CHROMA SUB-SAMPLING

What do the numbers mean, and why do it? Jim Peterson from Lumagen channels the answers.

When we find someone who clearly knows stuff, we like to ask them questions. And when we heard a talk by Jim Peterson, the President of Lumagen Inc., on the company’s Radiance Pro processor (pictured below), he mentioned in passing that they recommend 4:2:2 chroma subsampling over 4:4:4. So in addition to why that might be, we asked him to explain the whole idea of chroma subsampling, and why it happens. SOUND+IMAGE: What do the numbers in 4:2:2 and 4:4:4 actually mean? JIM PETERSON: All consumer video is encoded using 4:2:0 YCBCR, where Y is Luminance, CB is the ‘Chroma Blue’ channel and CR is the ‘Chroma Red’ channel. But note that CB is different to B, and CR is different to R in the red-green-blue (RGB) format. RGB is the colour format used in computers and to generate photo-realistic images in a computer.

The numbers in the YC YCBCR format indicate how many Luminance samples share a common C CB sample and a common CR sample. All consumer video is encoded as 4:2:0, which means that a 2×2 array of Luminance samples share a common C CR sample, and a common CB C sample. So that’s six pieces of data to encode four pixels.

For 4:2:2 a 1×2 array of Luminance samples share a common CB sample, and a common CR sample. So four pieces of data for two pixels. And for 4:4:4 there is a CR and a CB for each Luma sample — so three pieces of data for each pixel.

Many people think that the 4:4:4 format is better than the other formats, but this is often not true for video due to 4:4:4 having lower bit depth than 4:2:2 for some formats. For example, for 4K/60 and 4K/50 video at 18GHz [the current bandwidth limitation of HDMI], the maximum bit depth for 4:4:4 (or RGB) is 8 bits, but 4:2:2 supports a 12-bit depth. Since the video originated at quarterresolution Chroma (4:2:0), the 4:2:2 format (half-resolution Chroma) provides more than enough data to reproduce the 4:2:0 source, and supports more bits than the 10-bit pixel depth needed for HDR. But the 4:4:4 format can only support an 8-bit depth for 4K/60 and 4K/50 at 18GHz, which is not considered enough for HDR. S+I: But why do we choose to remove useful colour information through subsampling? JP: Reducing the number of colour samples for Chroma by half in the horizontal and vertical dimensions does not noticeably affect image quality. The human eye is about one-ninth as sensitive to Chroma as Luminance (roughly in horizontal and vertically). So down-sampling Chroma to one-half the resolution of Luma in the horizontal and vertical axis is an excellent way to reduce the data requirements without appreciably impacting image quality. By encoding Chroma at ¼ resolution, this allows encoding at a higher video quality for a given bit-rate. I like to describe 4:2:0 format video encoding as giving up something you cannot see — Chroma resolution — to gain something you can see — colour fidelity.

Lumagen is distributed in Australia by Cogworks: http://cogworks.io