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

[nhw_pulsar]

Hello,

Just a very quick update, so the industry and academia are not interested in NHW.That's also totally right that if I then ask sponsoring on the Internet, then I will have to spend a good amount of time on adapting to any image size, which after some thought I am not ready for for now.So NHW will stay a spare time hobby, I wanted to thank the few people who follow my work, I hope I will find the motivation to resume my work on NHW this year, because it is on pause since the last update.

Very quickly to finish, I am also always quite surprised when regularly some people tell me that they find that NHW is not visually good/pleasant, mainly because it does not retain details, but it's certainly a problem with me, but my eyes are not catched by this latter aspect, but really more by neatness... but again there is maybe a problem because I really seem the only one to have that opinion.-Anyway, I'll continue to develop for neatness, which will be a new approach for image compression, even if not considered.-

Cheers,
Raphael

[benwaggoner]

Originally Posted by nhw_pulsar

Hello,

Just a very quick update, so the industry and academia are not interested in NHW.That's also totally right that if I then ask sponsoring on the Internet, then I will have to spend a good amount of time on adapting to any image size, which after some thought I am not ready for for now.So NHW will stay a spare time hobby, I wanted to thank the few people who follow my work, I hope I will find the motivation to resume my work on NHW this year, because it is on pause since the last update.

Very quickly to finish, I am also always quite surprised when regularly some people tell me that they find that NHW is not visually good/pleasant, mainly because it does not retain details, but it's certainly a problem with me, but my eyes are not catched by this latter aspect, but really more by neatness... but again there is maybe a problem because I really seem the only one to have that opinion.-Anyway, I'll continue to develop for neatness, which will be a new approach for image compression, even if not considered.-

Yeah, there's a big impact of personal preference on psychovisual tradeoffs. People definitely tune for personal preference, and then prefer their own tunings. That's a big reason double-blind MOS testing with a variety of testers is industry standard.

[nhw_pulsar]

Hi,

Yes I definitely agree and I clearly think that in my codec I tune for personal/not-shared visual preference.On the other hand, I don't have the money to perform standard double-blind MOS testing with a variety of testers.I also suspect that most probably a result of such a testing would be to add more details, but it will be quite difficult with my algorithm I think, not to say a fundamental limitation of my approach (completely oriented toward neatness).

Cheers,
Raphael

[nhw_pulsar]

Hi,

Just very quickly, I have released the 0.2.1 version of NHW.

I have (just) rewritten the pre_processing function more properly and improve it.I find that this new version has then a little more neatness and precision.

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

Cheers,
Raphael

[nhw_pulsar]

Hello,

I am still not considering to spend all my weekend (and nights) at adapting NHW to any image size, very sorry.Some people however also told me that it would be nice now to remove aliasing in NHW.

Yes I think the next big thing in NHW now to add, as in the other codecs, is a post-processing enhancement filter, that will remove aliasing here and then enhance decoded image quality.

But my knowledge (and time) are very limited for such a thing.So far, I have seen very interesting results of the directional filters of CDEF filter in AV1, that seem to be able, if well tuned, to remove aliasing.Would there be experts of the CDEF filter that could give some of their time to test it on the decoded NHW images and see if it can be effectively tuned to remove aliasing? Also I find that aliasing appears in NHW only along low-contrast edges.So would it be possible to apply such a processing only on the blurred/low-contrast edges? Also in order to reduce the complexity/processing time, as I would like that the post-processing would be fast.Really any person who would like to work and have skills for such a post-processing enhancement filter is more than welcome!

To finish, machine learning can also be used and trained on aliasing patterns that are often the same in NHW images.But I would really need your insights here, as I am very out of skills here!

Cheers,
Raphael

[nhw_pulsar]

Hello,

I have improved the 0.2.1 version of NHW.This is a minor fix in the pre_processing, but this new version has a better precision.I did not touch the neatness, as I don't want to oversharpen.

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

Cheers,
Raphael

[nhw_pulsar]

Hello,

Just a very quick update about NHW, as I could test FPAQ compressor and I am quite surprised for now because it is very good and very fast! I tried FPAQ to compress the wavelet DC bitstream of NHW that is not optimal.For now, I still apply my extremely fast compression scheme of the DC parts which can be optimized (notably on the UV DC parts) as a first pre-processing stage, kind of pre-compression, and then I apply FPAQ.In average I can save another 1.5KB on the DC parts compression, so that's very good but where I am extremely surprised is that FPAQ is very fast, 3ms to compress and 1ms to decompress for a 512x512 image.I never would have thought that a PAQ series compressor would be that fast! So the objective to save 2.5KB per .nhw compressed files while staying very fast in the same time is totally reachable (for the huffman coding of the wavelet coeffs, can save bits by compressing the huffman codebook, select between different huffman tables and not only one, and other things..., also add chroma from luma for chroma compression,...).

Now the biggest challenge for NHW seems to see if we can design an efficient post-processing filter that will remove aliasing.Again aliasing seems to appear in NHW only along blurred/low-contrast edges, so it could seem a question of just detecting them, I am working ultra-ultra-slowly on it, but it seems more complex than it seems for now, at least for me... Any help is very welcome here!

Cheers,
Raphael

[nhw_pulsar]

Hello,

I have released the 0.2.2 version of NHW.This new version has a better neatness and a better precision thanks to an improved pre_processing.

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

Again it is so extremely hard to discuss with the industry, but some people tell me that the extreme speed of NHW should qualify it for the niche market.Again if you would know any company/organization that could develop NHW notably for the niche market, do not hesitate to let me know.

Cheers,
Raphael

[nhw_pulsar]

Hello,

Just a quick answer to few remarks I had.

That's right that I keep saying:"this new version has a better neatness and a better precision...", and this is very subjective and completely vague.So objectively, you can for example measure neatness with my neatness metrics and you can measure precision with PSNR or SSIM.You'll then understand that I completely develop for neatness and not for precision, which is not mainstream/recognized.

Some people also told me that since 2007, as NHW never interested anybody, why I did not abandon this project? In fact since 2009, NHW is completely just a spare time hobby.You can notice for example that since 2 years I am focusing on the pre_processing, but you can clearly see that I am extremely lazy because clearly all these improvements if they have been considered seriously could have been done in 1 or 2 weeks, and not so lazily in more than 2 years... More generally, I think if I had worked seriously on NHW, the latest 0.2.2 version really could have been done in 2010 or 2011... So in the end, it's also hard to accept all these years lost because I couldn't work profesionnaly on NHW but just on my very few spare time.

Currently, I am still ultra lazy at NHW and I hope I will keep the pace of 6 months for a new improved version, if it could be useful to the few people who follow my work/updates.

Cheers,
Raphael

[nhw_pulsar]

Hello,

For those interested, I have released the 0.2.3 version of NHW.Again and as usual (...), this new version has a better neatness and a better precision thanks to a better pre_processing.

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

Also very quickly, few people seem to test at extreme compression, just wanted to notice that NHW is absolutely not ready for extreme compression, for now it works up to -l12 very high compression setting.

Cheers,
Raphael

[nhw_pulsar]

Hello,

Just a message as I have visually evaluated this weekend the last 0.2.3 version of NHW (mainly on "good quality neat" images), because I little rushed to publish the 0.2.3 version without correctly visually testing it, and in fact I am really satisfied with the level of neatness reached now.Neatness is really the "innovative" aspect I want to put forward with NHW, and this new version could start to be convincing... I also started to work a couple months ago on a full-neatness version or a -neatness-boost switch, I have even more neatness level but the problem for now is that it also oversharpens the image which on the other hand visually hurts a little... So for now the latest 0.2.3 version has the most correct and good neatness, and it would be really great if some of you could reconsider NHW.As fewer and fewer people seem to follow my work, I wanted to let you know that if you tested NHW 2-3 years ago and was not convinced, hope you could give a second chance at the 0.2.3 version because for me it starts to be interesting... -but really don't know if the industry for example will change its mind...-

Also quickly as I previously said, the biggest challenge now for NHW is maybe to design a post-processing filter that will remove aliasing from NHW images, actually I have seen that anti-aliasing processings exist, I have notably read that anti-aliasing could be used in super-resolution, if someone could point me some documentation on this, it would be great!

Any feedback/opinion on the new 0.2.3 version is very welcome!

Cheers,
Raphael

[paul97]

hi, I did drag and drop with new nhw but it produced same file? could you indicate the new settings if there are any?

[nhw_pulsar]

Hi,

So sorry for my lack of explanation, actually the new version concerns -l4 setting to -l12 setting (mid compression to very high compression, -l13 to -l19 are not developed/ready at all for now).The setting you use is above -l4, and actually above -l4 it is high quality and low compression, and I did not find time to optimize this range for a long time, so it is unchanged for now... so sorry again.

Thank you very much for your time and for testing NHW.

Cheers,
Raphael

[nhw_pulsar]

Hello,

For those interested, I have released the 0.2.4 version of NHW.This new version has a better precision (and a little more neatness) thanks to an improved pre_processing.

For now, this improvement is a little too complex and memory consuming but this can be reduced, NHW is still extremely fast however.

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

Also thanks to skal, this new version has a CMake makefile and fixes the compilation warnings.

Cheers,
Raphael

[nhw_pulsar]

Hello,

I recently could talk with some profesionnals, notably to know why I have lost contact with AOM for example, and they told me that I don't realize that machine/deep learning image/video codecs are knocking very hard at AOM's door and they are announcing next-generation results, outperforming AV1 and VVC, and even outperforming AV2 and ECM... and so the industry can not be interested in my "little" wavelet codec... I just wanted to notice that it's a little sad (for me) that industry estimates quality of a codec with PSNR and SSIM, because NHW has a very better subjective quality notably thanks to its very good neatness... but professionals are completely not sensitive to this argument...

Some people however told me that a true advantage of NHW is its extreme speed and so I should clearly refocus now on niche applications.But it's clearly also very difficult to be in contact with the niche market.Just in 2015, an engineer told me MotionJPEG was mostly used in embedded IP cameras working in remote environment, MJPEG was not used because of its quality/compression but because of its very low complexity/power consumption, and so he told me that NHW could be then even more interesting because it has a better quality/compression than JPEG and especially it is faster than JPEG.So very quickly to finish, if you think you can help me to take contact with the niche industry, many thanks to let me know, it would be great!

Cheers,
Raphael

[BlueSwordM]

Bro, a ton of AOM and aomenc folks only care about PissNR as I like to call it and SSIM as well.
If you want to prove gains, use much more sounds psycho-visual metrics like ssimulacra2 and butteraugli.

[nhw_pulsar]

Hello,

Thank you for your advice.

I also think that AOM and MPEG are highly involved with DCT block-based technology because they invested a lot and a lot in it since many years, and that's however right that for the DCT-block-based technology evolution, PSNR and SSIM seem good indicators, as PSNR-wise: JPEG XR<WEBP<BPG<AVIF<VVC, (I do not include JPEG XL in this curve as I did not mesure PSNR with it), and that's right that the visual-pleasantness follows this PSNR curve for these DCT codecs, but with a new wavelet technology like NHW, it is not the same reasoning because NHW has very bad PSNR and SSIM, but for me it is visually more pleasant, mainly because it has more neatness, -and PSNR and SSIM only measure precision-...

For now I didn't really find convincing psycho-visual metrics (for me).I wrote a neatness metrics: NHW_Neatness_Metrics, that is not finished (actually much more work must be done on it...), but it can start to objectively measure neatness I want to put forward with NHW.

Cheers,
Raphael

[nhw_pulsar]

Hello,

For those interested, I have released the 0.2.5 version of NHW.Again this new version has a better neatness and a better precision thanks to an improved pre_processing.

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

I find that this new version starts to be interesting.

Cheers,
Raphael

[benwaggoner]

Originally Posted by BlueSwordM

Bro, a ton of AOM and aomenc folks only care about PissNR as I like to call it and SSIM as well.
If you want to prove gains, use much more sounds psycho-visual metrics like ssimulacra2 and butteraugli.

AOM seems to care about VMAF as well, which is the least-bad broadly available metric. But not even VMAF solves the problem of how to score temporal variations in quality, like I or B frame strobing. All our metrics are fundamentally per-frame metrics, and it isn't uncommon for two frames with the same metrics to look quite a bit different, even if the overall subjective quality is in the same ballpark. We commonly see this encoding grain, where the grain texture "jumps" between frames in an unnatural and distracting way that metrics are poor at detecting.

[Jamaika]

Originally Posted by BlueSwordM

Bro, a ton of AOM and aomenc folks only care about PissNR as I like to call it and SSIM as well.
If you want to prove gains, use much more sounds psycho-visual metrics like ssimulacra2 and butteraugli.

Interesting for me is that the AOM codec can be compiled without the object ssim.o. file. For me, is this dead function or unfinished project?