A level set method for image segmentation based on Bregman divergence and multi-scale local binary fitting

Image segmentation is an important processing in many applications such as image retrieval and computer vision. The level set method based on local information is one of the most successful models for image segmentation. However, in practice, these models are at risk for existence of local minima in the active contour energy and the considerable computing-consuming. In this paper, a novel region-based level set method based on Bregman divergence and multi-scale local binary fitting(MLBF), called Bregman-MLBF, is proposed. Bregman-MLBF utilizes both global and local information to formulate a new energy function. The global information by Bregman divergence which can be approximated by the data-dependent weighted L2 − norm, not only accelerates the contour evolution, especially, when the contour is far away from object boundaries but also boosts the robustness to the initial placement. The local information is used to improve the capability of coping with intensity inhomogeneity and to attract the contour to stop at the object boundaries. The experiments conducted on synthetic images, real images and benchmark image datasets have demonstrated that Bregman-MLBF outperforms the piece-wise constant (PC) model in handling intensity inhomogeneity and is more effective than the local binary fitting model and more robust than the local and global intensity fitting model.