Chroma subsampling

The asymmetry it exploits

Human vision has far more receptors for brightness than for colour, and resolves fine luminance detail much better than fine chrominance detail. A lossy encoder converts the image from RGB into a space that separates one brightness channel from two colour channels precisely so it can treat them differently.

Chroma subsampling then stores the colour channels at reduced resolution — typically half in each direction. That removes half the total data before any transform or quantization has happened, and on a photograph it is essentially invisible.

Reading the notation

The three-number notation describes a sampling region four pixels wide and two rows tall. 4:4:4 means no subsampling: full colour resolution. 4:2:2 means the colour channels are halved horizontally and kept full vertically. 4:2:0 means halved in both directions, so one colour sample covers a 2x2 block of pixels.

4:2:0 is the default in most JPEG encoders and in video. 4:4:4 is available in most encoders as an option, usually labelled something like "full chroma" or "no subsampling".

Where it becomes visible

Saturated colour against a contrasting background, at small scale. Red text on a white background is the classic failure: the letterforms are defined by a colour transition, and that transition is now stored at half resolution, so the edges smear and fringe. The same happens to thin coloured lines in a chart, and to fine coloured detail against a neutral field.

This is one of the reasons a screenshot should not be a JPEG. It is not only the quantization artefacts; the colour resolution has already been halved before quantization started.

If you must ship a lossy file containing coloured text or graphics, disabling subsampling costs perhaps 20% in file size and removes the fringing entirely.