Filter in the new axis?

Now there are invalid variations of filter since the gaussain filter introduced. usually we filter in the X and Y axis, if the filter factor in X and Y axis are the same, then the filter is the isotropic filter, or it is the anisotropic filter.

Recentl paper "coherent line drawing" Use the gradient ction and its vertical driection to do filter, One Direction use the dog filter, and other direction just the gaussain filter. then use the threshold to make black-and white image. in fact, the author maybe read the paper "Real Time Video Processing action", find dog filter maybe can use for line drawing. then programming for a try, and get some Not good results. Then author think can we make some varaibtion in dog filter? So produce a good paper
Can I make some extenstion to define strative visualization?

In fact Jack tumbinin 1999,200 3 propose anisotropic diffusion, and bilater filter shoshould in curvature, graident direction. but the difference is the jack think we shoshould directly modify the raw intensity, just make some filter in its graident direction. then use the Taylor series approximation to get the smooth intesnity results. however, Jack's trilateral filter depends on the bilateral filter results grdient. furthermore, the gradient traverse is difficult to implement in GPU.

Now my research problem is: how to improve bilateral filter and implement it in GPU?
By the way, in cgi2007. one zju paper also about improving the trilateral filter, however, in this paper, it just use a new filter or Taylor series to replace the preliminary gaussain to acclerate the computation. I don't think it have solve the problem.