Abstract:

We propose a nonlocal image segmentation model based on local signed difference (LSD) energy for images with intensity inhomogeneity. The model consists of a data driven term based on LSD energy and a nonlocal total variation regularization term, and possesses local separability and global consistency. Due to the convexity of the model, the splitBregman iteration algorithm can be used in the numerical implementation, thus having a fast speed. Compared with the classical local regionbased active contour models (ACMs), the proposed model has several advantages as follows: 1) it is less sensitive to the initialization; 2) it is more efficient by using the splitBregman iteration algorithm; 3) it can correctly segment images with fine textures and weak object boundaries. Experimental results show that the model can segment images with intensity inhomogeneity correctly and works more robust than other models.

Key words: image segmentation;nonlocal total variation;intensity inhomogeneity;LSD;splitBregman algorithm