Image Segmentation Based On Graph Cuts And Level Sets

Image segmentation refers to segment an image into a number of disjointed set of homogeneous regions. The degree of separation depends on the problem to be solved. and its substance is the process of clustering of properties in accordance with pixel attributes (gray, color, texture, shape, spatial location, etc). It aims to segmenting an image into a number of continuous, homogenous uniform regions; in addition, it locates as precise as possible at the edges of the regions. Thus it abstracts the Objects of Interest (OoI) from the complex scene in order to further analytical processing. It is a critical technology in computer machine vision field, and the segmentation result directly related to the image analysis, image understanding and subsequent image processing result.Graph cut theory is based on Markov Random Field to establish its image model, and it constructs an energy function with support of Gibbs random field. It uses maximum flow/minimum cut algorithm to be the intelligent optimization technical optimization theory. This theory can integrate a variety of knowledge; and it has the best overall optimization ability. It provides a uniform framework for solving computer vision problem, and it is the hot study spot based on energy minimization in recent years.The image segmentation method based on level set active contour model is widely concerned by domestic and foreign scholars for its superior performance, flexible structure, and diverse forms.This paper is supported by the Graph cut theory and level set model, and mainly did the job in following three respects:Firstly, the paper proposed an image segmentation method combined Fuzzy C-means clustering (FCM) and Graph Cuts to overcome the defect that Fuzzy C-means clustering has no consideration of the pixel space information. Based on Graph Cuts theory, considering the pixel space information, this method establishes a global energy function concerning labels, and takes the FCM clustering center as terminal to found a multi-terminal network, the network solutes the global minimum or the approximate minimum energy function corresponded to the label function f by the a expansion move algorithm. This method Re-demarcated all pixels among all classes to achieve the right segmentation. Experiments results show that the method has a greater improvement in accuracy, performance and noise immunity.Secondly, the paper proposed an image segmentation method based on variation level set and combining both boundary information and regional information to overcome the defect that the Li model only employees the gradient information of the image edges and have no consideration of the overall information. This method improves the energy function of Li model, introduces external energy terms of C-V model, combines the image edge gradient information and region information, and applies it to the ultrasonic image segmentation. The experimental results show that the new model not only inherits the advantages of the two models, but also performs better than the above models mentioned in the speed of segmentation, accuracy, noise immunity, etc. It can extract the objects better from ultrasonic image.Thirdly, this paper proposed an image segmentation method which combines improved variation level set with graph cuts, in order to overcome the defects that the Active Contours (GCBAC) can not segment concave target, and easy to fall into local minimum and at the same time to reduce the time complexity of level set model. This method firstly pre-splits the object region by GCBAC algorithm for its Rapid, robust, and strong anti-noise advantages, then the object regional approximate to the real boundary contour is remained. Secondly, this contour is served as the initial outline of the improved variational level set model, after the limited iterations of the level set model, and enables the contour to fast convergence and precisely positioned at the boundary of the region finally to achieve to segment the object precisely.