Bicubic sampling is pretty good for up sampling an image and making it larger, but is it a good choice for down sampling as well? Are there better choices?
$\begingroup$
$\endgroup$
1
-
1$\begingroup$ These two old articles from The Inner Product talk about filters for mipmap generation, which might be relevant to you: Link1, link2. $\endgroup$– glampertAug 6, 2015 at 19:38
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
When Sean and I wrote stb_image_resize we chose Mitchell for downsizing. Mitchell is similar to Cubic, you can read about the cubic class of sampling filters in Mitchell Netravali 1988. They are all pretty similar and will get you very similar results.
I can't find any written record between Sean and I of why we decided to go with Mitchell, but if memory serves we just resampled a bunch of images and used the algorithm that we thought looked best. I wouldn't say that there is one authoritative or best filter, you should use the one that looks best on your data.
Edit: Like joojaa says, a windowed sinc filter is also good, if not quite as cheap. You can find some implementations here.
-
3$\begingroup$ There is a authoritatively best filter, its a infinitely wide sinc filter. Its just not possible to use it. untill that time lanczos windowed sinc is a good alternative to michell $\endgroup$– joojaaAug 6, 2015 at 15:38
-
$\begingroup$ For reference, there's also the Nvidia image tools. $\endgroup$– glampertAug 6, 2015 at 19:48
-
$\begingroup$ If you are using a cubic-esque or lanczos filter, do those guys work equally well for scaling up as they do for scaling down? $\endgroup$ Aug 6, 2015 at 20:14
-
$\begingroup$ I've not tried lanczos so I can't speak to that. We chose catmull-rom for upscaling, which is a cubic, and it worked well. $\endgroup$ Aug 7, 2015 at 17:13