2011.06.05 12:47 "[Tiff] Questions on TIFF LZW compression", by Thomas Richter
would it be possible that some of the knowledgeable people here help me
understanding some details of the LZW compression
specified in the TIFF specs? I'm not asking about the compression itself
which is clear enough, but rather its integration into the
First, how does LZW compression work for images that are not 8 bits/pixel?
As far as I read the specs, the input to the LZW compression is the
"raw" TIFF stripe buffer, but this leaves a couple of corner
cases open. I see that this type of compression works for any bit depth
that divides eight, or is a multiple of eight, but what
happens for 10 bits per pixel images? Is LZW really applied to the
bit-packed(!) input, or is the LZW algorithm applied to 16 bit
data, where 10 bits are packed (left-justified? right-justified?) into a
16 bit word? In the former case, LZW will very unlikely give
much compression performance, of course.
What happens with 16 bit/pixel data? Is the input to the LZW compression
big-endian or little-endian? Or does it depend on the
endian-ness of the TIFF file? I suspect the latter, but couldn't find
any clear indicator of this.
What about the predictor mode (horizontal difference predictor). While
there is no problem with overflow with 8 or 16 bit per pixel data,
what happens with 10 bit/pixel data, for example? While the specs don't
spell this out, do I need to take the difference modulo 2^N, where
N is the bit-depths? Or modulo 2^M, where M is the bit-size of a
"container" word, say 16 bits for 10 bit data? What happens if I encode
10 bit/pixel data with a predictor? Do I apply the predictor on the
pixel values (sensible) or on the raw 8-bit values resulting from bit-
packing the 10-bit data into 8-bit containers (not very sensible, but
not ruled out by the specs).
Can the predictor also be specified without LZW compression? It probably
makes little sense then, but can - for example - the predictor
be combined with the Fax (CCITT, ITU) compression?
And finally, what about the predictor and tiles? Does the predictor
predict across tile edges? The specs seem to indicate this, but this
would also imply that tiles cannot be decompressed independently. Is
this really the purpose of prediction or are the specs incomplete
and should I stop prediction at tile boundaries (which makes sense if
you want to retain the independence of the tiles).
Thanks a lot for your time!