AreaResize

[Sakura]

Hello! I port AreaResize to VapourSynth, based on the original AviSynth verison by Chikuzen and the modified AviSynth version by Aktan.

AreaResize is an area average downscale resizer:

• Support 8-16 bit and 32 bit sample type.

• Support YUV and RGB color family.

• Downscaling in 8-16 bit RGB with gamma corrected.

Usage

core.area.AreaResize(clip clip, int width, int height[, float gamma=2.2])

# Parameter "gamma" is for gamma corrected, only using for 8-16 bit RGB input

Github: https://github.com/Kiyamou/VapourSynth-AreaResize
Download: https://github.com/Kiyamou/VapourSyn...esize/releases

There is known bug for special target size, such as 1920x1080 -> 1200x700 for 8~16bit YUV and 8bit RGB.

[Lypheo]

Could you perhaps explain the algorithm behind this? How does it differ from conventional convolution-based resizers?

[Sakura]

Originally Posted by Lypheo

Could you perhaps explain the algorithm behind this? How does it differ from conventional convolution-based resizers?

I have something, causing me reply slowly.

This algorithm is simple, just calculating the average of pixel. (So it is fast.)
For example, 1080p->720p, greatest common divisor: (1920,1280)=640, 1920/640=3, 1280/640=2.
So calculate the value of dst like this:

y_dst[0] = (y_src[0] + y_src[0] + y_src[1]) / 3
y_dst[1] = (y_src[1] + y_src[2] + y_src[2]) / 3
...

After calculating, the width of clip is changed from 1920 to 1280. And then use the same method to calcuate for height.

Compared other conventional convolution-based resizers, firstly it can be only used to downscale.
From the aspect of effect, I think AreaResize is better than Bicubic, while Spline and Lanczos is better than AreaResize.