Does downsampling an image twice result in a "worse" (or different) image than downsampling once

Question

I am working on a project where the input is a large high quality image (2560 x 1440 pixels), which is eventually scaled down a lot (256 x 256) and used as input to an ML model.

For reasons that I won't go into, I'd like to first scale the "Large" image down to something like 1024 x 768 ("Medium"), and then later scale it down a second time to the final "Small" size (256 x 256). I know the aspect ratios don't match here, in all cases I'd first crop to get the right aspect ratio, and then scale down.

What I want to know, is if I would be introducing any artifacts or quality degradations by downscaling twice (Large -> Medium -> Small), as compared to downscaling once (Large -> Small).

Answer 1

That depends on the methods and filters involved. To actually find out you could either

• do the math: take the impulse responses and everything else that matters and compare the two approaches.
• try out with a basic image: create a simple 2560x1440 image with e.g. some sine waves or rectangles. Then create the corresponding 256x256 image (not by processing) that would be the output of an ideal downsampler. Then do both processing approaches and just compare the SNRs.