Smoothing Large Blotches

[kthurner]

Hello,

I have a problem with some very indistinct blotchy discoloration in a clear sunset sky which has resisted my attempts at correction. The image is a 100 MB file so everything including these blotches, is very large in pixel terms. Using various blur techniques has not yielded good results primarily because of the large pixel dimensions of the blotches. I end up with waves of discoloration instead of blotches of discoloration. I can easily isolate the sky, and the best I've come up with is to resample down drastically, add Gaussian Blur, then resize back up with stair interpolation. But this yields some grid pixelization in the sky. Perhaps there's a better way? Here's a link to a small crop of the image: http://www.pbase.com/image/10142424/original

Scrolling the image or emphasizing with Levels will more clearly show the discoloration.

All suggestions are greatly appreciated.

[jeaniesa]

Hi kthurner - and welcome to RetouchPRO!

My first thought when I adjusted the levels was - why don't you turn this into a painting? Just joking, but the blotchiness looks like one of the effects we try to get when we're turning a photo into an artistic rendition.

More seriously, when I looked at your image at 100% and severely adjusted the levels, the blotchiness looked like JPEG artifacting to me. At less than 100%, it looks like clouds or variations in the atmosphere, but you indicated that it was a clear sky. You also mentioned that this image is 100MB - which would indicate no JPEG compression (to me at least), so perhaps the JPEG artifacts are the result of compression for you to upload the image to pbase? Or was this image enlarged from the original?

Regardless of the cause, I can see the problem. Unfortunately, I don't know if any good way to resolve it. I tried Alian Skin Image Doctor JPEG Repair, but the result didn't seem to be much different than a simple Guassian blur. (That said, I'm not real familiar with Image Doctor, so perhaps I just don't know how to tweak it correctly).

I did try using a full strength (radius 100) Median blur (rather than Gaussian blur). This produced slightly less "transition lines" than the Gaussian blur. Then I added about 3% monochromatic uniform noise and faded it to about 40%. (Adding noise can significantly reduce the transition lines caused by strong blurring - but it does add a "grainy" look to the image.) I could then use the healing brush to clean up any further lines that I could still see. (For some reason, when I tried using the healing brush to clean up the transition lines, it didn't work at all - made the situation worse.)

Hope this helps,
Jeanie

[Stephen M]

I have not played with things, but a quick look at the webpage shows both the colour variation and the naked eye can see slightly visible grain or noise, which as mentioned may be due to jpeg issues.

Dupe layer, set to colour mode - apply large median and gaussian blur to merge the colour variation.

http://members.ozemail.com.au/~binaryfx/IT_artifactremoval.html

Dupe original layer and place over the top of others, set to luminosity and small smart blur or other methods as in these links:

http://members.ozemail.com.au/~binaryfx/PSTV_links.html#G

Of course LAB mode edits produces similar results to the duped layers.

There seems to be luminance variation as well as colour - so the AB or colour blend heavy filtering may not do the whole job in this case.

Stephen Marsh.

[Doug Nelson]

It does look like compression noise. There are many threads here about various de-jpg-noisifiers, such as neatimage and Grain Surgery. I seem to recall one thread that had a link comparing a great number of them, but can't find it now.

Anyhow, it won't work on all images, but I had success with yours in going to each individual color channel and running a very slight dust/scratches filter. Whatever method you end up using, you'll find it enhanced by doing each channel separately, and to the minimum necessary for that channel. The blue channel is notoriously noisy, especially.

[DannyRaphael]

Shan Canfield observations (she got back to me via e-mail):

"As far as I can tell it looks like jpeg artifacts (but maybe it's the ozone!) Try blurring that particular sample's red channel directly at around amt 40, and the blue channels, amt 40. Then going back in Red Channel & Blue Channel and adding uniform monochrome noise less than 1, seemed to improve it.

Interesting thought... work on the individual color channels vs. all at once through vanilla Gblur.

(Thanks, Shan.)

~DannyR~

[kthurner]

Thanks for all your replies. Let me give you more background into the image. It is true that I had to post a jpg rather than the original tiff format to get the image on pbase. The visual effects however are not jpg artifacts. This image has already been through Neat Image (much attention paid for optimal results) and some Curves, Hue/Saturation adjustments. The blotches are a fairly common result of using Neat Image on very old and noisy Kodachromes that have gradual color gradients.

The separate channel approach as well as adding noise has given me more things to try. However, my initial attempts on separate channels still fall far short of my interpolation solution. Applying Blur to this image at original size always results in distinct waves of discoloration for me.

Thanks for your help, and keep the suggestions coming!

[Mig]

There isn't a grain removal-type filter I'm aware of that can quite fix this sample. Grain removal filters will just move things here and there and not smooth out blotches like this.
If you look at the channels after a levels adjustment you'll see the green channel is by far the worst, with almost nothing left of it.
As an experiment I took the picture to the extreme.
If you take a look at the picture attached the left panel shows the green channel after levels and it's histogram, which shows there's very little useful digital information left. The centre panel shows the orginal image after levels and it's histogram. The third panel is the image after each channel was blurred to the max, and it's histogram (which is an improvement), but the picture is now nothing but a gradient, not resembling a sky.

The bottom line is you can't improve the whole picture via levels/curves without making it appear worse to the eye, even though it would be "better".
Your instints have been good so far though, and as others have suggested, using noise will help with any banding you end up with as a result of your experiments. Strangely, your best bet might be to camouflage the blotches by doing nothing.

Mig

out of curiosity, if you could, please post a sample of the whole picture.

[Stephen M]

So it sounds like NeatImage is casing the problems.

Try filtering the colour component first.

Then run Neat in this section, or perhaps other methods as indicated in my grain links listed earlier.

Directly blurring the individual channels is a poor option when it is the first thing you try.

It is much better to indirectly filter via a layer set to colour blend or the AB of LAB.

Then, if there are remaining issues in the RGB channels, then direct filtering can be used.

Indirect filtering is easy, free and will affect all the channels at once. Unlike direct filtering it does not destroy detail and it can clean up individual channels like magic.

Stephen Marsh.

[Vikki]

Tough problem!
The only solution I found still left a bit of wave, but not as bad.
I duplicated the layer.
On the top layer, I did a motion blur, straight horizontal.
On the bottom layer, I did a motion blur, straight vertical
I set the layer blend opacity to 50%.

[kthurner]

Many thanks to all for your ideas. After much experimentation incorporating various of your suggestions I have come to the conclusion that a variation on my original solution is still yielding the best (though imperfect) results in this particular image. I am still open to new ideas. My current solution is as follows:

1. Create Alpha Channel of Sky (problem area)
2. Duplicate Image
3. Flatten copy of Image
4. Duplicate Layer, apply sky Alpha Channel, inverse selection, clear
5. Resample unconstrained from 5250ppi to 300ppi
6. Add Gaussian Blur 2.0
7. Resample unconstrained in steps (600ppi, 1200ppi, 2400ppi, 4800ppi, 5250ppi) (Check to make sure pixel dimensions are same as original file; they should be and were in my case)
8. Drag sky Layer from the resampled copy file to the original file
9. Apply sky Alpha Channel to sky layer, inverse selection, clear.
10. Set sky Layer opacity to 80%

I was at an advantage by having the whole image to study in my tests. The crux of the problem was in the pixel dimensions of the blotchiness, this is why I had to post a crop of the image at original size to portray the problem accurately. For those curious, here's a thumbnail of the whole image. Thanks again.

http://www.pbase.com/image/10351591/original

[jeaniesa]

Thanks for getting back to us with your solution. I'm afraid I don't have any other great ideas for you. I'll be tucking your solution away for further reference as I'm sure this is not an isolated incident!

BTW, nice image. Did you scan the image yourself or have it scanned prefessionally on a drum scanner? (Just asking because the digital size is quite large for a "typical" scan.)

Jeanie

[TheTexan]

Im sure this is simplistic since the background sky is probably more complex in the original photo but here on the right side of your blotchy jpg is my duplicate. I simply replicated the exact gradient and it is hard to distinguish from the original. The dithering you see is caused by the jpg compression to put it on the web. It is not present on my computer or on the print.

Tex

[kthurner]

Tex,

How exactly did you do the replication?

[TheTexan]

The way to duplicate the gradient is like this. (I will use for my illustration a linear gradient, that is, one that would fill a rectangle top to bottom with no curvature).

1. Take the original gradient image and divide the image into several sections using guides. I used a guide every 25% of the image height.

2. Then in the gradient editor place a color sampler at each of the 25% locations.

3. Sample the color for each of the gradient editor color tabs from the corrosponding location on the original image. This is duplicate the gradient in that image.

4. Then simply fill a layer with that gradient.

You can use as many samples as necessary to get a better replication. You will also have to adjust the mixer points between each color tab to duplicate the rate of gradiation change to match the original. However, the more you divide the original image the less important this becomes.

Now if the gradient has a curve to it as would a morning sky with the sun just below the horizon you can use the shear filter to add the slight curve to the overall gradient. To recreate the gradient portion of the curved gradient you have to take a narrow slice of the gradient from the focus point outward that represents the colors involved.

In the example I created the top gradient from PS then curved the gradient. Then I started over with a clean image. I duplicated the gradient taking samples at each of 8 12.5% points along the path (drawn). That gave me the gradient. Then I used the shear filter to make the curve. The bottom image is the recreated duplicate.

Here's how to divide an image evenly. First make sure snap is on.

1. Use the retangle marquee to make a selection around the entire image or even a retangular portion of the image that runs from the top to bottom extremes of the gradient you are trying to reproduce.

2. Then move a guide line to the mid point of the selection. Guess what! PS automatically finds the center point of the selection and will snap the guide to there.

3. Then make a new selection with the marquee tool from that guide to the top and repeat from the guide to the bottom. Each time move a guide to the center point (snap) of that selection.

By doing this over and over you can accurately divide the picture into portions that you can find on the location slider of the gradient editor.

Now that you have the gradient that duplicates the sky or whatever you can replace the bad background with the good background.

By the way, this method works with luminescence masks, or any other kind of mask. Let's say you want to reproduce the gradual gradiated darkening of a wallboard behind someone that needs retouching. Simply reproduce the texture of the wallboard then add the gradual change in color or shadow change by first extracting a representitive gradient of the luminence and then reproduce it with a gradient mask using the above method.

Hope this helps.

Tex

[jeaniesa]

Wow - what a great description Tex!! THANKS!!

Jeanie