[sane-devel] Color depth upsampling and compressibility

[Matto Marjanovic]

 >> On Mon, Jul 11, 2005 at 05:11:48PM +0200, Mattias Ellert wrote:
 >>> Even better would be:
 >>>
 >>>               *dst-- = ((*src << 4) & 0xf0) + ((*src)        & 0x0f);
 >>>               *dst-- = ((*src)      & 0xf0) + ((*src-- >> 4) & 0x0f);
 >>>
 >>> Then 0 would map to 0 and 15 to 255, i.e. white is white and black is
 >>> black.
 >>
 >> You're right. I've changed taht in sp15c.c in CVS.
 >
 >Some quick Googling doesn't provide dispositive information on this
 >strategy's impact on compressibility.  However I suspect that introducing
 >pseudo-random data in the low nybble will reduce compressibility beyond
 >simply squaring the symbol table.  (Hence, my initial ponderings on
 >tweaking the constant)

(Admittedly, I haven't been completely following this thread, but...)

To scale from a [0,15] value to a [0,255] value, you have to multiply
 by (255/15) = 17.

And the above nybble-duplication algorithm effectively muliplies by 0x11,
 i.e.... 17.

So, it's not pseudo-random data; it's the effect of correctly scaling the
 original value.  There isn't really any other way to do it without
 introducing junk/noise bits into the image.

(However, given that 17 is the most random number, this may have some
 effect on compressibility. :)

-matt m.