Improving the convergence of local binary fitting energy for image segmentation
Yangyang Song, Guohua Peng · 2019
This paper explores a Riemannian steepest method for fast converging local binary fitting model. The proposed method takes advantage of intensity information in local regions, which solves the intensity inhomogeneous images with satisfactory results. Furthermore, the Riemannian steepest descent method can be employed to local binary fitting model from exponential family and achieves convergence fast. The main contribution of this paper is that presents a general closed-form expression for the manifold’s Riemannian metric tensor of local binary fitting model, which makes the computation of Riemannian gradient flow possible. In addition, to ensure the accuracy of the segmentation results, we regularize the level set function by Gaussian smooth operator. Experimental results for synthetic and real-life images show satisfactory performances of proposed method.