Best Resampling Method

[phil146]

Trying to understand which resampling method would be best to use on landscape or wildlife images.  From the Help section, it looks like Lanczos 3(non separable) would give the best results.  The 4 choices in the Export menu(Linear, Bicubic, etc.) are only names to me, so any clarification about them would be helpful. (Just an amateur enjoying the hobby)

Thanks,

Phil

[James Ritson]

Hi Phil, the resampling methods are ordered by sharpness. For photographic purposes you can safely ignore Nearest Neighbour. Bilinear will produce the softest result, whereas Lanczos 3 non-separable will produce the sharpest (the difference between separable and non-separable is very slight). Bicubic is somewhere in the middle and will probably suffice for the majority of your images.

 

One reason to use a softer resampling method might be if your image has lots of high frequency content - from excessive noise to very fine detail like chainlink fences and difficult patterning in architecture. Using a sharper resampling method may exacerbate these parts of the image, and may even produce resampling artefacts (though they are more noticeable in video content to be honest - stills, not so much).

 

As always, experiment! But for a good safe bet with most of your images, try Bicubic. Hope that helps!

[phil146]

Thanks, James, for clarifying those methods.  Very helpful.

[TextusGames]

Just created 62 pixel circle in 64 pixel file and exported with all 5 methods. And the bilinear method produced most smooth result for me at least. 

[John Rostron]

4 hours ago, TextusGames said:

Just created 62 pixel circle in 64 pixel file and exported with all 5 methods. And the bilinear method produced most smooth result for me at least. 

Not really surprising. As @James Ritsonimplied, the Nearest Neighbour and Bilinear are not suitable for photo work, but they are good for line artwork such as your circle. Hence the choice.

John

[sudarmindamir]

is it working for images only or may have different about vector file?

[sudarmindamir]

On 5/22/2020 at 2:28 AM, John Rostron said:

Not really surprising. As @James Ritsonimplied, the Nearest Neighbour and Bilinear are not suitable for photo work, but they are good for line artwork such as your circle. Hence the choice.

John

 

On 1/11/2017 at 11:47 PM, James Ritson said:

Hi Phil, the resampling methods are ordered by sharpness. For photographic purposes you can safely ignore Nearest Neighbour. Bilinear will produce the softest result, whereas Lanczos 3 non-separable will produce the sharpest (the difference between separable and non-separable is very slight). Bicubic is somewhere in the middle and will probably suffice for the majority of your images.

 

One reason to use a softer resampling method might be if your image has lots of high frequency content - from excessive noise to very fine detail like chainlink fences and difficult patterning in architecture. Using a sharper resampling method may exacerbate these parts of the image, and may even produce resampling artefacts (though they are more noticeable in video content to be honest - stills, not so much).

 

As always, experiment! But for a good safe bet with most of your images, try Bicubic. Hope that helps!

is it working for images only or may have different about vector file?

[Alfred]

5 minutes ago, sudarmindamir said:

is it working for images only or may have different about vector file?

Vector drawings are resolution independent, so there is no need for resampling.

[Medical Officer Bones]

55 minutes ago, Alfred said:

Vector drawings are resolution independent, so there is no need for resampling.

That depends. Vector graphics are often converted to a bitmap at the end for various purposes - and this conversion can result in rather bad quality depending on the rasterization and resampling methods used.

For example, Illustrator produced (and still does) rather sub-par lower resolution exported bitmap versions. Compared to Designer, Lunacy/Sketch and other vector editors, the output of an (for example) 240px by 240px icon with simple shapes just looks pretty awful: too soft, and it often introduces pixels at awkward positions which affect the quality of rounded shapes.

In short: the base vector-->bitmap conversion in Illustrator is not very acceptable in my opinion.

A legacy trick to improve the overall final quality has always been to export a much higher resolution bitmap version and then use another tool (image editor or dedicated optimization tools) to scale the result to the required lower resolution. This results in a much more acceptable lower resolution variant.

True often even for other vector applications (Designer, Lunacy, etc.) which already produce (much) better output at the original required low resolution. Rounded corners look visibly more rounded: the base exported low resolution versions often cause slightly too straight looking rounded corners, for example.

The resampling method plays a major part in this last step. I tend to stick with Catmull-Rom myself when scaling down vector graphics that are exported as larger bitmaps.

That said, really, generally the output of Designer and most other vector apps is fine, and there is no real need to perform that in-between step. Unless, of course, you work with Illustrator

But if I have to work with high-resolution bitmap versions of what used to be vector art work, and need to produce lower resolution versions, I always avoid resampling algorithm such as bilinear, bicubic, lanczos, or Mitchell-Netravali. Those result in too soft looking downsampled results. Catmull-Rom or even Box work best in these cases. Catmull-Rom also preserves the most detail in scaling down photos to lower resolutions, by the way (since most software does not support this resampling method, I rely on ColorQuantizer for this last step).