Scaled Total Variation Flow Based Scale Space Construction Using Structure Tensor

The total variation flow is used to construct image nonlinear scale space by its structure preserved characteristics. The concept of divergence used in this model is mainly based on image gradient which is often computed in one order neighbor. This restricts the selection of scale parameters so that only diffusion time and weight of fidelity function can be used. Actually, the common sense of spatial scale is not applied in previous nonlinear filtering process. In this paper, structure tensor is applied in the analysis of local neighbor of more than one order, new extended divergence of image diffusion is built up on expanded pixel neighbor. As a result, the spatial domain can be introduced in scale parameter. Not only the effect of nonlinear filtering is obtained, but also the iteration time is reduced greatly.