High tap count Spline filters? (Spline36/64/100/144/196/256)

[Akira]

OK. That makes more sense. But for support 2=n, you should evaluate (2*n)^2 = 16 values (in one dimension?). But I got originally 202 values. Please clarify the correspondence.

In two demension. Spline2 (Spline16), Lanczos2, Bi-cubic are all 4x4. Spline3 (Spline36) and Lanczos3 are 6x6.

[magick]

Akira, what is your preference? -define filter:lobes={2,3,4} or -define:support={2,3,4}?

[Akira]

-define filter:lobes={2,3,4} should be righter if you use filter:lobes={2,3,4} for Lanczos2,3,4

[fmw42]

Makes sense to me to use whatever notation is consistent with the current notation for lanczos for example.

Correct me if I am wrong, but Anthony uses lobes for the number of lobes and support for how many lobes the filter truncates it to.

[magick]

IMv7 Beta is ready for testing. Use CubicSpline and define filter:lobes={2,3,4}.

[Akira]

Makes sense to me to use whatever notation is consistent with the current notation for lanczos for example.

Correct me if I am wrong, but Anthony uses lobes for the number of lobes and support for how many lobes the filter truncates it to.

That's why I think filter:lobes is more appropriate than filter:support.
Support of Spline16,36,64 are 2, 3, 4 because the lobes of them are 2,3,4, so the filter doesn't truncate it within the lobe. If you use filter:suppor for selecting spline16/36/64, you can't use the custom filter support option for them.

http://svn.int64.org/viewvc/int64/resam ... rnels.html
Samples in 3.0-4.0 area in Spline64 almost don't affect the final result, so 4-lobed Spline (Spline64) and the support size of 3 might be effective for example.

[Akira]

IMv7 Beta is ready for testing. Use CubicSpline and define filter:lobes={2,3,4}.

Thanks! It works beautifully!

This might be not a good test image. I can't tell any difference between Lanczos one and Spline36 one (or 16/64). Apparently, this NASA image just tries to tell us why we need to resize a picture in the linear light space.

https://visibleearth.nasa.gov/view.php?id=55167
https://eoimages.gsfc.nasa.gov/images/i ... _16384.tif

Spline36 - gamma space

Code: Select all

magick land_lights_16384.tif -filter CubicSpline -define filter:lobes=3 -resize 512 land_lights_spline36_gamma.png 

Lanczos3 - gamma space

Code: Select all

magick land_lights_16384.tif -filter Lanczos -resize 512 land_lights_lanczos_gamma.png 

Spline36 - linear space

Code: Select all

magick land_lights_16384.tif -colorspace RGB -filter CubicSpline -define filter:lobes=3 -resize 512 -colorspace sRGB land_lights_spline36_linear.png 

Lanczos3 - linear space

Code: Select all

magick land_lights_16384.tif -colorspace RGB -filter Lanczos -resize 512 -colorspace sRGB land_lights_lanczos_linear.png 

[fmw42]

Here are my results for the new beta. The filter values now have no long list of zeros and seem reasonable. The verbose data seems to indicate the correct filter name and support.

Code: Select all

magick logo128.png -colorspace RGB -define filter:verbose=true -define filter:lobes=2 -filter cubicspline -resize 400% -colorspace sRGB logo128_spline16.png
magick logo128.png -colorspace RGB -define filter:verbose=true -define filter:lobes=3 -filter cubicspline -resize 400% -colorspace sRGB logo128_spline36.png
magick logo128.png -colorspace RGB -define filter:verbose=true -define filter:lobes=4 -filter cubicspline -resize 400% -colorspace sRGB logo128_spline64.png
magick logo128.png -colorspace RGB -define filter:verbose=true -filter lanczos -resize 400% -colorspace sRGB logo128_lanczos.png
magick logo128.png -colorspace RGB -define filter:verbose=true -filter catrom -resize 400% -colorspace sRGB logo128_catrom.png
magick logo128.png -colorspace RGB -define filter:verbose=true -filter triangle -resize 400% -colorspace sRGB logo128_triangle.png

Code: Select all

magick compare -metric rmse logo128_lanczos.png logo128_spline16.png -format "\n" null:
1711.18 (0.0261109)

magick compare -metric rmse logo128_lanczos.png logo128_spline36.png -format "\n" null:
562.698 (0.00858621)

magick compare -metric rmse logo128_lanczos.png logo128_spline64.png -format "\n" null:
628.19 (0.00958556)


magick compare -metric rmse logo128_catrom.png logo128_spline16.png -format "\n" null:
609.158 (0.00929516)

magick compare -metric rmse logo128_catrom.png logo128_spline36.png -format "\n" null:
1541.94 (0.0235285)

magick compare -metric rmse logo128_catrom.png logo128_spline64.png -format "\n" null:
1686.94 (0.025741)

[Akira]

Code: Select all

magick compare -metric rmse logo128_lanczos.png logo128_spline16.png -format "\n" null:
1711.18 (0.0261109)

magick compare -metric rmse logo128_lanczos.png logo128_spline36.png -format "\n" null:
562.698 (0.00858621)

magick compare -metric rmse logo128_lanczos.png logo128_spline64.png -format "\n" null:
628.19 (0.00958556)


magick compare -metric rmse logo128_catrom.png logo128_spline16.png -format "\n" null:
609.158 (0.00929516)

magick compare -metric rmse logo128_catrom.png logo128_spline36.png -format "\n" null:
1541.94 (0.0235285)

magick compare -metric rmse logo128_catrom.png logo128_spline64.png -format "\n" null:
1686.94 (0.025741)

SSIM indexes:

Code: Select all

ffmpeg -i logo128_lanczos.png -i logo128_spline16.png -filter_complex ssim -an -f null -
SSIM R:0.946036 (12.678943) G:0.957871 (13.754209) B:0.970681 (15.328484) All:0.958196 (13.787820)

ffmpeg -i logo128_lanczos.png -logo128_spline36.png -filter_complex ssim -an -f null -
SSIM R:0.994943 (22.961192) G:0.995593 (23.558111) B:0.997249 (25.605576) All:0.995928 (23.902356)

ffmpeg -i logo128_lanczos.png -i logo128_spline64.png -filter_complex ssim -an -f null -
SSIM R:0.989257 (19.688591) G:0.991401 (20.655318) B:0.995461 (23.430614) All:0.992040 (20.990597)

I think evaluating different algorithms by RMSE, PSNR or SSIM is not a good idea, but SSIM index of around 0.992 to 0.995 is what you usually get when you compare an original PNG and a [Quality=95 YUV444] JPEG.