2010.06.14 07:51 "Re: [Tiff] Class Y images with alpha channels", by

Thomas,

is there a possibility to add an alpha channel to a class Y (YCbCr) tiff image? I don't see this in the rev.6 specs.

Nor does the spec mention 12 bit per channel YCbCr, nor does the spec mention RGB with alpha and two unspecified channels, nor does it mention 20 bit per channel grayscale, etc. We ought to understand the 'class' concept in the spec as a (misleading) enumeration of 'typical usage', whilst the tag system makes perfect sense as to allow just any variation.

Well, but then how it's done if it is not specified?

Here are two concrete questions I cannot answer from the specs:

*) Given by bits/component is not divisible by 8, how do I pack the samples? Given my understanding of the "RGB" class, I suppose I need to bit-pack the samples

as tightly as possible, that is, if I have a 9 bit YCbCr (let's assume this for a moment) with 420 subsampling, then I write 9 bits Y, 9 bits Y, 9 bits Y, 9 bits Y, 9 bits Cb, 9 bits Cr, making a total of 6*9 = 54 bits. Now, the specs furthermore say that I need to flush to the end of the byte at the end of a row. What is the end of the row here exactly? Actually, I'm writing not a single row, but two rows at once. I suppose I need to flush at the end of the "row pair" in 420?

Packing issue is as you expect. Pack as tightly as possible, leaving no padding bits. Pad lines, which is in this case not lines of pixels but lines of 'subsampling blocks', and pad them to a byte boundary. I believe the spec is clear on this, though it's been some years since I last reread it so I cannot quote the exact section(s).

*) Given I have an alpha channel, how is it interleaved with the remaining channels? Is the alpha channel subsampled (I suppose not, but I don't know). Is it

YYYYCbCrAAAA or is it YYYYAAAACbCr, or YYYYCbCrA?

This is another issue. Your original question was about adding alpha to YCbCr, and didn't mention subsampling (though I should have understood you meant to imply this). I'd strongly recommend against using subsampling on YCbCr plus alpha. I don't think there's a single codec on the planet that is going to make sense of it, my own proprietary codec included.

This being said, there is a single unambigious way to do it. Channel order is unambigious, alpha and undefined channels always follow the color channels and are defined by the TIFFTAG_EXTRASAMPLES tag. So in your case, the channel order should be Y-Cb-Cr-A. As to subsampling, it is clear from the subsampling concept it applies to chroma channels (the human eye having less resolution for chroma then it has for luma, subsampling is actually meant as a compression step, saving space by specifying Cb and Cr in lower resolutions). The single and unambigious way to do it, in my mind, would therefore be YYYYCbCrAAAA (assuming subsampling values [2,2]).

Best regards,

Joris