Scan Correct Color Smooth Texture Reduce Noise Remove Dust Repair Crop

Some photos are embossed (or imprinted) with regularly spaced patterns. Typical patterns consist of rows of thousands of very tiny dimples, all the same size and shape. Scans of dimpled photos will contain patterns embedded into the images matching the pattern of the photos.

An FFT noise filter is the best way to remove embedded patterns in scanned images. This kind of filter uses a Fast Fourier Transform algorithm which is used for a variety of applications including signal and image processing. For images, the FFT filter can be used to remove periodic noise (such as patterns). Periodic noise can be found in scans of halftone images found in newspapers and magazines, as well as scans of photos printed on embossed photo paper.

Most FFT noise filters work similarly. The FFT filter displays a box containing thousands of faint pixels plus a pattern of very bright spots or stars. These bright spots and stars represent the unwanted pattern of the textured print. Painting over the spots and stars and clicking Apply in the FFT filter removes most of the pattern.

FFT noise filters are available as Photoshop plugins. Also, Affinity Photo has an FFT noise filter (called FFT Denoise) as a standard feature.

Looking for a Pattern

There are several pages on the web that have instructions on removing pattern noise with FFT noise filters. These instructional webpages explain how to use dots or crosses to mark out the stars, but some of the finer details such as brush hardness or brush size are not discussed in detail. I wanted to test different settings to find out the best values to use for the brush.

Three photos in my collection had an extremely strong pattern from deeply embossed dimples in the photo paper. These photos were taken at different times and printed on different kinds of photo paper. The following table represents results of my preliminary testing with these three photos.

For comparison purposes, I used a control group of images with pre-determined settings and techniques. Settings I used for my control group include a brush hardness of 100 and a brush size of 10. Techniques for the control group include using a dot to mark stars and marking all stars indicated by the FFT filter. The FFT filter was completely zoomed out to see and mark all of the stars.

For each experiment, I varied a single setting or technique. I then compared each experimental result to the control result to find out how settings were affecting the FFT filter results.

Without any further mumbo jumbo, this table shows how settings & techniques affected the FFT results for the three test photos:
Setting "army" photo "cash" photo "feathers" photo
hardness=80
(instead of 100) 		very slightly worse for solid white areas no significant difference no significant difference
brush size=8
(instead of 10) 		very slightly worse noticeably worse on skin and other solid areas of color slightly worse in certain small areas
brush size=12
(instead of 10) 		very slightly better whites and sunlit neck just barely better no significant difference
using crosses
(instead of dots) 		much better than dots for texture removal much better texture removal noticeably better, especially the face
marking only central stars
(instead of all stars) 		a very tiny insignificant change in the pattern noise a very tiny insignificant change in the pattern noise a bit worse for background and part of the skin

before FFT filter


after FFT filter


paper deselected,
then FFT filter

Drawing a Blank

All photos were improved after any use of the FFT noise filter, but I noticed some pattern-free areas for two of the three photos developed a faint pattern. These areas consisted of bright solid colors such as white walls or sunlit skin. See the before and after images to the right. The white paper developed a definite pattern.

The two affected images were retested by deselecting the areas prone to developing a pattern before using the FFT filter. The retested images came out great. The deselected areas remained pattern-free while the selected areas had most of their pattern removed. See the image to the far right for the results of the retest.

So, if you use an FFT filter on a photo and this particular problem occurs, simply undo the FFT filter, deselect the problem areas, and use the FFT filter again.

Observations for the FFT Noise Filter

Brush hardness: I tested a hardness of 80 since Affinity Photo defaults to it. Using a brush hardness of 80 was either about the same or very slightly worse than using a hardness of 100. Since it doesn't take any extra time to use a brush with a hardness of 100, I recommend 100.


using dots
using crosses
Dot size: Using a larger dot size was generally slightly better than using a smaller dot size. More testing is needed to determine the best size.

Marking with a dot or a cross: Using a cross instead of a dot was much better for removing texture for all 3 test photos. See the example photo to the right for an example of how much better crosses can work.

Marking only the central stars vs. marking all stars: This was a test to see if marking just the stars near the center would be enough to remove the pattern. Interestingly, marking only the central stars was enough for two of the test photos, but the third test photo was noticeably a bit worse when only the central stars were marked. So it appears to be safest to take the time and mark all the stars before using an FFT filter.

Family Photo Feud: The Dots vs The Crosses

This test should determine the optimal size and shape of the brush. Ten photos were used in the test, taken at different times with various cameras and photo paper. All the photos had regularly spaced dimples in rows or columns.

Three different size brushes were used for both dots and crosses, for a total of six test images per photo. For each photo, the six images were compared with each other and sorted by quality of results.


size 10 dots


size 6 crosses, various shapes
The shorthand I use for results is as follows: So a result of "x10=x8=x6 > d10=d8 >> d6" for a hypothetical photo means: (1) Crosses worked just slightly better than dots; (2) The size of the cross didn't matter; (3) Dots of size 10 were no better than dots of size 8; and (4) Dots of size 8 worked noticeably better than dots of size 6.... In short, a cross of size 6 worked fine for this hypothetical photo.
Photo Result Explanation
Army
(test 1) 		x10 > x8 > x6 >>> d10 > d8 > d6 Most of the photo was massively improved by covering stars with crosses. However, the sunlit neck and solid white areas (which had no pattern before) developed a faint pattern, especially when using dots. Using crosses of size 10 worked best for these particular areas.
Army
(test 2) 		x10=x8=x6 >>> d10 > d8 > d6 This is a second test of the same photo, but the pattern-free areas that developed a pattern in the first test were deselected before using the fft filter. The sunlit neck and solid white areas were deselected and no longer developed a faint pattern after the fft filter. Crosses of size 6 now worked fine on the entire photo.
Buddy x10=x8=x6 > d10=d8=d6 A cross of any size worked best.
Cash
(test 1) 		x10=x8=x6 >>> d10=d8=d6 Using crosses worked vastly better than using dots for most of the photo. However, solid white pattern-free areas developed a faint pattern, especially when using dots.
Cash
(test 2) 		x10=x8=x6 >>> d10=d8=d6 This is a second test of the same photo, but the solid white areas were deselected before using the fft filter. After using the fft filter, white areas no longer developed a faint pattern. Also, using crosses still worked vastly better than using dots.
Feathers x10=x8=x6 >> d10=d8=d6 Using crosses worked noticeably better than using dots.
Flower x10=x8=x6=d10=d8=d6 Even a small dot worked fine.
Fur x10=x8=x6 > d10=d8=d6 There was a slight difference between using a cross and a dot, and the difference was only in the face. Otherwise, dots worked as well as crosses.
Picket x10=x8=x6=d10=d8=d6 Even a small dot worked fine.
Poodle x10=x8=x6 > d10=d8 >> d6 A cross of any size worked best.
Snowtree x10=x8=x6=d10=d8=d6 Even a small dot worked fine.
Zorg x10=x8=x6=d10=d8=d6 Even a small dot worked fine.
For all 10 photos, when the pattern-free areas prone to developing a faint pattern were deselected before using the fft filter, then nothing worked better than an appropriately shaped cross of size 6.

Common Sense Conclusion

Use an appropriately shaped cross using a brush size of 6. For bright stars with long vertical and horizontal lines, you may need a cross shaped like a plus sign. For smaller stars, a nubby little cross may be enough. For dim spots far from the center, a single dot might do.

If pattern-free areas develop faint patterns after using the fft filter, simply start over and deselect problem areas before using the fft filter again.
Don't Forget to Dot Your Spots and Cross Your Stars!

Who Cleans First? The Chicken Scratches or the Egg Bumps?

When removing textures from photos, it is necessary to use both an FFT filter (to clean pattern noise) and a noise filter (to clean random noise) to clean a photo. But which one should be used first?

This is a test to determine whether it's better to use FFT filters before using noise filters, or vice versa. Five photos with regularly spaced patterns of dimples were used. Two different noise filters were tested, both before and after using an FFT filter. Results are below.
Test army buddy cash feather fur
FFT first vs. Reduce Noise first FFT first much better about the same FFT first much better FFT first slightly better FFT first slightly better
FFT first vs. Topaz DeNoise first FFT first much better FFT first much better FFT first much better FFT first much better FFT first much better
Evidence is overwhelming. It is far better to use an FFT filter before using a noise filter. It appears that the noise filter changes the embedded pattern in the image too much for the FFT filter to work correctly.
Dump the dimples first. Then nix the noise.

Dimples are for Golf Balls

Descreening in scanning software is all or nothing. With an image editor, you can be more selective with which areas to filter, and what filters to use. This example demonstrates the difference between using descreening options in scanning software vs. using FFT & blur filters in image editors. These images are from a 600dpi scan of one of the most heavily textured photos in my collection. The images along with my subjective opinions are in the table below.
1: before

Looks horrible.
2: Epson Scan, General descreening

Looks better, but dimples are still noticeable on wall, and eyes are a tiny bit blurry.
3: Epson Scan, Newspaper descreening

Dimples are much less noticeable, but hair is slightly blurry, eyes and face are slightly blurry, and vertical lines in door trim are lost.
4: Affinity Photo FFT Denoise

Easier to see edges of objects (such as vertical lines in door trim), and almost completely free of pattern noise, but plenty of random noise and discoloration left.
5: Affinity Photo FFT + Topaz DeNoise

Almost all noise (dimples, texture, discoloration) gone, eyes and trim lines still sharp, but hair is slightly blurry and a few small spots have very faint lines leftover from the pattern noise.
1: before

(shown here again for easier before/after image comparison)
I think she looks better without dimples, don't you?