We have successfully done a non-linear color conversion from RGB to gray scale on the GPU.  To perform the non-linear conversion we first converted the image from RGB to Modified L*u*v space by using the set of equations in the document listed below. After converting to L*u*v space a neural network variant called a SOM(Self-Organizing Map) was trained to perform the non-linear conversion to gray scale.  Each node in the trained SOM corresponds to a level of gray scale 0-255.  After the SOM was trained the gray scale conversion was simply the index of the node with the smallest sum-square difference to the current L*u*v pixel.  The resulting gray scale image preserves the topology of the original color image and will be used for image segmentation and eventually disparity and distance calculations. Currently the program is implemented in CUDA and performs non-linear color conversion on a 1024×768 image in about 20ms on the TeslaC1060 GPU.  The results of an image conversion are posted below along with the article explaining the algorithm.

Original RGB Image

Non-Linear Gray Scale