Minimizing image detail loss while downsizing

I don't know the details of the algorithm for downsizing pictures. I am curious, would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:

1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

Would image detail loss be more pronounced in the second case? Is the downsizing ratio relevant at all?

Rhus wrote:

I don't know the details of the algorithm for downsizing pictures. I am curious, would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:

1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

Would image detail loss be more pronounced in the second case? Is the downsizing ratio relevant at all?

An easy way to satisfy yourself of the answer to that question, try it.

Brian.

Well, thanks for the glib answer. Of course I tried it. I am not talking about something that immediately jumps at you. I am talking about minute details that may not be immediately noticeable but would require careful and detailed study and would perhaps manifest itself under certain circumstances. I was asking someone who understood the algorithm, so I could benefit from that knowledge.

"Rhus" wrote in message

I don't know the details of the algorithm for downsizing pictures. I am curious, would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:

1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

Would image detail loss be more pronounced in the second case? Is the downsizing ratio relevant at all?

My understanding (which may or may not be correct -- I hope Chris or someone else jumps in) is that it depends on your source format.. e.g. JPEGs do better being reduced or enlarged on 16-byte boundaries, while for lossless formats (PSD, TIFF etc) it doesn't matter.

Rhus wrote

would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:
1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

For this case, where you're not downsampling very much, it shouldn't matter.

If you have CS use the 'bicubic sharper' option.

Sometimes if you are *really* downsampling, like say a 10,000 pixel per side scan down to say a web thumbnail of 75 pixels you'll be better off downsampling in steps, but for small percentages you likely won't see any difference.

And as Brian implied, sometimes you should do it both ways and compare since under some conditions you'll see differences.

"Rhus" wrote in message

I don't know the details of the algorithm for downsizing pictures. I am curious, would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:

1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

Would image detail loss be more pronounced in the second case? Is the downsizing ratio relevant at all?

The more you down-sample, the more you lose. The difference between integer or non-integer scaling amounts is small, depending on the method chosen. But you may possibly be more interested in detrimental effects (such as aliasing) that *will* occur when down-sampling. Although the effects may not be too pronounced with the small resampling ratio in your example, there are better and worse methods of handling:
http://www.xs4all.nl/~bvdwolf/main/foto/down_sample/down_sam ple.htm http://www.xs4all.nl/~bvdwolf/main/foto/down_sample/example1 .htm

Photoshop is probably not too bad when you use regular Bicubic resampling (perhaps preceded with a little Gaussian Blur or a Custom weighted averaging filter) and small size reduction, but there are several better methods than the Adobe recommendation of Bicubic sharper.

Bart

Bill Hilton wrote:

Rhus wrote

would the image detail loss be more pronounced if the ratio of downsizing was not easily divisible by 2, for example:
1. I have an image 2048 pixels wide and resize it to 1024 pixels wide
2. I have an image 2048 pixels wide and resize it to 970 pixels wide

For this case, where you're not downsampling very much, it shouldn't matter.

If you have CS use the 'bicubic sharper' option.

Sometimes if you are *really* downsampling, like say a 10,000 pixel per side scan down to say a web thumbnail of 75 pixels you'll be better off downsampling in steps, but for small percentages you likely won't see any difference.

And as Brian implied, sometimes you should do it both ways and compare since under some conditions you'll see differences.

I have a related question which I intended to ask for a while, but held back because I still use GIMP 1.2. Why is it that when down-sizing (scaling) an image, I get a slight blurring effect? Is it just me? Is it something I can (and should) expect?

Roy

--
Roy S. Schestowitz
http://Schestowitz.com

Bill Hilton wrote:

Sometimes if you are *really* downsampling, like say a 10,000 pixel per side scan down to say a web thumbnail of 75 pixels you'll be better off downsampling in steps, but for small percentages you likely won't see any difference.

I have read about incrementally upsampling for the best results, but this is the first mention of incrementally downsampling would work the same.

Sometimes if you are *really* downsampling, like say a 10,000 pixel per side scan down to say a web thumbnail of 75 pixels you'll be better off downsampling in steps

I have read about incrementally upsampling for the best results, but this is the first mention of incrementally downsampling would work the

same.

I first heard of this from Bill Atkinson of Apple Computer design team fame when I took a class with him a couple of years ago. He was scanning medium format at 5,000 dpi (basically ending up with 12,000 x 12,000 pixel files) and using these edited files for everything from printing to web use. I think he downsampled in 50% increments until close to 85 pixels for thumbs, then went directly to 85 pixels (this was in an action). This was prior to CS and 'bicubic sharper' so he was using straight 'bicubic'. Not sure if he still does it that way.

You can see some of his thumbnails here ...
http://www.billatkinson.com/CatalogIndex.html

Bill

In comp.graphics.apps.gimp Bart van der Wolf wrote:

"Rhus" wrote in message

<snip>

Would image detail loss be more pronounced in the second case? Is the downsizing ratio relevant at all?

The more you down-sample, the more you lose. The difference between integer or non-integer scaling amounts is small, depending on the method chosen. But you may possibly be more interested in detrimental effects (such as aliasing) that *will* occur when down-sampling. Although the effects may not be too pronounced with the small resampling ratio in your example, there are better and worse methods of handling:
http://www.xs4all.nl/~bvdwolf/main/foto/down_sample/down_sam ple.htm http://www.xs4all.nl/~bvdwolf/main/foto/down_sample/example1 .htm
Photoshop is probably not too bad when you use regular Bicubic resampling (perhaps preceded with a little Gaussian Blur or a Custom weighted averaging filter) and small size reduction, but there are several better methods than the Adobe recommendation of Bicubic sharper.

You might be interrested in that the Lanczos interpolation algorithm was added to the development branch of gimp back in january: http://bugzilla.gnome.org/show_bug.cgi?id=162250
--
Pelle

Hello Roy,

Roy Schestowitz wrote:

I have a related question which I intended to ask for a while, but held back because I still use GIMP 1.2. Why is it that when down-sizing (scaling) an image, I get a slight blurring effect? Is it just me? Is it something I can (and should) expect?

If there are sharp lines or edges, which are exactly at pixel boundaries then these will get blurred, because they are not at pixel boundaries after scaling.

So, if you want to sharpen the image, then do this /after/ scaling.

regards,

Peter

"Pär Forsling" wrote in message
SNIP

You might be interrested in that the Lanczos interpolation algorithm was added to the development branch of gimp back in january: http://bugzilla.gnome.org/show_bug.cgi?id=162250

Thanks for the link(s).
Now, if only the Gimp would natively support 16-bits/channel or more. Last time I looked at cinepaint, it wasn't mature enough. Photoshop is going to add 48-bit HDR support, so how about it ...

Bart

Bill Hilton wrote:

Sometimes if you are *really* downsampling, like say a 10,000 pixel per side scan down to say a web thumbnail of 75 pixels you'll be better off downsampling in steps

I have read about incrementally upsampling for the best results, but this is the first mention of incrementally downsampling would work the

same.

I first heard of this from Bill Atkinson of Apple Computer design team fame when I took a class with him a couple of years ago. He was scanning medium format at 5,000 dpi (basically ending up with 12,000 x 12,000 pixel files) and using these edited files for everything from printing to web use. I think he downsampled in 50% increments until close to 85 pixels for thumbs, then went directly to 85 pixels (this was in an action). This was prior to CS and 'bicubic sharper' so he was using straight 'bicubic'. Not sure if he still does it that way.
You can see some of his thumbnails here ...
http://www.billatkinson.com/CatalogIndex.html

I'm also need to downsample my high resolution film scans for web use. How is 'bicubic sharper' different from the generic 'bicubic'? Are you suggesting that using 'bicubic sharper' can eliminate the incremental downsampling? Or to use both for best results? Thanks.

How is 'bicubic sharper' different from the generic 'bicubic'?

Adobe added 'bicubic sharper' and 'bicubic smoother' with CS, one for downsampling and one for upsampling. Read the Help files (search interpolation, probably) for details.

I'm also need to downsample my high resolution film scans for web

use. ...

Are you suggesting that using 'bicubic sharper' can eliminate the incremental downsampling?

I was just repeating what Atkinson used but that was with Photoshop 7 two years ago when I took his class, prior to 'bicubic sharper' and I don't know what he's doing these days.

Or to use both for best results?

You should be able to pick ten or so representative scans and downsample both ways (write up an action for the stair version) and compare and decide what's best for your flow. 'bicubic sharper' does a good job, especially with photos and with moderate downsampling ratios, but there are vector-like geometric patterns that give it problems with aliasing (van der Wolf has some tests that show this clearly) and stair downsampling works better on those cases.

Bill

"Bill Hilton" wrote in message

How is 'bicubic sharper' different from the generic 'bicubic'?

Adobe added 'bicubic sharper' and 'bicubic smoother' with CS, one for downsampling and one for upsampling.

Adding to Bill's exactly correct statement; read it as "..., respectively. IOW: For images with detail, Adobe recommends Bicubic Smoother for upsampling and Bicubic Sharper for downsampling, but all standard disclamiers apply - for one, Bicubic Sharper can oversharpen. You really have to try each with your specific image to be sure. You can _see_ how the methods vary by making extreme cases.

In comp.graphics.apps.gimp Bart van der Wolf wrote:

"Pär Forsling" wrote in message
SNIP

You might be interrested in that the Lanczos interpolation algorithm was added to the development branch of gimp back in january: http://bugzilla.gnome.org/show_bug.cgi?id=162250

Thanks for the link(s).
Now, if only the Gimp would natively support 16-bits/channel or more.

Yeah, That'd be great. It's on the way, but it will take some time. Gimp is supposed to switch to a new image processing library called gegl, which is in an early stage of development right now. http://www.gegl.org/

Last time I looked at cinepaint, it wasn't mature enough. Photoshop is going to add 48-bit HDR support, so how about it ...

The last version of PS I used was v7. The 16-bit support was a bit of a dissapointment. Curves, Gaussian blur and unsharp mask were the only useful things you could do IIRC. How is the situation with more recent versions? Is there support for working with layers/adjustment layers in 16-bit mode?

--
Pelle

Forget my last question, google was my friend.
--
Pelle