Upscaling: I honestly want hear your opinion about

[luquinhas0021]

In a old thread, I comment about the use of edge detection for upscaling a image or a video. Someone here spoke edge detection is not a goal, because there's no edge detector that take all edges and only true edges (An ideal Edge detector). Honestly, I don't agree, by many reasons. But I wish hear you, my dear friens (LESS GROUCHO 2004): if edge detection isn't a goal to upscaling, what it is?

[pandy]

if edge detection isn't a goal to upscaling, what it is?

It depends... upscaling from mathematical perspective (DSP) is usually non human optimal (perceived quality is subjective and no mathematical model sufficiently universal exist yet)

[luquinhas0021]

You spoke upscaling by a mathematical approach! Does exist other approach?

[pandy]

You spoke upscaling by a mathematical approach! Does exist other approach?

Well - consider mathematical approach as something called signal processing (signal interpolation, extrapolation, signal reconstruction etc) - so theoretically perfect reconstruction filter is sinc and it should provide mathematically optimal results (based on Nyquist–Shannon sampling theorem and on Whittaker–Shannon interpolation formula) and sinc usually provide suboptimal picture quality (well no offense but if you are visually impaired then maybe sinc is OK).

[Sparktank]

Does exist other approach?

You only have two choices:
1) objective
2) subjective

You seem not to care about objectivity, so what does it matter what anyone else says?

You explicitly forbade Groucho to speak.
Leaving everyone to believe you only care about SUBjective views that match your own.

This is really a useless thread as you only to hear people who share the same subjective views as your own.
It matters not even if they CAN prove mathematically that your views are of pure heresy.

Sooner or later you have to pick something.
You're trying your best to convince yourself you are always right in a world of practicality.
Picking those to deny and ignoring mathematics does not make you right!
It only proves your arrogance and ensures no one cares to reply why you must reach such desperate measures.

That's just my opinion though.
This is "Geneeral Discussion". So you can't hold anything against me. Just how I generally see things.

[Groucho2004]

You explicitly forbade Groucho to speak.
I couldn't care less. This clown is actually quite entertaining.

[feisty2]

A universal truth: you get to convert something from one form to another, but you can't create things from "emptiness", how u gonna create those never actually existed missing details to get a native kind of HD copy? Guess GOD can help you with that.

[Keiyakusha]

with constantly improving resampling algorithms and things like nnedi now being available in most of the decent players out there, I find any kind of upscaling to be even more waste of time and space than it was before. Not to mention that keeping stuff as close to the source as possible allows any viewer to turn it into something (s)he wants it to be. And those who do not have enough knowledge about this stuff are unlikely to see the difference between upscaled and non-upscaled video anyway or won't care enough even if they'll see it.
The only real use for nnedi or its alternatives at encoding time is when you want to permanently go from interlaced content to progressive. Which is what they were designed for anyway.

[luquinhas0021]

Feisty2, deconvolution can "recover" "miss" details, and must be done after resolution increase.
Again, you are being down

[feisty2]

U just do whatever u like, I'm thru explaining

[Bloax]

The only real use for nnedi or its alternatives at encoding time is when you want to permanently go from interlaced content to progressive. Which is what they were designed for anyway.
And fortunately we have great solutions like QTGMC for just such an occasion.

[StainlessS]

I wish you all of the best upon your loveliest of lovelies quests, but they do seem a little repetitive.

EDIT:
My interpretation:

I want the bestest of the best best what you have ever seen, more best than is possible,
where can I find it ?


EDIT: Also, may I ask your age, you are starting to come across as about 14, ?.

feisty is a bit older than you and a well clever young dude, take note of every thing he says, (weeird way with words but well clever dude).

[Overdrive80]

Feisty2, deconvolution can "recover" "miss" details, and must be done after resolution increase.
Again, you are being down

Its impossible to recover anything that not exist.

[feisty2]

weeird way with words.

weird way like, how I use words or the spelling, I been doing both much better than Miley Cyrus at leeeeeeeeeeeeeeeast (typical "Cyrus" stuuuuuuff) I guess

[luquinhas0021]

I am 20 years old, StainleSS!

[Groucho2004]

I been doing both much better than Miley Cyrus at leeeeeeeeeeeeeeeast (typical "Cyrus" stuuuuuuff) I guess
She should not be your point of reference. :rolleyes:

[StainlessS]

Feisty, just reference to your text speak eg gr8 instead of great etc (you aint that bad really :) ).
I know nought of Miley Cyrus.

[pandy]

Its impossible to recover anything that not exist.
Recover not but you can guess (in educated or not way).
In audio there is something like spectral band replication - it doesn't mean original audio spectrum is reconstructed, it means that new spectrum can be perceived as better than without SBR.
So you can feed to reconstruction filter random data based on particular spectral information and you may (or not ) received better picture - for example adding noise to video/image will improve perceived resolution/sharpness...

[Motenai Yoda]

but sbr is based on data (reference) collected before the waveform is lowpassed, and encoded as well into bitstream.

[pandy]

but sbr is based on data (reference) collected before the waveform is lowpassed, and encoded as well into bitstream.

Not only - some blind method exist also...
Audio Engineering Society Convention Paper 6761 "High Quality Blind Bandwidth Extension of Audio for Portable Player Applications"

[raffriff42]

In audio, you can add additional harmonics to give the illusion of higher bandwidth; in video, this just looks like "ringing."

https://en.wikipedia.org/wiki/Spectral_band_replication
https://en.wikipedia.org/wiki/Ringing_artifacts

EDIT See below. Corrections gratefully acknowledged.

[foxyshadis]

The video equivalent would be sharpening after upsizing (generating high frequency components from low frequency patterns), kind of like PNS is the equivalent of FGM or just adding digital noise.

It wouldn't be ringing, because SBR's replication is in the frequency domain, not the time domain.

[pandy]

In audio, you can add additional harmonics to give the illusion of higher bandwidth; in video, this just looks like "ringing."

https://en.wikipedia.org/wiki/Spectral_band_replication
https://en.wikipedia.org/wiki/Ringing_artifacts

Nope ringing is made by something else (sinx/x is infinite and that's why you have ringing - usually you limiting infinite sinc by windowing it - you can always use window function as resampling kernel - work quite OK for downsampling - but still this is limited approach and usually suboptimal from subjective perspective).

Resizing in frequency domain (for example trough DCT) will give different artifacts.
So key is to develop method similar to blind spectrum replication/bandwidth extension subjectively free from artifact (it may be not mathematically optimal but this is not important as we talking about subjective quality).

[luquinhas0021]

Resize in time/frequency domain generates artifacts, this is it.
What about resize directly in every single one color channel (YCbCr or RGB), and then join it? Whithout transform to time/frequency domain?

[vivan]

What about resize directly in every single one color channel (YCbCr or RGB), and then join it? Whithout transform to time/frequency domain?What a novel idea!

Congratulations on reaching new low.

[feisty2]

What about resize directly in every single one color channel (YCbCr or RGB), and then join it? Whithout transform to time/frequency domain?

ALL resizers work like this (resizing each channel separately and get them back together when it's done)

the problem is, how you resize WITHIN each channel, and that's the point of "transform to time/frequency domain?" or whatever, sure you can do a nice work on the whole image if, you can do a nice work on each channel of the image

[Groucho2004]

http://s7.postimg.org/5pyezrls7/Image1.png
Just wondering what a "Narcizist" is - Dyslexic Self-Admirer?

[luquinhas0021]

Feisty2, I didn't know all resizers makes upscaling channel by channel, but that it wasn't my doubt. My doubt was if a resizer can do upscaling directly in pixel value (R, G or B), without any type of transformation, like frequency domain, time domain, gradient domain...

[colours]

"Time domain" is just jargon meaning that the signal has not been transformed. Also less misleadingly known as "spatial domain" in the context of image processing, where it actually has nothing to do with time.

And what do you think almost every resizing algorithm does? Of course they resize in time domain. Converting to frequency domain first is too costly for too little benefit; a fast Fourier transform might be fast, but not having to use it is even faster. There're situations where operating in the frequency domain can be useful (removing artifacts at certain frequencies, say), but resizing an image is not quite one of those.

[Parabola]

perhaps look at superresolution techniques...

[pandy]

perhaps look at superresolution techniques...

This imply time domain and spatial domain.