Downsizing methods

Bart van der Wolf wrote:

FYI

I’ve put up a first version of a webpage reviewing different resizing/downsampling methods:
http://www.xs4all.nl/~bvdwolf/main/foto/down_sample/down_sam ple.htm
Many original images, from scanners or digicams are too large for webdisplay. However, downsizing without proper precautions will produce resampling artifacts. The methods described allow to improve the image quality if downsizing needs to be applied.

Bart

Thanks, Bart, for an original and very thorough article. Not much has been done on this important topic and you’ve given us a lot to think about. —

Mike Russell
www.curvemeister.com
www.geigy.2y.net

"Mike Russell" wrote in message

Thanks, Bart, for an original and very thorough article. Not much has

been

done on this important topic and you’ve given us a lot to think about.

FWIW, CS has new sampling algorithms for downsampling.

From: "jjs"

FWIW, CS has new sampling algorithms for downsampling.

Bart checked both new CS bicubic methods (sharper and smoother) in his tests.

I’ve been using bicubic sharper to downsample since I got CS and it works well for me.

"Mike Russell" wrote in message
SNIP

Thanks, Bart, for an original and very thorough article. Not much has

been

done on this important topic and you’ve given us a lot to think about.

You’re welcome. I thought it was intriguing to realize that by reducing the size, we increase the artifacts. Theoretically simple to predict results, but still revealing to see it with your own eyes.

Bart

"jjs" wrote in message
SNIP

FWIW, CS has new sampling algorithms for downsampling.

Correct, so the CS results shown on the webpage may/will look different with older versions. That’s why I included a link to the target, so people can try it themselves.

Bart

"Bill Hilton" wrote in message
SNIP

Bart checked both new CS bicubic methods (sharper and smoother) in his

tests.

I’ve been using bicubic sharper to downsample since I got CS and it works

well

for me.

It will even work better by pre-blurring in relation to the amount of size reduction. Bicubic sharper will require less additional sharpening, although the latter provides more control.

Bart

"Bill Hilton" wrote

I’ve been using bicubic sharper to downsample since I got CS and it works well for me.

From: "Bart van der Wolf"

It will even work better by pre-blurring in relation to the amount of size reduction.

What works for me with film scans is to not do ANY sharpening on the scan so it’s still a touch soft due to blooming from the scanner, then downsample in 50% steps using ‘bicubic sharper’ as many times as needed to get close to the target size, then doing it one more time to the target size. Since my film scans are all pretty much one of three basic sizes, depending on whether I’ve scanned 35 mm or 6×4.5 cm or 6×7 cm film, I have actions that do the steps and it’s pretty quick.

I think not sharpening the initial scan takes care of the pre-blurring 🙂 With ‘bicubic sharper’ I can see artifacting or a kind of crunchy look sometimes when I’ve downsampled sharpened files.

At any rate this method seems to work well for me on actual images, giving better results than I got pre-CS with a variety of techniques. I can see how it won’t work well with a target like you used though. For grins I’ll download your test pattern and see if downsizing in increments (probably smaller than 50% increments since the file is so sharp … maybe 10% steps?) gives better results than resampling in one fell swoop.

Bill

For grins I’ll download your test pattern and see if downsizing in increments (probably smaller than 50% increments since the file is so sharp … maybe 10% steps?) gives better results than resampling in one fell swoop.

Hi Bart,

I downsampled using bicubic sharper in CS in 90% steps (15 repetitions in an action to get to 206×206 pixels) and the results are somewhere between ImageMagick’s Triangle and Lanczos results, much better than resampling with one step. This is without a pre-blur.

Interesting experiment … thanks for posting. I think when I have images with plenty of fine details I’ll downsample in smaller increments using ‘bicubic sharper’ … perhaps at even finer steps it would be even better.

Bill

"Bill Hilton" wrote in message
SNIP

What works for me with film scans is to not do ANY sharpening on the scan

so

it’s still a touch soft due to blooming from the scanner, then downsample

in

50% steps using ‘bicubic sharper’ as many times as needed to get close to

the

target size, then doing it one more time to the target size. Since my

film

scans are all pretty much one of three basic sizes, depending on whether

I’ve

scanned 35 mm or 6×4.5 cm or 6×7 cm film, I have actions that do the steps

and

it’s pretty quick.

That would work a bit better with unsharpened (Bayer CFA) digicam images and most Flatbed scanners. A good filmscanner (assuming a top-notch film) will have real detail down to the single pixel.

I think not sharpening the initial scan takes care of the pre-blurring 🙂

With

‘bicubic sharper’ I can see artifacting or a kind of crunchy look

sometimes

when I’ve downsampled sharpened files.

In print that would probably become invisible, but downsizing is more often done for web publishing, so artifacts are not welcome.

At any rate this method seems to work well for me on actual images, giving better results than I got pre-CS with a variety of techniques. I can see

how

it won’t work well with a target like you used though. For grins I’ll

download

your test pattern and see if downsizing in increments (probably smaller

than

50% increments since the file is so sharp … maybe 10% steps?) gives

better

results than resampling in one fell swoop.

The benefit of the target is that it’ll represent a worst-case scenario. If the method used behaves well on the target (no artifacts beyond the radius as a reduction percentage of the diagonal), there’s no need to question less critical subjects. Although pre-blurring may seem counter productive for increasing the quality, remember that any pre-blur introduced will also be reduced in size.

In Photoshop CS, an 8-b/ch Gaussian blur extends to no more than the following number of pixels:
Radius Pixels
0.0-0.1 0
0.2-0.5 1
0.6-0.8 2
0.9-1.2 3
1.3-1.6 4
1.7-2.3 5
2.4-2.6 6
2.7-2.9 7
3.0-3.3 8
3.4-3.6 9
3.7-3.9 10
Given the shape of a Gaussian curve, I’d say that e.g. a 5x reduction (to 20%) would tolerate up to a 1.2 radius before losing resolution that can’t be reliably restored by post-sharpening. It is also possible to apply a self-defined averaging filter with the Other|Custom filter.

Bart

"Bill Hilton" wrote in message
SNIP

I downsampled using bicubic sharper in CS in 90% steps (15 repetitions in an action to get to 206×206 pixels) and the results are somewhere between ImageMagick’s Triangle and Lanczos results, much better than resampling with one step. This is without a pre-blur.

Pretty good, although that may/will work out differently in other software.

Interesting experiment … thanks for posting.

That’s the purpose. Especially if fixed reduction factors are used, one can optimize the action quite well with that target.

I think when I have images with plenty of fine details I’ll downsample in smaller increments using ‘bicubic sharper’ … perhaps at even finer steps it would be even better.

Yes, try finer and coarser because it depends on the software algorithms used. Let the worst-case target be your guide.

Bart