Corner-Preserving Anisotropic Diffusion and Junction Detection Using the Structure Tensor

Abstract

The restoration of noise-corrupted images using the PersonaMalik approach suffers from remaining noise near edges, whereas models using anisotropic diffusion parallel to edges like the mean curvature motion suffer from rounding-off of corners. We use the structure tensor, which can be considered as a combination and generalization of both approaches mentioned above, to dampen the diffusion at corners. This leads to noise remaining only at corners. In order to eliminate all noise, we propose a directional diffusion approach for corners. Apart from controlling the diffusion the structure tensor is also used for junction detection.