Re: [darktable-user] Wishlist: Upscaling via Machine Learning

2017-08-10 Thread Michael
Sometimes when I keystone I need to zoom in to get a proper vertical line.
I think the infinity symbol needs to zoom as well. I always thought that
but only just recently did it interfere with my work.


darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org

Re: [darktable-user] Wishlist: Upscaling via Machine Learning

2017-08-10 Thread Tim Rolph
Have you tried using neat?
On Thursday, 10 August 2017 00:04:26 BST Andy wrote:
> What you really want is deep nets trained for specific sensors/ISO
> combinations
> for denoising. Looks like neat doesn't use deep learning (if they would,
> I'm sure they'd mention it in their explanation).
> So it's unlikely to be state of the art (last time I checked).
> 
> On 07/10/2017 12:05 PM, Tim Rolph wrote:
> > Hi, if I have to upscale an image I usually use Reshade from
> > www.reshade.com it is a free app the works very well although its only
> > for windows but works fine under wine. Oh and its also free. If I ever
> > have a problem with noise I use Neat Image https://ni.neatvideo.com/ they
> > have a free linux version that is limited to save only in jpg but has
> > support for opencl / cuda GPU processing and it gives unbelievably good
> > results.
> > 
> > Tim.
> > 
> > On Sunday, 9 July 2017 17:16:12 BST Michael Below wrote:
> >> Hi,
> >> 
> >> last week I took a couple of images at a concert, and it turned out
> >> that only a small part of each image was interesting. I was too far
> >> away, with a wide-angle lens, so the band I wanted to photograph was in
> >> a small part in the center of the frame with lots of other stuff around
> >> them, stage, audience etc.
> >> 
> >> Now this can be solved by taking better pictures, coming closer, being
> >> prepared with a telephoto lens etc. - but there also seems to be a
> >> solution that could find its way into darktable.
> >> 
> >> There have been a number of media reports about machine learning
> >> experiments by Google etc. to add missing detail to images during
> >> upscaling. It seems like the results are often quite convincing. Now I
> >> stumbled upon a Github project for this that seems to offer a hands-on
> >> solution which might be a basis for implementation in darktable:
> >> 
> >> https://github.com/lucasdupin/ml-image-scaling
> >> 
> >> What do you think? I imagine this would be useful...
> >> 
> >> Cheers
> >> Michael
> >> _
> >> ___ darktable user mailing list
> >> to unsubscribe send a mail to
> >> darktable-user+unsubscr...@lists.darktable.org> 
> > __
> > __ darktable user mailing list
> > to unsubscribe send a mail to
> > darktable-user+unsubscr...@lists.darktable.org
> 
> darktable user mailing list
> to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org



darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org



Re: [darktable-user] Wishlist: Upscaling via Machine Learning

2017-08-09 Thread Andy
What you really want is deep nets trained for specific sensors/ISO 
combinations

for denoising. Looks like neat doesn't use deep learning (if they would,
I'm sure they'd mention it in their explanation).
So it's unlikely to be state of the art (last time I checked).

On 07/10/2017 12:05 PM, Tim Rolph wrote:

Hi, if I have to upscale an image I usually use Reshade from www.reshade.com
it is a free app the works very well although its only for windows but works
fine under wine. Oh and its also free. If I ever have a problem with noise I
use Neat Image https://ni.neatvideo.com/ they have a free linux version that
is limited to save only in jpg but has support for opencl / cuda GPU
processing and it gives unbelievably good results.

Tim.

On Sunday, 9 July 2017 17:16:12 BST Michael Below wrote:

Hi,

last week I took a couple of images at a concert, and it turned out
that only a small part of each image was interesting. I was too far
away, with a wide-angle lens, so the band I wanted to photograph was in
a small part in the center of the frame with lots of other stuff around
them, stage, audience etc.

Now this can be solved by taking better pictures, coming closer, being
prepared with a telephoto lens etc. - but there also seems to be a
solution that could find its way into darktable.

There have been a number of media reports about machine learning
experiments by Google etc. to add missing detail to images during
upscaling. It seems like the results are often quite convincing. Now I
stumbled upon a Github project for this that seems to offer a hands-on
solution which might be a basis for implementation in darktable:

https://github.com/lucasdupin/ml-image-scaling

What do you think? I imagine this would be useful...

Cheers
Michael

darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org



darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org




darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org



Re: [darktable-user] Wishlist: Upscaling via Machine Learning

2017-07-10 Thread Lorenzo Bolzani
These are the most impressive results I'm aware of:

https://github.com/david-gpu/srez
https://github.com/alexjc/neural-enhance
https://arxiv.org/pdf/1609.04802.pdf

but both gives the best results with specialized training (faces for
example). The one Michael proposed seems a more generic one.
No idea about performance but "execution" time is usually fast, memory may
be a much bigger requirement.

This one is about noise reduction:

http://webdav.is.mpg.de/pixel/files/neural_denoising/paper.pdf

I think in a couple of years machine learning will be the only way to
achieve state of the art results for about anything. If not today already.

Noise, white balance, clipping recovery, blurry/out of focus shots, color
casts, lens correction, defringe, sharpening, inpainting, etc. Up to this
(pose edit, content generation):

https://github.com/nightrome/really-awesome-gan
https://www.slideshare.net/Artifacia/generative-adversarial-networks-and-
their-applications (slide 24, 27)


It's going to be fun :)


Lorenzo


2017-07-09 18:14 GMT+02:00 David Vincent-Jones :

> Along the same line: Some years ago Cliff Reiter (Lafayette College)
> (using Jsoftware) demonstrated and published 'lossless edge' image
> rotation using fractals.
>
> David
>
> On 07/09/2017 08:16 AM, Michael Below wrote:
> > Hi,
> >
> > last week I took a couple of images at a concert, and it turned out
> > that only a small part of each image was interesting. I was too far
> > away, with a wide-angle lens, so the band I wanted to photograph was in
> > a small part in the center of the frame with lots of other stuff around
> > them, stage, audience etc.
> >
> > Now this can be solved by taking better pictures, coming closer, being
> > prepared with a telephoto lens etc. - but there also seems to be a
> > solution that could find its way into darktable.
> >
> > There have been a number of media reports about machine learning
> > experiments by Google etc. to add missing detail to images during
> > upscaling. It seems like the results are often quite convincing. Now I
> > stumbled upon a Github project for this that seems to offer a hands-on
> > solution which might be a basis for implementation in darktable:
> >
> > https://github.com/lucasdupin/ml-image-scaling
> >
> > What do you think? I imagine this would be useful...
> >
> > Cheers
> > Michael
> > 
> 
> > darktable user mailing list
> > to unsubscribe send a mail to darktable-user+unsubscribe@
> lists.darktable.org
> >
> 
> 
> darktable user mailing list
> to unsubscribe send a mail to darktable-user+unsubscribe@
> lists.darktable.org
>
>


darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org

Re: [darktable-user] Wishlist: Upscaling via Machine Learning

2017-07-10 Thread Tim Rolph
Hi, if I have to upscale an image I usually use Reshade from www.reshade.com 
it is a free app the works very well although its only for windows but works 
fine under wine. Oh and its also free. If I ever have a problem with noise I 
use Neat Image https://ni.neatvideo.com/ they have a free linux version that 
is limited to save only in jpg but has support for opencl / cuda GPU 
processing and it gives unbelievably good results. 

Tim.

On Sunday, 9 July 2017 17:16:12 BST Michael Below wrote:
> Hi,
> 
> last week I took a couple of images at a concert, and it turned out
> that only a small part of each image was interesting. I was too far
> away, with a wide-angle lens, so the band I wanted to photograph was in
> a small part in the center of the frame with lots of other stuff around
> them, stage, audience etc. 
> 
> Now this can be solved by taking better pictures, coming closer, being
> prepared with a telephoto lens etc. - but there also seems to be a
> solution that could find its way into darktable.
> 
> There have been a number of media reports about machine learning
> experiments by Google etc. to add missing detail to images during
> upscaling. It seems like the results are often quite convincing. Now I
> stumbled upon a Github project for this that seems to offer a hands-on
> solution which might be a basis for implementation in darktable:
> 
> https://github.com/lucasdupin/ml-image-scaling
> 
> What do you think? I imagine this would be useful...
> 
> Cheers
> Michael
> 
> darktable user mailing list
> to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org



darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org