Image resizing

[ppppenguin]

I've done a fair bit of practical work with 8 bit video. It works pretty well for gamma corrected signals. Except on test patterns designed to reveal it, there are no obvious artefacts. But there's no real headroom for processing so 10 bit became the standard. For non-gamma corrected signals you need more bits otherwise the dark areas start to look bad.

Reducing the number of bits during processing is a common thing to do. You've done calculations and the number of bits has grown. You need to get back to 10 for the interface. Truncation isn't good. Rounding is a little better. Adding dither will remove artefacts at the expense of noise. Quantel invented Dynamic Rounding which left the pixel alone if was an integer value at the output bit depth. Anything to the right of the binary point was rounded using a pseudo-random sequence. Easy to implement and very effective.

Add enough noise to dither the signal and you can use an arbitrarily small number of bits. I've played with 1 bit video and the pictures are easily recognised but they are very noisy.

It's getting even further off topic but I'm sure Mark could explain dithering of audio signals much better than I could.