A regularized restoration model based on geometrical features and noise evaluation

In this paper, we propose a new method for designing the variation restoration model which uses the noise evaluation to decide the approximation term and the information of geometrical structures in the blurred and noised images to choose the regularization term. We adjust the measurement for the approximation term based on the noise variance in the degraded image. By computing the mean curvature which is a local geometrical feature of a image surface, we can use the geometrical information to determine the regularization term effectively. This kind of regularization term can obtain a better proportion between de-noising and keeping edges and texture while avoiding piecewise constant because this model diffuses anisotropic. And it is a restoration model that can adjust the measurement adaptively according to the degraded image instead of using the single measurement to restore all different images. In addition, by using the inherent geometric features, we do not need to take any laborious work to choose an energy functional any more. Our experiments show that our idea is on the correct way and our method can preserve the details of the image while removing noises.