Enlarging (Interpolating) in Steps Degrades Digital Images

[Sam North]

Some time ago I read in a magazine that although it was popular to interpolate (resample) digital images in increments, this was in fact the wrong way to go about it.

I've done tests of my own and discovered that enlarging a little at a time does indeed degrade the image. I've updated my site with the infomation if anyone is interested in interpolation methods.

http://www.theimageplane.net/tip2_018.htm

Sam

[Jesper]

Interpolating in multiple small steps (10% at a time, for example) seems to be popular and people claim that it gives better results, but I'm still sceptical.

For example, nobody seems to be able to give an explanation as to why it works so well. I've seen one guy give a very complicated mathematical explanation once which would prove that doing it in steps was better, but I wasn't convinced. I've also seen people explain that Photoshop CS has a better algorithm for resizing images which makes interpolating in multiple smaller steps unnecessary.

I've tried comparing the multiple steps method and one step method myself, but didn't see any significant difference between both methods in the results.

[CyberDyneSystems]

I've allways been skeptical too.. although thus far I've never tried it. I have very lite reason to interpolate the actual file. Whwn I need to enlarge.. I have found the output printer driver to be more than sufficient for the sizes at which I work. In fact I have found the results to be nothing short of flawless... Thus my skeptisism about the multi stage approach.

[chris.bailey]

I tried it for a large (A1) size picture annd could honestly see no difference between it and a single stage upsize. Extensis Smartscale does a much better job.

[Sam North]

Well, we are right to be sceptical.

Just for easy reference:

http://www.theimageplane.net/1step.jpg
http://www.theimageplane.net/20step.jpg

[Scottes]

I wanted to try CS's options - Bicubic & Bicubic Sharper & bicubic Smoother - on image data rather than text. I cut a 100x100 block from a flower, upsized it in 10% increments for 15 steps, then did a 1 step upsize. I then cut a 200x200 block from the top left corner of each:

http://www.itsanadventure.com/postimages/upsizetest.jpg

Several are right out, but I had a tough time comparing the 1-step Bicubic to the 15-step Bicubic Smoother, so here they are side-by-side:

http://www.itsanadventure.com/postimages/upsizetest2.jpg

Overall, the 1-step seems noiser. But look at the 2 faint dots in the lower left - the 15-step has lost some detail there due to the smoothing algorithm. Is the noise too apparent? Can I deal with the loss of detail due to the blurring?

So I performed an unsharp mask approximating what I would use if I were printing this image. Considering that we're upsizing I have to assume it's for print, right? So here's the sharpened for print version:

http://www.itsanadventure.com/postimages/upsizetest3.jpg

And I'd say the winner - for this image at least - is the 15-step Bicubic Smoother. The noise in the 1-step is disgusting, whereas the 15-step could handle even more sharpening than I gave it.

However, I'm still worried. So I took the edge of the flower and compared 1-step bicubic to 15-step bicubic smoother, with the same sharpening:

http://www.itsanadventure.com/postimages/upsizetest4.jpg

Again, the clear winner is 15-step Bicubic Smoother.

Thanks for a very interesting exercise. I would still be using 15-step Bicubic if this topic hadn't been posted, and I wouldn't have been sharpening enough.

I've got some stuff going to the printer in a couple/few days, so I'll run a couple 8x10s through - one with 1-step Bicubic and the other with 15-step Bicubic Smoother with more Sharpening. I'll let you know how the results are in print.

[Sam North]

Someone on another forum told me:

You can improve the results further by upscaling larger than what you intend, sharpen slightly and then downsampling to final size. In fact, I wrote a set of actions that do all this automatically to different print sizes for this reason.

[slejhamer]

Thanks for a very interesting exercise. I would still be using 15-step Bicubic if this topic hadn't been posted, and I wouldn't have been sharpening enough.


Thanks Scottes for this followup. This is clearly the way to go.
And thanks to Sam for giving him the inspiration!
:D

[Scottes]

I played around with Extensis SmartScale last night. I do not like the end result - on screen. It gives the image a "painted" look, as if it were run through a watercolor filter or something. In some things it looks like it would do an excellent job, but it totally mucked up some fur on a wolf - it was so bad looking on screen that I deleted it immediately. However the results on a painted turtle looked very intriguing. I have a shot of the turtle going to the printer - it's for my niece who doubtfully will notice if it's less than perfect, so it's a good test shot. We'll see.

I also spent some time playing with upsizing since I was getting several shots ready for the printer. Although my test of 15-step Bicubic Smoothing looked good, I was bothered at the lack of sharpness. Extensive sharpening was required on the final product, but once things get that blurred it's tough to get things back. We'll have to see how the print looks.

Bothered by the blur, I tried another test. I did 15 steps, but every third step was Bicubic Sharper. So I upsized 10% at a time - Bicubic Smoother, Bicubic Smoother, Bicubic Sharper - until I reached the desired size. The results looked very promising. Noise got more prominent, but it was greatly reduced from using all Sharper. And much more detail was retained than staying pure Smoother. It seems a good mix. Again, we'll see how the print comes out.


Thinking about this even more, I was worried about the multiple sharpening steps along the way. I don't like losing control of the sharpening process. Bicubic Sharper has no options - it just does whatever Adobe programmed, and this does not take image size into account. Correct sharpening is heavily dependent on image size. It drives me crazy when I see books mention specific sharpening formulas with no mention of image size.

Since sharpening increases contrast on edges, multiple sharpening steps have a possibility of blowing out contrasty edges. Multiple sharpen steps can increase the contrast far past acceptable and create some ugly halos. I saw this on one image of a house against a sky. The halos were revolting. So I'm not so crazy about uncontrolled sharpening.


And I'm not so crazy about uncontrolled blurring for that matter. So this got me thinking about upsizing while controlling both the smoothing and sharpening along the way. With this method I could blur based on image size at the time, and sharpen using either Luminance or LAB methods and preferably keep these to midtones with a fade to both shadows and highlights. That is, sharpen more in middle-lightness area, and sharpen less and less as one goes towards either complete highlight or complete shadow. Final sharpening would highlight the laight & dark extremes in a controlled manner.

And this got me to thinking that I am completely nuts. This could take hours to get an image from 2000x3000 up to a 300dpi 20x30 print. (Where's that Cray when I need it?) It would be an interesting exericse to do for that one image that I really really want to be perfect, but the time involved stops a process like this from becoming commonplace.


So in the long run, I'm not really ecstatic with any of the above upsizing techniques. Though I haven't seen any prints yet, so it remains to be seen. I've got 2 more large ones to do tonight, and I think that I'll try out Genuine Fractals and see what that does. Then everything goes to the printer for the final test. Chances are very good that every image will be completely acceptable, but I'm a bit crazy when it comes to squeeking out something just a little bit better.

[Scottes]

I've allways been skeptical too.. although thus far I've never tried it. I have very lite reason to interpolate the actual file. Whwn I need to enlarge.. I have found the output printer driver to be more than sufficient for the sizes at which I work. In fact I have found the results to be nothing short of flawless... Thus my skeptisism about the multi stage approach.

Care for a bit of a challenge CDS?

Find an image that you have upsized your way - one that you've found to be "completely flawless." Give me the original RAW/JPG and I'll produce a JPG doing things my way(s), and then you can print and compare.

Since I'm sure we'll get together again sooner or later I'd like to compare, too.

[Ferdinand]

Well, if you have PSCS and don't want to spend the money on another 3rd party software to do your enlargement, I find that the 15 step, 10% Bicubic Smoother is the best way to go, I just did a test tonight and yup the Bicubic Smoother is indeed better in terms of photo quality. Thanks again, if you didn't post about it, I would still be stuck with a 1 step, Bicubic method *shudders*

Regards,
Ferdinand

[Brian Tilley]

Scottes,

Your tests have confirmed the suggestion given in Scott Kelby's "Photoshop CS book for digital photographers".

In it he suggests increasing a photo in 10% increments using the Bicubic Smoother method.

Thanks for the work you put into the examples that actually show the effect of this method compaired to other techniques.

[Scottes]

Well I'm not a great fan of always using Bicubic Smoother, though it depends on the image. With high detail, like fur, all Smoother loses some crispness. I like 3-step Smoother, Smoother, Sharper for this type of image.

From my tests and prints, it really depends on the image. High detail, no noise? Alternating Smoother & Sharper may be better. High detail but with noise? Smoother, Smoother, Sharper. And so on.


I've been thinking that a 10% Bicubic Smoother increase followed by a *very* small bit of sharpening on midtones would be ideal. I don't like losing control of either the sharpening or smoothing steps. But the process would not be easy at all. Off the top of my head I'm thinking upsize 10%, new mask, find edges, invert, gaussian blur, USM at like Amount 20-30% / Radius 0.5 - 2.5 depending on current image size / Threshold somewhere betwen 0-5, then delete the mask. Now do that 15 times to get a 400% upsize.... Seems insane.

Which means that I probably try it this weekend if it rains...

[Belmondo]

I've been thinking that a 10% Bicubic Smoother increase followed by a *very* small bit of sharpening on midtones would be ideal. I don't like losing control of either the sharpening or smoothing steps. But the process would not be easy at all. Off the top of my head I'm thinking upsize 10%, new mask, find edges, invert, gaussian blur, USM at like Amount 20-30% / Radius 0.5 - 2.5 depending on current image size / Threshold somewhere betwen 0-5, then delete the mask. Now do that 15 times to get a 400% upsize.... Seems insane.

Which means that I probably try it this weekend if it rains...

Oh my lord, Scottes. I'm praying for rain. :lol:

Please let us know how that turns out.

[Scottes]

Well, Tom, the fact that you and I agree on this insanity really bothers me.

I highly doubt that it surprises anyone, though.
:)