Research On Image Segmentation Technique Based On Deformation Model

Image segmentation is the key step from image processing to image analysis, and it is a classical difficulty between image processing and machine vision. Although many techniques and methods have been proposed, this problem doesn't have a general theory yet. Image segmentation technique based on deformation model, which utilizes areas and edges of images and combines geometry physics and approximation theory, is the most widely investigated and applied method.First, we introduce the basic idea of the level set method, the Mumford-Shah model and Active Contours model. Then we rewrite the original models in the level-set formulation and build the Euler-Lagrange equation. This work emphasizes on the piecewise constant level set methods proposed by Lie etc. The time-consuming re-initialization procedure is not necessary during iterations and it thus speeds up the curve evolution and the segmentation. Moreover, the level set function can be flexibly initialized with a piecewise constant function rather than a signed distance function in practice. Thereby the consuming time to compute a signed distance function from an initial curve in irregular shape is saved.We do much research on the methods of solving models, we find that although the gradient descent method doesn't need good initial values, the segmentation speed is very slow; Newton method can segment quickly, but it needs good initial values. Based on the Mumford-Shah model, we propose a fast implementation for segmentation. The new method is robust to noise. We use primal-dual method and Newton method for numerical computations. We program in Matlab and give the segmented results later.After studying on the sine-sinc model proposed by Y.M.Jung etc, we propose a new multiphase segmentation model. For numerical computations, we introduce a function to choose values of the parameters, and solve the problem with the gradient descent method. Numerical experiments show that new model can process more complex images, extract each individual segment or phase exactly, and the restoration preserve much more details.