FFT and Golf Balls

[Cameraken]

This photograph is a small part of a long panoramic group photo.
It has been scanned following all the suggestions / tutorials on this site and this is the best I can get.

I am not asking for anyone to restore this but I hope to get some suggestions on the golf ball texture.

I think Neat image blurs this picture to much.

FFT which has always worked well for me in the past has let me down with this picture
When I run FFT I can get rid of the grain but I am left with lines on the print which is just as bad.

I suppose my question is
Does anyone understand the Stars that FFT gives? I usually remove all except the big one. Is there perhaps a way of removing some of the stars to avoid getting lines on the picture? When I remove just the 4 stars nearest the big star I seem to get the best results from this picture.

This picture clearly needs some Level / Colour / Sharpen adjustment but this also improves the golf balls so I’ve posted the best original scan.

I did complete this image by eliminating every golf ball on the face with the Clone Stamp and my friend is happy with the finished picture. I’m just hoping to find some suggestions for an easier way next time.

[Gary Richardson]

Did you scan this, or was it given to you as is.

If you have the original to scan, try scanning once, then rotate image by 180 degrees and scan again.

In photoshop, copy both scans to seperate layers.
Set blend mode of top layer to either lighten, or darken (whichever gives best results). Because the shadows caused by the bumps are different, due to the change in scanned lighting, they cancel out, and the bumps should be less noticeable.

Hope this helps a little.

[Flora]

Hi Cameraken,

Welcome to RP!

The result I got with FFT was acceptable ... It is a bit more blurred than I expected but the resolution of the image posted is very, very low ....

[Cameraken]

Gary this is scanned from about ¼ inch of the original print. I followed all the tutorials about rotating the image and setting blend modes etc.
This image is already a combination of that process.


Flora.
You have done a wonderful job.
I had to increase the jpg compression to get the file uploaded to retouch pro (100kb max)
I tried again on the compressed file and I still don’t get the results you have.
I am still getting lines across the image.
See attached image


All
If anyone reading this is interested I will move this to a ‘Reverse Tutorial’
As it seems that only a few people here understand it (Duv, Flora,Byro)
Can someone can please tell me where to post it?

The best information I found was here
http://www.retouchpro.com/forums/sho...59&postcount=6
But following this still does not give me the results Flora got.

I think this texture removal is really important to making good images as many old photographs have texture.

FFT has worked OK for me in the past and I can’t understand why I am having problems this time

[Duv]

You might have some success trying the following:
Make 3 copies of your picture. Adjust levels.
On the first, select a small portion of the pic and clean it up by cloning out the balls. Go to Filter:Pattern Maker and click on Generate. Save the pattern with the button at the bottom. Cancel out of Pattern Maker. Go to Healing Brush and set it to pattern. Pick out the one you created and heal out the balls.

On second copy, run FFT. You should end up with a B/W image. Shift drag onto Healed image, change to Luminosity and change opacity to suit. High Pass sharpen and fade back in Soft Light.

Cheers
Dave

[jcr6]

Since I sell FFT code, Doug gave me a head's-up on this. The UI for my software is completely INAPPROPRIATE for photography or people that don't want to know about FFTs, and a separate discussion will need to be held over what you'd like for tools (since I'm a plugin writer).

In any case, here's what I was able to do with removing the pattern noise from the Power Spectrum manually (attachment).

I have not attempted to go after the random noise in the image and there is evidence of the top layer of the photograph flaking off and yellowing in the blue channel.

This should give you an indication of what's possible.

-Chris Russ

P.S. there is a 3-week demo of the FoveaPro software available if anyone wants to play with it. I do NOT expect any of you guys to buy it -- you're not the intended audience. If anyone is interested in playing I'll be happy to provide a URL for download.

[jcr6]

And, I suppose I could write up a discussion on what you're actually seeing when you look at a PowerSpectrum (the "FFT Stars" that someone was referring to, above).

Gluttons for punishment.

[Duv]

Thanks for the info Chris. Personally, I'd like to play with the FoveaPro if you could provide the URL.

Also, do you have any tips on how much of a star to remove. What I mean is that often, horizontal and vertical lines emanate out from the centre of the star, sometimes for quite a distance. Is it recommended just to remove the centre star or should we be cloning out the lines as well? Or is it, try it and see what happens?

Cheers
Dave

[jcr6]

Demo for FoveaPro can be downloaded from:

Mac Version
PC Version

(Runs for 3 weeks.)

There is a tutorial on using it for Fourier processing at:

Chapter 4 - Processing in Frequency Space

But rather than make this a blatant ad for a product that was designed for a different group of people than y'all, let me instead talk about reading Fourier transforms and some of the things you can look for in them.

#1 - The Fourier Transform. We convert an image (no loss of data) into a summation of sine waves. Each possible sine wave has a strength and those are plotted in the Power Spectrum. This is a map of the direction of the different sine waves and their period/phase -- longer sine waves are in the center of the plot (DC is the center and represents the mean value of the image) with shorter sine waves around the outside.

The center cross-shaped radial lines are a combination of the extra frequencies that it takes to wrap around from the left side of the image to the right side of the image (vertical part) and the top side to the bottom side of the image (horizontal part).

Here is an example Power Spectrum:

http://reindeergraphics.com/tutorial...fig_4_13_2.jpg

In it you can see the center cross (caused by the edges of the image not matching each other) and a spike in the center which corresponds to the overall image brightness and the gross gradient in the image. As we come out further from the center, the amount of power starts to back off. (This plot shows power as black so the "background" will tend to lighten as we come further out radially from the center.)

You will also notice a number of other spikes. These are caused by a halftone pattern in the original picture:

http://reindeergraphics.com/tutorial...fig_4_13_1.jpg

If we create a mask to eliminate just those spikes and apply it to the Power Spectrum (like this one):

http://reindeergraphics.com/tutorial...fig_4_13_3.jpg

We get the following result:

http://reindeergraphics.com/tutorial...fig_4_13_4.jpg

Actually, to get a very good result, and to blur the image some, we also apply a low-pass filter (in this case a Butterworth) and make the Power Spectrum look like this:

http://reindeergraphics.com/tutorial...fig_4_13_6.jpg

Notice that it now lightens up a lot toward the outside edges and there are holes where the spikes used to be.

Here is the result image:

http://reindeergraphics.com/tutorial...fig_4_13_7.jpg

That low-pass filter actually filled in the gaps in the original halftone dots. There are a host of other filters that can be done, but this should give you a small idea of what's possible.

It turns out that you can do very high speed convolution in this mathematical space (especially with large filters), remove pattern noise, make some measurements for periodic shapes, determine the real vs. empty resolution of an image, construct really interesting filters (like this one), and otherwise confuse people.

Enjoy.

P.S. When I was a college student in the early 80's we were challenged to perform a periodic noise removal problem. At that time (less politically correct than today) Cheryl Teigs had posed in a fishnet bathing suit that left little to the imagination. Since that suit was fishnet, and thus periodic, we were challenged to remove the suit. (Obviously there were no women in the class.) Needless to say, the technique works.

[jcr6]

You asked whick parts of the "FFT Star" or the cross to remove. I would say NONE. That cross represents the frequencies that it takes to do the step from the top to the bottom and the left to the right -- if you remove it, you'll mess up your image up significantly.

Now, if there is a spike sitting on top of the cross, you can remove THAT. And, lines that are not at the 12, 3, 6, and 9 o'clock positions that do not run through the center of your image you can safely remove.

Other stuff might be problemmatic. Perhaps you should post an image and we'll deconstruct it.

Also, let's start with b&w. Color information can be different in the three channels, so lets deal with one channel at a time.

[philbach]

Well I ran the median filter at 3 then Neat Image on the picture.

[byRo]

I couldn't let another FFT thread pass by without sticking my nose in....

Originally Posted by Cameraken

FFT which has always worked well for me in the past has let me down with this picture ...... FFT has worked OK for me in the past and I can’t understand why I am having problems this time

FFT is a mathematical process - as you can see by jcr6's post. It is great for detecting well-defined repetitive frequencies in an image.

To our eyes the "golf balls" seem to be exactly repetitive and using FFT you will be able to remove all the repeating frequencies. What you have got left after that is noise that was superimposed on those frequencies.

Although we can see some sort of pattern in this noise, it doesn't repeat any more - there are some white dots, then a few black ones......

Why is this one different? Because other paper textures are usually less embossed than this one and don't pick up so much dirt (=noise).

Originally Posted by Cameraken

The best information I found was here
http://www.retouchpro.com/forums/sho...59&postcount=6

I left a note there advising that there is an easier version available now which doesn't need the HSB stuff.

For those (few, OK) mathematically inclined.. there is actually a way to calculate where the stars will be in the FFT. Something like.. measure the pixel distance between the texture ridges, divide the total image dimension by this number, then count out from the centre. I did have it figured out once and it did work.

Rô
(Yeah, I know - 1.496Km to go)

[Ken Fournelle]

Duv,

Can you come up with a report on the FoveaPro at some point. Something for non-numbers people to appreciate? Ro?

k

[byRo]

OK, I'm on it
But it's a 22MB download - so only tomorrow.

Rô

[jcr6]

I never intended for this to be an ad for Fovea -- it is a general purpose image analysis system for scientific imaging.

Now, some of the things in there (like the Fourier suite) *might* make an interesting product at a more acceptable price, but I need to see what you guys want/need/are willing to learn.

My basic inclination is to completely hide the FFT under-the-hood, whether for deconvolution (Optipix->Refocus), using an ideal inverse (a very neat tool for image enhancement), halftone removal, etc.

The learning curve, for one, is huge. Two, the Fourier suite works in b&w. Three, there are 177 plugins in Fovea. Four, the "tutorial" is 500 pages.

Then, again, it might be interesting to see how he feels after going through it. <evil laughter>

[Doug Nelson]

That's how Chris and I first connected. A couple of years ago I was looking for focus and pattern solutions and came across Fovea Pro. We exchanged some emails and I even printed out the manual and made an attempt at figuring it out. But it was written from the perspective of a place I've never been (technical image analysis), and I was quickly in way over my head.

But I do feel there is a valuable restoration tool waiting to be designed, here. Something without painted stars, with sliders (and a realtime preview). Perhaps something that would internally rank the patterns and let you page through them ("no, don't remove that one, yes, remove that one and that one, but leave that one").

The Cheryl Tiegs assignment was inspired

[Ken Fournelle]

My guess it already exists, in Langley, Virginia, USA

k

[byRo]

Originally Posted by jcr6

Then, again, it might be interesting to see how he feels after going through it. <evil laughter>

I'm a kid in a candy store.

(I had already read through most of your site, and numbers and formulas don't scare me - I almost majored in maths)

Thanks,

Rô

[jcr6]

Phil showed an example of using a radius=3 median filter to remove stuff. If you think about a median filter, let's start with the simplest radius=1 (3x3 neighborhood), it will remove lines that are 1 pixel wide or clusters that are less than 9/2 =4.5 pixels. So a block that is 2x2 will be completely removed as will the thinnest of lines.

As this neighborhood radius gets larger (he used an example of radius=3 - 7x7 neighborhood, lines that are 3 pixels wide or narrower will get removed, blocks that are less than 49/2 or 24.5 pixels will be removed, and corners on objects will be rounded off. This is why you were able to remove the pattern with a median filter. It is a really harsh filter, though, in that it removes all kinds of useful stuff.

There is a variation on the median called a "Hybrid Median" that is nowhere near as harsh. We've implemented it in FoveaPro, but also in Optipix as "Safe Median." Basically it is much more computationally intensive, but can preserve lines and corners. It still does a good job at removing noise (and, regretfully, fine detail).

We probably should have a discussion about noise removal and what is noise and what isn't. It might be fun to write a couple of plugins for you guys to play with to see what we can learn. (I like having guinea pigs!)

-Chris

[Cameraken]

To Philbach
Thanks for the suggestion of median filter. This is quick and easy and has removed all the texture but has sacrificed detail (look at the eyes)

Duv
You have retained all the detail but have left a little texture (or is this noise byRo)
If FFT had worked perfectly your first step would not be necessary
Your image has more contrast than Jcr6

Flora
Your picture still looks good. You have removed all the texture

Jcr6
Thanks very much for your excellent contribution.
Your picture is great. Although there is still a little texture/noise you have retained all the detail (not that there is much in this image)
Duv has got a similar result but with a little more contrast
I’ve downloaded FoveaPro trial and will read Tutorial5.pdf

What you have written here has already taught me a lot

OK here are my new steps

1) Set background colour to grey 127,127,127 (for Padding)
2) Image > Canvas Size > 512,512 in Pixels (In this case - square with dimension equal to an exact power of two)
3) Image > Adjust > Desaturate
4) Filter > FFT RGB
5) On the Red Channel > clone (using Darken) out the Stars except middle one
6) Blur step here? (Butterworth?)
7) Filter > IFFT RGB


Is this OK? My results Have improved

I have a few Questions
I am (now) Using FFT_RGB_PlugIns_13April_2005. Is this the version everyone else is using? And is this the correct version?

In http://reindeergraphics.com/tutorial...fig_4_13_6.jpg
The corners are missing. Is this significant?

Step 6) Any suggestions?

You suggest starting with a B&W image. Do you mean Greyscale or RGB>Desaturated. My Fourier Transform Folder is unavailable with a Greyscale Image


In the UK Ilford used to make a paper called Velvet Stipple. This was one of the most textured papers available. Does anyone have a sample for Jcr6 to deconstruct.

Ken

[jcr6]

I have a few Questions
I am (now) Using FFT_RGB_PlugIns_13April_2005. Is this the version everyone else is using? And is this the correct version?

In http://reindeergraphics.com/tutorial...fig_4_13_6.jpg
The corners are missing. Is this significant?

Step 6) Any suggestions?

You suggest starting with a B&W image. Do you mean Greyscale or RGB>Desaturated. My Fourier Transform Folder is unavailable with a Greyscale Image

In the UK Ilford used to make a paper called Velvet Stipple. This was one of the most textured papers available. Does anyone have a sample for Jcr6 to deconstruct.

As you may have guessed, I'm not using FFT RGB... Where do I get it?

The corners are "missing" but aren't really missing. It turns out that they're considerably lighter because they've been reduced so much in power. This is a side-effect of the Butterworth filter. The banding in the image is a result of compression.

The reason I suggested B&W was because there is a different signature in the different color channels -- this is more evident if you're looking at a scanned image from a magazine where different colors have different screen angles.

You might also try scanning a blank piece of Velvet Stipple and performing an FFT on that -- this should give you your definitive map of bad frequencies.

Fovea works on one channel at a time.

[byRo]

Originally Posted by jcr6

As you may have guessed, I'm not using FFT RGB... Where do I get it?

This is the Alex Chirakov filter, you'll find it here.
It's called RGB because the earlier verison encoded the FFT information into HSB, now it´s a lot easier because the FFT is in the Red and Green channels.

OBS: (for jcr6) it will take any size of input image

Rô

[Cameraken]

byRo

Thanks for answering jcr6. I was just about to reply
When I follow your link there are 2 Options
Option 1)
Download new version of FFT/IFFT Photoshop plugin - 235Kb (April 2005)
Option 2)
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)

I have downloaded the second option which seems to follow what you have been saying. I this correct?

It is important here that we are all talking about the same filter.

I had added steps 1 and 2 (and what is now 4) because of what I read from jcr6. But from what you are saying they are not needed with Alex Chirakov filter (and I did not notice any improvement from this)
So. New Steps

1) Image > Adjust > Desaturate
2) Filter > FFT RGB
3) On the Red Channel > clone (using Darken) out the Stars except middle one
4) Blur step here? (Butterworth?)
5) Filter > IFFT RGB



Proxy has submitted what looks like Velvet Stipple Here
http://www.retouchpro.com/forums/image-help/10633-photo-paper-texture-problem.html

I was a Professional Photographer for over 20 years and have copied hundreds of photos. Before digital this was easy I put a print on my copy stand got my trusty H/Blad and made copies without worrying about texture as my lighting eliminated it. The problem is caused by scanners which have directional lighting. Why can’t they invent a scanner with a light around the sensor?

Ken

[Caitlin]

Sorry for being a bit dense here - but can I just confirm something that is probably bleeding obvious:

6th step - overlay with original image set to color mode? Or is there some other way of getting the colour back?

[Gary Richardson]

Ro, tried downloading the RGB version, but can't open it. (its a rar file, can't find any info on it).

[Cameraken]

Caitlin
Put the repaired B&W layer above your colour layer and set blending mode to Luminocity (thanks Flora)

Gary
A rar file is a compressed file you will need a program called Winrar to open it
There is a demo available here
http://www.rarlab.com/download.htm

[byRo]

Gary, didn't quite understand if you don't know about RAR files or if this one was corrupted.

Anyway - general instructions for all...

- You will need WinRar to open the ".rar" files, homepage here. It's shareware, but doesn't stop working after the time limit (just nags).

- If it's the second, better download again.

- When you've got it open in WinRar, open the "bin" folder and extract the two ".8bf" filters to your PS filters folder.

Rô

[jcr6]

I just realized that you're only correcting the brightness information, not the color channels.

(One of those "Aha!" moments!)

In order to do this correctly, you really have to correct ALL THREE CHANNELS. And, there can be a different pattern in each. In the case of the "Golf Balls" they're refracting light, so the PowerSpectrum for each (R, G, B) will be different.

If you don't do this to all of the channels, your colors will be screwy -- and will retain the golf ball pattern.

[byRo]

Morning, Camaraken - beat me by three minutes!

Originally Posted by Cameraken

6th step - overlay with original image set to color mode? .......Put the repaired B&W layer above your colour layer and set blending mode to Luminosity

They are in fact exactly the same thing, the Lumosity is more efficient because you don't need another layer.

Originally Posted by Cameraken

1) Image > Adjust > Desaturate

I wouldn't do this. When you do the FFT you will be working on a greyscale image (in RGB mode). After the FFT process, as you rightly said, you mix in Luminosity. It would be much better then if you separated the Luminosity and not Desaturated (they are very different). Then you process FFT and mix back in the clean Luminosity.

To get Luminosity: New Layer (<Ctrl><J>), Edit>Fill - 50% grey, color, 100%. (There are other ways, this is the quickest)

Originally Posted by Cameraken

4) Blur step here? (Butterworth?)

Don't think this would help any.

So summing up...
1) Duplicate original image, right click on title bar (that's a new image, not just a layer);
2) On the new image, get Luminosity (as above);
3) Flatten all, <Alt><L><F> (FFT does not understand layers);
4) Run FFT RGB;
5) Clone out stars in the Red Channel;
6) Select all channels again (easy to forget this bit), and run IFFT RGB;
7) Slide this layer on top of the original image (with <Shift> pressed, to align);
8) Set this new layer's mode to Luminosity.

Rô

[byRo]

Originally Posted by jcr6

In order to do this correctly, you really have to correct ALL THREE CHANNELS. And, there can be a different pattern in each....... If you don't do this to all of the channels, your colors will be screwy -- and will retain the golf ball pattern.

The theory is solid, OK. But in practice you'll be doing more than three times as much work (including channel separation and recombination) and the difference in the final result jsut isn't worth the trouble.

I did it and, in fact, I found that the Luminosity way gave a slightly better looking result (but then again, I may be biased because I expected it to)

[Geekstuff]My litte theory - The eye responds much better to luminosity than to colours (JPEG compression reserves 4 times more space for Luminosity information). If you do FFT separately in each channel then you run the risk of introducing Luiminosity artifacts that weren't in the original[/Geekstuff]

Rô