Research And Improvement Of The Segmentation Algorithm

In past almost twenty years, image segmentation, as an important research field in computer vision, attracts the attention of more and more researchers. Because it plays more and more important roles in many fields, many methods about it come out. The first important method is active contour model. It is a top-down processing with prior knowledge and provides a theoretically uniform framework to a series of problems, such as contour extraction, stereo matching and object tracking. So the method has been successfully applied to image segmentation, medical image processing, human-computer interaction and many other research and practical fields. The second is geometrical method based on active contour model. It is first proposed by Osher and Sethian to solve the changing process of blaze following the thermodynamics equation because of the high dynamic of blaze and the random of the topology changing. So the second method has much advantage in image segmentation about the object with complex topology changing and has been applied in medical diagnosis. But these methods still have many problems and need to study these models thorough and propose new model or algorithm, for some inherent disadvantage.The paper illuminates the foundation of active contour model, and studies the improved algorithm of energy minimizing of traditional parametric active contour model. First because active contour model cannot solve boundary concavities better, we introduce the area energy, do some experiment and get better results. But the problem that original contour must be placed close to the real boundary of object is still not resolved. To solve the problem, we introduce the gradient vector flow (GVF). It can solve the concavity problem better than the active contour model. But the deep boundary concavities problem is still not solved. On the basis of analysing the force field of gradient vector flow model; we normalize the force vector of GVF in the diffusion operations. This makes the external forces have no influence from the distance between the points of the contour and the boundary of object. So this method can solve deep boundary concavities better. At the same time introducing direction vector in the GVF model, in the region specified by this vector we can shield the force fields of the objects that we don't want, and then can segment the interested objects more precisely. But the snake model and the other traditional parametric active contour model both have some disadvantage. 1) can not segment object with complex structure and branch; 2) three dimensions can not come true easily; 3) the segment result is sensitive to the initial position; 4) can not ascertain the proper termination rules. So we study the geometric method and on the base of it we introduce the Mumford-Shah model. After the improvement of the Mumford-Shah model, we segment some colourful images and some medical images and solve the above problems.The experiment results show that the active model, the GVF model and the Mumford-Shah model have their advantage and disadvantages. But we should utilize their advantage in different occasions to serve us.