A segmentation model that combines the Mumford Shah(M S) model and narrow band scheme of level set was presented. The disadvantage of Mumford Shah model is computationally time consuming. In each step of its iteration, the data of whole image have to be renewed, which is unbearable for segmentation of large image or 3D image. Therefore, a fast segmentation model was introduce, which combines the M S model and narrow band scheme by a new initialization method. The new initialization method is based on fast marching method, and the computing time decreases to O(N) . In each step of iteration, the new segmentation model only deals with the data in a narrow band instead of the whole image. The experiments show that the two models can obtain almost the same segmentation result, but the computing time of new narrow band M S model is much less than that of M S model.