Institute for Data Analysis and Visualization

Most of the recent work on inhomogeneous diffusion in image filtering focuses on diffusing the isotope curve. We present a less familiar approach to the development of inhomogeneous diffusion algorithms in which the image is regarded as a surface in three-space. The magnitude of the surface normal controls a diffusion that evolves the image surface at the speed proportional to its mean curvature. A discrete algorithm to implement this proposed diffusion is introduced and we show experimentally that the algorithm develops singularities that reveal, preserve, and enhance the underlying signal. If the imput signal is an isolated noisy edge, this leads to complete noise removal and enhancement without affecting the edge locality.