Chroma subsampling

Why colour is sampled less finely

The human visual system dedicates many more receptors to brightness than to colour. A fine pattern of light and dark is perceived clearly, but the same pattern rendered only as alternating hues is perceived much less sharply. A reference that spends equal resolution on luminance and chrominance is therefore spending colour detail that the viewer cannot resolve. Chroma subsampling is the deliberate reduction of chrominance resolution to match the eye's own limit, and it is the single largest saving in an uncompressed video pipeline.

The notation is compact but easily misread. The three numbers describe, in order, the number of luminance samples in a reference block and the number of chroma samples taken at two positions within that block. In practice the figures are read as a ratio against the first number: 4:4:4 keeps every chroma sample, 4:2:2 keeps half, and 4:2:0 keeps a quarter. The two chroma channels are affected together, so the total chroma reduction in 4:2:0 is to one quarter of the full-resolution chroma count.

The three grids, cell by cell

It is easier to trust the ratio when the grids are drawn. In the diagrams below each small square is a luminance sample and each marked disc is a chrominance sample. The pattern repeats horizontally across the picture.

The visual weight of the difference is modest in ordinary viewing and becomes obvious only under scrutiny. Fine red text on a grey panel, a saturated edge between complementary colours, or a single-pixel-wide coloured line will show softness or slight bleeding under 4:2:0. For archival masters and for any stage that will be edited repeatedly, 4:4:4 or 4:2:2 is preferred, because each round of subsampling and upsampling loses a little colour precision permanently.

Bandwidth consequences

The saving can be stated exactly. Full colour data for a picture of W by H samples requires three channels, so 3 × W × H samples. Under 4:2:2 the two chroma channels are halved horizontally, giving 2 × W × H samples, a one-third saving. Under 4:2:0 the chroma channels are halved in both directions, giving 1.5 × W × H samples, a one-half saving. For a 1920 by 1080 picture that is 6.22 million samples at full colour, 4.15 million under 4:2:2 and 3.11 million under 4:2:0, before any compression at all.

These figures describe the raw samples that enter the compression stage, not the final bitrate. A well-designed encoder removes far more redundancy than subsampling ever could, but subsampling has the advantage of being unconditional: it reduces the sample count before a single transform coefficient is computed, and it costs nothing in decoder complexity. For that reason it has survived every generation of coding standards and is still specified in the newest schemes.