Marti Maria schrieb:
The computations are done in 16 bits of precision, unless of course, you recompile the engine with USE_FLOAT. But even in this case temporary results are stored in 16 bits. This is enough for most imaging operations were visual accuracy is about 11 bits at most.
Hi Marti,
this number corresponds with an experiment I've done recently:
I selected 5000 colors, distributed approximately evenly within the sRGB gamut. Then I computed the gradients dXYZ/dE94 for each color. The lowest value I got was 0.023, which is approx 100/4347. This means, in the worst areas of the sRGB gamut, an dX, dY or dZ step of 0.023 corresponds to one de94 unit, and in all other areas, even a larger XYZ step is necessary to become noticeable. Log2(4347) = 12.1, thus approx. 12 bits (in linear XYZ space) are sufficient, in order that (in the worst case) the size of a one LSB step drops below 1 dE94 unit. Since the quantization error is only 1/2 LSB, we need even one bit less, if our goal is a max. quantization error below 1 dE94.
For 4750 out of the 5000 considered colors (i.e. 95% quantile), the gradient was above 0.079, which is 100/1265, corresponding to 10.3 bits, to achieve a one LSB step size of 1 dE94. Thus 9.3 bits are sufficient for 95% of all considered colors, if I want the quantization error drop below 1 dE94.
So yes, with 16 bits of precision for the internal computations of the transform there should be still some room in the "normal" case, where the image is already normalized to the desired white point luminance, before the transform is applied.
But it looks like people also desire to apply transforms to HDR images. So if one applies a transform to an image, and THEREAFTER wants for instance to scale up the luminace of the transformed image by e.g. a factor of 1000, in order to extract very dark areas of the HDR image, then he has rather lost with a 16-bit transform, even if the original HDR image (before the transform) would have contained enough accurate shadow detail information for the desired luminance upscaling.
Regards, Gerhard
------------------------------------------------------- This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170 Project Admins to receive an Apple iPod Mini FREE for your judgement on who ports your project to Linux PPC the best. Sponsored by IBM. Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php _______________________________________________ Lcms-user mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/lcms-user
