Gerhard Fuernkranz wrote:
The model of a "lut16Type" for the A2B tables is RGB -> matrix -> shaper -> CLUT -> shaper -> PCS. So I'm not sure, whether using this matrix makes sense for non-linear RGB. If we assume linear (gamma 1.0) RGB (e.g. raw data from the sensors), and if we populate this matrix with an approximate RGB -> XYZ transformation, then it would likely shrink the coverage of the CLUT, resulting in the need for a CLUT with an even higher grid resolution in order to achieve the same accuracy. More appropriate would be IMO rather a matrix, which maximizes the CLUT utilization for the set of all valid RGB triples which can be returned by the scanner, while still producing non-negative numbers after applying the matrix for all valid RGB triples which can be returned by the scanner. BTW: The ICC spec also sais that the matrix shall be the identity matrixunless the input color space is XYZ.
oh ... I've overlooked that the /input/ space has to be XYZ, sorry.What I had in mind was ultimately something like refining the results of a matrix profile with an (abstract) Lab CLUT profile, but I didn't think through the details.
Maybe "refine" can do the trick? -- I'll see. Klaus