View Full Version : In 4:2:2 color sampling, how much of green is sampled?
In 4:2:2, the first digit means Y' (luma) is fully sampled, the second digit Cb' means that Blue minus out the luma is sampled at 50%, and the third digit represents Cr' (red minus out the luma) is also sampled at 50%, is that correct?
What about green, there's no mention of green in 4:2:2, how much of green is sampled?
Finally, if most of the color is green, followed by red, and the least is blue, then why sample more green versus red, and more red versus blue?
In 4:4:4, luma, blue, and red are completely sampled, but again there's no mention of green. I assume green would also be fully sampled??
Thanks for clearing up this mystery for me. :)
03-14-2011, 01:16 AM
Y is made from RG&B.
G is made from Y, Cb, and Bc.
Traditionally, RG&B would be 4:4:4. You'd transcode (with multiplies, additions, and possible limiters) to Y, Cb, Cr, filter Cb and Cr and throw away half the Cb and Cr samples. When reconstructing RG&B, you should fill in the blanks of Cb and Cr and make RG&B 4:4:4 again.
After that round trip, RG&B will show any black and white transitions at full bandwidth, but the color will be a bit sloppy. You can see it if you look at color bars closely. There will be a sharp brightness transition, but the color transition is slower. It's especially apparent between green and magenta, which are opposites. You get no color right between the bars.
Cameras can make this more complicated. CMOS sensors generally use a Bayer filter, which is a pattern of R, G & B dots with twice as many G dots as R or B dots. Reconstructing the output of a Bayer sensor to RGB or Y, Cb, Cr is a real art.
So in 4:2:2 you would filter the Cb (blue) and Cr (red) and throw away half, so that it's sampled at 50%? Does that mean nothing is thrown away with green?
Obviously, 4:4:4 is the best, but how much are we losing compared to 4:2:2? If we can quantify it, are we losing 20%?? And even if we are in real world terms, not a big deal??
Then going to 4:2:0, is that better than 4:1:1? With these 2 we are definitely losing color information (resolution), but how much compared to 4:4:4?
Thanks for any info!
03-16-2011, 01:47 AM
Visually, 4:2:2 looks great - except on certain test patterns. The eye is not as sensitive to detail in color shifts as it is sensitive to details in light level changes. Something to do with rods and cones in the eye...
So, is green filtered? Yes and no. Same for blue and red. Yes and no.
When red, green, and blue move together, they are not filtered at all. They are 4:4:4 when in unison. When they move in a way that signifies a color change but not a light level change, they are all filtered (2:2:2). The bottom line is that luma changes are crisp and chroma changes are smeared, and the eye is happy with that.
In a way, asking if G is filtered is like asking if a section of freeway operates on gas or diesel. When gas cars drive by, the freeway is operating on gas. When diesel trucks drive by, it's operating on diesel. So... it depends.
The one place where true 4:4:4 is clearly beneficial is in chroma keying. Being able to detect a crisp color change even when the luma level stays the same allows one to key a crisp, clean, accurate, low-noise edge. With 4:2:2 you can still get a good key, but you need some extra processing to sharpen the edges and that sharpening can make the edge susceptible to noise.
Anyway, when photographic material is viewed at a normal distance, 4:2:2 looks great. It's low-tech compression. In fact, back in the early '50s when color NTSC was invented, they reduced the color bandwidth in order to get as much information in the 6MHz TV channel bands as possible. Just as we have digital compression today (MPEG-2, h.264, etc), color bandwidth reduction was the analog compression of the day.