Using the FFT filter to remove photgraphic paper texture

[bart_hickman]

BTW, your remaining pattern is manifesting itself as stars along the center axis. The attachement shows the red channel with the contrast enhanced and black marks showing where they were. I'd try smaller dots for actually removing the stars to minimize information loss--I used big dots just to illustrate. The pattern may look different on a full resolution image (I'll check it out tomorrow)--this reduced size image has probably created aliasing.

Bart

[garazon]

Great tutorial Rô! Thanks, I have a few scans I'd just about given up on but now I've got the perfect tool to get them done. I've been playing for the last few days with this thing and it's amazing how much info is stored in just the cental part of the red and green channels. Been toying with manipulating parts of the green channel and was able to increase the contrast/clarity in some images also! (but only very slightly so far, if I get it down to something useful, I'll post back)

Also wanted to share a couple of brushes for PSP 8,9 or X users who might want them so they're attached here.

Thanks again for the tut. Glad I found RetouchPro

[Bmud]

1. Sorry about the confusion surrounding the fft brush. I fixed the file. (finally)

2. Thanks for all the attempts to figure out the circle pattern. We've been using FFT in the way that you suggested in the tutorial for all images where a 150dpi was acceptable because they were full pages. It still worked after a small median filter. In the case of the image I posted on page 4, the image was so small to begin with, well, you get the idea.

3. Now I'm curious how well the FFT would do at de-interlacing. Probably poor. It most certainly did NOT work when I tried to get some blockiness out of bloating an image to 150%. Just a huge grid of stars.

[Cameraken]

Great Discussion Here, Thanks Rô, BMud and Bart.

It’s not surprising that there is a lot of confusion with this filter.

When you go to
http://www.pages.drexel.edu/~avc25/archive.htm#FFT

There are two filters available for download.

2D Fast Fourier Transform plugin for Adobe

There are two versions
The top one
Download new version of FFT/IFFT Photoshop plugin - 235Kb (April 2005)
This is the HSL version

The bottom one
Download RGB version of FFT/IFFT Photoshop plugin - 156Kb (April 2005) (Red and Green channels are used instead of Hue and Light to store complex phase and amplitude information)
This is the RGB version which Rô’s tutorial is written for.

All the discussion in the forum regarding FFT prior to mid 2005 relates to the HSL version. Discussion after mid 2005 relates to the RGB version

Very often we have tried to make the distinction by calling the new version FFT-RGB


Ken

[bart_hickman]

I'm using PSP which doesn't have the RGB channel palette mode found in PS, so I have to use a script that breaks the FFT into it's RGB components in a manner that mimicks the PS channel palette. I was wondering how other PSP users are doing it.

Bart

[garazon]

Hi Bart
Using PSP9 here, and the only way I've found to do it is to split the FFt image into RGB channels, then when done editing, recombine the channels, run IFFT on it, copy and paste as new layer over the original. Setting blend mode to Luminance. Btw, I use Luminance (legacy) mode rather than plain Luminance, looks a bit better to me. It's a couple of extra steps more than what PS appears to use , but not a major thing. I can't see any other way to do it. If you know of something else, please pass it on.

Unrelated to PSP: In regards to using FFT, I was reading up on it's applications somewhere (might have been the Aragon Systems site, I'll have to look it up), I noticed all the examples they and other places used of any application to image manipulation were done on square images, and specfically where the dimensions are multiples of 32 (32, 64, 128, 256.......) and having tried it on a couple of images I can see a slight difference when the FFT technique is applied to a square image. Every little bit helps! I didn't recall seeing this discussed previously, so apologies if it has been.

[byRo]

Originally Posted by garazon

........square images, and specfically where the dimensions are multiples of 32 (32, 64, 128, 256.......) ........ I didn't recall seeing this discussed previously, so apologies if it has been.

No need to aplogize! Yes, we did discuss the subject. You might find the "golfball thread" very interesting.
There, thanks to some great input form Chris (of Reindeergraphics), we went into a few more details of FFT.
The FFT plug-in that we're using here (Alex Chirakov) is easy to use and free. However there is a price - it is less than perfect.
A perfect FFT tool, such as supplied by Reindeergraphics, will work on square, multiples of 2 and the conversion process will not introduce any new noise.
With the Chirakov version there is some noise, BUT seeing as we're usually working with less than pristine images the new noise has un-noticable effects.

Rô

[garazon]

Thanks Rô

I read the thread with great interest, and it answered a lot of questions I had. Also followed chris' link to optipix and downloaded the free custom filter, looks intriguing also. This will all prove to be invaluable when it comes to cleaning up a lot of scans I have made where a paper texture needs to be remove. Results have been remarkable so far

Thanks again!

Chester

[byRo]

Using the FFT process has become (even ) easier!!

Our friend, Stefan has put together an action for the FFT process (7.0 and CS2 tested) which can be found together with the "how to do.." at his site here.

Now with Stefan's action and Bmud's star brush we've got ourselves a pretty good toolkit for removing paper texture from scanned photos.

Rô

[Bmud]

Greetings once again. I've been noticing every once in a while that images I run the FFT on have completely black red channels. This was a mystery to me until today.

Apparently the plugin doesn't like images over a certain size. I've been trying to pinpoint the behavior as to reverse engineer some sizing rules or to make a proper bug report, but FFT i so slow while I'm at work that I'm having no luck even with a binary method. Just losing lots of time on the clock. It might also be a RAM problem.

Somewhere around the thus far, i've gotten..

bigger than 2662x2991
smaller than 2862x3191

I'm guessing that this means that 1024*3 = 3072 is the magic number. I'll do some more tests at home with my dual core and 2gb of RAM and see if I get similar results. I've got a 512 stick here.

[Bmud]

Guess what. Let's say you're like me and you just scanned a 1x2 inch image at 1200dpi and the moire/patterns are ridiculous. After the FFTRGB filter there are too many stars to count, and the significant patterns are difficult to pick up. Good news. There's a handy guide right under your nose to finding the right stars, the "Key Stars" as I call 'em.

Look at your green channel at 50% zoom.

We've been talking a lot about the red channel but not much about the green one. Turns out that dropping a black star where you see a "cup lid straw hole" is right where you need to be. Use you imagination and discretion to figure out if you need to repeat those stars out further by eyeballing the distance to the main star's center or cardinal axis depending on the behavior of your pattern.

Just stay away from putting black on either axis of the main star without being really accurate with your clicking -- it's what creates that crazy black smoke.

By the way, if you're using my star brush make sure that the Flow is set to 100%. I think I have it saved in the preset as 50% which will mess things up. I'd appreciate if someone let me know if it's OK straight out of the cereal box.

[Cameraken]

Hi Bmud.

I could be wrong here but I don’t think you can save the brush opacity or Flow. When I load your FFT brush the opacity stays as it was previously set. I think this happens with all brushes so I don’t think you can save the Opacity/Flow.

Some interesting ideas there could you post an example of a "cup lid straw hole"

By the way, have you seen this thread?
http://www.retouchpro.com/forums/sho...g+stripes+ugly

There are only 2 stars but they are on the axis. I guess that’s what you mean by Black Smoke.


Ken

[Bmud]

This post is just to let you know that I've recieved everyone's comments about my posts being hard to follow. I'm spell checking and revising and attaching screenshots riiiiight nooow.

[bart_hickman]

This might have been mentioned before, but in the interest of completeness, I thought I would mention that making sure the texture aligns perfectly in the horizontal/vertical directions will simplify the spur pattern and thus improve the final result. That means if you have texture pattern that is somehow at an angle with respect to the photo, it will ease the job if you either rotate the photo in the scanner or rotate it in Paintshop/Photoshop prior to doing the Fourier transform.

I've occasionally seen texture that was slightly non-linear. In that case, see if you can warp the photo to make the pattern as linear as possible prior to the FFT. Then remove the spurs. Then undo the warping on the cleaned photo. If the warping is too severe, then the trick of scanning twice at 0 and 180 degrees might be to only option.

Bart

[Taco]

I've searched for a while for something that would remove the moire (or halftone dot) pattern that is found in magazine prints. I think this has a lot of potential. One complaint that I have is that it's not able to handle a 600 dpi scanned image (at least not a full page). Another is the added "noise" that comes with it. I've been considering buying one of the commercial versions, but can anyone tell me if it can handle larger images?

I think a giant leap forward would be if it were automated. Sure, there's a script, but that is only part of it. I'd like to apply filter and process complete. I wouldn't hesitate to pay for a commercial version if this were the case.

Nice site, and thanks for the help.

[bart_hickman]

Seems like just using the median filter (bare minimum needed to soften the pattern), followed by a good NR tool (Neat Image, Paintshop Pro, etc...) seems to work pretty well for that. Sometimes those magazine patterns are very irregular--the FFT will have a ton of harmonics and thus be time-consuming.

Bart

[Bmud]

I found today that FFT really is just fine the original way we were doing it with the FFT and median filter. I split a 600dpi image up today and did the CMYK channels seperately only to find that it comes out more harsh and messes up the colors -- so I had to use a median layer to correct color anyway.

Maybe the trick to getting color halftone removal to look better is redesigning a brush with a linear gradient to help with the circles of different sizes.

I agree though. this i all very fascinating, but I'm about fed up with the work involved. A full blown pattern removal filter would be nice.

If this discussion got any more technical we'd have enough to figure it out ourselves :P

[Taco]

Originally Posted by bart_hickman

Seems like just using the median filter (bare minimum needed to soften the pattern), followed by a good NR tool (Neat Image, Paintshop Pro, etc...) seems to work pretty well for that. Sometimes those magazine patterns are very irregular--the FFT will have a ton of harmonics and thus be time-consuming.

Bart

I've been doing the median filter/noise reduction for a while and it's pretty good. As well as scanning at 600 dpi and reducing to 300 dpi. but I'm always on the look out for something better.

[Bmud]

More recently, with pictures that are big enough, I've just been running the despeckle filter to kill halftone patterns.

[Taco]

not a big fan of despeckle myself, but the new noise reduction filter in psp X has a lot of potential. A friend of mine swears by it to remove the regular pattern without harming the detail. I can't match the results yet unfortunately, but I'm still playing around with it.

Anyway, didn't mean to get this subject off topic. I'm still interested in an automated fft filter such as this. I set up a script is psp that uses the poster filter (set to 4) and then the threshold filter (set to 152) to seperate the brighter stars from the darker background. cut out the center, then color it gray, overlay it to a copy of the fft image and use ifft. the results work and are pretty quick, but it could be very much improved. I left out a lot of details, but it was just something I was toying with.

[bart_hickman]

Here's a handy script I use to split the active layer into RGB components so I can edit them directly. For the FFT, you can paint on the red layer. When you're done, just merge the group and you're ready for the IFFT. The colors are put into groups so it's very easy to quickly add a curve to the red channel to make the stars more vivid while you erase them. Just delete or turn off the added layers prior to merging the RGB group.

In PSPX, it can run in the restricted folder. It won't run in PSP9 or earlier yet--I have that on my to-do list.

http://home.comcast.net/~zumbari/Scr...pRGB.PspScript

Bart

[Bmud]

I just recently found the magic of the new Reduce Noise filter circa Photoshop CS2. I havn't run FFT in quite a while.

[bart_hickman]

Hi Klassy,

Is there a chance you used the wrong FFT? There are two FFT plugins--you want to use the RGB FFT. Which image did you do the FFT on? Can you post the result of the FFT?

Bart

[bart_hickman]

Your attached image is too small for me to see the texture. Could you post either a larger size or a link to the original?

Bart

[bart_hickman]

The FFT technique works on pattern noise. Pattern noise is concentrated at a few discrete frequencies so when you do a Fourier transform on the image, the pattern noise shows up as a star pattern--the stars are the frequency-domain representation of the noise. (The technical term for noise like this is spurious noise and the stars are called spurs.)

In the image you attached, it looks like a combination of aliasing (due to image size reduction) and jpeg compression has caused the pattern noise to be converted into essentially random noise and thus no longer removable with the FFT.

If you have textured photos, you want to be sure to scan at a high enough resolution to not just capture the image details, but capture the details of the texture as well.

Bart

[videosean]

Originally Posted by Bmud

Apparently the plugin doesn't like images over a certain size. I've been trying to pinpoint the behavior as to reverse engineer some sizing rules or to make a proper bug report, but FFT i so slow while I'm at work that I'm having no luck even with a binary method. Just losing lots of time on the clock. It might also be a RAM problem.

Somewhere around the thus far, i've gotten..

bigger than 2662x2991
smaller than 2862x3191

I'm guessing that this means that 1024*3 = 3072 is the magic number. I'll do some more tests at home with my dual core and 2gb of RAM and see if I get similar results. I've got a 512 stick here.

Your post is getting close to what my problem has been I think. I have a scanned image that is 2878x3996 and I get a Red channel that is nothing but black (flat, 0 grey value) after running FFT-RGB and a Green channel with alot of noise and a cross pattern but no stars and the Blue channel is a flat 128 grey value.

I resized my image so that it was 2253 x 3072 (using your 3072 as the largest dimension - I had also added a 50% grey border before downsizing as per other user tips in this thread) and got the results I was expecting... stars and whatnot For a free filter I guess this is acceptable. The texture I'm trying to remove is a fairly typical mid to late 1970's matte paper surface from a portrait studio. The color in the image is completely faded/shifted to red/orange and there isn't any kind of watermark on the back of the print nor a studio's name on the front.

edit - because I'm lazy, I used the magic wand to select all the stars for fixing.

1. Magic wand tolerance = 8, uncheck contiguous, click in the middle of a star to select the brightest point.
2. Use the lasso tool, deselect all the areas you don't wish to correct - the large cross, for example, I deselected as well as a few small points that didn't look like stars.
3. Select -> modify -> expand by 5 pixels
4. Image -> Adjust -> Brightness/Contrast, I just reduced the contrast all the way down and the made a slight adjustment to the brightness so it looked like it matched.

This image has been a personal project that I've left sitting on my HD (it's been maybe 2 years since I scanned this image LOL) waiting for the day I found a good way to kill the paper's texture. Thanks RetouchPro and Alex

[sattva]

Originally Posted by Taco

I've searched for a while for something that would remove the moire (or halftone dot) pattern that is found in magazine prints. ... I'd like to apply filter and process complete. I wouldn't hesitate to pay for a commercial version if this were the case.

Taco,
There are a special Descreen filter to eliminate the halftone dots (the screen) and prevent the moire pattern:
http://www.descreen.net/eng/soft/descreen/descreen.htm
This plug-in employs the FFT.

Attachments:
1) The original image.
2) The same image after screen removing. You may see the prominent halftone moire pattern.
3) The same image after screen removing and moire reducing.

[chiko321]

I had come across another technique to remove the texture if you are having trouble getting software to work. It is quite simple. If you think of the texture that you see as a result of shadows and highlights on what is basically glossy paper, all you need to do is 1 of 2 things. Get a scanner that has dual bulbs, that is to say, one before and after the scanning head. If you can't find a scanner with dual bulbs, you can also do a neat trick by scanning the photo upright and then scanning the photo again (precisely 180 degrees) and scanning the "shadows and highlights" in reverse. Once the 2nd scan is rotated upright and aligned to the first scan at 50% transparency, pretty much all of the texture can be eliminated. This works best if your scanner head does not have any problems with variable skewing, otherwise you'll be "fudging" and stretching different zones of the photo to blend well with the first scan. In the meantime i'll definitely be looking into this other app as time allows! It could be a great timesaver! Especially if you know how to create actions or JSX scripts for CS3!

[oppor2nity]

First of all, thanks for posting this tutorial, as it seemed to be the solution to my current problem, and most likely will be. Alas, right now I cannot make this work.

No matter how I try, I can't achieve the final result as shown in the tutorial.

I'm starting out with an RGB8 scan which is yellow-ish as has a heavy embossed dot pattern all over the image. I have also tried desaturating this first, but that hasn't changed anything.

I copy, apply FFT, select the red channel only, see the stars and paint them out. I apply IFFT and end up with an image in which the pattern seems to be not gone, but more like cancelled out. The problem is, it resembles a negative. I drag that over the original and change the blending mode to Luminosity, but no colour can be seen through the greyish layer.

Where am I going wrong?

Cheers,

/Ulf

[chillin]

You could use ImageJ instead. It requires less steps.