Intrinsic image decomposition from a single image via nonlinear anisotropic diffusion

Intrinsic image decomposition has become a hot topic since the ground truth dataset was proposed by Grosse and his colleagues in 2009. In this paper, we present a simple but effective approach to intrinsic image decomposition based on nonlinear anisotropic diffusion. The procedure originates from the iterative Retinex algorithm for illumination estimation, which can be interpreted as a nonlinear isotropic diffusion. By introducing a novel edge-stopping function incorporating intensity derivatives and color differences, the nonlinear anisotropic diffusion is quite effective in preserving intensity edges with little color change while calculating the shading image. With this respect, the shading is estimated from its neighbors with similar color, and is efficiently propagated across the image by the diffusion process. Experiments show that the proposed method produces good results on the benchmark, and has better performance according to the established principles compared with the state of the art single-image based methods.