Designing spectral sensitivity curves for use with Artificial Color

Artificial Color uses data from two or more spectrally overlapping sensitivity curves to assign class membership to pixels and ultimately to images. The usefulness of Artificial Color for various scene segmentation tasks has been shown in several recent papers, but those demonstrations all used sensitivity curves not optimized for the particular task, i.e. the R, G, B filters of commercial color cameras. This paper explores means to evolve suitable spectral sensitivity curves suited to any specialized task and illustrates that method with synthetic data chosen to be very hard to discriminate spectrally. Two special cases are illustrated. In one, a single Gaussian curve is used for a dichroic beamsplitter, so that the curve and its complement are used for discrimination. In the other case, two essentially orthogonal curves are utilized for the same task. The single Gaussian curve leads to poorer discrimination but better light efficiency relative to the two curves. Both do quite well on the difficult target problem.