hello,

For those interested, I have released the NHW v0.3.0-rc2 version.

This new version presents a "micro core improvement", now I better compute the "weights" for the nhw_kernel in the pre_processing, resulting in a little visually better version.

More at: http://nhwcodec.blogspot.com/

-This visual improvement seems more pronounced on some images than others, but I maybe did not test on enough images at that time...-

Cheers,
Raphael