An advanced level set method based on Bregman divergence for inhomogeneous image segmentation

Abstract

Intensity inhomogeneity often occurs in real images. Local information based level set methods are comparatively effective in segmenting image with inhomogeneous intensity. However, in practice, these models suffer from local minima and high computational cost. In this paper, a novel region-based level set method based on Bregman divergence and local binary fitting, hereafter referred to as Bregman-LBF, is proposed for image segmentation. The proposed method utilizes global and local information to formulate a new energy function. The Bregman-LBF model enjoys the following advantages: (1) Bregman-LBF outperforms the piece-wise constant(PC) model in handling intensity inhomogeneity. (2) Bregman-LBF is more effective than the local binary fitting (LBF) model and more robust than the global and local intensity fitting (GLIF) model. The relationship between the Bregman-LBF model and the existing models, e.g. the Chan-Vese(CV) model, is discussed. The experiments conducted on synthetic and benchmark image datasets have shown that the proposed Bregman-LBF outperforms the piece-wise constant (PC) model in handling intensity inhomogeneity. The experimental results have also shown that the Bregman-LBF is more effective than the local binary fitting (LBF) model and more robust than the global and local intensity fitting (GLIF) model.