A Local Region-based Level Set Method For Image Segmentation

Image segmentation is one of the fundamental problems in the fields of digital image processing and computer vision. It is the process of dividing images into disjoint subsets. However, intensity inhomogeneities often occur in vast majority of real-word images, and it may cause considerable difficulties in image segmentation. To solve this problem, many researchers have done great efforts and proposed a wide variety of methods for image segmentation.This dissertation mainly introduces the development of some classic level set methods, analyzes these segmentation algorithms, and then proposes a new segmentation method based on these models. C-V model is less sensitive to initial contour, and does well in segmenting noisy images. But it can not get a satisfactory result in dealing with images with intensity inhomogeneity. Based on C-V model, RSF model utilizes local intensity information in its energy functional and it can get better segmentation contour when coping with images with intensity inhomogeneity. However,RSF model is largely dependent on the initialization of the contour. If the initial position of contour is set far away from the actual boundary, RSF model is prone to get stuck in local minimum as a result of the non-convexity of its energy function. In classical level set approaches, n level set functions are utilized to define up 2nphases. The level setf needs to be constantly reinitialized as the signed distance function, so that it has some nice properties. The energy functional is non-differentiable. All of these may raise difficulties in minimizing the energy function.In order to overcome these shortcomings, we introduce a piecewise constant level set function, and each phase is represented by a unique constant value. Our method needs only one level set function, no matter how many parts a image is segmented into.Our model avoids different segmentation contours caused by different initializing position. We only need one random matrix to define the initial level set for all images.This model transforms the minimization problem of energy function with an equality to an unconstrained problem by using the augmented Lagrangian method. The energy functional for our model is locally convex and differentiable because we do not use the Delta and Heaviside functions which are non-differentiable. A large number of comparative experiments demonstrate that our method is more computational. Inaddition, our algorithm is robust with the destructive noise.