CUDA accelerated illumination preprocessing on GPUs

In this paper we develop a parallelized implementation of the anisotropic diffusion image preprocessing algorithm for illumination invariant face recognition proposed by Gross and Brajovic. Our implementation employs Red-Black Gauss-Seidel relaxation running on inexpensive Graphics Processing Units (GPUs) programmed with Nvidia's CUDA framework. We are able to achieve a 20X speedup over a multithreaded implementation running on a quadcore CPU. Additionally a comparison to an open-source implementation of anisotropic diffusion in the Torch3vision library is performed, demonstrating a GPU speedup of greater than 900X over this commonly used machine vision library.