Downsampling resolution without full trans-coding: Possible ?

[Alexey Eromenko]

Hello,

I have a new idea: Downsampling resolution without full trans-coding (using fast trans-code). Today Google YouTube encodes video at multiple resolutions, from 240p up to 1080p. This takes up a *LOT* of space. Can it be eliminated?

Full (traditional) trans-coding re-computes DCT, quantization,  motion predictions intra predictions, is really slow and results in obvious quality loss. Bitrate output of a full recoding can be anything, and the output resolution can be anything.

Fast re-coding: (or trans-coding)
Basically I speak about power of 2 spatial compression of DCT blocks plus quantizer of all channels:
The obvious benefit: reduces server disk space usage for YouTube. (and other video services)
Output Resolution will be integer-factor. For  example 1080p video, results will be 960x540 (1/2), or 480x270 (1/4)
Bitrate output is exactly 1/4 or 1/16 of the original.

Here is the procedure of fast recoding flow:
Just like color space conversion from YUV4:4:4 to YUV4:2:0 is ultra-fast spatial reduction of the chroma channels, integer factor,  this method is very similar in principle, but extends to ALL channels (including Luma channel).

assume 8x8 blocks, all blocks get fast re-encoded on the fly to 4x4 blocks on the server-side, with pixel values averaged, with DCT and quantizer table also get downsampled like that.
Theoretically this can be ultra fast because you don't need to do search for motion predictions,  or intra predictions and all the time consuming stuff.
This trans-coding can happen server-side in real-time with the user getting downsampled video on the fly, with the server having only 2 videos, say 1080p and 720p. (Or 4K and 1440p original)

This will require a new server-side has a special Apache module for fast re-encoding.

What do you think?  Can it work?

-Alexey