Learning non-homogenous textures and the unlearning problem with application to drusen detection in retinal images