scanned images are going pixelated

Faced the very same (or similar?) problem while trying to resize photos made by digital camera. Very similar "grids" appear on nearly-flat surfaces with a bit of noise (like on sky). Surprisingly, faced the very same grids not only in GIMP, but in some other programs as well (in Image Viewer in Ubuntu 14.04, for example), so there may be a problem with underlying library (GTK?), not with GIMP itself. Here is an full-size example image:
https://drive.google.com/file/d/0B2nZFc31UFnoNUVTcGpoUHFSbm8/view?usp=sharing And it's resize (width 1920 px) - grid is easily visible: https://drive.google.com/file/d/0B2nZFc31UFnoT1ZvM0xxQW8xSU0/view?usp=sharing Tests were performend with gimp 2.8.10-0ubuntu1 on Ubuntu 14.04, and with gimp 2.8.14 on Windows 7 Home.

* renat [04-19-15 05:26]:

scanned images are going pixelated

Faced the very same (or similar?) problem while trying to resize photos made by digital camera. Very similar "grids" appear on nearly-flat surfaces with a bit of noise (like on sky). Surprisingly, faced the very same grids not only in GIMP, but in some other programs as well (in Image Viewer in Ubuntu 14.04, for example), so there may be a problem with underlying library (GTK?), not with GIMP itself.
Here is an full-size example image:
https://drive.google.com/file/d/0B2nZFc31UFnoNUVTcGpoUHFSbm8/view?usp=sharing And it's resize (width 1920 px) - grid is easily visible: https://drive.google.com/file/d/0B2nZFc31UFnoT1ZvM0xxQW8xSU0/view?usp=sharing Tests were performend with gimp 2.8.10-0ubuntu1 on Ubuntu 14.04, and with gimp 2.8.14 on Windows 7 Home.

Attachments: * http://www.gimpusers.com/system/attachments/195/original/gimp-ubuntu-scaled.png

DL'l your "original" and viewed on openSUSE Tumbleweed in geeqie, gimp and imagemagick (display) and do not see the "grids" even after pushing to 200%. Scaling "method" appeared to make no difference, nor did resolution. My installed gimp package is from openSUSE Build Serice.

kde4 4.14.6 gtk3 3.16.0
gimp 2.8.14

(paka)Patrick Shanahan       Plainfield, Indiana, USA          @ptilopteri
http://en.opensuse.org    openSUSE Community Member    facebook/ptilopteri
http://wahoo.no-ip.org        Photo Album: http://wahoo.no-ip.org/gallery2
Registered Linux User #207535                    @ http://linuxcounter.net

On Sun, 2015-04-19 at 01:26 +0200, renat wrote:

scanned images are going pixelated

Faced the very same (or similar?) problem while trying to resize photos made by
digital camera. Very similar "grids" appear on nearly-flat surfaces with a bit
of noise (like on sky). Surprisingly, faced the very same grids not only in
GIMP, but in some other programs as well (in Image Viewer in Ubuntu 14.04, for
example), so there may be a problem with underlying library (GTK?), not with
GIMP itself.
Here is an full-size example image:
https://drive.google.com/file/d/0B2nZFc31UFnoNUVTcGpoUHFSbm8/view?usp=sharing And it's resize (width 1920 px) - grid is easily visible: https://drive.google.com/file/d/0B2nZFc31UFnoT1ZvM0xxQW8xSU0/view?usp=sharing

Yes, there;s a Moiré pattern in the resized image.

How was the image created exactly? It looks like you photographed a printed image and then used rawtherapee to convert to a jpeg for gimp to open?? If so, the low frequency grid is from the four-colour print process.

I see this pattern in gimp scaling with linear but not with cubic -- see Tool Options.

Liam

Tests were performend with gimp 2.8.10-0ubuntu1 on Ubuntu 14.04, and with gimp
2.8.14 on Windows 7 Home.

Attachments: *
http://www.gimpusers.com/system/attachments/195/original/gimp-ubuntu-scaled.png

On 04/19/2015 07:34 AM, Patrick Shanahan wrote:

* renat [04-19-15 05:26]:

scanned images are going pixelated

Faced the very same (or similar?) problem while trying to resize photos made by digital camera. Very similar "grids" appear on nearly-flat surfaces with a bit of noise (like on sky). Surprisingly, faced the very same grids not only in GIMP, but in some other programs as well (in Image Viewer in Ubuntu 14.04, for example), so there may be a problem with underlying library (GTK?), not with GIMP itself.

[...]

Attachments:
* http://www.gimpusers.com/system/attachments/195/original/gimp-ubuntu-scaled.png

DL'l your "original" and viewed on openSUSE Tumbleweed in geeqie, gimp and imagemagick (display) and do not see the "grids" even after pushing to 200%. Scaling "method" appeared to make no difference, nor did resolution. My installed gimp package is from openSUSE Build Serice.

I see the "grid" in Firefox and the GIMP. Here's a sample, scaled up 300% for clarity:

http://pilobilus.net/xfer/gimp-ubuntu-scaled-detail.png

I have never seen anything exactly like this, so I am confident that it is not an artifact introduced by scaling an image in the GIMP. It looks to me like Moiré pattern noise, smoothed out by some kind of dithering process. Example:

https://en.wikipedia.org/wiki/File:Moir%C3%A9_grid.svg

I would look at the path the image took from sensor to camera memory to hard drive: It seems likely that, some step along the way, a Bad Thing is happening. I would look into saving RAW format files in the camera, for conversion to PNG or some other lossless format in a tool like UFRaw. A few photos of subjects chosen to make that checkerboard likely to appear would show whether that works.

Meanwhile, it's not too hard to fix in existing images like the sample: Load the image into the GIMP, make a duplicate of the base layer, and apply Gaussian blur with a 10px radius. Add a layer mask, fill it with black, and paint on the mask in white to "erase" the artifact by making those bits of the blurred layer visible.

Example:

http://pilobilus.net/xfer/sample-before.png

http://pilobilus.net/xfer/sample-after.png

If the too-smooth blurred regions look unnatural, try running Filters > Noise > HSV Noise against it, with the Hue setting at zero (no color noise) and the Saturation and Value sliders tweaked as necessary to restore a bit of grain.

:o)

Steve

On 19/04/15 22:11, wk_ wrote:

Faced the very same (or similar?) problem while trying to resize photos made by digital camera. Very similar "grids" appear on nearly-flat surfaces with a bit of noise (like on sky). Surprisingly, faced the very same grids not only in GIMP, but in some other programs as well (in Image Viewer in Ubuntu 14.04, for example), so there may be a problem with underlying library (GTK?), not with GIMP itself.

But you said, you tested on Win 7 also. Was there same problem or not?

Tests were performend with gimp 2.8.10-0ubuntu1 on Ubuntu 14.04, and with gimp 2.8.14 on Windows 7 Home.

I tried with your image too and can confirm that Moire appearead (with linear, cubic and sinc methods). It is serious problem, I think.

Moir patterns are due to spatial frequency folding. They can normally be dealt with by blurring the picture before scaling down (in other words, cutting off the high frequencies that won't be in the final picture anyway). Typically you apply a Gaussian blur of X pixels when you scale down by a factor of X.

The frequency response of the various interpolation methods make them sensitive to different frequencies, so without blurring one method can give better results than others, but that doesn't mean it will work every time.

On 04/19/2015 04:20 PM, wk_ wrote:

Steve wrote:

I have never seen anything exactly like this, so I am confident that it is not an artifact introduced by scaling an image in the GIMP.

I don't understand your way to this conclusion. In my original post I have source scan, which I scaled down with 3 different tools (Gimp 2.8, Gimp 2.6 and ImageMаgick). With first one Moire pattern appears, but not with others. How it could be not realated to Gimp 2.8? The process is easily reproducible.

OH - in that case I missed an earlier post. I will go back and see what I can make of it. Making scaling come out "right" vs. being able to fix the noise would be a Good Thing.

On 04/19/2015 10:10 PM, Steve Kinney wrote:

OH - in that case I missed an earlier post. I will go back and see what I can make of it. Making scaling come out "right" vs. being able to fix the noise would be a Good Thing.

That's bizarre. Scaling _MG_4282.jpg to 50% or 25% of its original size did not produce the noise, but scaling it to a width of 1920 px did produce the pattern. Same result with linear, cubic and sinc interpolation. This is what made me believe that there must be some noise introduced by the camera's conversion algorithm, that survives dithering to produce a Moiré pattern when the resulting image is scaled to dimensions close to your target one.

I got the original image you posted, and enhanced the contrast with duplicate layers set to Overlay mode, plus some color curve tweaking:

http://pilobilus.net/xfer/_MG_4282-grain-enhanced.png

I see a consistent rectilinear pattern in the nose, not obvious in the unmodified original. I still think this comes from some interaction of the sensor's noise floor and whatever algorithm the camera uses to convert sensor data into JPG format, with the same possible remedy - save to RAW format if possible and try converting that to a more useable format on your workstation.

:o)

On Sun, 2015-04-19 at 22:20 +0200, wk_ wrote:

Steve wrote:

I have never seen anything exactly like this, so I am confident that it is not an artifact introduced by scaling an image in the GIMP.

I don't understand your way to this conclusion. In my original post I have
source scan,

If it is a scan of a printed document is is to be expected that there may be moiré patterns, even if the scanner is set to apply a "descreen filter".

If that's the case, the "grid" is liable to appear at any time on scaling down the image, or possibly with other image editing operations, both in gimp and in other programs, especially with 8-bit per channel colour. It's a function of the image, not of the software.

I deal with these often in processing scans. You can use a frequency decomposition to remove them, or you can do a guassian blur as I think others have suggested, before scaling down.

which I scaled down with 3 different tools (Gimp 2.8, Gimp 2.6 and ImageMаgick). With first one Moire pattern appears, but not with others. How it
could be not realated to Gimp 2.8? The process is easily reproducible.

It depends on the interpolation method that's used - in a recent gimp 2.9 snapshot I found the pattern appeard with one interpolation method but not another. The default interpolation method I think changes from time to time in different GIMP releases.

Liam

Wbr,

Gunnar

See https://en.wikipedia.org/wiki/Aliasing

So, if you scale down to 51% (which is about 2x) you first do a 2px Gaussian blur. Result here:

http://imgur.com/iZB3lmw

On 20/04/15 09:25, wk_ wrote:

Steve wrote:

I got the original image you posted, and enhanced the contrast with duplicate layers set to Overlay mode, plus some color curve tweaking:

This is not my original post you are talking about, it is already renat's posting to this topic. So for interest of clarity i post my OP again, because there are described simple steps I did:

Hi!

After installing new Gimp 2.8 (on Lubuntu 14.04 64bit) for our employee, she noticed, that scanned images are going pixelated (like moire effect) much more than with previous Gimp when scaling images down. I made a little set to comparision.

This is clip of original scanned (300dpi) image: http://pbrd.co/1CTKhHB

This is how it looks after scaling down (Sinc algorythm) to the 51 percent http://pbrd.co/1CTKHO7

Same procedure with Gimp 2.6: http://pbrd.co/1CTL9vR

And scaled down with ImageMagick: (convert -geometry 292x360 100precent.png 51percent_imagemagick.png) http://pbrd.co/1CTLgYg

I like most the Gimp 2.6 version, it has no pixelation, but black characters have still sharp edges. ImageMagick is also not bad. But Gimp 2.8 version is not for my . What is wrong here? Is there something to make it better?

TIA,

Gunnar

On Wed, 2015-04-22 at 21:11 +0200, wk_ wrote:

Liam wrote:

If it is a scan of a printed document is is to be expected that there may be moiré patterns, even if the scanner is set to apply a "descreen filter".

I feel that there is something lost from my original post. Like it never reached
to the list. I see still it in archive, I point it for reference:

http://www.mail-archive.com/gimp-user-list%40gnome.org/msg08063.html

Have you looked my original scan I linked there?

Yes, I looked at it in detail.

Yes, it is scan of printed
document. But this is not problem. I have more than 20 years experience of
scanning printed docs.

Please remember that we don't know your background.

Problem is, when I use same settings in Gimp 2.6 and 2.8 for scaling down the same sample I got completely different results.

This is probably because the default downscaling method changed. The new method is better for some things and worse for others.

And unfortunately, 2.8 is so much worse. So we must use additional processing before
scaling. I like to have full workflow foolproof and simple, so I can delegate it
whomever I need. Adding additional levels of processing is bad practice in my
environment.

You can change the default downscaling method.

I tried with two different computers (both Gimp 2.8) and got same ugly result.
It seemed unbelievable, that new version may have such comedown and no one has
noticed such behaviour, so I asked here for others experience.

If that's the case, the "grid" is liable to appear at any time on scaling down the image, or possibly with other image editing operations, both in gimp and in other programs, especially with 8- bit per channel colour. It's a function of the image, not of the software.

So, if I take same image and scale it down with 3 different tools (Gimp 2.8,
Gimp 2.6 and ImageMagick), I got 3 different results and it is function of
image??? How?

It is a combination of the image and the tools, of course.

I deal with these often in processing scans. You can use a frequency decomposition to remove them, or you can do a guassian blur as I think others have suggested, before scaling down.

I do, if needed. In 2.6 it was not necessary, generally. In 2.8 it is not
avoidable, that's my problem.

It depends on the interpolation method that's used - in a recent gimp 2.9 snapshot I found the pattern appeard with one interpolation method but not another. The default interpolation method I think changes from time to time in different GIMP releases.

I explicitly used Sinc interpolation in both cases, with 2.6 and with 2.8.

Well, you didn't say that before. However, I do believe the code changed. I think in 2.6 the dialogue was misleading as Cubic was used for downscaling in that case.

Wbr,

Gunnar

On 23/04/15 15:19, wk_ wrote:

So, if you scale down to 51% (which is about 2x) you first do a 2px Gaussian blur. Result here:

http://imgur.com/iZB3lmw

Do you agree that Gimp 2.6 version looks still out much better?

http://pasteboard.co/27ib9Dvx.png

Wbr,

Gunnar

I do agree that this specific implementation/version of the algorithm on this specific part of this specific photo looks better. Period(*)

But the point is, the frequency response of a specific algorithm and its implementation is not linear (especially with potential round-off errors in 8-bit) so you cannot tell in advance if the combination algorithm+image+scaling factor is going to produce artifacts or not. Maybe 2.6 is much worse at another scaling factor... Only the pre-blur will work in all cases.

Most digital cameras have a low-pass filter in front of the sensor to avoid this very phenomenon (hardware blur, so to speak), only a handful of DSLRs work without this (a specific Canon model for astrophotography, and now a Nikon).

(*) An engineer, a physicist, and a mathematician are in car in Scotland, and come across a sheep.
The engineer: in Scotland, sheep have a black head The physicist; more accurately, it seems that in Scotland, most sheep have a black head, but there can be exceptions The mathematician; sorry chaps, but the only thing we can say is that in Scotland, there is at least one sheep that has at least one side of its head black.

Wow, that's a huge tread. Sorry for delayed response. I absolutely agree with Liam R. E. Quin, Steve Kinney, wk_ and Ofnuts - from my POV it's definitely a moire.

There were many questions about real original file and it's post-processing, so here is a RAW file: https://drive.google.com/file/d/0B2nZFc31UFnoc2RxemYwTlVuS28/view?usp=sharing And a TIFF file (converted in RAWTherapee 4.2, default AMAZE demosaic method), where we don't have to deal with all this nasty JPEG artifacts. However, if you'll try to scale it down, you'll see, that moire pattern is here. https://drive.google.com/file/d/0B2nZFc31UFnoMTJLNk9qd2JLV0E/view?usp=sharing

How was the image created exactly? It looks like you photographed a printed image and then used rawtherapee to convert to a jpeg for gimp to open?? If so, the low frequency grid is from the four-colour print process.

No, it's a real landscape shoot by an old Canon 350D. I haven't faced such patterns earlier while processing pics from that camera, so it looks like algorithms were really different in GIMP 2.6.

I see this pattern in gimp scaling with linear but not with cubic

Would be great, if you'll provide exact settings you've used in GIMP - I, of course, have tested all three interpolation methods, but the grid is easily visible in each case.

Meanwhile, it's not too hard to fix in existing images like the sample: Load the image into the GIMP, make a duplicate of the base layer, and apply Gaussian blur with a 10px radius. Add a layer mask, fill it with black, and paint on the mask in white to "erase" the artifact by making those bits of the blurred layer visible.

Thanks for your suggestion, but it's million times easier to prevent this grid from appearing - I have just used graphicsmagick lib in pair with Node.js, and result is more than fine: https://drive.google.com/file/d/0B2nZFc31UFnoNHlTZEJFdWJpb2M/view?usp=sharing That's why I don't think there is some problem with my camera or image processing outside GIMP.

But you said, you tested on Win 7 also. Was there same problem or not?

Yes, I got the very same grid on Win7. And what about your case and Windows?

They can normally be dealt with by blurring the picture before scaling down (in other words, cutting off the high frequencies that won't be in the final picture anyway). Typically you apply a Gaussian blur of X pixels when you scale down by a factor of X.

Yes, I know it, as I'm an astrophotographer, but I'm not sure it's the case for all users of the GIMP. Wouldn't it be good to have at least one (default, of course) interpolation method, which will blur original image itself before scaling it down?

That's bizarre. Scaling _MG_4282.jpg to 50% or 25% of its original size did not produce the noise, but scaling it to a width of 1920 px did produce the pattern. Same result with linear, cubic and sinc interpolation.

In fact, it's not that strange - you can even see similar things by a naked eye, if you'll do some experimentation with, say, two textile nets. It's just how moire pattern appears.

I see a consistent rectilinear pattern in the nose, not obvious in the unmodified original. I still think this comes from some interaction of the sensor's noise floor and whatever algorithm the camera uses to convert sensor data into JPG format

For me that looks like usual noise and JPEG compression artifacts (by the way, level of compression was set to lowest in RAWTherapee). Of course that noise is the source of the problem, but well, real images are never free from noise.

To sum up, this problem is not a problem for me, I've just used another piece of software to downscale my images for my TV. It's more about bug reporting, because I agree with topic starter - such behavior would be considered a regression by many GIMP users. I'm well aware that's it's not always perfect to blur image, but we have different interpolation methods anyway, so it would be good to have one suitable for general public.

ofnuts wrote:
And renat's case show's that this is not only a matter of scanning printed docs.

Moreover, I faced the similar grid while downscaling many other landscape pics with huge flat - but not absolutely flat - surfaces, be it sky hit by noise a bit, or a sea-view with a small waves. So this problem may be pretty common.